Next best solution after diesel-carrying donkeys: long (mostly copper and aluminium) cables that transfer electrons instead of energy dense hydrocarbons. But, this means Capex investments in (you guessed it) copper, and transformers, and towers, plus Opex through maintenance, power management, etc. A number of recent reasons why this is painful and costly:

1. 13.5% of all aluminium stranded cable with steel core (ASCR) which is the standard conductor for overhead power lines

2. 12.6% of all aluminium wires

3. 11% of all copper wire

4. 8.8% of all circuit breakers

5. 7.9% of all refined copper

goes through the Strait of Hormuz (or used to).

To no surprise to anyone, our existence is fragile and dependent on our ability to keep global trade alive. Technology is our damnation and savior. There are ways to transfer power from A to B without physical connection. Perfect for where generation is in one place but usage somewhere else. This is called power beaming, or wireless power transfer.

Subscribe now

There are some recent developments that bring us closer to reality, with DARPA achieving a distance record in 2025 of 800W delivered over 30s across 8.6km; Starcatcher Industries transmitted 1.1 kW to COTS solar panels at ~1km distance with a multi-wavelength laser; Japan did a space-to-ground beam transfer of around 1 kW in Dec 2024.

Deployment scenarios

Just like with data centers in space, the argument goes that there’s a lot of sun in space, so why not harvest it and send it down to earth. Taken to it’s extreme we end up with Dyson Spheres which I’m all for, at some point. Climb the Kardashev Scale and so on. But, we can’t start at the end; it doesn’t make sense to light the candle before we made the cake. So before Dyson Spheres, there are simpler use cases in discussion on governmental levels, like NASA and ESA who analyzed Space Based Solar Power (SBSP) quite extensively. And on the private side, companies like TerraSpark, Aetherflux (though I heard they’re also doing data centers now), or Overview Energy. There’s a great article from Space Ambition on SBSP so I won’t go too deep into details here.

Where could power beaming be done:

• Space to space

• Space to atmosphere

• Space to ground

• Atmosphere to atmosphere

• Atmosphere to ground

• Ground to space

• Ground to atmosphere

• Ground to ground

Out of these modes, we would like to figure what is most feasible and which won’t work for a long time. That analysis is dominated by one key metric: end-to-end efficiency as a percentage of energy available for usage at the receiving end, vs. energy at the transmitting end.

The core technical challenge is that current best-demonstrated figures range from 10-20% wall-to-wall, meaning 80-90% of input energy is lost to heat and conversion losses across the transmitter, beam propagation, atmosphere, and receiver chain.

Usually there are five sequential conversion/transmission steps that incur losses:

1. Power source - solar panels in space, or grid/generator on Earth

2. Transmitter conversion - electricity becomes photons (laser) or microwaves

3. Beam propagation - the beam crosses the distance, fighting divergence and whatever medium it passes through

4. Beam capture - the receiver collects what arrives (aperture and tracking efficiency)

5. Receiver conversion - photons or microwaves become electricity again

In these 5 steps, there’s one fundamental choice to make: lasers or microwaves to do the transfer.

Laser beaming is precise and compact. You can hit a small receiver from kilometers away. It dominates in vacuum and thin atmosphere: space-to-space, space-to-stratosphere, and controlled ground environments. But fog, rain, and clouds can eat up to ~99% of your beam, making it unreliable for anything that needs to work in all weather.

Ah, you might say, but what about cloud-clearing, essentially shooting away the clouds with lasers!? Well, let’s do the math shortly. There are two approaches:

1. another laser beam that evaporates the water droplets of clouds—the hard way as you’re boiling the column of water

2. or a filamentation technique pioneered by Wolf’s group at Geneva, where a femtosecond Ti:Sapphire pulse Kerr-self-focuses into a plasma filament

Wolf/Geneva is much less energy demanding (back of the envelope 3-10x less power vs. raw boiling) and has been experimentally tested for laser communications, not yet for power transfer. So let’s consider that: the optical overhead to maintain a 10-30cm cone of ~500m cloud of water content of 0.3g/m³ is something like 4-40kW. There’s no experimental data on how much of the water was cleared, so let’s assume something like 90%. This means our new transmission efficiency is unfortunately still only 5%. Even if we clear 99% of the water tx efficiency is just ~75%.

In addition, the efficiency of the filament pumped lasers aren’t amazing, plus payload laser, so to get to 4kW you’d need to supply around 30-80kW. And that assumes we cna achieve a multi-filament array over 500m. Still unproven AFAIK.

Once the payload is 100 kW–MW CW through a 30 cm column, thermal blooming inside the cleared channel becomes the limiting physics, not the cost of clearing. So Wolf/Geneva opens a window between "too small to be worth it" (sub-10 kW, clouds are a minor annoyance you can route around) and "blooming-limited" (hundreds of kW+, the cleared channel heats up and defocuses your beam).

Microwave beaming punches through weather but requires enormous receivers. It’s the only viable option for space-to-earth at scale, and the safer bet for any scenario that demands all-weather reliability. The tradeoff is physical footprint - a space-to-ground rectenna is measured in kilometers. Even the best beam forming won’t solve diffraction fundamentally.

The max gain is approximately proportional to aperture divided by the wavelength squared. Additionally, there are limits to how much power can be transmitted per unit bandwidth per unit solid angle. Discrete arrays are constrained by λ/2‑type spacing and spatial Nyquist; exceeding this creates unavoidable grating lobes.

You can trade aperture size, frequency, scan angle, sidelobes, bandwidth, and power, but you can’t design a system whose far‑field beam is narrower or more intense than allowed by these constraints for a given aperture, wavelength, and power.

End-to-end efficiency

Taking the above steps and applying some common efficiency numbers to it, we get compounding steps of losses.

Without talking about the economics of any of this, we see quite clearly that very few use cases are robustly viable. Unless LCOE is close to zero. Operationally however, the goal must be to have weather independence on these power, at least to the degree that any power transfers can happen with sufficient accuracy.

Operational restrictions

The physical constraints start at the generation and end at the receiver end. In between it’s mostly safety and regulation issues that we’re facing. Turns out beaming loads of power through lasers or microwaves in free space might harm people, animals or equipment.

Generation is the same no matter what you need to do with the energy, but for space based specifically, or for air to ground beaming, you need to fly/orbit large structures of solar panels. In GEO it might be alright though, and GTO could be optimized for high perigee. On the ground, power could be coming from anywhere and more likely just use whatever mix the grid is on.

Transmission and DC conversion will likely not cause any larger issues, as it’s fairly standard infrastructure. We’re fairly good at turning power into photons/microwaves and back again. But for microwave transmitters the size is hard to avoid, meaning the infra will be fixed and large. For laser systems we get fairly high losses but infrastructure is physically smaller though not necessarily cheaper or simpler.

Microwaves use fairly developed systems like magnetrons or klystrons and can hit 80%+ efficiencies. Klystrons operate at very high voltages and gyrotrons (the MW-class cousins used in fusion heating) operating at high frequencies might jump between cavity modes. The supply chains are in construction in parallel with the fusion industry. They’re normally quite complex, and decently sized (telephone pole for ~1MW) so on the ground they aren’t limited, but in space or the air it might be an issue. One big constraint of magnetrons is that they can’t be phase-locked, hence phased array transmission will only work with klystrons/gyrotrons/solid-state.

Rectenna/receivers aren’t large issues for laser beams, with multi-junction or GaAs and the beam divergence being small means it doesn’t have to be a big PV. For microwave rectennas they will be large though, with power received growing linearly with surface area until the rectenna exceeds the beam footprint, at which point additional area catches nothing.. The beams aren’t uniform though, they are Gaussian, with most power in the center and less at the edges, which helps us.

At GEO, rectenna area is essentially fixed by physics at tens of square kilometers regardless of power level below the per-element ceiling. The 1979 reference design’s 10×13 km footprint is set by diffraction: at 2.45 GHz the wavelength is 12.2 cm, so from 36,000 km through a 1 km transmit aperture you get a minimum spot of λR/D ≈ 4.4 km diameter. Area is bounded by physics, not power level.

This is why SBSP economics are so brittle: you pay for ~100 km² of land, civil works, and array fabrication whether your satellite delivers 1 GW or 5 GW. The fixed land cost dominates the LCOE at small scales. You need multi-GW output per station to amortize the rectenna infrastructure, which in turn demands km-scale transmit arrays in orbit, which is where the launch-cost and in-space-assembly problem comes in.

For drones and aircraft, the main problem is really economical and operational:

• the physical infrastructure needed to beam power (high voltage grid connection+transmitter/antenna) becomes expensive and complex

• for lasers, weather conditions will dominate (e.g., in the UK which is the most overcast developed country in the world according to my research) and destroy any reliability. Not even in the desert is it reliable, as dust storms will kill transmission efficiency

• beaming microwaves at high power through the sky is not great for birds and bees, not to speak of other electronics or airplanes passing through, it would have to be a cleared airspace to be safe most likely (see more below)

Safety Considerations and Regulatory Landscape

Eye Safety

Near-infrared wavelengths (800-1400 nm) are particularly hazardous because they are invisible and do not trigger the blink reflex, yet can cause retinal damage below the threshold of pain. The Maximum Permissible Exposure (MPE) for continuous near-IR laser exposure is orders of magnitude lower than the intensities used in power beaming systems. This means we need:

• Automatic beam shutoff systems triggered by obstruction detection

• No-fly and no-go zones around active beam paths

• Operational altitude limits (systems designed to beam upward avoid low-altitude human exposure)

You can also introduce closed-loop systems that turn off automatically should something disrupt the path.

FAA and Aviation

The FAA requires advance notification (30+ days) for outdoor use of Class 3B or 4 lasers (above 5 mW), with details on power, location relative to airports, and control measures. Pointing a laser at aircraft is a federal crime with fines up to ~$32k per violation. In 2024, pilots reported >12,000 laser strikes - demonstrating the regulatory sensitivity around aviation laser safety. Power beaming systems operating near airports or flight paths face significant coordination requirements.

Regulatory Gap

There is currently no comprehensive regulatory framework specifically for high-power beam-path operations in public airspace. Each major demonstration (DARPA at White Sands, Star Catcher at Kennedy Space Center) has operated on military or controlled NASA land, avoiding the public airspace regulatory complexity. Commercial deployment in civilian areas will require new frameworks from the FCC (microwave frequency licensing), FAA (airspace coordination), and potentially OSHA (workplace laser safety).

LTA safety

For lighter-than-air aircraft, i.e., balloons/aerostats and airships, we need to consider the heating of the envelope and the contained gas inside. This could either harm the structural integrity of the aircraft, or make it change altitude. Some back-of-the-envelope calculations, make laser the higher risk here, as harming the envelope would need 3-4x OOMs less energy to cause damage. If the LTA is helium-filled, then the only way to heat up the gas would be by heating up the envelope or structure. Even 3kW into 1,200 kg helium would only add ~2 mK per second. And given the volume there’s a large surface area with cold air around it. For a laser though, it’s enough that it misses the receiver slightly and hits the LDPE film. It would likely only need 5-20s of 3kW laser to structurally harm the envelope.

So when does this actually make sense?

If your expectation is a general-purpose replacement for the grid then we’re far off, and Dyson Sphere’s aren’t close yet. The physics stack doesn’t compete with copper cables, battery storage, or even diesel generators on $/kWh. It competes with them only when they’re infeasible: when you can’t run a cable, can’t carry a battery, or can’t convoy fuel through contested territory.

That’s a smaller market than we might like, but it’s not zero. Four tiers, in descending order of how much I’d bet on the timeline working out and there being any economic case for them:

Tier 1 - Military drone persistence (2026-2028). PowerLight is under contract with CENTCOM to demonstrate “infinite flight” for Group-2 UAVs in 2026. DARPA’s POWER program is targeting 5 kW at 120 miles by ~2028 via relay drones. So it’s interesting to someone. The military doesn’t care about $/kWh the way a utility does; they care about not sending fuel convoys into contested territory and about keeping eyes in the sky without landing. A remotely-powered drone that never lands is worth more than the energy it wastes.

BUT: the real technology need isn’t power beaming itself; what would pay here is persistent infrastructure. Caveats on power beaming for this:

• this type of power beaming is mostly proven for low altitude, clear skies.

• Deploying this in e.g., Ukraine or the High North where conditions are often overcast or foggy will have very different performance parameters, even for microwave.

• Not harming the envelope will be hard, and setting up massive (20m+ diameter) antennas in contested environments with huge power supply is an issue

If we can keep eyes-in-the-sky or cell towers-in-the-sky permanently without having to connect huge microwaves on the ground (which can be taken out) then that’s preferable.

Tier 2 - Orbital power relay (2026-2030). Sunlit satellites powering eclipsed ones. Star Catcher is building toward this. Best physics of any scenario (vacuum, no weather, no diffraction hell), and the economics work because the alternative is bigger batteries or accepting blackouts in shadow. Not a consumer market, but a real one.

Tier 3 - Disaster and remote power (2027-2032). After an earthquake flattens the grid, you can’t run a cable. NTT and MHI have demonstrated ~15% efficiency through real atmospheric turbulence - enough for emergency comms, water pumping, medical equipment.

Tier 4 - Space-based solar power (2035+). The grand dream. Collect solar 24/7 in orbit, beam it down. Physically demonstrable (JAXA OHISAMA did it), economically brittle: you pay for a km² rectenna regardless of the power level running through it. ESA’s SOLARIS models suggest €88-155/MWh by 2040 - competitive with peaking plants but not with terrestrial solar at $20-60/MWh. I’d bet on it eventually. I wouldn’t bet a fund on the timeline.

Special shout-out to Big Space Lamp (TM), companies wanting to deliver lighting from space, e.g., like Reflect Orbital running constellations of huge mirrors. This sounds like the worst of all use cases and I’m not enlightened (hah) when it comes to the investment strategy there. Professional, mobile lighting towers might be a market of about $0.6-$2.5B but not sure what the EBITDA look like when you need to run a constellation to unlock it…

The Hormuz data at the top of this post is troubling, but wireless power doesn’t save anyone from a global copper-logistics problem at any realistic scale this decade. What it does change is the arithmetic at the small end: a forward operating base doesn’t need a diesel convoy if it can pull power from a rear-area laser. A post-disaster hospital doesn’t need to wait for the grid to come back. A satellite doesn’t need a bigger battery for every eclipse.

Considering the alternatives though, like tethered drones, VLEO and stratospheric airplanes, there seem to be better and more economically feasible solutions to the problems.

The shortest path (ground-to-ground) has the worst physics. The longest path (GEO-to-ground) has the best business case in theory and the worst capex in practice. The one in between (orbit-to-orbit) is where I’d look for the cleanest near-term outcome.

In our perception, public support for major civil movements, companies, technologies, nations and leaders have swung wider in the last 5 years than in the previous 20 years.

Defense was taboo just 2 years ago, now it’s all the rage, even though the russian invasion of Ukraine started 12 years ago. Natalism is having a comeback, industrialism is cool, dropping out of university to join YC is the new middle class path; the examples go on.

Maybe we’re in an opinion bubble as founders, investors and techies, but it feels like nothing really shocks us anymore. What most people think is edgy feels like something a character in a Silicon Valley reboot would say.

We want to examine the Wheel of Heresy: what does the machinery look like that turns heresies into orthodoxies, and orthodoxies back into heresies, and what is even heretical anymore?

Why? Because venture is built on paradigm shifts, not incrementalism. We also feel Europe is severely lacking in the public dialogue on valid but non institutionalized topics.

We are hosting the first Prometheus Forum this year, as an experiment to have heretics exchange ideas, and hopefully change some minds.

Vindicated Heretics

The canonical examples are Galileo and Copernicus, but these are too distant and too clean. In reality, there are too many stories to tell, and those are only the documented ones. There may be hundreds of historical truth tellers that were burned as witches without anyone to write down their words.

Ignaz Semmelweis — Handwashing (1847)

Ignaz Semmelweis discovered in 1847 that doctors washing their hands between autopsies and deliveries could reduce maternal mortality from 18 percent to under 2 percent. The medical establishment’s response was not to investigate his evidence but to take offense. Many doctors were insulted by the suggestion that a gentleman’s hands could be unclean. In 1865, his colleagues committed him to an asylum. He died fourteen days later from a beating by guards, of a gangrenous wound on his right hand. He died of the kind of infection he had spent his life trying to prevent. The eponymous “Semmelweis reflex” is the automatic rejection of new information that contradicts established norms. Adoption of handwashing took another twenty years.1

Dan Shechtman — Quasicrystals (1982)

Dan Shechtman observed a diffraction pattern in 1982 that violated what every crystallographer believed was a fundamental law: crystals must have periodic atomic arrangements. His lab notebook read “10-fold??”, acknowledging that what he was seeing was supposed to be impossible. The head of his research group told him to “go back and read the textbook.” He was asked to leave the group for “bringing disgrace” on the team. Linus Pauling, two-time Nobel laureate, dismissed him publicly.

“There are no quasicrystals, only quasi-scientists."

Pauling maintained this position until his death in 1994. Shechtman received the Nobel Prize in Chemistry in 2011. He received it alone.2

Lynn Margulis — Endosymbiosis (1966)

Lynn Margulis argued in 1966 that the cells in your body are the result of an ancient merger between separate organisms, and that mitochondria were once free-living bacteria. Her paper was rejected by approximately fifteen journals before anyone would publish it. She was called "science's unruly earth mother." The theory is now the orthodox understanding of how complex cells evolved.3

Stanley Prusiner — Prions (1982)

Stanley Prusiner proposed in 1982 that infectious diseases could be caused by misfolded proteins. No DNA, no RNA, no virus, no bacteria. Just a protein that recruited other proteins to misfold. This violated the Central Dogma of molecular biology at the most fundamental level. He was ridiculed. He won the Nobel Prize in 1997. The subsequent discovery that prion-like mechanisms underlie Alzheimer's, Parkinson's, and ALS has made this one of the most consequential heretical ideas of the twentieth century.4

The pattern is consistent enough to be taught as a law: the more consequential the idea, the more violent the initial rejection. And the rejection is not random. It follows predictable sociological lines: career incentives, institutional prestige, the economics of existing treatments and theories that would become obsolete. If a conclusion is highly uncomfortable, then it's more likely to be rejected. Big, important ideas tend to lead to discomfort.

The Ant Mill

There is a phenomenon in army ants called a death spiral. When a group of ants becomes separated from their colony, they begin following the pheromone trail of the ant in front of them. Since they are walking in a circle, each ant reinforces the trail for the next. They march in a continuously rotating circle until they die of exhaustion. Collective action resulting in tragedy. Jesper Grimstrup, a disillusioned former theoretical physicist, published about the Ant Mill phenomenon in 2025.5

Thomas Kuhn had a word for the productive version of this: "normal science". During normal science, researchers solve puzzles within an accepted paradigm. They don't question the paradigm itself, that would be unproductive, even career-ending. The paradigm defines which questions are legitimate, which methods are acceptable, which answers count. This is not a bug. Most of the time, constraining the search space is how you make progress. Kuhn called it "the essential tension": science needs both the discipline of tradition and the disruption of innovation.6

But Kuhn identified a structural vulnerability. When anomalies accumulate and results don't fit the paradigm, normal science doesn't abandon the framework. LHC not finding supersymmetric particles didn't lead to a major revision of theory. It generated ad hoc modifications, auxiliary hypotheses, epicycles. This continues until a crisis point, when an alternative framework emerges and enough practitioners switch allegiance to tip the balance. Kuhn observed that it resembles a "conversion experience" more than rational persuasion, which is why the old guard rarely converts, and why Planck was right that science advances one funeral at a time.

The ant mill is what happens when normal science loses its self-correcting mechanism: the anomalies pile up but the phase transition never comes. Through “almost correct” incentive structures, the field veers off into an unproductive direction.

The young solo researcher, Grimstrup notes, has almost entirely disappeared from theoretical physics. Half a century ago, it was the norm. Today, the field has coalesced into a small set of large research communities where technically skilled “technicians” focusing on narrow specialties outcompete visionary, conceptual thinkers.

The ant mill turns.

If science fails this badly, public discourse fails even faster and worse.

The Spinning Corridor

The same machinery operates outside science, faster, and with less pretense of objectivity.

The Germans call it Meinungskorridor. In 1990, 78 percent of Germans felt free to speak politically. By 2023: 40 percent. The Overton window used to move slowly: gay marriage took 25 years from unthinkable to settled law. Now the window spins. UBI went from fringe to near-mainstream in five years. The free trade consensus collapsed in a decade. The rules-based order is in question. Nuclear energy has flipped moral valence four times in living memory and now polls at 61 percent, with its strongest support among Republicans. Same technology, same physics, different tribal signals.7 8

The lab leak hypothesis is a good example. Misinformation in 2021, organized against via a conflicted Lancet letter, privately acknowledged as plausible by the same scientists who publicly dismissed it, endorsed by the FBI by 2024. The frame shifted completely. No accountability for those who enforced the wrong one.

The Overton window is oscillating, about to explode like an Isar Aerospace launch. Keynesianism to monetarism took a generation; now intellectual fashion cycles like fast fashion. Each generation overturns the previous one’s orthodoxy out of identity, not evidence. As usual we want to rebel against our parents, but accidentally take on our great-grandparents’ opinions to do so.

The Selection Problem

Most heretical ideas are just wrong. Social structures are reasonably good at tolerating harmless wrong beliefs and correcting dangerous ones toward correct-enough norms. I’m sure the Artemis II crew agrees.

Unorthodox beliefs are a mixture of genuine insights and crankery, and the ratio heavily favors crankery. Flat earthers are heretical. Anti-vaxxers are heretical. The person who thinks they have disproved Einstein in their garage is, in the vast majority of cases, simply wrong. For every Shechtman there are ten thousand cranks. For every McClintock, a thousand confused amateurs.

The hard problem is not holding heretical ideas. Anyone can do that. The hard problem is designing selection mechanisms that filter for productive unorthodoxy without filtering for conformity. This is what Prometheus attempts: presenting heretical ideas openly as a way of breaking social barriers. These are experiments in selection mechanism design.

The ideas that matter most are often the ones the machinery is designed to suppress. Not because suppression is evidence of truth, but because the machinery does not distinguish between the heretic and the crank. The ratio of world-changing insights among the uncomfortable ideas is higher than zero.

With Prometheus, we’d like to invite those of dissenting opinion to present their cases for why they’re right and most others are wrong about their area of expertise. This can range widely from philanthropy to AI, from public health to statecraft. The main point is that we want to hear from people who have thought deeply about important topics, and we want to spread infectious ideas in an accepting forum focused on discourse instead of judgement or norms.

We will filter out for “cranks” and provocateurs who are driven by the attention, as well as not-that-unorthodox ideas that are widely accepted in the tech world but maybe not yet in wider society.

The exact design of the forum is still in the making, but we’d like to focus on very few talks/shows/demonstrations, and small, curated interactive workshops around specific topics.

If you are open to challenging the paradigm, if you want to experience the weird ideas of tomorrow, or if you want to convince others of your weird ideas, please join us at the inaugural Prometheus Forum in Munich.

Sign up here

Maritime persistent surveillance is a problem defined by three numbers: 361 million km2 of ocean surface, ~100,000 merchant vessels in transit at any time, and the speed of a ship at 15 knots - 667 km per day. Any vessel unobserved for 6 hours could be anywhere in a 160 km radius. Underwater vessels are even harder to detect or track due to physics.

We spent some time figuring out what it actually costs to watch the ocean. The short answer: it depends enormously on how you do it. A manned frigate costs about 10,000x more per km2 of coverage than a satellite. No single sensor works everywhere. The physics won’t let you look at everything at once in high resolution - so you have to stack cheap wide-area systems on top of expensive close-range ones.

The market for autonomous maritime surveillance is somewhere between $2-5B/year and growing fast, driven by submarine cables getting cut, offshore wind farms needing protection, and NATO suddenly realizing that 4-6 frigates can’t monitor the Baltic. The opportunity isn’t in building another drone boat but in selling persistent ocean awareness as a service, priced against what navies currently pay for frigate patrols. That reframes a small hardware market into a much larger recurring one.

Backstory (skippable but adds to the mood)

I grew up swimming in the Seas of Kattegat and Skagerrak. Snorkeling before I could swim on my own, exploring the green murkiness of what felt like an infinite, secret world. I felt at home among the cliffs, the cold salt water spray, and the uninhabited islands as the only barrier to the real Ocean. The real Ocean featuring 30m waves, unknown isolation, creatures and distances unheard of. This world belonged to me and I to it. It was my own Narnia, just to dream of this.

What made it magical was how restricted it was. To enter this world you needed to prepare, to dare and then to be comfortable about not knowing and not being in your own element. Your perspective in the water shifts from eye-level to floor level, assuming there are no waves, but in return you are flying on top of a pillar of glassy substance that carries you wherever you go. It’s flexible enough that if you chose to you can split through it and let negative buoyancy work. Once you pass the equilibrium between compression and trapped air, you are pulled down and don’t need to fight anymore to gain speed. Aerodynamics are now hydrodynamics and the way you hold yourself has immediate feedback on your speed.

When you dive, you don’t see the same way, colors are distorted and light reflects in weird places, if there is any at all. Looking up you see a mirror of the world you normally belong to. It’s bright and warm, yet the pressure under water, and constant touch makes air seem so lightweight and fleeting. Just getting out to the place where there might be something interesting to see underwater meant either swimming or navigating on a boat, which in the Swedish Archipelago is quite treacherous (pre widely available GPS). Water is, as we know, conductive. Not just to electricity, but also to waves in the electromagnetic spectrum. Or more precisely, water molecules get easily excited by incoming photons. This makes a little water transparent and, slightly more, very opaque. In fact, so opaque they use it to great success to contain radiation in nuclear reactors, and also why you can’t easily detect submarines.

Of course, as a 6 year old, detecting submarines wasn’t my main priority. I just wanted to learn the secret passage to the sea. So I decided to become a marine biologist and started learning everything I could about dolphins. This didn’t exactly pan out, as I later went into a much drier (pun) field of engineering physics. Fast forward about 20 years and I was snorkeling in one of the most frequent shark-attack waters on earth to spot wild dolphins. But that’s another story.

We moved away from the sea, I started diving in pools and flooded silver mines (probably in hindsight not super healthy), until I moved back to the sea on my own as a 16 y.o.. I felt entirely at home in the water, so much so that my new family thought they’d lost me the first day at the water, during a precursor for the autumn storms that were to thrash the North Atlantic coast in Bretagne. I was just exploring the moving sand and rolling waves after a decade of calm waters. Finally coming up for air and putting my hand to my head, just to say “tout va bien”.

60,000 Nautical Miles under the sea

In the years away from the sea, my favorite book growing up was “20,000 Leagues under the sea” by Jules Verne, which tells the story of a secretive submarine captain taking a professor captive and exploring the worlds oceanic environments. Unbeknownst to me, I was moving not far from Jules Vernes native Nantes about 180 years after he was born on a tiny island in the Loire. He excelled at mémoire (remembering things by heart and reciting them, which was a core skill of that time), latin, and geography. Captain Nemo and Professor Aronnax explore creatures and treasures hidden in the sea unheard of until then. Likely based on myths and seafarer stories from the docks. It was written in 1870, before submarines were really a thing, and when diving equipment was limited to diving bells and surface air supplied through tubes to massive metal helmets with glass windows.

The vessel that they travel in, the Nautilus, was an electric sub, cigar shaped and about 70m long x 8m in diameter. Verne intentionally or accidentally actually didn’t violate too many physical laws or engineering principles when designing it. It’s surprisingly feasible. It would have a displacement of about 1,500 m3 sea water meaning it would weigh around 1.5e6 kg. Nautilus traverses the Mediterranean (~2,000 km in Verne’s geography) in about forty‑eight hours, implying an average speed on the order of 20–25 knots depending on exact endpoints. It’s fully electric and probably has to produce around 5MW in shaft power at that speed. It's said to run on some sodium-mercury batteries, with days and days of range without charging. These batteries or engines didn’t exist at the time of Verne’s writing, one might add. But to produce such power, you’d likely need O(200m3) worth of batteries, and about 200t worth of batteries. Which could plausibly fit inside a 70x8m cylinder and could be carried weight-wise! You’d need about 300-600Wh/kg on a system level, and about 10x what that type of batteries could do at the time.

Now the fun part is that a nuclear submarine has virtually unlimited range and about 3-5x the shaft power of Nautilus. We’ve outdone Verne’s imagination. The sea monsters are human made, once again.

Maritime surveillance economics (skip to here)

If you skipped the preamble about 19th century sci-fi and dolphins, welcome back! You’re alright.

Recent conversations I’ve had, as well as naval warfare events I’m planning (as you do), as well as recent events in the Strait of Hormuz, plus pending events in the GIUK gap, Baltic Sea infractions, a pending South China Sea conflict, etc. had me think:

What is the most cost-effective way to detect what’s going on everywhere on the oceans, and maybe below?

Why this matters

Maritime monitoring and surveillance was around ~$24-26B in 2024, but it only takes fairly little disturbance at sea to significantly disrupt entire industries. A few pirates with drones. A couple of jerky misjudgments on demand/supply of household electronics during a pandemic. A single nuclear sub. A few “fishing” vessels dragging their anchors a few hundred times over fiber optic cables.

That last one isn’t hypothetical. There’s currently something like 1.2 million km of submarine telecom cables in the water, and another $13B worth planned for 2025-2027. This is driven by the hyperscalers and Meta mostly, with insane communications demand. The bandwidth is simply unbeatable, and we’re getting better and better at it. There are people working on building underwater power generators to avoid having to install land-based repeaters, and it looks like a large market. (If you are working on this or something related, come talk to us!)

Beyond cables, historically the sea has been a fairly high leverage way to conquer new areas. This funnily enough also has a physics explanation, as transporting heavy, bulky things on water requires much less energy than on roads or rails. In fact more than 80% of the worlds goods use maritime transport today! With developing economies and Asia growing massively in that domain, while the rest are declining.

Who’s building

The US just backed a maritime-only focused VC called “Mare Liberum” after the 1609 treatise by Hugo Grotius arguing for free access to the South China Seas. Back then over the Portuguese, not presumably over the Chinese themselves.

On the defense side, there’s a growing wave of naval startups - Saronic (~$700M raised), Kraken Technology Group (UK/NATO contracts), UForce (whose Magura drones have been instrumental in taking out the Russian navy in the Black Sea), plus HavocAI, Vatn Systems, Toloka, Nautrik, Polar Mist, Ray Systems, Regent Craft, and others. Helsing and Anduril have their own products too. Lots of capital deployed, unclear who’s made money yet.

The proof point that these matter: at a recent NATO exercise, REPMUS, Ukrainian USVs were pitted against allied naval vessels armed to the teeth. NATO forces didn’t even have time to notice the opponent was there before they were hit. A real confrontation would not have ended well for NATO.

On the civilian/dual-use side, companies like Hawkeye360, Seasats, and Saildrone started with commercial missions but have been pulled toward defense. Newer ones are going after seabed mapping, hull cleaning, or offshore platform maintenance. Saildrone set up its EU headquarters in Denmark after the Danes invested ~$25M. The dual-use trajectories are converging.

Subscribe now

Our approach

So assuming there is demand for persistent monitoring of maritime environments, we need to optimize for one variable: cost per km2-hr of monitoring. In our findings, these vary with over 3 orders of magnitude across platforms. Mainly due to operating constraints and clear superior reach of certain technologies. Manned frigates are just insanely expensive to build and to man and operate. It feels like absolute insanity that these are still being built to the extent they are, and still people complain about our ship building capacity. At the same time, we need something to carry all our drones and USVs and fancy unmanned systems as they won’t have the range needed to fight somewhere in the middle of the pacific.

We built an interactive page to test our own assumptions: Maritime Economics. Check it out, and if you’re extra nerdy, go to the physics section.

If you’re against vibe coded things by principle, feel free to skip it. Now back to the carefully handcrafted text:

The big physical problem

Is that the seas are huge. Like 361 million square km or so big on the surface. Naval forces probably cover <1% of that at any given time. Which gives boats and sneaky dolphins way too much space for mischief and play. Enter AIS (automatic identification system), which is an identification system for ships. Ironically it’s not so automatic as it should be, as it can easily be turned off. And even if something is detected, the sea is so huge that doing something about it, intercepting or so, becomes an even greater challenge.

So the sea is vast, isn’t nice to electromagnetic fields, and generally deteriorates most mechanical systems. In summary, this is why it’s hard. There’s an information theoretic “principle” which becomes apparent when you try to design a system that looks everywhere at once at high resolution. You can’t look everywhere at once, with high resolution. The explanation is on the physics page of the Maritime Economics dashboard we built!

What about radar?

Radar at sea becomes an even bigger pain than normal radar, because all the water starts making the radar waves bounce, leading to low signal-to-clutter (SCR) ratios. It’s “clutter-dominated” (insert joke about my tidiness) meaning the amount of background noise is determining if you find something rather than receiver sensitivity, on a whole. Unless it’s a very calm day at sea, it becomes impossible after some distance to detect anything with a low false positive rate. Finding a periscope at sea is like finding a sparrow in a thunderstorm.

The radar range equation governs all active microwave sensor performance. The range scales as R^4, which means that detecting something 2x further away requires 16x the power. So just making bigger antennas quickly becomes impractical. The maximum detection range R_max for a monostatic radar is:

Where:

• Ps = Transmitted peak power (W) - typical maritime surveillance radar: 25 kW to 1 MW

• G = Antenna gain (dimensionless, typically expressed in dBi)

• λ = Wavelength (m), inversely related to frequency: λ = c/f

• σ = Radar cross section of target (m²)

• P_Emin = Minimum detectable received power (W), set by receiver noise floor

• L_total = Total system losses (atmospheric attenuation + hardware losses)

The R^4 dependence is the defining constraint: doubling detection range requires a 16x increase in transmitted power or antenna aperture. This creates a fundamental size/weight/power tradeoff for airborne and space-based platforms. Switching from Gaussian to K-distribution statistics - which occurs at high resolution or heavy sea states - degrades detection performance by up to 12 dB when using standard pulse-to-pulse integration. This means systems designed for calm-sea performance can catastrophically underperform in heavy weather.

There are interesting ways around some these limitations, and ways that engineers try to limit the radar cross section (RCS) of their vessels. My favorite type of radar is probably so-called over-the-horizon radar (OTHR), which uses the earths ionosphere to bounce radar waves off of to extend the range and avoid some sea surface scattering. This means ionospheric variability - diurnal, seasonal, solar cycle, and storm-induced - creates significant temporal uncertainty in system performance. Frequency management systems must continuously adjust operating frequency (typically sweeping 4-30 MHz) based on real-time ionospheric sounding. It also means OTHR have blind spots due to the multi-hop nature.

There’s a non weather-dependent type where we instead bounce polarized RF off of the sea surface, called Surface Wave Radar (HFSWR), which is useful for persistent coastal detection.

On a whole, though, these radars are massive and enormously expensive. For reference, a single OTHR cost OOM hundreds of $M, in capex and annual opex in the tens of millions. And their coverage is good but resolution is not great, e.g., for small vessels. Canada recently bought one $4B OTHR system from Australia to cover the Arctic…

What about optical sensors?

About 67% of the sea is covered in clouds at any given point in time. So pure optical is hard/inefficient. This is why SAR (Synthetic Aperture Radar) has become the real workhorse for maritime surveillance from space - it works through clouds, at night, and produces images good enough to detect ships. Companies like ICEYE, Capella, and Umbra are deploying constellations of small SAR satellites at ~$3M per unit. At 0.25-1m resolution you can identify vessel type, and wide-area scan modes cover 100x100 km per scene. Sentinel-1 does 400 km swaths at 40m resolution. The tradeoff between resolution and swath width is physics, not engineering - same information-theoretic constraint from above.

For the visible/IR spectrum, things get more interesting but also more limited. Water vapor absorbs certain IR wavelengths aggressively, which makes it difficult. Interestingly, MWIR (3-5 um) actually outperforms LWIR (8-12 um) in humid maritime conditions, because of precisely this. And it works best for things that have a thermal signature, like ship engines. Using the Johnson criteria and NVTherm modeling, we can figure out practical detection ranges for maritime vessel targets using MWIR (640x512, 330 mm lens):

• Large vessel (heat signature, 10 m target extent): Detection to 15-30 km

• Small vessel (5 m target extent): Detection to 8-15 km

• Human in water: Detection to 2-4 km

• Range degrades sharply in fog, rain, or high humidity - often to <1 km in bad conditions

Thermal is interesting, but of course there are countermeasures. Submarines running hot engines cool down their exhausts before ejecting them so no hot bubbles rise to the surface.

From space, optical constellations like Planet (~200 satellites, 3-5m resolution, up to 12 passes/day) and Maxar WorldView Legion (30 cm resolution, 15 revisits/day) can image ships in clear weather. But “clear weather” is the catch - two thirds of the time you’re looking at clouds. This is why SAR, not optical, is what actually finds dark ships today. The intelligence companies doing AIS anomaly detection (Windward, Global Fishing Watch) correlate SAR detections with missing AIS signals to flag vessels that have gone dark. It’s clever, but it’s only as persistent as the satellite revisit rate allows.

We’re still limited by the information theoretic principle of swath vs. resolution tradeoff from above. If we’re closer to the earth, we can have higher resolution but can’t get good coverage. If we’re using satellites, high resolution forces narrow swaths. No amount of engineering budget changes this - it’s the space-bandwidth product doing what the space-bandwidth product does.

Using water as a lens

Importantly, since radar doesn’t work under water, we need its slightly cooler cousin who only listens to vinyl (analog only baby), sonar. It’s effectively like radar, but instead of RF emissions, you emit sound (vibrations). Since speed of sound in a medium depends on the temperature and pressure of the medium, and those two things differ in water depending on depth, there’s a phenomenon called SOFAR, or SOund Fixing And Ranging. It’s essentially using a depth range where the sound bounces cylindrically instead of spherically, leading to a ~30dB advantage. It’s usually 600m-1000m but depends on the environment. There it’s easier to detect subs or things under water, but also much easier to communicate between vessels.

Still, sound degrades very quickly under water, which means detection at distance becomes a real problem, especially when someone tries to avoid detection by going electric or generally masking their sound signature. So if you’ve played around in the sonar detection section of the Maritime Economics page, you’ve probably found it pretty hard to detect a sub.

Still, in the SOFAR, sound can travel thousands of km underwater, whereas radar is horizon limited, at least. Unless it’s OTHR of course.

Active sonar transmits a pulse and listens for the echo. The sonar equation for active detection:

The critical difference: 2 ⋅ TL - the signal must travel to the target AND back, doubling the transmission loss in dB. This means active sonar detection range is always shorter than passive for the same source level. Active sonar also faces reverberation - acoustic energy scattered by the ocean surface, bottom, and volume. In shallow water, reverberation often exceeds ambient noise, making active sonar reverberation-limited rather than noise-limited.

Share

The fancy methods

If we think specifically about detecting modern submarines, there are a bunch of fun ways people have thought about to detect them. Creativity is the limit here, really. A submarine is something very very heavy, moving quite a lot of water, making some sounds, and probably disturbing whatever is around it. How can we detect it?

Modern diesel-electric and AIP submarines have reduced source levels by 30-40 dB compared to Cold War boats:

A 30 dB reduction in source level translates to a 1,000x reduction in acoustic power. Detection range scales roughly as

for cylindrical spreading - so a 30 dB quieter submarine is detectable at roughly 1/30th the range… Not good for detection.

Subs are heavy! So that means we could sense an increase in gravitational force near it, right? This is called Gravity Gradiometry. A sub weighing 2,000-8,000 tonnes does create a gravitational anomaly, but unfortunately that anomaly is about 1,000x smaller than the sensitivity of modern gravimeters. So either subs need to get a lot bigger, or we need to make quantum gravimeters a reality.

Subs are metallic chunks. That should be magnetic, right? A submarine’s steel hull does create a magnetic dipole anomaly in Earth’s field. The dipole field falls as 1/R^3:

For a submarine with magnetic moment ~10^6 A*m^2 and a magnetometer sensitivity of ~0.01 nT we get:

• Detection range: ~500m-1km laterally

• Requires direct overflight at low altitude (<300m)

• Used by P-8A, P-3C, and ASW helicopters for final localization

I’ve heard that there are places dedicated solely to the demagnetization of submarines though, so this might not work as well.

This could be avoided if submarines were built outside of a magnetic field completely, like in space, but that has different difficulties attached to it. Even so, when the metal moves through Earth’s magnetic field, it accumulates a temporary magnetic component that changes with direction. Modern submarines also have built-in degaussers, which are essentially coils built into the hull to counteract magnetic fields.

Submarines should make waves underwater, right? A submarine moving at depth creates detectable surface signatures:

1. Kelvin wake: V-shaped surface wave pattern (half-angle 19.5 degrees) from pressure wave at hull

2. Internal waves: Displacement of density layers, potentially visible in SAR as modulated surface roughness

3. Bernoulli hump: Local sea surface height increase above the submarine (~mm scale)

SAR satellites can theoretically detect internal waves from submarines at periscope depth (30-50m). Operational capability is claimed by several programs but unconfirmed in open literature. The physics is marginal - the signal is comparable to natural internal wave amplitudes.

When it moves it should stir up organic matter from depths, right? In waters with high dinoflagellate concentrations, a submarine’s passage mechanically stimulates bioluminescence. This creates a glowing wake visible to sensitive airborne EO sensors at night. Detection ranges: 1-5 km from directly above. Limited to specific ocean regions and seasons. Not operationally reliable. Same applies to thermal mixing from different temperatures at different depths. Not a big enough temp diff to measure.

All of these non-acoustic methods share the same problem: they’re either too short-range to search with (MAD), too insensitive to work (gravity), or too dependent on perfect conditions (wake, bioluminescence). Acoustics will remain the primary way to find subs for at least another decade.

None of these methods alone solves the problem. So instead of asking which sensor is best, the right question is: what does each one cost per unit of ocean monitored?

So what does it cost?

All of these physics constraints translate into wildly different price tags for the same output: knowing what’s going on in a patch of ocean. We normalized everything to cost per km2-hour of persistent coverage and the range is absurd. More than three orders of magnitude:

Reading this table top-to-bottom reveals that it isn’t a market failure we don’t have persistent surveillance - it’s all physics. Cheap persistent coverage tells you very little (OTHR: “something is somewhere”). Expensive close-range platforms tell you everything (boarding a vessel) but cover almost nothing. The design question isn’t “which platform” but “how to stack them.”

The optimal architecture is likely nested: wide-area cueing from satellites and OTHR triggers HALE UAVs, or HAPS or SAR tasking, which cues USVs or MPA to investigate, which cues a surface vessel to intercept, which could cue UUVs. Each layer is using physics it’s good at and handing off to the next. The alternative - patrolling a million km2 with frigates - costs roughly $5B/year in operating costs for the US surface fleet alone, and still leaves >99% of the ocean unobserved.

The small boat market that isn’t small

Austin Gray, a former Navy officer and now co-founder of Blue Water Autonomy, wrote a piece that’s making some very good points. His thesis: the small USV market is tiny (~$120M/year) and already saturated. He counts 20 chokepoints, estimates ~1,190 USVs needed (330 sentries + 860 speedboats), subtracts the ~600 already under contract, and concludes there’s only ~$117M/year left if you refresh every 5 years.

He’s right about the sea denial market. If you’re building speedboats to blow up (a la Ukraine), MAPC already has the Replicator production contract at 32/month, Saronic has hundreds of millions in bookings, and there isn’t room for a fifth entrant. That market is spoken for. Assuming those companies are able to fulfil the requirements of working products...

But he misses the monitoring market because he counts USVs like inventory rather than like patrols.

A renewably-powered sentry USV (Saildrone, SeaSats, OceanAero) cruises at 2-3 knots. It covers about 50-70 km/day. To persistently monitor the GIUK Gap (~300 km wide, 200 km deep), you need about 48 USVs on station simultaneously. But they rotate - transit to station, time on station, transit back, maintenance. At 67% utilization, you need ~72 in the fleet just for one gap. Gray’s model puts maybe 10-15 there.

Scale this across the six NATO priority theaters (GIUK, Baltic, Mediterranean, Black Sea, North Sea, Arctic) and you get ~336 sentry USVs for NATO alone. Add the Indo-Pacific, Gulf, and allied waters and you’re at 650-700 globally. That’s 2x Gray’s estimate, and the refresh rate isn’t every 5-10 years like hardware - it’s continuous. Sensors need upgrading every 3-5 years, hulls take a beating, and you lose 5-10% annually to weather and the occasional unfriendly interaction. Effective annual fleet replacement: ~22%.

So the monitoring hardware market is maybe $77M/year globally. Still not huge. But that’s the wrong frame.

Don’t sell boats. Sell awareness.

The right comparison isn’t “what does a boat cost” but “what does it cost to know what’s happening in this piece of ocean.”

Today, NATO pays for this knowledge with frigates. A Constellation-class frigate costs $63-130M/year to operate. Its radar covers maybe 1,250 km2 around it. For persistent coverage of the Baltic cable corridors (~10,000 km2), you’d need 8 frigates on continuous rotation. That’s $500M-1B/year. To monitor cables.

50 sentry USVs covering the same area as a service? $25-40M/year. One-twentieth the cost, genuinely persistent (24/7/365), and you don’t need to find 2,000 sailors to man them.

The pricing implication is clear: don’t price against other USVs. Price against the frigate. If your monitoring service costs $15,000-25,000/day for 5,000 km2 of persistent coverage, you’re charging 10-15% of what a frigate costs for coverage a frigate literally cannot provide. The customer saves money AND gets better coverage. Those are the kinds of trades that close.

This reframes the addressable market from Gray’s $117M/year hardware business to a $200-350M/year recurring service business anchored against NATO maritime patrol budgets. Different game entirely.

Does NATO have the budget?

Yes. And it’s getting bigger fast.

NATO defense spending hit 2.76% of GDP in 2025 - first time all 32 allies met the old 2% floor. At the June 2025 Hague Summit, they set a new target: 5% of GDP by 2035. That’s not a typo. Five percent. For context, the US is around 3.4% and Europe was at 1.4% as recently as 2014.

European allies + Canada collectively spent $482B in 2024. EU equipment investment grew 42% in a single year to EUR 106B. Germany went from a geopolitical punchline to spending $107B on defense. Norway is dropping $55B through 2036 on 5 new frigates, 5 submarines, and 28 other vessels. The UK’s Royal Navy equipment budget jumped 41%.

The money is there. The question is where it goes.

Most of it is still going to traditional platforms. Germany’s F126 frigate program is EUR 10B for 8 ships. France just greenlit a EUR 10.25B aircraft carrier. These are necessary but they’re not solving the monitoring problem. You can’t have a EUR 1.25B frigate babysitting a fiber optic cable 24/7.

The encouraging signal is that NATO is already building autonomous monitoring at small scale. Task Force X Baltic launched in late 2024 with 8 nations contributing USVs. By mid-2025 they had 50-60 deployed with a target of ~100. Baltic Sentry launched in January 2025 specifically for undersea cable protection. The US stood up its third USV squadron. NATO DIANA selected 150 innovators with maritime autonomy as an explicit challenge area.

The spending wave from 2.76% to 5% GDP creates roughly $100-150B/year in new allied defense spending over the next decade. If even 1% of that increment flows to autonomous maritime monitoring, that’s $1-1.5B/year in new market by 2030.

The awkward part about selling to the US Navy

There’s a catch, and Gray & Gray document it painfully well in “Startups & Sea Power“: the US Navy is structurally hostile to startups. Only 4% of top defense startup revenue comes from the Navy (vs. 31% Army, 37% Air Force). The Navy has spent $7B on flexible OTA contracts vs. the Army’s $53B. 71% of NAVAIR contracts are sole-sourced. Only 8% of NAVSEA contracts are even theoretically accessible to a company without an existing relationship and a security clearance.

Navy FY25 budget for USVs? $172M. For shipbuilding? $36B. That’s a 200:1 ratio of old to new.

So if you’re building a maritime monitoring company, the US Navy is probably your worst first customer. Not because they don’t need it - they clearly do - but because the institution is optimized to buy expensive, exquisite, manned platforms from incumbents who’ve been doing this for 50 years.

European NATO allies are the better wedge. They’re scaling budgets 40-70% from a lower base, they have acute monitoring gaps (ask anyone in the Baltic states about cable security), and their procurement systems - while not perfect - don’t have the same institutional antibodies to new entrants. Denmark invested $25M in Saildrone’s European HQ. Norway is planning 28 new vessels. The demand signal is there. If you get in front of it, you might not need the US Navy at all for the first few years.

What it means

The ocean is vast, hostile to sensors, and increasingly contested. The physics is hard but well understood. The economics are clear: manned platforms are 10-100x more expensive per unit of awareness than unmanned alternatives. The budgets are growing. The technology works - Ukraine proved that small USVs change the calculus at sea, and Saildrone proved you can keep a drone on station for months.

The opportunity isn’t in building another speedboat. It’s in building the company that makes “what’s happening in this patch of ocean” a commodity service priced at a fraction of what a frigate costs. The winner will own the operational complexity - keeping 50-300 USVs at sea, processing the data, delivering actionable intelligence - and price it as surveillance-as-a-service against naval patrol budgets.

We’re keeping a close eye on this space. If you’re building something here, or thinking about it, and have a product or technology, you should reach out. We’re also still building the interactive tools - the Maritime Economics dashboard and physics page are live and probably have bugs. Have a poke around and let us know what you think.

Leave a comment

Subscribe now

And if you ever happen to be snorkeling off the Swedish coast, keep an eye out for submarines. They’re sometimes easier to find than you’d think.

Note: If you’re new to the most powerful way to organize your knowledge base, check out the book Building a Second Brain. You can find a quick summary video here.

The Problem With Every Note-Taking System

Humans think in associations, not hierarchies. You connect a semiconductor supply chain memo to a defense strategy note to a book passage about national sovereignty - because they share structural logic. Traditional tools force this associative thinking into trees. The graph gets flattened into a filing cabinet.

The second brain movement solved the capture problem. Millions of people now have vaults, notebooks, and databases full of highlights, clippings, and notes. But retrieval stayed manual. You had to remember what you wrote, where you put it, and why it mattered. The more you captured, the harder it got to find anything. Past a few hundred notes, most vaults become graveyards - well-organized, rarely revisited.

Three things had to converge to change this:

• Graph-native knowledge tools (Obsidian, Anytype, Roam) that treat links as first-class citizens, not afterthoughts

• Embedding models that find connections you never explicitly made, based on meaning rather than keywords

• Agentic AI that can traverse, query, and synthesize across the entire graph in a single reasoning chain

The knowledge tool stores structure. The embedding model finds hidden similarity. The agent reasons over both. Together, they turn a passive archive into an active thinking partner.

What This Looks Like in Practice

One implementation of the above is Cornelius, an open source library you can check out here. A Claude Code agent configured to operate my Obsidian vault (or Anytype or Roam vault) as a second brain. Think of it as a layer cake: me at the top, Claude Code as the general-purpose AI, and Cornelius as a specialized layer underneath with purpose-built tools for knowledge work. It manages notes, extracts insights from what I read, discovers connections between ideas, and synthesizes across months of accumulated thinking.

The system currently runs across 298 notes with 2,354 edges in the knowledge graph - roughly half explicit wiki-links I created manually, half semantic connections the AI discovered on its own.

This is a real time glimpse into my second brain:

Subscribe now

This stack enables a few things that haven’t been possible before. Here is one out of hundreds of examples from the last few months:

1/ Cross-domain connection discovery. Three portfolio companies - one doing encrypted computation, one building cryptographic silicon, one working on undetectable networking - were written up in separate memos years apart. The system found they form a single architecture. Fully homomorphic encryption is too slow for real workloads, which creates the need for hardware acceleration, but encrypted data still travels on observable networks where metadata alone reveals everything, which creates the need for dark networking. Each company’s binding constraint is the next company’s thesis. The AI traced a constraint cascade across several unrelated investment memos and surfaced an emergent “zero trust compute stack”. Obviously we were aware of those connections when we took the investment decisions but still I was very surprised to see a machine “reason” through seemingly unrelated notes where such connections weren’t made explicit at all.

2/ Synthesized recall across months of work. One question - “what do we know about post quantum cryptography” - pulled together a technical notes on the state of quantum computing, several deal calls with companies working on quantum hardware and software, and book passages about national security infrastructure. Finding those building blocks of knowledge and connect them through years of time and siloed information scattered across different emails, folders, gdrive and CRM tools would have been impossible before. The system reconstructed my accumulated position on the topic in seconds.

3/ Portfolio-level pattern recognition. Once the system identified the zero trust stack, it kept going. It found that the constraint cascade pattern - where one company’s limitation creates the next company’s opportunity - repeated across other portfolio clusters. It surfaced a bridge between the encrypted compute thesis and some of my public market semiconductor positions. That emergent portfolio coherence was invisible until the tool traversed the full graph.

4/ Unsolicited expert network mapping. While researching the encrypted computation space, the system returned the obvious domain experts in my network. Then it added people I didn’t ask about. A professor whose papers were already in my research library, and an engineering lead my partner had met at a conference months earlier who had worked on exactly the problem space I was diving into. The AI mapped my network against my research question and found relevant nodes I had forgotten existed.

The Tool Unification Problem

Knowledge does not live in one app. My research is in Obsidian, Anytype and Gdrive. My relationships and dealflow are in a CRM. My venture portfolio positions sit in Carta. My public market positions in a tracker. The breakthrough here is MCP - Model Context Protocol - which lets the AI agent query across all of them in a single reasoning chain. Any tool with an MCP server becomes part of the second brain. A query can run across our CRM, board decks, founder conversation notes and the knowledge graph simultaneously. Instead of having ten tools I have one intelligence layer that reasons across ten tools. Context switching and knowledge synthesis is done in one pass.

Subscribe now

Why the Compounding Effect Matters

A traditional note-taking system has linear returns. Note number 300 is about as useful as note number 30. You still have to find it, read it, and connect it yourself. An AI-enhanced system has compounding returns. Every note you add increases the surface area for future connections. Note 300 can be cross-referenced against all 299 that came before it - semantically, and across domains you would never think to check.

This changes the economics of note-taking. The quality of notes still matters enormously (garbage in, garbage out) but the cross referencing of other notes became exponentially better. Write more, organize less. Capture your thinking in your own words, link where connections are obvious, and let the AI find the rest. In the old world you needed to “build a perfect system.” Now you only need to “feed a good-enough system consistently.”

What This Means for Knowledge Workers

What I’m trying to describe is not just a little productivity hack. It is a structural shift in how individuals can memorize reason at scale. Before long we will see the same patterns emerge on an organizational level too.

A venture investor can maintain deep, synthesized positions across dozens of companies and sectors without losing threads. A researcher can surface cross-disciplinary connections that would take months of literature review. A writer can draw on years of accumulated thinking without re-reading everything they have ever written. A founder can maintain institutional memory as the company scales past what fits in one person’s head.

The key design choices, for anyone building their own version:

• Pick a graph-native tool. Obsidian, Logseq, Anytype (portfolio), Roam - the specific tool matters less than its ability to represent links between ideas as first-class objects.

• Write atomic notes in your own words. One idea per note, in your voice. Copy-pasted highlights are raw material, not thinking. The AI needs your reasoning patterns, not someone else’s.

• Link aggressively. Every explicit link you create strengthens the graph the AI traverses. Over-linking is almost impossible.

• Point an AI agent at the result. Use MCP servers, vector search, or agentic workflows to turn the vault from a reference library into a reasoning partner.

The second brain was always a powerful idea I found deeply fascinating. The missing piece was the ability to think with it rather than just store static knowledge in it. That piece is here now, and the gap between people who build these systems and people who do not will compound faster than most realize.

Be aware that many of the technologies mentioned in this post are highly experimental. We set up our system with the support of seasoned engineers and best practices for security and data protection in mind.

How are you building and maintaining your second brain?

The Fragility of Orbit

LEO faces several risks, some more detrimental and more probable than others. Let’s explore them one by one.

Kessler Syndrome

In 1978, NASA scientist Donald Kessler described a feedback loop: collisions in orbit create debris, which creates more collisions, which creates more debris, until an orbital band becomes unusable. For decades this was treated as a distant concern. It is now an operational reality.

The numbers are past the threshold. A March 2025 analysis by Lewis and Kessler using the latest population data found that the current number of intact objects exceeds the runaway threshold at nearly all altitudes between 520 km and 1,000 km. Planned mega-constellation deployments will push even more altitude bands beyond stability limits.

The operational evidence. SpaceX’s Starlink constellation - 9,300+ satellites as of late 2025 - performed 300,000 collision avoidance maneuvers in 2025 alone. That is roughly 820 maneuvers per day. And each new satellite added to orbit increases collision probability with every existing satellite.

The CRASH Clock tells the story. Researchers developed a metric called the Conjunction Risk Assessment for Space Highways - which tells us how quickly satellites would start colliding if they lost the ability to avoid each other. In 2018, it stood at 164 days. By January 2026, it had collapsed to 3.8 days. The safety margin eroded by 97% in eight years. IEEE Spectrum’s analysis notes an important nuance: the CRASH Clock measures time to the first collision, not a full Kessler cascade. But a first collision is exactly how a cascade starts - and 3.8 days without active avoidance is the margin that stands between operational orbit and the beginning of a chain reaction.

ESA’s 2025 Space Environment Report tracks over 40,000 objects, with models estimating 1.2 million objects larger than 1 cm and 130 million larger than 1 mm. The 2009 collision between Iridium 33 and Cosmos 2251- a single event - created over 1,800 trackable debris fragments, many of which will remain in orbit for decades. The collision doubled the risk in the 700-800 km altitude band.

Now imagine this with 15,000-18,000 satellites in LEO by end of 2026, and China filing for a 200,000-satellite constellation.

Space War

Kessler Syndrome is an accident. Space war is a choice - and the arsenals are ready.

**Kinetic anti-satellite weapons (ASAT): Russia’s November 2021 test destroyed its own Cosmos 1408 satellite, generating 1,500+ trackable debris fragments at 480 km altitude - directly threatening the ISS and Starlink’s orbital band. China demonstrated kinetic ASAT capability in 2007 by destroying a weather satellite, creating debris still tracked today. The US, India, and likely others have operational capability.

Directed energy weapons: China has likely fielded ground-based lasers capable of dazzling or damaging low-orbit satellite sensors. Russia has deployed the Peresvet system with mobile ICBM units. The U.S. Space Force is deploying electronic satellite jammers and accelerating counterspace weapons development. Every major debris removal laser has inherent dual-use capability - the ablation mechanism works identically on debris and functional satellites.

Cyber warfare: The 2022 Viasat/KA-SAT attack at the start of Russia’s invasion of Ukraine disabled satellite broadband across Europe in hours - a single cyberattack that affected tens of thousands of terminals. The CSIS 2025 Space Threat Assessment reports that cyberattacks, jamming, and spoofing have “become commonplace and rarely trigger an escalatory or retaliatory response.” GPS jamming is now routine across the Baltic, Middle East, and South Asia.

Nuclear EMP: Russia is reportedly developing an orbital nuclear anti-satellite capability. A nuclear detonation in LEO would not just destroy nearby satellites - it would create an electromagnetic pulse and a persistent radiation belt (similar to the 1962 Starfish Prime test) that could degrade electronics on every satellite passing through the affected region for months or years.

The Secure World Foundation’s 2025 Global Counterspace Capabilities report grew from 148 pages in 2018 to 316 pages in 2025, tracking 12 nations with active counterspace programs. The domain is militarizing faster than governance can respond.

The cascading scenario: A kinetic ASAT strike during a Taiwan Strait conflict could deliberately trigger a Kessler cascade in specific orbital bands. Even a “limited” exchange - targeting 10-20 reconnaissance satellites - would generate thousands of debris fragments, each capable of destroying other satellites. The debris does not distinguish between military and civilian infrastructure. GPS, Starlink, weather satellites, and the ISS all share the same orbital neighborhood.

Solar Weather

The Sun is indifferent to human infrastructure planning.

In February 2022, a moderate geomagnetic storm during the rising phase of Solar Cycle 25 increased atmospheric drag in LEO, causing 38-40 newly launched Starlink satellites to lose altitude and burn up before reaching their operational orbit.

We are currently near the maximum of Solar Cycle 25. Research published in Frontiers in Astronomy and Space Sciences (2025) shows that reentry rates for LEO satellites have increased significantly during this solar maximum, with atmospheric drag becoming harder to predict as solar activity fluctuates.

The Carrington scenario: A Carrington-class event (the 1859 solar superstorm, the strongest recorded) directed at Earth would be categorically different from a moderate storm. Simulations suggest it could disable or degrade a significant fraction of LEO satellites through radiation damage to electronics, dramatically increased atmospheric drag, and disruption of ground control links. New simulations warn that a Carrington-level event could “wipe out all our satellites.”

The probability is not negligible. Solar physicists estimate a 1-12% probability per decade of a Carrington-class event. Over a 50-year infrastructure planning horizon, the cumulative probability becomes significant.

Economic and Regulatory Failure

LEO’s commons problem is economics.

The bankruptcy scenario: Starlink reached profitability in 2024 - barely, with $72.7 million net profit on $2.7 billion revenue. Project Kuiper has invested $10B+ and must launch 1,600 satellites by July 2026 just to retain FCC license rights. China’s constellations are state-backed. OneWeb went bankrupt once already. If a mega-constellation operator fails, thousands of uncontrolled satellites remain in orbit with no entity responsible for their disposal. At a 1-2% annual failure rate, even a healthy Starlink generates 20-40 derelict satellites per year that cannot self-deorbit.

No enforceable global space traffic management exists. ITU spectrum allocation operates first-come-first-served. China’s 200,000-satellite filing is a spectrum land-grab, not a real deployment plan. The FCC regulates US operators but has no jurisdiction over Chinese or Russian constellations. Active debris removal technology exists (ClearSpace, Astroscale) but remains entirely government-funded. There is no commercial market because there is no mandatory liability framework. A proposed EU Space Act (June 2025) covers only EU-licensed operators.

Insurance barely covers the risk. Only about 6% of satellites carry in-orbit insurance. The space insurance market cannot price catastrophic correlated events (Kessler cascade, solar superstorm) because they would simultaneously affect all insured assets. This is not insurable risk. It is too systemic. A PNAS analysis found that a harmonized orbital-use fee of ~$235,000 per satellite-year would correct incentives and increase long-run industry value from ~$600B to ~$3T. No such mechanism exists.

The Correlation Problem

This is the factor that transforms LEO vulnerability from a sectoral concern into a civilizational one. All LEO-dependent services share the same physical environment. A debris cascade, solar storm, or ASAT campaign does not selectively degrade broadband while leaving GPS intact. It does not spare weather satellites while destroying ISR platforms. The failure mode is correlated across every service category.

A modern economy losing GPS simultaneously with satellite broadband and weather forecasting faces compounding failures that no sector-specific contingency plan addresses. Financial markets lose timing synchronization. Aviation loses precision approach capability. Precision agriculture loses guidance. Military forces lose ISR and secure communications simultaneously. Maritime shipping loses AIS and navigation.

LEO is a single point of failure for civilization-scale infrastructure. No equivalent single-point-of-failure exists in terrestrial infrastructure - no one earthquake can simultaneously disable the internet, GPS, weather forecasting, and military communications globally.

The Stratospheric Alternative

At 20 kilometers altitude, above weather and conventional air traffic but far below the orbital debris field, there exists a layer of the atmosphere that is largely empty, physically benign, and operationally proven.

The stratosphere is a hedge against LEO’s correlated failure modes - and for several applications, it is the superior solution regardless.

The Physics Case

The comparison between a satellite at 500 km and a stratospheric platform at 20 km comes down to distance - and the physics of distance are unforgiving.

Signal strength: Free Space Path Loss (FSPL) at 10 GHz from 20 km altitude is 138 dB; from 500 km it is 166 dB. The 28 dB difference sounds small until you remember that decibels are logarithmic- every 10 dB represents a 10x change in power. So 28 dB = 10^(28/10) = ~630x more signal power reaching a ground receiver from a HAPS than from a LEO satellite. A stratospheric platform talking to a standard 4G handset has physics working in its favour by nearly three orders of magnitude.

Latency: Round-trip latency from 20 km is approximately 0.13 milliseconds. From 500 km LEO orbit, it is 3.33 milliseconds. From GEO, it is 600+ milliseconds.

Imaging resolution: The Rayleigh criterion for diffraction-limited optics means that achieving 10 cm ground resolution from 500 km requires a 3.05 meter mirror. From 20 km, the same resolution needs a 12.2 cm lens. 25x times smaller. This is why satellites top out around 30 cm commercial resolution while HAPS can achieve sub-10 cm with modest optics.

Direct-to-hand-held-device connectivity: The April 2025 HAPS Alliance whitepaper establishes that HAPS are compatible with existing 4G/5G handsets without modification. Satellite direct-to-device (Starlink, T-Mobile partnership) requires specialized protocols and delivers a fraction of the throughput. In early 2025, Aalto’s Zephyr demonstrated the first wireless connection from a fixed-wing HAPS at 60,000 feet to a standard 4G mobile device on the ground.

Technology Maturity

In April 2025, Airbus subsidiary AALTO’s Zephyr completed 67 days, 6 hours, and 52 minutes of continuous stratospheric flight - a world record. Solar-powered, battery-sustained through the night, operating at 60,000+ feet. It is an operational endurance demonstration exceeding most military deployment cycles.

The HAPS ecosystem in 2025-2026:

Radical (Seattle, US): A fixed wing form factor; solar powered; flew their full systems maiden flight in autumn 20254; (an Inflection portfolio company)

Aalto/Zephyr: 67-day record. Commercial entry-into-service in Japan planned for 2026, backed by $100M consortium (NTT DOCOMO, Space Compass, Mizuho Bank, DBJ)

Kea Aerospace (New Zealand): Historic first stratospheric flight, February 8, 2025. Atmos Mk1b reached 56,284 feet, flew 8 hours 20 minutes

Sceye (US): NASA/USGS partnership for methane monitoring. SoftBank investment. Pre-commercial HAPS services in Japan 2026

BAE Systems PHASA-35: Presented at DSEI 2025 as a defense product for persistent ISR and communications. Months-long endurance without landing

India DRDO AS-HAPS: First flight trials May 2025 at 17 km altitude. Defence Acquisition Council approved

Russia Barrage-1:Tested February 2026: - a stratospheric aerostat developed explicitly as a Starlink alternative after SpaceX restricted access to Russian military user. The proliferation is telling. When Russia builds a battlefield stratospheric relay within weeks of losing Starlink access, the dependency - and the remediation speed - are clear.

The Resilience Case

When LEO degrades - through debris cascade, ASAT strike, or solar storm - the stratosphere remains intact. A HAPS at 20km sits below the orbital debris field, is shielded from radiation by the atmosphere, presents no viable ASAT target, and can be replaced in days rather than years. It is an independent infrastructure layer whose failure modes are uncorrelated with orbit.

The applications that matter most in a LEO-denial scenario are precisely where HAPS excels:

For connectivity, a single platform covers a 50km+ radius with direct-to-device service to standard handsets, deployable from a portable ground station in hours - Japan’s SoftBank/Sceye program is explicitly designed for earthquake disaster recovery, and Taiwan demonstrated 14+ days of balloon-based emergency comms in 2025.

For navigation, a HAPS pseudolite at 20km delivers a signal roughly one million times stronger than GPS at 20,200km, making it much harder to jam while providing sub-meter accuracy over sovereign territory that can be repositioned in hours.

For ISR and earth observation, sub-10cm resolution with indefinite dwell time and real-time video replaces the satellite model of daily revisits at 30cm with minutes of dwell per pass - the difference between a photograph and continuous surveillance. Defense analysts are converging on a three-tier aerial architecture - aircraft, stratospheric HAPS, and space - treating the stratosphere as an existential building block.

Building the Mitigation Stack

The five failure modes described above are not inevitable. They are engineering problems. Three companies in Inflection’s portfolio are building specific capabilities to extend LEO’s operational life and mitigate its risks.

If the CRASH Clock reads 3.8 days, the first question is: can we track what we need to avoid? Foundational builds the precision tracking layer for space situational awareness using Space Laser Ranging (SLR). Where radar tracking provides meter-level accuracy, SLR provides sub-centimeter-level precision - 1,000x more accurate. The difference matters when you are trying to predict whether two objects at 7.5 km/s will miss each other by 10 meters or collide.

Tracking threats is necessary but not sufficient. Something must physically intervene. Lodestar is building autonomous satellite defense capabilities in contested space environments. Their system provides onboard decision-making - using game theory to predict adversary intent, detect threats, and execute orbital responses without ground control.

Radical is building autonomous solar-powered aircraft that fly perpetually in the stratosphere at 20km altitude. An aircraft can deliver what a constellation of thousands of satellites struggles to match. Two verticals: (1) connectivity - beaming 5G-speed internet directly to handheld devices over a 50km+ radius, targeting 3 billion underserved people, and (2) persistent imagery - real-time sub-10cm resolution earth observation including live video, physically impossible from orbit due to lens size and revisit constraints.

Subscribe now

If you’re building related solution to mitigate infrastructure fall outs due to LEO failure, please get in touch.

As SemiAnalysis put it:

“Claude Code is the inflection point for AI Agents and is a glimpse into the future of how AI will function.” - Semi Analysis, Claude Code is the Inflection Point

We agree. 4% of all GitHub public commits are now authored by Claude Code. At the current trajectory, that number reaches 20%+ by year-end. Something fundamental is shifting - and coding is just the beachhead. What we are experiencing is a canary in the coal mine for a profound reorganization of the software stack. The boundary between application and operating system is dissolving. The interface between human and machine is collapsing from visual to verbal. And the implications extend far beyond software development into every information business.

The Old Era

The technology stack that dominated computing for four decades has four layers.

For forty years, humans adapted to this stack. We learned keyboard shortcuts. We memorized menu hierarchies. We followed workflows that developers prescribed for us with marginal customization opportunity. We created accounts, accepted terms of service, and organized our work around the constraints of each application.

Value concentrated at the application layer because building complex software was expensive and switching costs were high. Once you stored your leads in Salesforce, your documents in Google Drive, your creative assets in Adobe, and your messages in Slack - that data became hard to move. Software companies optimized for this lock-in. It produced a generation of businesses trading at premium multiples on recurring revenue and feature moats.

The human brain was the only integration layer. We were the ones copying data between applications, context-switching between fifteen tabs, manually synthesizing information scattered across silos. The applications didn’t talk to each other. We did the talking for them.

The New Era

The stack is collapsing from four layers to three.

Hardware remains at the base - unchanged, still the physics constraint. Operating system still manages resources but now includes the model runtime. AI and interface merge into a single layer. The AI is simultaneously the reasoning engine and the interface. Natural language in, generative output out.

Applications vanish as a standalone layer. Their logic gets absorbed into the AI layer. Lead scoring, image editing, financial modeling, document formatting, data visualization - all of this becomes commodity inference. The AI doesn’t “use” applications. It replaces the need for them by synthesizing directly from data.

The output is format-agnostic. The same reasoning engine can produce a chart, a piece of code, an audio summary, a video, a formatted document, or a fully functional piece of software. Applications were containers - standardized, rigid, same for everyone. The AI layer is a synthesizer. Think of the difference between IKEA furniture and having a carpenter in your house who builds exactly what you need from raw materials, on demand.

This is what we experience daily. When we ask Claude Code to analyze a dataset, it doesn’t open Excel. It writes a script, runs the analysis, generates the visualization, and returns the result. When we need a report, it doesn’t open Google Docs. It reads the source material, reasons about it, and produces the output in whatever format we need. The application layer is absent.

What it means in the big picture of info tech

Zoom out far enough and you see this shift as the fourth inflection point in how humans handle information.

Each leap solved the previous era’s bottleneck but created a new one. Printing solved storage. The internet solved distribution but drowned us in noise. AI solves synthesis. But has several bottlenecks. The most imminent is context: to reason, machines need access to everything.

Current application design is fundamentally at odds with this requirement. Information is organized in vertical silos by function. Google Drive stores documents. Superhuman handles email. Signal carries messages. Calendar manages time. Attio tracks business relationships. Substack publishes writing. Each application is a walled garden with its own login, its own data model, its own way of holding your information hostage. A human can context-switch between fifteen apps and mentally stitch the picture together. An AI needs unified access to reason across all of them.

The most obvious quick fix is connecting these silos through machine-readable protocols. APIs provide standardized doors into each silo - they’ve existed for decades but adoption remains uneven. MCPs (Model Context Protocols) are a newer standard that lets AI models connect to external tools and data sources through a universal adapter format. Together, they are the duct tape holding the transition together. They let AI reach into existing silos without tearing them down.

This works. For now.

Subscribe now

What It Means for Software Businesses

The instinct of every software incumbent is to bolt AI onto the existing product. Preserve the interface, preserve the subscription, preserve the revenue model. Adobe added Firefly inside Photoshop. Figma embedded AI design assistants into its canvas. Microsoft launched Copilot across Office 365. Salesforce built Agentforce on top of its CRM.

If the interface collapses into language and output is generated directly by AI, adding AI features to your interface doesn’t save the interface. It’s like adding a GPS to a horse-drawn carriage after cars arrive. It won’t save software as we know it.

Not every software company is equally exposed. The critical variable is whether the moat exists beyond software layer. Some examples: Uber looks like a software company but is actually a physical marketplace - 5 million drivers, regulatory licenses in 10,000+ cities, real-time liquidity density that took a decade and billions in subsidies to build. The app is the thinnest layer. Making software free makes Uber cheaper to operate, not easier to disrupt. Wolters Kluwer sells regulatory truth with legal standing. AI makes their data more queryable - it doesn’t replace it. Google’s moats sit below the software layer in cloud infrastructure, search index, and proprietary data assets. Salesforce is the exception that proves the rule - it’s moving down the stack with MuleSoft (API infrastructure), Data Cloud (derived data assets), and Agentforce (governed gateway for AI agents accessing enterprise data).

Public markets are pricing this in - aggressively.

The paradox is striking. Earnings are growing. Margins are stable or expanding. Yet stocks are cratering. Markets are repricing the terminal value of software moats. Current earnings are fine - but future earnings are worth less because the moat is dissolving.

Are markets overshooting? Possibly, in some cases. AI adoption won’t happen overnight. In regulated industries - healthcare, accounting, law - institutional inertia is enormous. Employees need retraining. Compliance workflows are sticky. Liability frameworks haven’t adapted. Legacy tech is protected by friction, not innovation. This buys time but the terminal direction is clear - the life cycle ends.

In private markets the repricing hasn’t even started. Most VCs have significant portfolio exposure to enterprise SaaS. The existential threat to interface-centric software isn’t reflected in book values. Adding AI features on top of a traditional SaaS product is cosmetic - it doesn’t address the structural shift. Most VCs didn’t hear the shot. Their LPs are looking at over-bloated book values that will quietly collapse over the next few years as markdowns catch up with reality.

What’s Next

APIs and MCPs are transitional. They solve the access problem by breaking open data silos, but they create a new one: when one AI provider connects to your email, files, CRM, calendar, and messages, you’ve built a single point of failure. Compromise one integration layer and the attacker gets everything. It’s a security nightmare.

The long-term stack needs to be fundamentally different. Some developments we follow closely: Fully homomorphic encryption allows computation on data that stays encrypted throughout - the AI reasons without ever seeing the plaintext. Zero-knowledge proofs let you verify a computation was done correctly without revealing the inputs. Apple’s Private Cloud Compute extends on-device security into the cloud through stateless, zero-trust enclaves where data is processed and immediately deleted. Local-first architectures keep data on your device and sync peer-to-peer, eliminating the central server entirely. On-device models handle sensitive queries without data ever leaving your hardware. None of these are mature yet but they are laying the plumbing of the next era. Plumbing is harder to replace than applications.

Inflection has barely any SaaS exposure and concentrated on companies that address critical bottlenecks of the emerging stack. We invest thematically and focus rigidly on where the new architecture breaks down. Some examples: Trust and verifiability: [[Anytype]] (local-first, encrypted collaboration where users own their data), Fabric (custom silicon accelerating encrypted AI workloads), Ubitium (reprogrammable silicon for edge AI workloads). Data bottlenecks: Deep Earth (subsurface mapping / geospacial AI). Connectivity resilience: Hedy (an alternative networking stack that makes connected devices invisible and provides failover in contested environments. Physical-world integration: Ark (robotic fleet control - vertically integrated yet modular), NAD (drone interception network - vertically integrated yet modular), Levtek (modular industrial robotics), Stealth (positioning data through space laser ranging - physical sensor networks addressing data bottlenecks). None of these are pure software plays. None can be replicated by an LLM with API access. When software production costs go to zero, these companies become more valuable - they are anti-fragile in this era.

If you’re building at the bottlenecks of the new stack, we’d like to talk. And if you disagree with any of the above - we’d like to hear that too.

Subscribe now

Along the way, I’ve noticed a recurring pattern: most engineers joining VC come from the outside, and there’s no real onboarding guide for the role.

They’re talented people who haven’t worked in venture capital before, don’t yet know how funds operate, and often spend their early months figuring out the basics: what data matters, what tools exist, where to focus first. I was one of them back in the day.

So I wrote the guide I wish I’d had: Building for Venture Capital.

Subscribe now

What It Covers

The guide is organized into three parts:

Part 1: Understanding VC

How venture capital funds actually work, from the perspective of someone building technology for them. Fund structures, LP relationships, investment processes, and the common mistakes technical hires make when they don’t understand the domain.

Part 2: The VC Tech Stack

A tour of the tools and systems VC funds use: research platforms, sourcing tools, CRM and deal flow, fund operations, portfolio support, fundraising, and external presence. For each category, I cover what the tool does, when to build vs. buy, and real examples from working funds.

Part 3: Technical Foundations

The deeper technical work: choosing your technology stack, working with data providers, data modeling for VC, entity resolution, data quality, warehousing, integrations, security and compliance, and emerging trends like MCP and AI agent orchestration.

There’s also a resources section with curated newsletters, books, research papers, and a prompt bank for common VC workflows.

Why Open Source?

At Inflection, we believe the VC industry benefits when funds share knowledge about how technology is shaping the way we operate. Too often, this expertise stays siloed within individual firms. By publishing this guide, we want to create more conversation about how new technologies are changing what's possible for VC funds of all sizes.

The whole thing is on GitHub. If something’s wrong or missing, you can open an issue or submit a PR. If you’ve got prompts, resources, or perspectives that would help others, contributions are more than welcome.

Acknowledgments

This wouldn’t exist without the people I’ve worked with at Inflection and EQT who shaped how I think about this work. Special thanks to the external contributors who shared their perspectives.

Check it out at buildingfor.vc. Feedback welcome.

// AP

Subscribe now

The ocean of incremental sameness

The venture industry sits at the intersection of technology, politics and culture. Over the last few decades all of those areas have become more homogenous, more compliant, more complacent and more boring. Let’s go through some examples.

Music used to evolve alongside other cultural phenomena in a specific, local context and Zeitgeist. It polarised between old and young. It needed to overcome repression by establishment institutions before it could spread into the world.

From RUN DMC’s Raising Hell Album (1986) by Bob Gendron:

Few albums change the world. Raising Hell did so in the face of bitter resistance and prejudice. (…) Still, nothing and no one could stem the tide unleashed by Raising Hell. Not the regressive Parents Resource Music Center whose co-founder maintained that “angry, disillusioned, unloved kids unite behind heavy metal and rap music, and the music says it’s OK to beat people”; not irresponsible city mayors who threatened to ban Run-DMC concerts; not pundits who sought to preserve the status quo by falling back on the type of ignorant contentions that nearly muted Rock ‘n’ Roll in the late 1950s. Raising Hell disrupted tradition, dared listeners to reconsider what they thought they knew and provided a platform for art, sound and discourse that lent power, strength and identity to marginalised voices.

When was the last time you listed to fundamentally new, radically different music?

Contrast this to the emergence of hyper-popular artists and types of music who master the industry and fan-interactions to an unprecedented level of perfection. Who doesn’t like Taylor Swift?

Furniture and interior design seem to have converged immensely. No matter if you book an AirBNB in Berlin, Cape Town or San Francisco, chances are very high that you find the same plants, design furniture, exposed walls and minimalist art decoration in all of them.

The post–Cold War period saw a notable institutional convergence in political systems. Several countries adopted multi‑party elections, similar constitutions, independent central banks, regulatory agencies, and market‑oriented policies. The European Union pushed for regional integration. China innovated around an authoritarian-capitalist system and more recently the US started to resemble such a system increasingly: both rely on powerful executives - China through plans and directives, the US through executive orders. Both moved from laissez-faire towards state-shaping of strategic sectors. Both are fusing national security with their economies through an interlock. Europe is years behind but started the fusion process as well.

Apparently, technology startups and the venture ecosystem around them are not immune to the trend of convergence as we will explore more in the next section.

Why?

What is driving all this? Complex systems and (single factor) causalities don’t match well. Therefore, I won’t make any bold claims on what the definitive drivers are. Rather I share some speculations on what might be contributing ones.

1/ Incentives: through the internet we learned to track consumer preferences and responsiveness to any kind of content. Music, design, movies, political views. In fact consumer preferences can be predicted and manipulated at scale. This is flipping the script for creators. Instead of taking high conviction bets on creating something from scratch and iterating over it with a local community they can just listen to their audience. In order to boost their reach and distribution they rather create something incremental that is consumed by the masses than something genuine that is consumed by a tiny minority but holds the potential to set a trend. Faster horses instead of cars.

2/ Mimetics: Mimetics studies how ideas, behaviours, and aspirations spread by imitation. People copy what others signal as desirable. In markets, this copying creates clusters of shared beliefs about what “good” looks like. When many chase the same signals of success, competition narrows around a few accepted patterns. Rivalry intensifies and differentiation collapses. There are fewer outliers because deviation carries social and professional risk. Mimetics are massively amplified by social media and global connectivity. While memes have newfound potential to reach a wide audience much faster than in the past, they are also much more short-lived. We get bored faster, or get used to new ideas faster, to phrase it positively.

3/ AI training and governance: AI model training is an increasingly homogenous process where incrementally different LLMs are fed with mostly the same data. The selected input data is subject to selection bias in terms of religious beliefs, political views, cultural norms and legislation amongst others. This type of monopoly control over the mind is amplified by the terms of service applied by the big tech platforms deploying said models. My friend Erik Voorhees had a few things to say about this in his incredible piece The Separation of Mind and State. Here is the introduction:

Hundreds of years ago, through tremendous sacrifice, the institution of religion was incrementally removed from the government’s sphere of authority. Today, religious or not, we recognize the importance of that separation of church and state.The original cypherpunks sought to separate language from state through encryption—for a while declared a “munition,” where certain types of math were banned from export. Their descendants, the early Bitcoin pioneers, sought to similarly separate money and state.But if monopoly control over god or language or money should be granted to no one, then at the dawn of powerful machine intelligence, we should ask ourselves, what of monopoly control over mind?

This is not something AI researchers aren’t aware of, so we see people working on custom data sets, expert-curation, and alternative data types in order to make models spikier, or even just have more common sense.

What happens when the venture ecosystem loses its diversity and becomes one with the vast ocean of incremental sameness?

Subscribe now

The rise of consensus capital and the venture industrial complex

The characteristics of venture capital changed profoundly over the last few decades:

1/ Fund sizes ballooned, capital concentrated: In the early days, fund sizes used to be around $100M with return expectations of 5x and more for the top performing ones. Small funds took a high number of absurdly ambitious moon shot bets with extreme variance in outcomes. Today, large, multi stage funds are dominating while emerging managers struggle to survive. In 2024, a16z raised $7.2 billion—over 11% of all US VC. The top 30 firms captured 75% of the venture market. With large fund sizes come different expectations of return predictability and certainty. Venture became inherently risk averse.

2/ State-economy fusion: with billion dollar fund sizes the game changed. Instead of funding high risk, high reward experiments, the venture industry started to merge with the state by getting increasingly involved in lobbying. This might have started with a16z’s initiatives around operation choke point, a coordinated attempt of the US government to shut down the crypto industry without legislation by leveraging de-banking and de-platforming practices. It is ironic that venture’s push back against executive over-reach to protect an inherently libertarian industry spilled over into a full blown fusion of private venture with government.

In “Honey, We Need To Talk About Venture Capital“ captures the essence of the above developments with more nuance and countless examples. Her conclusion for the role of what she describes as little venture is the following:

If you’re a venture fund that isn’t doing any of this (no policy alignment, no industrial positioning, no narrative or regulatory leverage), then you are not in the same business as Big Venture. You’re in venture capital, the original game of chasing founders, markets, and luck. Said another way, you’re in the nostalgia business. While Big Venture manufactures outcomes, you’re literally just buying lottery tickets, and getting paid 2% a year to do that. (…) Big Venture shapes the game and collects steady wins. Venture capital thinks it’s finding alpha, but really it’s just getting played.

In a similar vain my friend and colleague wrote in his piece Consensus Capital:

Being right in isolation, insofar as it ever was actually true, is a dead strategy. If you have an insight but cannot coordinate follow-on capital, you lose. If the state floods your sector with subsidies favouring different players, you lose. If megafunds deploy $500 million rounds and dilute your position to irrelevance, you lose. The era of the lone wolf VC who finds the overlooked founder and waits for the world to catch up has largely passed. (…) The best VC today is not the one who sees the future first. Or writes the best blog. It is the one who convenes the consortium that funds it. The future that is, not the blog.

The dark side of consensus investing

I share many of the observations made in the above articles but I disagree with the conclusions for several reasons. Consensus investing is unlikely to be the dominant, alpha-generating strategy in venture over extended time horizons.

1/ Central planning has a history of failure: despite a massive deployment of resources, centrally planned technologies historically rarely worked. In Why Greatness cannot be planned by Kenneth O. Stanley and Joel Lehman there are countless examples explaining why the world’s most significant discoveries and innovations arose from serendipity, exploration and open ended searches, rather than from pursuing pre-defined goals. E.g. novelty search in robotics: Agents learn to move or “walk” better when optimized for behavioral novelty rather than a direct “walk farther/faster” objective. By rewarding new, interesting behaviours, they accumulate the prerequisites to walk. Top down objectives can be a false compass.

2/ Financial Bubbles: When large funds must deploy vast sums, they herd into consensus categories and companies. Prices are pushed above fundamentals. Escalating valuations driven by cross over funding rounds look like guaranteed demand. They are not. Public-market discipline eventually re-prices growth stories at some point. Looking at Palantir’s 600x PE ratio that might not be tomorrow though. Extreme concentration isn’t limited to private markets as late cycle dynamics favour big techs structurally.

3/ Policy capture is brittle: Aligning with procurement and subsidies can raise certainty but it also concentrates political risk and single buyer exposure. Trump Jr.’s Vulcan Elements and associated 1789 fund portfolio companies won several $600M+ government contracts. This looks like an unfair advantage (in the literal sense, looking up the definition of “corruption”). However, things can change through election reversals, program cancellations, government shut downs and shifting security doctrines. Long term capture in frontier tech seems easier said than done long term.

4/ Founder incentives: As soon as you can put a label on it the alpha is gone. With labels and established categories comes fierce competition. Talent wars break out. Margins collapse. Such an environment is not an ideal place for breakthrough innovation. As my colleague David Peterson was writing in Startup Stagnation:

The advantage of an overcrowded consensus is simple: the field is wide open for things that are actually different. As consensus ideas soak up attention and capital, the founders building in weird markets have zero competition. They have time to be illegible, to iterate without the spotlight and to build credible monopolies. The cover bands will keep playing the hits. But somewhere, quietly, the new music is being written.

This anecdotal wisdom is supported by some data. A recent piece of MIT Research found that Disagreement Predicts Startup Success: Evidence from Venture Competitions. The paper finds that higher dispersion in judges’ scores strongly correlates with better outcomes—more funding, higher revenues, and greater exit likelihood. The mechanism is that unique and hard-to-evaluate ideas naturally spark disagreement; common opinions rarely confer advantage. The big caveat of this research is a selection bias as the underlying data has been gathered only between 2011-2020. I would love to see someone continue this research over extended time periods and throughout the era of consensus capital.

Where the heretics go

If vast parts of the market are soaked up by consensus capital - where do heretics go?

Now is a good time to define what I mean by heretics in this context. I would describe them as people who prioritise first-principles over playbooks, accept delayed validation and higher variance to unlock non-obvious breakthroughs.

1/ Scarcity bottlenecks: The consensus herd typically looks at a problem and backs obvious solutions. AI training and inference needs more compute, so they back data centers and accompanying infrastructure. Heretics think 2, 3 steps ahead by anticipating shifts in future scarcity bottlenecks.

In an interview with Floodgate Founder put it this way:

(…) maybe genuine heresy requires all three. You need the perceptual difference to notice something meaningful, the relevant experience to know when that perception matters, and the courage to act on it. Which might explain why heresy is so rare. (…) Every major technology shift creates a new scarcity, and that scarcity is never the enabling technology itself. When computation got cheap, software became scarce, because it was the layer that made computers useful. When communication got cheap, organizing layers that connected people to content, commerce and community became scarce. Now AI is making cognition abundant.

What might be the next scarcity bottlenecks?

In high stakes, high friction fields (nuclear reactor design; drug molecule design; autonomous robotics) cheap cognition amplifies the risk of failure. We need to be able to verify and trust the results put out by machines. Provenance and auditability (logs, reproducibility) will play important roles in those areas and very few people are working on it. Our portfolios Fabric and Flashbots are some of them.

AI’s next frontier is energy. Physical delivery energy markets are the achilles heel of AI scalability going forward. They are opaque, inefficient and unreliable. Fixing them is a vast, structural challenge some heretics are working on already.

2/ unregulated domains: Regulatory capture is only feasible in areas where legislation is in place. Heretics go where there is no regulation in place that could be captured by the venture industrial complex. The space domain comes to mind. Space security services as provided by our portfolio Lodestar or stratospheric satellites for remote sensing, connectivity and imagery built by our portfolio Radical come to mind. Ocean autonomy, deep sea mining or geo engineering fall into the same category. Network states might be an alternative to over-regulation and create a space for innovation.

3/ short consensus: bold heretics might consider to short consensus narratives and ventures. Are we really expecting vibe coding startups to build sustainable moats against established dev tooling and infra companies? Do we really believe that Perplexity can win against Google’s full stack infrastructure moats and integrated services landscape? Do we really think that AI native marketing and customer care products can win against Salesforce’s agent platform? Besides buying some deep out of the money put options on some trends, heretics might consider to use prediction markets as a tool of choice to short private market consensus in some areas. This is an area we started exploring outside of our venture funds.

4/ other weirdness: we still believe there will be undiscovered talent (e.g., consensus today is 20-something “based” drop-out with fierce velocity, but who would back a biology high-school teacher?) and that there will be ideas that don’t work with incentive structures of big venture firms. Which venture fund would back a company in an industry that historically distributes a large chunk of their earnings as dividends? Common beliefs need to be questioned but people are usually not promoted if they bring in 19 strange ideas underpinned by non-common beliefs. Studies on peer review and grants in the scientific community show that the truly novel, orthogonal ideas (like mRNA therapies) rarely make it to publication or receive awards. We need more improbable bets with a high variance in outcomes. Some of them will be massive.

Subscribe now

At Inflection, we continue to back companies at the heretic frontier. If you’re in our camp and build something weird - reach out.

Why won’t you fall out of a rollercoaster?
Well, you are usually strapped in.

And why not, more reasons?
I guess if you’re going really fast in a loop then you’re pressed against the seat anyway by the centripetal force and won’t fall down.

But how can it hold you?
Hm, the seats are really durable and they are held by really strong metal bars that also support the wagon.

Is it harder than teeth?
Uhhh, I think so? In some ways…

And what about dinosaur teeth?
Good question…

Because of the nature of our work, I’m usually on the kid’s side of this dialogue with founders or researchers. (Hopefully with slightly less non sequiturs involving dinosaur teeth.) We try to understand the tech as quickly as possible, have good conversations about the business at hand, follow our own curiosity and the specific complexities of the area. Ideally, we have a prepared mind and can skip the first couple of layers of questions and dive into the more interesting stuff faster. In some cases we’ve written about the topic publicly before, or we’re close to companies who have indicated where the interesting problems of that space exist. But there comes a time when the response amounts to “Good question…” in some more or less strategically phrased way.

Why does the depth matter?

(now it’s nested list time, get ready)

We care about how deep we can go because it shows us something about the founder's way of thinking. To give some examples, it shows us:

a. How much the other person thought about their business, which is a proxy for:
i. How obsessed are they?
ii. Are their level of curiosity greater than ours?
iii. Are they stuck in a particular way of thinking or can they change perspectives?

b. What the founder’s angle-of-ideas (see illustration below) for the future of the company looks like, which shows:
i. Priorities and certainty in predictions
ii. What dimensions of the plan are fixed vs. flexible?
iii. Understanding of getting big, ambitious projects done

c. How the person deals with not knowing, this can cause:
i. Avoidance, misdirection, conversational “tactics”
ii. Nervousness and awkward silence (side note: it doesn’t make us uncomfortable)
iii. Making stuff up, acting like you know when you don’t
iv. (this is the worst) Actually thinking they know when they don’t
v. Acceptance of lack of understanding as a common part of going deep into something

Bad reasons and people who use them

The most frustrating situations we face aren’t the ones where the people we talk to simply don’t know and we wouldn’t expect them to know more, or even where people make stuff up (doesn’t happen so often). Many business conversations end up becoming rhetorical battles. Some companies try to fight it systematically (see Bridgewater’s real-time meeting scoring system, and Amazon’s memo practice), but before there’s much of a company we need to go for the individuals shaping the culture. If someone is overusing techniques from “How to make friends and influence people” in early conversations, then we have a hard time believing they’ll hire the right people and instill them with a culture that cares about progress over politics.

The counter argument to this is someone who seems overly sales-y to us, might actually be great sales people in certain settings. Since both hiring and fundraising are sales games, besides selling the product, it might be the right person for that company. But authenticity needs to be present.

My absolute worst pet peeve is the “you don’t need to know that, and so I didn’t look into it”-style response, unless it comes with very good reasoning for why we’re asking the wrong question (it happens, see dinosaur teeth). Arrogance isn’t necessarily bad in a founder; to some degree you need to believe in yourself to an extraordinary amount to start a company; but if it’s paired with hiding insecurities and going on the defensive when faced with some limit of their own knowledge, then that rather shows a lack of open-mindedness and curiosity. The feeling we get is “trust me, bro” and when we don’t trust, the counter is “Oh, so you don’t have conviction”. Or inversely, we hear of other investors committing on very short timelines, and the founders saying “They had very high conviction”. But on what exactly?

Conviction capital

To us, having high conviction doesn’t mean that we think the company has a particularly high probability of success, conviction means that we’ve done the work to understand what we are underwriting, and importantly what we’re not. Conviction is multi-dimensional, at the earliest stages, it’s 80%+ dimensions in the founder traits, but what might not be obvious is that every aspect of the company is a reflection of the founder. The founder is the soil and the company is a flower whose petals and roots we want to observe to understand what brought such a thing to life.

We need to be the truth-seeking biologists and horticulturists, not florists; we need to be Georg Mendel taking careful note of cause and effect rather than a superficial observer. I’m saying this on an individual company-level, as well as a market level. Only by studying something carefully can we know it. This assumes that we believe in something being knowable, of course. The obvious conclusion to seeing investing as a knowledge-enhancing activity is that in order to do great investments, we need to operate on the fringes of what is known. Some things should be more certain, like the physics governing the fundamental technology, or that someone will want to buy it, should it work (aka. “Big if true”). Other things can be more uncertain. What we don’t want, are the companies moving entirely within the scope of human knowledge—the playbook companies (note: my own em dash). We are not interested because those companies will be funded by other pots of money, so our dollars are not really value-adding.

We build conviction around a founder, and their ability to be the best at building exactly that company they want to build. This conviction-building requires us to understand everything around the company possible, up to some level where the uncertainties on the predictions become too vague to have an impact. This includes technology, market dynamics, other players, funding landscape, etc. It also includes personal judgement on the founder as a person, spending time with them, getting to know them over calls, walks, coffees, and awkward silences.

These are all variables in the conviction equation.

Taste is quite clearly expressed in how we communicate, including visual, textual and verbal cues. We sense it subconsciously - in a pitch deck’s title page, or a company’s landing page. How do you explain what your company is? Do the details matter to you (font choice, I’m looking at you), what imagery do you use? For the stage where we invest, we don’t think you need to have a designer put together an advanced corporate identity, that would be putting the horse before the cart. But everything is a choice that tells us something. Black on white vs. white on black. Attention to detail or fast and loose. Scientific versus corporate versus artsy. Bold or boring.

How you express the identity of your company informs who you attract, as employees, investors, and customers, and is a reflection of the thing that unites all the people working towards your company's future. There is a clear dilution as companies get bigger and get more responsibilities towards shareholders. From attracting alien talent and smart capital in the early years, many identities dilute just to not offend anyone. But if you are like us, you are likely in a position where you would rather attract the top 1% of your target demographic and repel the rest.

We created Inflection in 2019 to back founders others wouldn’t. The urge to create such a firm felt overwhelming at a time when most venture dollars chased the next SaaS or market place business. Powerful ideas and alien founding teams were overlooked, especially by European venture investors. Our ambition was to reclaim venture’s origins by backing engineering moonshots that matter. We wanted to create a firm that is rooted in craft and character, not another soulless startup factory chasing AUM.

Subscribe now

Looking back

Raising a fund from scratch is the hardest thing I’ve done in my life thus far. Being hit by a global pandemic after getting the first commitments 9 months into the journey didn’t help. We had no resources whatsoever, hence I drafted the first websites myself. Over time we got some serious support - shout out to Futur Drei (v1), Herd, Maksim and Obys (v2, 3).

In hindsight our previous iterations feel very zeitgeisty. We wanted to stand out as a firm doing things differently. Bold, alien shapes and themes rooted in the navigation of new frontiers. Gradient colours to represented a continuous process of transition into the new. This was great for some years but as soon as we started getting too many compliments for the design (especially from LPs) and others started copying the look and feel (you know who you are) we felt motivated to move on.

We wanted to focus more on the craft and timelessness of our business. Getting closer to the essence of who we are and what we do. Something that felt more connected to humanity’s past and future.

Subscribe now

Looking forward

We took a few steps back and got to first principles. A shout out to Lundgren Lindqvist, a Swedish design agency that helped us understand ourselves better and communicate it to the outside world.

Absorbing the patterns of progress lies at the core of what we do. Breakthrough innovations aren’t developed in a silo. They are closely tied to a network of previous inventions. Studying those is a necessity for getting a sense of where we are and what might be possible. It is a study of history in a way. Our approach is to spot inflections, deep but not always obvious shifts in technology, geo politics or culture that enable a change of behaviour at a large scale. I’ve been writing about it in depth here.

The farther backward you can look, the farther forward you are likely to see.

Winston Churchill

We also were in search for something that associates us closer with the divine act of creation. Building something from nothing with passion, dedication and skill.

Lundgren Lindqvist found a perfect representation of such ideas in Janus - the roman god of new beginnings, transitions and endings. His two faces empower him to look into the past and the future at the same time. Domains he governs are (1) time: the shift from past to future, (2) space: thresholds, doors and gates and (3) actions: the start of new journeys, projects and decisions.

From this basic idea, a new Logo was born. A shape that resembles the Janus representations while also reminding of an adoption S-curve that flows from bottom left to top right. It also reminds the shape of an “I”, the first letter of our firm’s name.

Building on this foundation we wanted to use timeless, crafty imagery that has a somewhat retro sci fi touch and dreamy touch to it while also representing boldness and reality. We experimented with various AI prompts to get it right. Here are some examples:

Our colour palette reflects optimism and calmness. We went with a modern, minimalist sans serif type font.

We also found a new home for our website. It should represent better who we are and the platform we are building for founders and investors. Check it out:

https://inflection.fund/

As always, we’d love to hear your thoughts.

Subscribe now

Ifa Reist, Lee Moser and Tal Kelman - Venture and General Partners at Protego Ventures

Ifa, Lee and Tal invited me to moderate a panel discussion around European Defense in Munich last week. They were doing a little roadshow for the first private Israeli defense venture fund. As former diplomats and 2nd in charge of the IDF Airforce, there were some good discussions in the room.

🚀 Companies

MagnetFab - High-precision Micromagnets

Status: Grant funded

Source: CAVI

Founders: Milan, Abdellatif

Why it’s cool:

Provider of high-performance micromagnets for semiconductor and microelectronics applications. Manufactures wafer-integrated permanent micromagnets at microscale for MEMS devices, biomagnetic sensors, and quantum computing components.

Voltrac - Autonomous, electric tractors

Status: Pre-Seed

Source: Network

Founders: Fran and Tom

Why it’s cool:

Fran and Tom are not part of the Inflection portfolio, but we are big fans of them and their work since 1y+. Fran built cruise missiles and hypersonics at Destinus previously, while Tom co-founded Extropic, the AI chip company out of SF.

With Voltrac they are rethinking how to build large, electrified autonomous ground vehicles, from the drive train upwards. One important impact of their platform, not covered as much in the publicity they received recently through The Next Web, is the potential impact on the battlefield of the future. With our investment in Ark Robotics, we’ve observed first-hand the impact of UGVs in Ukraine, and specifically the coordination of all kinds of autonomous systems on modern warfare. Ground robots have a very important place there. Just today, I was in Prague at our 10th European Defense Tech Hackathon in 11 months, and I spoke to a medic from the Ukrainian frontlines who said they have to use UGVs for medical evacuations and casualty evacuations because the front is controlled by Russian drones. ItIt’s too dangerous to send an ambulance or any type of personnel to retrieve the wounded, as the Russians specifically go after any medical support in order to increase their own lethality. This in itself is a violation of the Geneva convention, by the way. In Ukraine, they’re being modified according to needs, with brigade-level engineering teams welding and re-programming according to the mission needs. In the future, once we have a robust supply chain of crystalized designs that make sense, we’ll likely see modular platforms, similar to how modern car manufacturers have a “vehicle platform” that they share across models and brands. E.g., MQB (Modularer Querbaukasten) is one used by Volkswagen Golf, Škoda Octavia, Seat Leon, and Audi A3. One can vary wheelbase, track widths, and body style while maintaining the same mounting points.

Talking shortly about the difficulties of building useful UGVs, they need to be versatile, incredibly robust and modifiable. The versatility means having one platform that can be used for multiple use cases, in terms of range, conditions like mud, snow, low/high temperatures, etc.. Specifically for Europe, that means being able to handle Arctic conditions (-40 C and snow/ice) in the north as well as muddy swamps or rocky hills. Not necessarily all in one configuration, but at least it should be clear what conditions requires which adjustments. On the software side, we need everything from the mission planner to the world understanding and action planning to work, long-term. Short-term, it might be enough to have a robust enough fiber optic cable system for UGVs so that they can be operated in EW-denied environments.

With Voltrac, I think we’re seeing the early signs of the next generation of APCs, IFVs, maybe even main battle tanks, and certainly recovery and support vehicles. The simplicity of the components make sense, the approach to the software makes sense. Autonomy will come, Tom has worked on it most of his career. We just need to incorporate the feedback loops with military.

Share

[REDACTED] - Brain organoid replacement

Status: Seed

Source: CAVI

Founders: [REDACTED]

Why it’s cool:

Growing brain organoids is difficult but could be very useful for replacing damaged parts of the brain, disease modeling, or just studying how the brain develops. Long-term, we might even be able to copy-paste our brain structures.

💡 Ideas & Science

How Big Things Get Done

Big projects go over budget and time almost always. It’s insane. Especially nuclear power plants, nuclear decommissioning projects and rail projects. Big solar, wind and road projects are modular and thus are easier to estimate, plan and execute. There’s an amazing researcher, Bent Flyvberg, who built up a data base of BIG projects globally to study them. One could argue a startup is a big project, or it should be at least…

Praha byla centrem inovací v obranném systému. Hlavní roli sehrály drony

I also don’t know what that means, but it’s what Czech media had to say about our most recent hackathon. Please check out the video to see a dubbed version of me speaking Czech! The other Czech TV-channel, CNN Prima and Czech Reuters picked up the story too.

Subscribe now

🤼 People

Vishaal Hariprasad - CEO at Resilience

Vishaal helped us out as an expert in cybersecurity, and government—something we’re looking closely at right now. (Probably those are two of the most hated words in trad VC). He’s been very a founder for 10+ years, and prior to that was a supreme being of cybernetic warfare in the service of the US Airforce, presumably preventing the Chinese from installing little kill-switches into powerplant inverters and such.

🚀 Companies

Heron Power - Better power electronics

Status: Seed

Source: CAVI

Founders:

Why it’s cool:

Electrifying everything and renewables means a huge shift in grid balancing and massive strains for the infrastructure that goes along it. Transformers, inverters, rectifiers, even the high-voltage cables weren’t meant to stand up to the volatility and frequencies we’re about to put them up to. Two questions: 1. What’s the timeline on this shift and how is demand distributed? 2. What’s the venture business in selling relatively low cost components?

Stealth - part investment firm, part semiconductor co

Status: Seed

Source: Network

Founders: 1

Why it’s cool:

How do we build an Intel today, and why should governments be interested in funding not just the incumbents, but new large scale projects like Stargate and the data center buildouts in the Gulf region? Probably with massive ambition and realisation that it’s not something you finance through VC only. The tides should have turned with the CHIPS act in the US and similar investments in Europe, but critical industry components are still sourced through Taiwan, China and South Korea. Somewhere in the space between specialized chips, on-shore trailing edge fabs and sovereign wealth funds there’s a massive business to be built.

Subscribe now

💡 Ideas & Science

High-Precision Navigation: The future is near

Mostly basic info on how GNSS works, but good reminder of how it’s not perfect, yet. There are some ways of using tricks to get to high-precision, but nothing magical.

The Re-Industrialization paradox

Felix Neubeck of Playfair published a great thesis on re-industrialization and what types of companies will truly succeed in this space. This type of content is much better suited for humans to write and consume. Why read AI slop from someone else’s prompt? Much prefer to read well-founded opinions and future looking conclusions than primers I could interact with myself.

Seven Brief Lessons on Physics

Recommended read for everyone, even those who actually studied physics. Carlo Rovelli explains complex topics in simple ways, without lessening the beauty of the ideas.

Overview

The “agent economy” refers to a new computational paradigm driven by AI-powered autonomous agents that can independently perform tasks, make decisions, and engage in transactions on behalf of individuals and businesses.

This fundamental change in how computational systems interact may reshape how work is coordinated, decisions are made, and value is created.

This article examines the evolution from isolated AI systems to an interconnected agent ecosystem, exploring how standardized communication protocols, specialized hardware, and emerging technical capabilities are converging to create what may become one of the most significant economic and computational transformations of our era.

A Brief History of Agents

The concept of autonomous agents has deep philosophical roots, beginning with Alan Turing’s 1950 paper on machine intelligence and John McCarthy’s 1959 “advice taker” concept that explored computer systems capable of using logic to deduce new information and take actions.

The theoretical foundation for modern agent systems emerged in the 1970s with Carl Hewitt’s Actor Model, which proposed computational entities that could communicate asynchronously.

The 1990s saw practical implementations through expert systems and agent communication languages like Knowledge Query and Manipulation Language (KQML). Microsoft’s Office Assistant (”Clippy”) represented an early consumer application attempt, despite its limitations.

Clippy: The O.G. Agent?

Academic research on multi-agent systems continued through the 1990s-2010s with frameworks like Swarm, Repast, and JADE. While these platforms pioneered standardized communication protocols and modular architectures, most applications remained in simulations rather than real-world deployment, due to factors such as market readiness, limited computing resources of the era, and insufficient reasoning capabilities within the agents themselves.

The 2000s brought web services and service-oriented architecture with protocols like SOAP and REST, breaking down monolithic applications into independently deployable services and setting the stage for distributed computational models.

Voice assistants emerged in the 2010s with Siri (2011), Google Assistant (2016), and Amazon Alexa (2014), offering natural language interfaces but with agency limited to pre-programmed functions.

The 2020s marked a fundamental shift with transformer-based large language models, beginning with the “Attention is All You Need” paper (2017). Early experiments like AutoGPT and BabyAGI in 2023 demonstrated potential for LLMs as reasoning engines, but were primarily focused on prompt chaining and task management rather than true agent functionality.

By 2024, dedicated agent frameworks emerged (LangGraph, CrewAI, and AutoGen), alongside Anthropic’s Model Context Protocol (MCP), a proposed standard for managing access to additional context, moving beyond simple prompt chaining toward more sophisticated agent functionality.

Beyond Isolated Agents: Communication Standards Emerge (2025)

Interest around Google’s A2A is already seeing an incredible growth rate (reaching 10k stars on Github in 5 days vs. 12 weeks for MCP, indicating substantial community interest).

The introduction of agent communication protocols like the Agent-to-Agent (A2A) standard in 2025 represents an important milestone in this evolutionary path. These emerging protocols are facilitating the transition from isolated agents to an integrated agent ecosystem for several reasons:

1. Industry Collaboration: The support of over 50 major technology companies for standards like A2A demonstrates widespread industry recognition of the agent paradigm and a collective commitment to interoperability.

2. Addressing Integration Challenges: Enterprise software has long been plagued by integration problems. Open communication standards provide standardized ways for autonomous agents to work across organizational boundaries and software systems.

3. Enabling Composition Without Central Control: By establishing protocols for agents to discover and leverage each other’s capabilities, these standards enable the composition of increasingly complex agent behaviors without requiring centralized coordination.

Just as earlier protocols like SOAP and REST enabled the API economy, emerging agent communication standards will likely form the foundation for the agent economy, allowing agents to discover each other, coordinate activities, and work together across organizational boundaries in ways that were previously impossible.

Inflections

A question that we ask ourselves at Inflection every day is “What are the profound shifts in technology, science and markets that are enabling novel behaviors at a vast scale?”

Cultural Inflections

https://hai.stanford.edu/ai-index/2025-ai-index-report/public-opinion

Global sentiment data reveals a profound shift in how AI is perceived. From 2022-2024, the Ipsos survey shows most countries increasingly believe AI products and services offer more benefits than drawbacks, with nations like France and Germany seeing 10% increases in positive sentiment. This transition from skepticism to acceptance represents a foundational inflection enabling novel behaviors at scale.

This inflection is most evident in enterprise adoption of autonomous AI agents. Johnson & Johnson now employs agents to optimize pharmaceutical solvent switches in drug discovery, while Moody’s has developed 35 interconnected agents with distinct “personalities” that can reach different analytical conclusions on complex financial matters [WSJ Article]. Similarly, Inflection’s own team of agents assist with deal sourcing and help with research. These examples demonstrate how organizations are evolving from viewing AI as mere tools to treating them as semi-autonomous team members with defined roles and responsibilities, creating human-AI collaborative ecosystems that will restructure knowledge work across industries.

Technical Inflections

https://ourworldindata.org/grapher/artificial-intelligence-parameter-count

The technological foundation for the agent economy rests on several key inflections.

1. We’ve witnessed an exponential growth in model capability, with systems now exceeding 1 trillion parameters, delivering unprecedented performance in generating text, images, audio, and video. This scale-up has been accompanied by democratization through open source, allowing secure deployment of powerful models virtually anywhere.

https://objectbox.io/the-rise-of-small-language-models/

2. Interestingly, 2024 has seen the emergence of a countertrend with “Small Language Models” (SLMs). Despite being larger than nearly all pre-2020 models, these SLMs represent a pivot away from ever-increasing parameter counts.
   

https://hai-production.s3.amazonaws.com/files/hai_ai_index_report_2025.pdf

This trend toward optimization rather than raw scale enables local-first agents that can run directly on consumer devices. The performance trajectory of these smaller models is impressive: some 8B parameter models now achieve over 60% on challenging benchmarks like Massive Multitask Language Understanding (MMLU), approaching the capabilities previously exclusive to flagship models but with dramatically reduced computational requirements.

3. The agent paradigm has been further accelerated by specialized architectures like Large Action Models (LAMs). Unlike traditional LLMs optimized for text generation, LAMs are specifically designed for tool use and action execution. This architectural specialization allows LAMs to outperform general-purpose LLMs of comparable size when it comes to completing multi-step tasks, making them ideal for autonomous agent applications. The work by Salesforce on xLAM exemplifies this trend toward purpose-built model architectures for specific agent capabilities.

https://groq.com/inference/

4. Complementing these model advancements is the rapid evolution of inference hardware. Companies like Groq and Cerebras have developed Application-Specific Integrated Circuits (ASICs) achieving inference speeds up to 3x faster than traditional GPUs. This hardware acceleration fundamentally changes the economics and responsiveness of agent interactions, enabling real-time performance even with complex reasoning chains.

These technical inflections, trillion-parameter models, efficient SLMs, specialized action-oriented architectures, and accelerated inference hardware, create the perfect storm of capabilities needed for the agent economy to flourish. When combined with standardized Agent-to-Agent (A2A) communication protocols, these technologies enable collaborative agent networks that can work across organizational boundaries at a scale and sophistication previously unimaginable. The following sections examine each layer of this stack in detail, exploring both their current state and potential evolution by 2035.

The Agent Technology Stack: Current State and Future Evolution

The following analysis examines five critical components of the agent technology stack, contrasting their current state with projected developments over the next decade, culminating in our vision for 2035. This timeframe allows for the necessary technical advances and market adoption cycles to realize the full potential of these technologies.

1. Intent Expression Systems

Current State: Today’s programming paradigm remains rooted in explicit instruction-giving. Specialized frameworks like DSPy and LMQL provide abstraction layers for LLM integration, while IDE-integrated assistants (such as Cursor, Windsurf, and Github Copilot) supplement traditional programming. Tools such as Lovable, v0, and Bolt are pioneering natural language interfaces that enable software creation with minimal coding knowledge; they still primarily serve traditional software development workflows: generating, modifying, and maintaining conventional code bases.

Future Evolution: By 2035, computation itself will be fundamentally reimagined around human intent rather than explicit programming. These new intent expression systems will transcend traditional software development and become universal interfaces for directing both digital and physical systems:

• Scientific research will be guided through high-level experimental design. Systems will understand protocols, manage lab automation, and autonomously adapt procedures based on emerging results.

• Creative professionals will express high-level goals while systems handle technical implementation: Architects will specify buildings that automatically conform to structural and zoning requirements.

• Manufacturing systems will understand intent throughout the entire process. They will translate high-level product specifications into detailed hardware optimizations, bridging human goals and physical production.

• Business processes will be defined through intuitive interfaces. Systems will orchestrate data collection, analysis, and process automation while maintaining compliance and operational constraints.

These systems will understand more than syntax or semantics. They will incorporate deep domain expertise, including regulatory requirements, safety constraints, best practices, and complex interdependencies. This will enable true intent-to-execution workflows where human experts can focus entirely on their domain goals instead of computational implementation.

The distinction between programmer and domain expert will vanish as these intent expression systems become the primary interface for directing both computational and physical work. This represents a fundamental shift: computation will adapt to human thought patterns rather than humans adapting to computational thinking.

2. Infrastructure Primitives

Current State: Multiple well-capitalized durable compute startups/scaleups (Temporal, Inngest, Modal, Trigger.dev) provide essential foundations for agent workloads by addressing fundamental limitations of traditional serverless functions. These platforms enable long-running processes, persistent state management, and workflow continuity. In parallel, specialized services like OpenRouter have emerged to help developers overcome reliability challenges in LLM routing and model selection. Despite these advances, current infrastructure solutions remain largely centralized and don’t adequately address the needs for fluid operation across different computing environments.

Future Evolution: By 2035, agent infrastructure will evolve into a distributed system with three distinct capabilities:

• Adaptive Execution Location: Agents will intelligently shift processing between edge devices and cloud resources based on contextual factors (network conditions, privacy requirements, battery life), maintaining state during transitions.

• Resource-Optimized Model Selection: Infrastructure will dynamically allocate computation across on-device, edge, and cloud resources. This includes using lightweight models locally for privacy-sensitive tasks, edge computing for regional needs, and cloud resources for complex reasoning. The system will optimize performance, efficiency, and responsiveness based on task requirements.

• Decentralized Data Architecture: Local-first synchronization engines (AnyType*, Zero) will enable peer-to-peer data transfer between devices without requiring cloud centralization, allowing agents to maintain data consistency even with intermittent connectivity.

While realizing this vision will require solving complex challenges around maintaining consistency and reliability across highly distributed systems, the foundational building blocks are already emerging. This distributed infrastructure will support continuous agent operation across diverse computing environments, balancing privacy and security with performance while optimizing resource utilization for each specific task.

3. Memory Systems

Current State: Today’s agents face fundamental memory limitations stemming from limited context windows. Current solutions include Retrieval Augmented Generation (RAG) systems for incorporating external knowledge, and early implementations of hierarchical memory models (Zep, Mem0). These systems typically run on conventional hardware using standard DRAM and high-bandwidth memory (HBM) designed for general GPU/TPU workloads rather than agent-specific memory patterns. The combination of software architectures not optimized for associative recall and off-the-shelf memory hardware restricts agents’ long-term coherence and reasoning capabilities.

Future Evolution: By 2035, specialized hardware-accelerated memory systems will transform agent capabilities through:

• Neuromorphic Memory: Hardware architectures inspired by the brain’s neural structures that enable efficient associative memory and pattern recognition. Unlike traditional memory that requires exact addressing, neuromorphic systems can retrieve information based on similarity and context, enabling more human-like recall.

• In-Memory Computing: Technology that performs calculations directly within memory units rather than transferring data to a separate processor. This eliminates the bottleneck of moving data between storage and computation, dramatically accelerating similarity searches and pattern matching operations essential for agent memory systems.

• Persistent Memory Technologies: Advanced storage mediums that maintain state without power while offering speeds approaching RAM. These technologies bridge the gap between volatile memory and permanent storage, preserving agent context across sessions without time-consuming serialization and deserialization processes.

Though significant engineering challenges remain in translating these biological inspirations into production-ready systems, the potential benefits justify continued investment and development. These memory advances will enable agents with more human-like memory characteristics: contextual recall, intelligent forgetting, and the ability to form connections between seemingly unrelated concepts, without requiring constant reloading of context or expensive computation to access relevant information.

4. Compute

Current State: Agent computation relies on general-purpose GPU/TPU infrastructure optimized for model inference rather than agent-specific workloads. Current architectures struggle with the distinctive compute profile of agents: long periods of minimal activity punctuated by intensive computation spikes. This mismatch creates economic barriers as cloud pricing models optimized for predictable workloads cannot efficiently handle the bursty-with-long-tail pattern of agent operations.

Future Evolution: By 2035, specialized compute architectures designed specifically for agent workloads will emerge:

• Heterogeneous Compute Orchestration: Systems that intelligently route different agent tasks to optimal hardware based on their computational profile, seamlessly bridging between AI operations and traditional code execution.

• Custom Inference ASICs: Application-specific integrated circuits optimized for agent inference workloads running directly on edge and mobile devices, enabling sophisticated agent capabilities without cloud dependency.

• Agent Communication Processors: Purpose-built silicon optimizing message parsing and routing between agents, dramatically reducing the overhead of inter-agent communication.

These specialized computing architectures will substantially reduce the computational costs of agent operations while enabling more sophisticated capabilities, making always-on agents economically viable for a broader range of applications.

5. Networking

Current State: As described earlier, the Agent-to-Agent (A2A) protocol has established a standardized way for agents to discover capabilities, coordinate activities, and work together across organizational boundaries. While this represents a significant advance in agent interoperability, the current JSON-based approach creates substantial overhead at scale, with verbose text-based protocols consuming excessive network resources during multi-agent communication.

Future Evolution: By 2035, agent networking will evolve from today’s verbose text-based protocols to more efficient communication paradigms optimized for machine-to-machine interactions at scale:

• Optimized Binary Protocols: Like the evolution from XML-SOAP to JSON-REST to gRPC in web services, agent communication will follow a similar optimization path, dramatically reducing bandwidth consumption and parsing overhead.

• Hardware-Accelerated Authentication: Specialized circuitry for high-throughput credential verification will enable secure agent interactions across organizational boundaries without the computational overhead of constantly verifying credentials and evaluating complex permission trees.

• Standardized Intent Schemas: A shared vocabulary of intentions and capabilities will enable more efficient negotiation between agents, allowing them to coordinate complex activities with minimal communication overhead.

• Privacy-Preserving Communication: Secure information sharing mechanisms will enable agents to communicate effectively without exposing underlying sensitive data, balancing collaboration needs with privacy requirements.

These advancements will make agent-to-agent communication not only more efficient but also more secure and trustworthy, enabling truly autonomous collaboration across organizational boundaries.

Investment Opportunities

The agent technology stack presents a wealth of investment opportunities across multiple layers. From a deep-tech, pre-seed investment perspective, here’s what excites us the most:

Infrastructure Optimization

• Purpose-built infrastructure optimized for agent workloads: will we see an agent-first cloud the same way EC2 got AWS started? (Agentuity)

• Agent orchestration systems that distribute tasks optimally across different types of compute

• Edge deployment orchestration systems that enable disconnected operation

Memory Systems

• Neuromorphic memory architectures for efficient associative recall

• In-memory computing solutions eliminating data transfer bottlenecks

• Persistent memory technologies preserving agent state without serialization overhead

Compute Optimization

• Dynamic model switching frameworks optimizing for specific workloads

• Specialized inference chips for edge and mobile devices (Ubitium*)

• Energy-efficient processors designed for agent operation patterns

Security and Information Sharing

• Technologies enabling secure agent-to-agent communication without exposing underlying data or meta-data (Tune Insight*, Hedy)

• Local computation frameworks that keep sensitive data within secure environments (AnyType*)

• Context-awareness “bubbles” constraining information to certain areas and actions

• Privacy-preserving techniques like federated learning, secure enclaves, and homomorphic encryption (Fabric*, Blyss)

Agent Discovery and Interoperability

• Agent registries and discovery protocols enabling agents to locate, assess, and invoke one another (Synergetics)

• Intent extraction and summarization techniques enabling more efficient coordination

• Distributed trust verification systems allowing secure cross-organizational interactions

The most compelling opportunities lie at the intersection of these domains, where technological advances in one area can unlock capabilities across the entire stack.

If you’re building something in this space, please let us know!

*) Inflection portfolio company

Conclusion

The emergence of agent communication protocols like A2A signals the beginning of a new phase in computing. This era is defined not by composable services, but by composable intelligence. Just as APIs unlocked the software economy, agent protocols may reshape how work is coordinated, decisions are made, and value is created.

This shift from programmer-defined interfaces to autonomous agent negotiation represents a foundational change in how computational systems interact. Its economic impact could rival or even exceed previous paradigm shifts such as personal computing, the internet, mobile, and cloud.

The agent technology stack emerging today includes LLM-native programming languages and specialized silicon. It represents a complete reimagining of our computing infrastructure for an autonomous future. Whether you’re an investor, builder, or observer, now is the time to engage with the infrastructure, protocols, and models shaping this emerging layer of computation and economic activity.

Tremendous thanks to Lele Cao for notes and comments.

🤼 People

Alexander Schubert - Partner at SciFounders

At a techbio dinner in Munich this week I was lucky enough to sit next to Alexander, we discussed West Coast vs. East Coast biotech funds, artificial embryos, and epigenetics. We also talked about the high signaling value of MD PhDs, and SF mindset.

🚀 Companies

Stealth - Sampling biological reality through parallelized experiments

Status: Pre-seed

Source: Network

Founders: Stealth

Why it’s cool:

Precision fermentation is growing by ~46% CAGR, it produces alternative proteins, enzymes, fats and molecules for pharma. However, scaling production and optimizing yield isn’t easily done. This company runs thousands of micro-experiments in parallel on bio-chips, essentially mL bioreactors, to sample from reality.

Share

Bi-pedal stealth co - Long-range bi-pedal robotics

Status: Pre-seed

Source: EDTH

Founders: Stealth

Why it’s cool:

Ostriches are pretty good runners… Bi-pedal robots are usually not that fast or energy efficient, so people put them on a wheeled platform. This team thinks they’ve figured out a mechanism for making them more scalable and flexible. Applications could be abound.

💡 Ideas & Science

How prediction markets create harmful outcomes: a case study

Prediction markets: the beloved use case of crypto, though it’s questionable how much of crypto is needed to run e.g., Polymarket. This is a pretty good criticism of some of the “harmful outcomes” of prediction markets by a high-performing forecaster. Incentivizing fraud, the production and spreading of misinformation and generally acting more like sports betting sites than futarchies. Co-incidentally, I earned my first bitcoin by participating in a forecasting contest in the early 2010s!

The AI energy bottleneck - Complex Systems podcast

Patrick McKenzie interviews Institute for Progress’ Tim Fist about how AI is fundamentally bottlenecked by energy supply, both in chip production, but also in operation. My favorite idea from this interview was that even if AI progress stagnates here, the massive economic pressure in the AI buildout for achieving cheap energy likely leads to a full green transition. It might be solar+batteries, it might be geothermal, or something else, but it’s being paid for by data centers.

In Pursuit of ‘Space Control,’ USSF Gets First Upgraded Jammer

Space is being weaponized. It’s clearly a strategic priority, with a fairly consistent upwards trend since the 1950s. We’re now seeing the first signs of Russian-designed Starshield (military Starlink)-targeting devices and likely capability of the Chinese to harm Western space assets. Luckily there is Lodestar Space and Radical Aero!

Subscribe now

As imagined by Midjourney: R2-D2 and a small droid floating in space, a technical drawing. Mid century sci fi art inspired.

A brief history

Physical AI refers to artificial intelligence embodied in physical agents (robots, smart machines) that can sense and act in the physical world​. Their intelligence arises from continual interaction with their environment via sensors and actuators besides data and algorithms in the cloud.

The robotics revolution has long been in the making. Since the 1960s, we’ve seen waves of interest in robotics: from Unimate’s industrial arms to Rodney Brooks’ embodied cognition principles to DARPA’s autonomous vehicle challenges. Each era pushed boundaries—but also exposed limitations due to hardware cost, brittle AI, or unreliable sensing.

That is changing now. The convergence of cheap, powerful edge compute, high-fidelity simulation, algorithmic breakthroughs (think foundation models, general purpose robot intelligence), and macroeconomic shifts (like aging populations, reshoring, labor shortages) enable the next wave of physical AI. We expect this wave to unfold over the next 5-10 years. Autonomous drone swarms (from bee swarms to tank sized UGVs) for transport, warfare, surveillance all the way to industrial robots (micro factories, additive manufacturing, autonomous science) are underway today. The current stage of intelligent robotics is comparable to the 1950s of the information age - when machines could only run one specialised algorithm at a time like decrypting messages or steering missiles and before CPUs for general compute were invented in the 1980s. While it took humanity decades to unlock general purpose compute we have reason to expect to achieve general purpose robotics faster thanks to compounding flywheel effects across the stack.

“The material economy could autonomously make and assemble the parts required for more key parts of the material economy — extracting materials, making parts, assembling robots, building entire new factories and power plants, and producing more chips to train AI to control the robots, too. The result is an industrial base which grows itself, and which can keep growing over many doublings without being bottlenecked by human labour, visibly transforming the world in the process.”

• Will MacAskill & Fin Moorhouse in Preparing for the Intelligence Explosion, March 2025

How much faster we will get from specialised services to a general purpose robotic economy is very hard to tell. In their AI 2027 Report, the AI Futures Project, a non profit think tank forecasting the future of AI expects economic growth to accelerate by about 1.5 orders of magnitude within a few years after super intelligence based on some historical precedence and trends. They also recognise that:

“Obviously, all of this is hard to predict. It’s like asking the inventors of the steam engine to guess how long it takes for a modern car factory to produce its own weight in cars, and also to guess how long it would take until such a factory first exists.”

With that off our chest, let’s explore what might be feasible today, what the underlying inflections and bottlenecks are before discussing some opportunities.

Subscribe now

Inflections

A question that we ask ourselves at Inflection every day is "What are the profound shifts in technology, science and markets that are enabling novel behaviours at a vast scale?”

Technology

The hardware cost curves and AI capabilities have finally aligned. Critical components like LiDAR and edge computing have plummeted in price – LiDAR sensors that cost $75,000 in 2015 now cost under $7,500 (90% drop), with some automotive LiDARs targeting <$500. On board processing units (like NVIDIA Jetson) declined from $3,000 to $399. Ubiquitous connectivity (5G) and IoT infrastructure further enable distributed robots.

We expect the developments in software to positively affect developments in hardware, hence the hardware capability curve started to steepen. E.g. generative design, where AI models optimize mechanical parts based on goals like strength, weight, or thermal resistance. Foundation Models (like GPT-4 or Claude) are beginning to assist in hardware workflows—writing firmware, generating CAD models, or summarizing engineering specs to enable a shorter, cheaper and more iterative hardware loop. Katie Vasquez has been writing up a great summary piece going a bit deeper, AI and Physics-Based Modeling: A Force Multiplier for the Future of Hardware.

Over-simplified schematic illustration of how increased software capabilities (generative design tools, simulations, world models etc.) enable faster and cheaper hardware improvements, thereby steepening the “hardware capability” curve over time.

Meanwhile, algorithms have leapt forward from basic perception to deep learning and large “world models” that give machines a form of common sense about physical environments, all the way to neuro inspired approaches to learning. The software toolchain for developing and testing hardware has matured, with high-fidelity simulators and better developer tools lowering the barrier to entry. We will dive into this more under “The Physical AI stack” below.

Macro

After decades of outsourcing productivity to the East, the West has a productivity issue and with that a (industrial, military and compute) sovereignty issue. Productivity levels are the input for industrial capacity. Industrial capacity is the input for military capacity.

Source: World Bank

Due to declining populations across western countries (and most of the globe actually), additional productivity levels can only be rooted in technology growth. The “New Labor Economy” - manufacturing, logistics, healthcare, agriculture, construction, etc. – represent a $100+ trillion market (the “atoms” economy), vastly dwarfing the ~$11T digital economy of software and internet services​.

In terms of industrial robotic installations China is leading and it controls the global supply of industrial robots with 47% (decreasing though for now). The US is leading in physical AI models and woke up to the risk of missing the boat. Europe needs to catch up on all front but has an exceptionally strong industrial base with deep process knowledge, especially in Germany.

https://ifr.org/img/worldrobotics/Press_Conference_2024.pdf

“The impact of this in robotics will be exponential compared to their last strategic industry captures. These will be robotics systems manufacturing more robotics systems, and with each unit produced the cost will be driven down continuously and the quality will improve, only strengthening their production flywheel.”

https://semianalysis.com/2025/03/11/america-is-missing-the-new-labor-economy-robotics-part-1/

Unblocking the future

Despite the strong tail winds mentioned above there are several non-technical bottlenecks robotics innovators are facing. Overcoming them will be cumbersome and time consuming, especially the ones outside the entrepreneur’s sphere of influence.

Cultural resistance: Robots often face resistance due to job displacement fears and cultural discomfort. Labor unions, policymakers, and conservative industries (like healthcare or education) often push back on automation. Even military adoption of autonomous systems faces deep ethical scrutiny.

Regulation: liability is often unclear - is it the manufacturer, the operator or the software provider? The complexities faced by Tesla and Wayve around insurance and liabilities are similar in other verticals, depending on exposure and risk profiles.

Standards & Interoperability: There is no equivalent of USB or TCP/IP for robotics as the industry relies on proprietary SDKs, control stacks or custom comms protocols. Components are rarely interoperable, software needs to be re-written from scratch for each hardware platform.

Distribution: unlike for mobile apps or books there is no clear distribution layer for robotics in place. Industrial sales cycles are long, trust-intensive, and vertical-specific.

Deployment: is complex and requires specialised, scarce knowledge around simulations and edge AI operations.

There might be many more bottlenecks to be overcome. At the end of the day we believe that the inflections and tail winds will out-weigh the resistance in terms of culture and consequently regulation. All other challenges can and will be solved by entrepreneurs, not bureaucrats.

Subscribe now

Rise of the physical AI stack

Physical AI Stack – an overview from the foundational hardware at the bottom to high-level applications at the top:

This is how technology maturity looks like today (2025) and how it might look like in about a decade from now (2035). Finding proxies for technology maturity and tipping points is hard. TRL needs a refresh as pointed out very coherently by Nathan Mintz. We tried to synthesize maturity levels from proxy data on manufacturing readiness, adoption levels, integration maturity, commercialisation and open standards adoption. Open AI’s Deep Research Model helped alongside feedback from various researchers and practitioners. To be enjoyed with a huge pinch of salt.

Here are the definitions:

1. Concept Only: Idea under academic or speculative discussion.

2. Early Research: Limited experiments, no real-world implementation.

3. Prototype in Lab: Exists in testbeds or research demos.

4. Lab Validated: Components tested under realistic, but controlled conditions.

5. Pilot-Ready: Used in small-scale field trials, not broadly reliable.

6. Deployed in Niche Use-Cases: Market exists, but only in verticalized settings.

7. Interoperable + Scalable: Standardized and modular; integrates with existing infrastructure.

8. Ecosystem-Integrated: Supported by dev tools, APIs, and external vendors.

9. Industry-Standard: Robust, widely deployed, with supply chain maturity.

10. Invisible Infrastructure : Ubiquitous, trusted, plug-and-play (e.g. like Wi-Fi, USB).

Sensors & Actuators (2025: 7 → 2035: 9)

Mature industrial base: depth cameras (Intel RealSense), LiDAR (Ouster, Hesai), force-torque sensors, and servo motors are widespread and commoditized. Bottlenecks remain in multi-modal fusion (sensor latency and integration) and in cost/power constraints for mobile platforms. Actuators are robust in industrial robots (ABB, Fanuc), but underpowered or inefficient in other areas.

Edge Compute & Control (2025: 6 → 2035: 9)

Edge AI hardware (e.g., Jetson Orin, Google Coral) is powerful and affordable, but thermal, power, and ruggedization (hardening systems to operate reliably under harsh environments, e.g. extreme temperature, vibration, dust etc.) are still issues. Real-time control stacks (e.g., ROS2 + DDS) are improving but remain brittle at scale. Hard real-time industrial controllers are mature, but they lack flexibility for general-purpose autonomy. Balancing compute loads between onboard and cloud remains a challenge with regards to latency and reliability. Analogies can be drawn from Daniel Kahneman’s Thinking Fast and Slow where different systems are responsible for different types of decision making, just like in human brains and nervous systems. NVIDIA has been pioneering the three computer framework spanning AI training, simulation and on board resources but many questions remain open.

AI Algorithms & Autonomy (2025: 5 → 2035: 8)

This layer is the most software centric. It caught a lot of attention and funding recently and spans key functions of perception, localization, decision / planning and learning / adaptation. The below is an attempt to break down a vast category with overlapping problem spaces.

Perception: Interpreting raw sensor data to understand the environment. Perception tells the machine about the world as it currently is given its sensor experience. For example, computer vision models detect and classify objects from camera images (using CNNs or now even vision transformers), while SLAM (Simultaneous Localization and Mapping) algorithms build 3D maps from lidar or camera data. Modern robots often have a perception stack that fuses multiple sensors to output a coherent representation of the world.

Localization: Determining the robot’s own position and orientation in the world. This can involve sensor fusion of IMU (inertial measurement unit), GPS and visual cues. Accurate localization underpins autonomy – whether it’s a vacuum cleaner knowing its position in a home, or an autonomous car pinpointing itself on a map.

Decision and Planning: Given a goal (like “move from A to B” or “pick up that object”), the AI must decide on a sequence of actions. This involves motion planning algorithms, e.g. to find a route for a self driving car as well as higher-level task planning, e.g. deciding in what order to pick items in an order fulfillment task. Low-level control algorithms ensure the planned actions are executed by the actuators.

Learning and Adaptation: What sets “AI” apart is the ability to learn from data and improve. Many Physical AI systems use machine learning models – from vision to control – that are trained on data in simulations.

World models are an attempt to leverage neural nets to model the physics of the world by simulating environments. They enable accurate predictions about how the world will change given a choice of actions. This is critical for proper learning and reveals a technology gap because most dominant approaches are policy or imitation based, neither of which doing anything that resembles planning. NVIDIA’s Cosmos or Google’s Genie teams (see a great overview here) have been working on this problem for years.

Another frontier are embodied foundation models focused on general purpose AI for multi task robotics, e.g. Google’s RT-2, Physical Intelligence’s π0). They output task specific actions (like grasping objects) and serve as base models for downstream applications.

In their Welcome to the Era of Experience paper David Silver and Richard S. Sutton explain how agents will acquire superhuman abilities by learning predominantly from experience (think reinforcement learning) as opposed to simulations or imitation learning which are limited by data bottlenecks and and don’t really make machines intelligent:

“The era of human data offered an appealing solution. Massive corpuses of human data contain examples of natural language for a huge diversity of tasks. Agents trained on this data achieved a wide range of competencies compared to the more narrow successes of the era of simulation. (...) However, something was lost in this transition: an agent’s ability to self-discover its own knowledge. In the era of experience, agents can inhibit streams of experience rather than short snippets of interaction. Their actions, observations and rewards will be grounded in the environment rather than human dialogue.”

Welcome to the era of experience, Fig 1 (highlighted by the author)

Following this line of thinking we expect to see entirely new, experience based models rise that depend less on static simulation or human interaction. One of them is called decentralized sensory learning and is inspired by living nervous systems. Every signal a “sensor cell” receives (think a change in temperature) is a problem to be solved. When the machine acts in a way that makes those signals smaller by moving towards lower temperature for example, it learns. This approach is discussed in A Foundational Theory for Decentralized Sensory Learning and implemented by Intuicell. Similarly, Noumenal is combining an “experienced physics” approach with a library of behaviours robots can pick from (on edge or loaded from the cloud) to dynamically interact with their environment.

Despite all this progress generalization across form factors/tasks remains very challenging. Unpredictable environments (mind the slippery floor!) are an unsolved problem. Cobot’s Brad Porter wrote a great in depth piece for more context: The Business of Robotics Foundation Models

DevOps & SimOps (2025: 4 → 2035: 8)

Robotics is still in its “DevOps infancy.” Simulation engines like Isaac Gym, MuJoCo, Brax, and Unity Robotics are powerful, but workflow tooling is fragmented. No GitHub/Hugging Face-style hubs for simulation versioning, scenario benchmarking, or real-world model validation exist yet. The support layer to handle cloud management, Continuous Integration analogous to MLOps is lacking. Until recently the sim-to-real gap seemed to be a critical bottleneck to overcome as it has been more art than science: AI models have been first tested in simulation, then deployed on physical robots; real-word data and outcomes have been collected to refine the models and the simulator to improve the next cycle’s performance. sim → train → deploy → learn → improve sim. This approach could turn into a self-reinforcing flywheel if managed well (think “physics-as-a-software” feedback loop). This problem might be solved by novel approaches to learning and adaptation as discussed above. However, sensory accuracy and messy environments (dust, fog, noise) might still be challenging to overcome.

Apps & Services (2025: 3 → 2035: 7)

There’s no “robot app store” or SDK ecosystem akin to iOS/Android available yet. Most robots are closed systems or require firmware-level dev work. Modular skill deployment (e.g., “fold laundry,” “pick lettuce”) is rare. Commercial APIs exist (e.g., for drone fleet ops), but only in tightly verticalized domains.

The TLDR is that the physical AI stack is still in its infancy. Basic hardware components like sensor and edge compute infrastructure made huge leaps over the last few years with AI algorithms and autonomy catching up quickly. Higher levels of abstraction like DevOps / SimOps or Apps & Services are under developed but expected to take off as the lower levels mature.

Opportunities

Over the last few years physical AI is seeing a narrative shift based on the above inflections, media presence and significant funding rounds especially in the autonomy software category.

Yet, there is a delta between insider and outsider perception creating an arbitrage opportunity for venture investors and entrepreneurs alike - at least for those who know what they are looking for. Jordan Nel put it well in Robotics: a product selection problem:

It seems the consensus outside-robotics take is “capex heavy, hard to scale, small TAM”, the consensus inside take is “scaling laws hold, full autonomy, humanoids, El Segundo, lab-spinouts, 1:1 domain-transfer from LLM learnings”.

At Inflection we like to take high convex positions where we assign a higher probability of success to an opportunity than the market consensus while optimising for fat tail outcomes. We also take our first-check mandate into account which keeps us disciplined and focused on specific opportunities.

Robotic AGI

As discussed above (AI algorithms & autonomy) this is the frontier in physical AI these days. It is dominated by rock star robotics entrepreneurs straight out of the leading labs with backing from large, multi stage funds. Therefore, a hyper competitive opportunity set very few micro funds like us should compete in.

Besides strong market signals we currently don’t think that this category will be particularly lucrative. Drawing analogies to the model wars in LLM foundation models we expect to see a very fragmented landscape of hierarchically organised models to power the robotic brains of the future. Depending on the task at hand a combination of different models will be used to optimise for various trade offs. Further, we struggle to imagine how those businesses can create real moats, parallel to Google’s “We have no moat, neither does OpenAI”. Defensibility could be increased through (1) control of distribution (e.g. marketplaces for collections of behaviours) or (2) vertical integration of robotic AGI (what NVIDIA seems to go after).

We’d be very keen to explore novel approaches bringing the experiential era to life.

Vertically integrated services

Based on the constraints discussed above we believe that vertically integrated specialist companies have more appeal in the near term. Here are some of the high level patterns we are looking for:

Hair on fire problem in a blue ocean / fragmented market: The problem set at hand should be urgent and existential for customers with no or only insufficient alternative solutions available. Critical industries that are typically fragmented and didn’t benefit much from automation over the last decades might hold more attractive opportunities for start-ups. Reducing human exposure to hazardous or dangerous locations increases urgency in general. Some examples:

• Military: ARK* for fleet control in autonomous drone warfare; NAD* for autonomous counter UAV; Laelaps for autonomous physical security; Radical* for cell towers and eyes in the stratosphere.

• Construction: Built Robotics for solar construction, Cosmic for critical infra maintenance, Shantui bulldozers, MAX for rebar tying. Raise Robotis and Monumental for on site construction.

• Inspection, Maintenance: Koks for silo cleaning; Nautica for under water inspection.

• Space: Lodestar* for autonomous object manipulation to secure the space domain, Motive Space Systems for in space servicing, assembly and manufacturing.

• Science: Trilobio for whole-lab automation in syn bio, LabLynx and Sapio for Laboratory Information Management Systems).

Capability centric: most customers aren’t interested in buying robots (those who do are large industrials) but full blown capabilities. They want to buy a service that can be easily integrated into their operations. E.g. the Military has no interest in buying hardware components from X and software from Y to put them together. Instead they need fully fledged, working solutions.

Lower cost: common pitfalls for robotics companies have been high R&D and up front CAPEX spending. As described in the Inflections section, those cycles started compressing significantly. Off the shelf hardware components, additive manufacturing techniques and intelligent simulation and CAD software drive down costs and increase velocity. Simple hardware form factors play a large role because hardware is expensive and complex hardware is very expensive. Humanoids (Figure, Tesla) don’t make sense for most use cases (high center of mass, wheels are simpler and cheaper than legs). From a customer’s perspective the service should be significantly cheaper than the next best alternative.

Defensibility: economies of scale can create strong moats but require deep integration and thereby time. Software and or data enabled network effects should be actively pursued as hardware alone will commoditise quickly. Robotics data is still a critical bottleneck to be overcome, e.g. environmental (presence sensing for collaborative robots, air quality, temperature, 3D spatial maps to avoid collisions or other accidents) or robot internal (joint angles, velocity, pressure of grippers, maintenance logs, balance etc.).

Micro Factories

Microfactories are highly automated, small-to-medium-scale manufacturing facilities designed to produce low volumes of products with high flexibility and efficiency. Unlike traditional factories that rely on mass production and large-scale infrastructure, microfactories use advanced technologies—such as robotics, artificial intelligence (AI), and digital fabrication tools (e.g., 3D printing, CNC machines)—to enable agile, on-demand, and often localized manufacturing.

Micro factories can provide an alternative, horizontal infrastructure to manufacture machines of all kinds as they allow for adaptive process management and the fast redirection of materials and distribution if needed. Their modular and distributed design increases supply chain resilience and flexibility.

This category is still very early in its development. Companies like Bright Machines (end to end automation suite for manufacturing), Isembard (franchise network for machine shops) or Arrival (electric vehicles) are early pioneers. As indicated in the introduction we might see a “industrial explosion” accelerated by embodied AI feedback loops.

“During World War II the United States and many other countries converted their civilian economies to total war economies. This meant converting factories that produced cars into factories that produced planes and tanks, redirecting raw materials from consumer products to military products, and rerouting transportation networks accordingly. (...) Roughly speaking, the plan is to convert existing factories to mass-produce a variety of robots (designed by superintelligences to be both better than existing robots and cheaper to produce) which then assist in the construction of newer, more efficient factories and laboratories, which produce larger quantities of more sophisticated robots, which produce even more advanced factories and laboratories, etc. until the combined robot economy spread across all the SEZs is as large as the human economy (and therefore needs to procure its own raw materials, energy, etc.)”

AI 2027 Report, Exhibit U - Robot Economy Doubling Times; (highlighted by the author)

Timelines remain very hard to predict but we expect fast adoption particularly in NATO defence applications where urgency is very high and local procurement laws often require “local manufacturing” of critical components. Think drone micro factories near front lines etc.

Picks and Shovels

The complementary approach to vertically integrated services would be horizontal services and components like sensors, chips, simulators, data platforms or “SimOps” infrastructure that those doing the deployment will need (Cogniteam for robotics cloud), Marketplaces for robotic data sets and behavioural libraries might fall into this category. Specialised edge AI silicon (Ubitium*) and resource orchetration or next generation sensing (Xavveo for synthetic aperture radar; Singular Photonics and Pixel Photonics for single photon detection; Qurv for wide spectrum image sensing) etc.

As the ecosystem matures, we expect some consolidation: shovel companies partnering with or being acquired by vertically integrated services companies who need their tech, and vice versa (a robot company open-sourcing some tools once they build their own, commoditizing that layer).

Subscribe now

*) Inflection portfolio company

If you are working on related companies solving any of the hard problems mapped out in this piece, we’d love to hear from you!

A big *thank you* to Prof. Jeff Beck from Noumenal, Felix Neubeck from Playfair, Viktor Luthman from Intuicell and Sophia Belser from Laelaps for critical feedback and ideas. My gratitude also goes to the many dozens of founders and researches who spent time with the Inflection team discussing the above themes over the last few years.

In the meantime, we’ve been doing some deep research on a couple of topics, which we’re going to be publishing soon on Kepler.

🤼 People

Jamie Croucher - Founder of Society for Technological Advancement

Jamie is now also an Officer for capabilities with NATO DIANA, but prior to that he started the SOTA, which is focused on advancing society through technology and science. Shouldn’t be a hard sell but surprisingly it’s one-of-a-kind org. We talked about high agency people and cultures, and how to further progress and abundance. He’s working on an important hackathon around rebooting state capacity in London on the 26-27th of April.

Danielle Strachman - Founding team member of the Thiel Fellowship, and Founder and GP of 1517 Fund

Danielle is a teacher who started an own charter school, then continued helping young talent through the Thiel Fellowship, now 15 years ago. Then she launched the 1517 Fund. We had a great conversation about [REDACTED] and how to identify truly exceptional people. In a podcast interview she said VCs wouldn’t give the Thiel Fellows the time of day, until they launched the 1517 Fund and showed that it was interesting from a venture perspective.

🚀 Companies

Axonic - Ozempic for sleep

Status: Pre-seed

Source: Network

Founders: Helena Rosengarten and Isaak Freeman

Why it’s cool:

Imagine you don’t have to sleep as much, but still feel replenished. Without negative effects! What a productivity gain for humanity his could be, basically reducing our unnecessary down-time from 7-8h to 4-6h! Helena and Isaak are on to something here, great examples of people who just can’t not do something.

DeepGate - Bit-level intelligence

Status: Unknown

Source: CAVI

Founders: Christian Taylor

Why it’s cool:

Deepgate is building energy-efficient AI systems by embedding intelligence at the logic gate level, aiming to drastically lower energy use and enhance inference speed. Targets significant reduction in energy consumption required for AI systems while also increasing the speed of AI inference.

The Electric Plant Co - Plant intelligence

Status: Seed

Source: CAVI

Founders: Lee M. von Kraus, Justin Mast

Why it’s cool:

Plant intelligence technology company developing bioware systems to capture and interpret electrochemical signals from plants. Creates AI models to decode plant data for environmental monitoring and climate adaptation applications.

New concepts unlocked: Cyborg Botany, Techno-Biophilia and Hildegard of Bingen.

💡 Ideas & Science

Violence and the Sacred: College as an incubator of Girardian terror

Mimetic contagion, as a concept from Rene Girard, is the worst at University. Young, impressionable minds within a context of zero-sum, top-of-class, scholarship students all choosing approximately the same paths. Of my engineering physics class, a very large amount went to consulting, almost no one founded a company. How can we make young people choose more for themselves?

“I’ll end with a quote from I See Satan Fall Like Lightning: “Mimetic desire enables us to escape from the animal realm. It is responsible for the best and the worst in us, for what lowers us below the animal level as well as what elevates us above it. Our unending discords are the ransom of our freedom.””

The State of Photonic Computing<2025>

Lawrence of Lunar ventures has published another great overview article on the state of photonic computing. Tl;dr: memory is still in its infancy, but we are already seeing early signs of adoption. It’s also increasingly interesting due to European sovereignty and applications in contested EM environments.

America Is Missing The New Labor Economy – Robotics Part 1

Semianalysis goes deep on robotics and the supply chain of manufacturing. China has continued to invest in automation way beyond any other country. “[…] China pivoted and increased their robotics installation by 44% from 2020-2021 in order to implement automation to make up for the lack of a workforce.”

Subscribe now

It’s gotten rediscovered recently, and we wanted to update it with some events, thoughts and experiences since then. Needless to say, quite a few things have changed. But fundamentally, the conclusion remains the same. It’s just easier for everyone to see. We’re not a defense tech fund, we invest into Sovereign Compute companies, some of which are highly relevant to defense.

This post is an abbreviated version of a longer research piece on our beta-launched research platform—Kepler. Beyond the research we're conducting internally, we're also excited to share that the public can now access our research findings through the platform. This post is the first in a series of articles that will explore our research on Sovereign Computation. Here’s Alex Patow’s run down of some early features:

A Brief History

The war in Ukraine is not an isolated geopolitical flashpoint but a continuation of centuries of conflict rooted in the geography and power dynamics of Eastern Europe. Russia’s invasion must be understood in the context of a 500-year history of invasions, occupations, and shifting alliances. This historical perspective underlines the enduring strategic importance of the buffering function states like Ukraine perform, to the benefit of all of European democracy, and highlights the long-term fragility of peace on the European continent. In light of growing instability, Europe can no longer rely on U.S. military guarantees as it did throughout the post–World War II era. Transatlantic cracks, fueled by political uncertainty in Washington and a pivot toward Asia are forcing Europe to re-evaluate its security posture, defense capabilities, and industrial resilience.

New Defense

We would like to introduce the concept of New Defense, which is a generational shift in how Europe must build, fund, and deploy defense capabilities. New Defense refers to a new breed of technology-first, venture-backable defense startups that operate with speed, adaptability, and technical excellence. Unlike legacy defense primes, New Defense companies are driven by rapid iterations from the field, modularly designed, and software-centric. Enabled by inflection points in sensor tech, autonomy, distributed manufacturing, and digital infrastructure, these startups represent the future of defense innovation.

Subscribe now

Challenges

Structural challenges persist. Europe’s defense ecosystem is fragmented across national lines, leading to inefficient procurement practices, protectionist tendencies, and a limited market for early-stage companies. Legacy defense primes dominate contracting, while startups struggle with burdensome regulations, opaque requirements, and the infamous “valley of death” between prototyping and scaled deployment. Moreover, cultural and political divergences among EU members hinder the formation of a unified defense-industrial strategy.

Opportunities

Yet significant change is underway. Defense spending across Europe is rising sharply, with over €1.3 trillion mobilized for defense, infrastructure, and strategic resilience. Public sentiment is shifting in favor of military investment, particularly in Eastern Europe (esp. those with a border to russia). Institutional support for cross-border collaboration and SME access is growing, exemplified by programs like EDIRPA. Simultaneously, battlefield dynamics in Ukraine are demonstrating how cheap, attritable systems (e.g., drones) can outperform legacy systems, provided they are developed and deployed with urgency.

What A Venture-Backable Defense Company Looks Like

Against this backdrop, venture capital has a critical role to play. Traditionally government-financed projects are not able to move at the same speed as private capital paired with elite talent. Startups can move faster, innovate at the edge, and build systems that are good enough, cheap enough, and scalable enough to meet the evolving needs of 21st-century warfare. We identify three types of venture-backable defense companies: (1) vertically integrated system builders that replace primes with faster, modular platforms; (2) horizontal component companies that supply critical subsystems such as AI targeting, GNSS-denied navigation, and propulsion; and (3) deep tech dual-use companies that spin out frontier technologies with civilian and military applications.

This thesis is also a call to action: Europe must embrace a defense innovation culture, one that rewards speed over perfection, iteration over inertia, and collaboration over siloed nationalism. Each nation can’t build their own top performing drone company. In the case of Ukraine, agile procurement models and field-led innovation have proven decisive. Similar models must be embraced across Europe. The convergence of geopolitical urgency, industrial potential, and technological progress offers a rare opportunity to reshape European defense.

Conclusion

New Defense is not only a market trend, it is a strategic imperative for European sovereignty, peace, and prosperity. By building the technological and industrial foundations of a resilient defense ecosystem, Europe can ensure it is not only protected but also prepared to lead in an increasingly contested and multi-polar world.

“Defense tech is peace tech”.

Read the full length piece on Kepler

Thanks to Toby Stone, Benjamin Wolba, Larysa Visengeriyeva⁩, and Eveline Beer for notes and comments.

🤼 People

Max Samuel - General Partner at Snö

Max moved from managing legal and investments at Thiel Capital in LA to a town in Norway with about 100k inhabitants and is now investing with the generalist, early-stage fund Snö (meaning snow in most Scandinavian languages). Norway saw a massive exodus of founders last year, so it’s mostly Sweden, Denmark and Finland where entrepreneurial endeavors happen.

Cyrus Yari - GP at Rational VC

Cyrus moved to Munich! He left the big city of London to enjoy the deep tech, Weißwurst and outdoors scene in Europe’s sixth largest GDP, Bavaria. He immediately set up a Founders running group, together with Elika, which I joined this week. A good mix of ex-Lilium engineers, AI founders and a few VCs.

Subscribe now

🚀 Companies

Portia - predictable, controllable and authenticated agents

Status: ?

Source: CAVI

Founders: Emma Burrows, Mounir Mouawad

Why it’s cool:

Not very counter-consensus I guess, but very likely to be useful. Agents will lead to an exponential compute explosion. Pre: human triggers program; Post: human triggers agent who triggers programs. We should ideally know how and what they do, on whose behalf.

Clairvoyant - Securing software supply chain

Status: 3rd year

Source: CAVI

Founders: Douglas Schultz, Gautam Altekar

Why it’s cool:

Zero trust is a great model! But cumbersome to implement. Clairvoyant helps you gain visibility into your IoC, faster.

Integral - implanted brain modulation devices for treating disorders

Status: ?

Source: CAVI

Founders: Philip Sabes (Neuralink cofounder), Milan Cvitkovic

Why it’s cool:

If we want to really make compute efficient, then we need to get closer to our own hardware - aka wetware, aka grey matter. If we can’t enhance our own brains it’ll be very hard to increase bandwidth enough to keep up with compute ex vivo. Humans will be outpaced rapidly.

💡 Ideas & Science

Chinese Deep Sea Cable Cutter

If you ever wondered why it looked like certain russian trawlers were going back and forth over the location where a subsea cable carrying vital internet infrastructure hundreds of times, it’s not because the fishing is so good there. It’s because they don’t have the Cable Cutter Pro (CCP), Xi’s newest gadget. More reasons to have redundant infrastructure, surveillance and counter-measures at sea of these assets.

Book review: Abundance

Noah Smith reviews Ezra Klein and Derek Thompson’s new book: Abundance. Still in the mail, so haven’t read yet, but based on the review, there might be many explanations for why we’re not getting things that we want, or at least the policy isn’t being made to enable that.

DNA is the new oil

M42 is Abu Dhabi’s biggest healthcare company, and they have one of the largest DNA collections in the world. They’ve sequenced data from over 800k people, 700k of which are Emiratis. Now they’re looking to monetize it.

Space-Based Solar Power And Its Current Bottleneck

We’ve covered a few SBSP companies here in the weekly digest, EMF Space and Reflect Orbital to name two, as well as some research on it (see the first edition ever). Space Ambition published a good overview of the basic concepts you need to understand to accurately judge viability. SBSP is all about economics, grid dynamics, EM, and orbital dynamics. Easy-peasy.

🤼 People

Martin Schilling - CEO at Deep Tech Momentum

Deep Tech Momentum is a deep tech conference and network in Berlin, and Martin founded it after Techstars Berlin. We discussed what the European deep tech ecosystem needs, what current topics of relevance are (surprise they intersect with our interests), and how to collaborate.

🚀 Companies

[REDACTED] - offline p2p money

Status: Idea/PoC

Source: Network

Founders: [REDACTED]

Why it’s cool:

Bitcoin is great, so are stablecoins. It’s clearly superior technology to traditional digital money. However, it’s not truly peer-to-peer. It always goes through a “node” and requires a global state update. What if we could have cryptographically secured, offline, digital cash? It would be private by default, work in all settings and be instantaneous. We could make it programmable without complex smart contracts or other scaffolding. All it requires is some very cool cryptography…

[REDACTED] - Wetlab on semiconductor

Status: pre-seed

Source: Network

Founders: PhDs in microfluidics

Why it’s cool:

We need to speed up the physical part of the learning loop. Especially in the biological and chemical realm. Why there? Because of the massive lag in physical loop vs digital loop. We can simulate and compute all we like, but if we don’t go closer to the actual physical reality and work on accelerating that part of the loop we’re not going to achieve much.

EMF Space - Space-based power to earth

Status: pre-raise

Source: Discord

Founders: Dan CaJacob, Eric Haengel

Why it’s cool:

Normally, the efficiency of reflecting sunlight onto earthly solar farms is iffy. EMF however, want to collect solar in high orbit 24/7, convert it to laser, and beam it in to earth, supplying solar farms with illumination. If this works, it’s very cool. Scale will be hard.

💡 Ideas & Science

Meta’s brain-to-text tech is here

Meta and the Basque Center on Brain, Cognition and Language have been working on decoding brain signals, training an AI model on it and then using it to better predict thoughts/words. It’s interesting, doesn’t seem ready yet, but along side a couple of other interesting initiatives we’re seeing it’s heading towards an inflection point in human-machine interfaces.

Microsoft’s topological qubit that wasn’t, maybe?

You may have read/heard about Majorana, Microsofts 1-qubit chip. It made splashes, Satya Nadella came on the Dwarkesh Podcast and the results looked promising. However, there are reasons to be skeptical. Henry Legg (University of St Andrews) argues that Microsoft’s topological gap protocol for detecting Majorana zero modes is vulnerable to false positives, undermining the foundation of their qubit design. Peer reviewers of Microsoft’s Nature paper also noted it provides no conclusive proof of MZMs, only "findings that might enable future experiments", which you know, surprise! Media overstated research implications and company didn’t correct them!

Microsoft also had to retract a Nature paper on this in 2018 due to data manipulation…

Space Force general: ‘Golden Dome’ missile shield requires Manhattan Project-scale effort

Remember Space Force, it’s not just a Netflix show, they actually exist and do… real things. Presumably. Just kidding, we know the Space domain will continue to be very important for national security. Partially hence our investment in Lodestar. A Golden Dome is not related to Trumps adventures in Moscow in the 80s, but rather an American Iron Dome (of course it’s golden) to protect against any aerial threat to the CONUS. And it will be expensive and difficult, says guy in charge of requesting budget for it and building it, General Michael Guetlein.

Subscribe now

Either way, now that the dust has settled, we can look back at the hackathon with pride. In the last year, we’ve built a tech-driven “do tank” for European defense. It feels more relevant today than ever before. Coming in as outsiders to the defense industry, we’ve been warmly welcomed and received such encouragement that it’s impossible not to continue. Is this what product-market fit feels like?

During my opening speech at the Munich defense tech hackathon we ran in parallel to the Munich Security Conference (no endorsement or affiliation either direction to be clear), I wanted to drive home three reasons “Why” and one ask for “How” hackers and members should participate.

The “Why” is a very clear story made up of long-lasting trends reinforced by headlines I pulled from 24h before the talk:

1/ We are already at war (whether we’ve realized or not) and it’s likely to continue

2/ The numbers are not on our side - Russia outspends and outmans non-Ukraine Europe any day and are replenishing the arsenals

3/ We can not count on the US to save us - in fact, as highlighted this past week, relying on the US is a liability

The “How” is a request and an imperative. Nothing else will work. We can only use our technological superiority and adoption thereof - we can not control much else. And for this innovation to matter, it must be built at a monumental scale. In fact I refer to a local Monument in Munich - Bavaria, lady of the state. She watches of the place of the Oktoberfest, wearing a bear skin from a bear she killed, she’s carrying a sword, next to her is a pet lion and she’s triumphantly waving a wreath. Pretty bad ass. It was built to remind the people of Bavaria of the bravery and achievements. To unite them against a potential enemy.

Our modern monuments will need to look differently. Symbolism and memes will matter, but achievements of what it represents will matter more. Hence, we need to create the achievements to be proud of, and create that memetic matter, the genetic code from which pride is built. This can only arise from insanely ambitious endeavors.

That was my plead to the 400 participants in Munich, and to you.

If you decide to take this up, and build a monumental idea with tech. Let me know and I will do everything in my power to make it come true.

As guests to the event we had the CTO of Helsing - Robert Fink, who’s been supporting the mission in action since the beginning, members of the Ukrainian parliament, representation from 3rd Assault Brigade of the AFU, veterans from other European militaries, Florian Seibel, founder of Quantum Systems, and literally hundreds of other extremely talented people. We have more work to do, much more, but I am extremely proud of what we’ve achieved. Thanks to Benjamin Wolba and the EDTH team, as well as our partners who make it possible.

Sifted wrote about the event here, and Anne quoted me here on something i said during the event.

🤼 People

Ifa Reist - Venture Partner at Protego Ventures

Ifa has been running business and investing into them for the last decade and currently acts as venture partner with Protego Ventures, a new VC in Israel investing mainly in defense tech. She and I spoke about the recent Munich Security Conference, as well as developments in Israel within the tech and VC community.

🚀 Companies

Check out the winners of the hackathon and their brief explanations here → they’re all cool for obvious reasons. But there were also 39 other teams who were cool for other reasons. We’ll try and support the ones who wish to continue post-hackathon and get them in contact with talent, customers, users.

💡 Ideas & Science

2034

A terrifying, easy to read, entertaining story of the next world war. It’s written by former Admiral of the US Navy, James G. Stavridis and Elliott Ackerman so has the appropriate level of military jargon if you’re into that. Thanks Claudio Flores for the recommendation! I’ve started 2054, the sequel already!

Why TRL is a flawed concept

Nathan Mintz runs a great blog, go read his piece on hiring veterans. Here’s a quote from this piece on Technology Readiness Levels and why the concept doesn’t work that well in implementation.

“There are several reasons why the TRL scale no longer make sense. Building hardware today, thanks to technical innovations like reprogrammable ICs (FPGAs, GPUs), 3D printers, rapid simulation tools and CAD/CAM techniques, has never been easier. It’s about as hard to build hardware today as it was to write software when I started my career 20 years ago. Software is even easier: nowadays there are so many libraries, open source tools and APIs, not to mention AI copilots that are just starting to see widespread adoption, that your average software engineer is 10x more productive than they were then”

The Gigabit Wall: No More Need for Speed

Spicy take: we don’t need faster than 1Gbps mobile internet!

Subscribe now