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.

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.

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.

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.

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.

Inflection theory

In mathematics, an inflection is where a curve changes its direction of bending. We called our firm that way because we believe that human progress is nothing but a continuous flow of inflections in a metaphorical sense.

A strategic inflection point is a time in the life of a business when its fundamentals are about to change. The change can mean an opportunity to rise to new heights. But it may just as likely signal the beginning of the end.

Andy Grove in his book Only the paranoid survive.

There have been other attempts of defining inflection points in a broader context than business strategy. For example in Pattern Breakers, Mike Maples Jr. & Peter Ziebelman systematically explore inflection theory as a framework to spot outlier ideas. Their TLDR; is that pattern breaking companies are rooted in (1) one or more inflections which led the founders to develop (2) a unique insight that is then manifested in (3) a product and scaled through (4) a movement. We will only look into 1-3 briefly but focus on inflections only. At Inflection (we created the firm long before the book came out btw) we apply parts of this framework amongst many other things to assess companies and drive our research.

Inflections

Inflections can be described as pronounced shifts in science, technology, culture or (geo) politics that hold the potential to change human behaviour at scale.

They are hard to see at the time of their arrival and very easy to spot in hindsight. Getting the timing right is existential for the creation of venture scale companies. If you're too early you end up with a science project (left). If you're too late you end up with intense competition and compressed margins (right). Most builders and venture investors are leaning towards the 'too early end of the spectrum'.

Ideally, inflections are not very obvious and consensus but nuanced or over-looked by most to build

Inflection examples:

Insights

Maples and Ziebelman go on to describe insights as a non obvious or contrarian truth harnessing one or multiple inflections to create a breakthrough product. They make the case for startups competing on being different, not better, faster or cheaper than incumbents. Changing the rules of the game to win.

Applying inflection theory to some companies:

As venture investors we might stumble upon some insights occasionally but our primary work is to spot inflections. Ideally the ones that are over-looked or under-estimated by most.

Products

Products in this context are manifestations of inflections and insights. Typically, in early stage startup land the first products don’t find product-market-fit and need various iterations, sometimes pivots. The companies innovating through them usually have been right about the underlying inflections and insights but wrong about the manifestation of them. A famous example is twitch, a company built on inflections around smartphone cameras and internet streaming becoming feasible but manifested in a first-person-infinite reality show called Justin TV.

With this basic framework in mind we can now look a little bit deeper into what inflections are and what makes them so tricky to spot for founders and venture investors alike.

Contextual awareness

The most powerful venture companies are rarely underpinned by just one inflection but by many. In their inspiring book Why Greatness cannot be planned (shout out to Alex Obadia from Flashbots for recommending it) AI researchers Kenneth O. Stanley and Joel Lehman think about it this way (strongly paraphrasing based on my flakey memory): all powerful ideas are already out there and want to be discovered. To reveal them we can use a tool called novelty search - an algorithm helping us to explore new frontiers in a serendipitous way. To understand what is novel we need a framework of reference because something can only be novel compared to something else, to something that already exists. Novelty cannot be defined for itself. The same applies to inflections - they never come alone and can only be spotted and understood in the context of what is today.

Matt Cohler from Benchmark put it this way:

Our job is not to see the future, it’s to see the present very clearly.

A wonderful illustration of the networked nature of inflections is the interactive tech tree from Calculating Empires. Watch the 5min audio tour, it's absolutely brilliant. (Shout out to Anton from Anytype who shared it with me).

Time resolution

By nature, inflection's aren't dots. In Technological Revolutions and Financial Capital: The Dynamics of Bubbles and Golden Ages Carlota Perez established a model for disruptive innovations using S-Curves (left). Somewhat related Gartner (right) established a hype cycle model capturing similar ideas.

Those frameworks work pretty well to think through technology cycles. The right way to apply them would be to factor in time resolution. That is a tricky undertaking.

Let’s observe the facts: Humans took more than a hundred thousand Earth years to progress from the Hunter-Gatherer Age to the Agricultural Age. To get from the Agricultural Age to the Industrial Age took a few thousand Earth years. But to go from the Industrial Age to the Atomic Age took only two hundred Earth years. Thereafter, in only a few Earth decades, they entered the Information Age. This civilisation possesses the terrifying ability to accelerate their progress. On Trisolaris, of the more than two hundred civilisations, including our own, none has ever experienced such accelerating development.

Trisolarian scientist in Liu Cixin‘s 3 Body Problem

One might be right seeing inflections at the horizon early but that doesn't help to build a company with a 2-5 year time horizon to find product market fit if the inflection is too early in its s-curve journey. Often, crucial building blocks are still missing to unlock the next breakthrough. We couldn’t have built the first computers without electricity or light bulbs (none of which have been invented with computers in mind though). Many novel ideas explored by previous innovators have been directionally right but wrong in terms of timing: Henry Ford envisioned energy money to "substitute gold with units of energy to end wars" 100 years before Bitcoin was invented. In the 1830s the first electric cars were invented but never took off because of lacking battery density and price pressure from alternative design. The list goes on.

A simple analogy to think of is a microscope with variable resolution to look at inflections of different scale, moving on vastly different time scales. The charts are AI generated and very rule-of thumb to illustrate the idea, simplicity over nuance.

From here we could dive deeper to explore the inflections underpinning the development of the mobile web like e.g. broad band connections, small and powerful enough chips and batteries to power them amongst many other things. You get the idea.

Conclusion

To start (or fund) a company rooted in non-obvious inflections and unique, contrarian insights requires the right intuition with regards to time resolution and contextual awareness. Figuring them out by studying history and observing what is today will be time and energy well spent. We are looking forward to open sourcing some of our work on that soon and cannot wait to collaborate on it with you.