The full thesis
Building the photon platform the next energy economy runs on.
Own the engine, drive the curve, make high-power light abundant.
Energy, defense, space, and fusion converge on high-power light, and its bottleneck is manufacturing. AI drives that bottleneck twice. The winner won't be a national lab or an incumbent; it'll be whoever owns the fastest-iterating wafer ecosystem. We're building to be that owner, and the lead compounds.
- 01One primitive, four trajectories, and it doesn't exist yet.
- 02Why now: four limits, and AI pushing on all of them.
- 03Why the carrier decides the bet: only one rides a wafer curve.
- 04Capital can buy fabs, not the learning rate.
- 05The end game: what the loop compounds into is a platform.
- 06Back us: the whole thesis rests on one chip.
The world claim
One primitive, four trajectories, and it doesn't exist yet.
Energy that crosses vacuum, carries no mass, focuses at range, and scales with electricity.
Energy, defense, space, and fusion are converging on the same primitive: energy projected as high-power light, across vacuum, focused at range, scaling with electricity. It doesn't yet exist in a form the world can build at scale. The reason isn't physics; it's that light sources are still hand-built, one bespoke system at a time.
One primitive, four trajectories: energy, defense, space, and fusion are converging on high-power light: energy projected as light, which crosses vacuum, carries no mass, focuses at range, and scales with electricity. What does not exist yet is a form the world can build at scale; the reason is not physics, light sources are still made by hand.
- Converging fields
- Energy, Defense, Space, Fusion
- Shared primitive
- high-power light, not yet buildable at scale
- Why light
- it crosses vacuum, carries no mass, focuses at range, scales with electricity
Why now
Why now: four limits, and AI pushing on all of them.
Four independent trajectories are hitting their limits at once: the grid can't distribute AI-scale load, defense marginal cost has to fall to zero, $/kg-to-orbit is phase-changing, and fusion reached ignition in 2022 and now needs an efficient, rep-rated driver. The public markers are in the field: a directed-energy interceptor, an LEO optical standard, a new government surface-emitter program, hyperscalers signing for firm clean power. And AI drives it from both ends: demand, as compute makes energy the binding input; and design, as AI outpaces what anyone can physically build.
The National Ignition Facility crossed target gain 1 (ignition) in December 2022, reached a record target gain of 4.13 in April 2025, and has now ignited eleven times; the physics question is closing, while the driver (efficient, repetition-rated, buildable at volume) is not.
- Ignition threshold
- target gain 1 (fusion yield divided by laser energy delivered)
- First ignition
- December 2022, target gain 1
- Record
- target gain 4.13, April 2025
- Ignitions to date
- eleven: the line is not crossed once, it is lived above
- Range shown
- target gain 0.001 to 10, from 2012 to 2026
Why now, as a dated field report rather than a forecast: every marker below is public, and each is demand committing ahead of supply.
- Space · Jul 2024: an LEO optical standard is set; optical crosslinks become a mandatory interface, not an option. (SDA Optical Communications Terminal Standard v4.0.0)
- Energy · Jan 2026: hyperscalers buy firm clean power; 6.6 GW contracted in a single announcement. (Meta nuclear agreements, up to 6.6 GW)
- Defense · Jan 2026: a directed-energy interceptor enters service, from program to fielded system. (Iron Beam entering Israeli service)
- The source layer · May 2026: a government surface-emitter program opens, seeking orders of magnitude more power from a photonic-crystal surface emitter, gated on manufacturability and domestic production. (DARPA AMPED solicitation DARPA-PS-26-119)
The carrier
Why the carrier decides the bet.
High-power light can be built two ways; only one rides a wafer curve.
A surface emitter is tested on the wafer, stacks power by area, sets its wavelength by lithography, and has no facet to fail. The edge emitter that supplies high power today is serial: it must be cleaved to test, and its cost floor is labor, not wafer area. So surface emission is not a device preference; it is the one carrier a manufacturing learning curve can grow on at all. That makes the carrier and the moat a single bet. The wall in front of it is not physics, which is largely solved; it is how fast manufacturing experience accumulates.
The bet
Capital can buy fabs, not the learning rate.
The moat is manufacturing speed, and AI makes it compound.
Capital can buy fabs; it cannot buy the rate at which manufacturing experience accumulates. That rate is set by ecosystem density, and Taiwan's compound-semiconductor cluster is the densest on earth. We aim it at high-power light and compound it into an iteration loop no one can copy, closing the data between design, the line, and the field. The archetype is TSMC: what compounds is learning speed, not any one product, so we are built to be the maker of this curve, not its beneficiary. AI is why it generalizes beyond photonics: AI now designs and simulates hardware faster than anyone can validate and build it, so the loop jams at the one step still physical. Own the slowest step, and you own the loop.
Generative Hardware runs two coupled loops: a fast AI search loop nested inside a slow physical loop, and Holonomy occupies the validation-and-manufacturing interface where they meet: the AI half scales without limit; that interface does not.
- Search loop (the AI half)
- Objective → Generate → Evaluate, many turns per outer turn; scales without limit
- Physical loop (the rate limiter)
- slow, expensive: one turn of it for many turns of the search loop
- The interface Holonomy occupies
- Simulate and Validate: every candidate the search loop produces must still come through validation, and the simulator stays correlated with reality only by what validation sends back
The end game
The end-game is a platform.
And what the loop compounds into is a platform. When manufacturing-experience speed makes high-power light abundant, from hand-built and scarce to something you buy off a curve, the next generation of energy, defense, space, and fusion gets built on it. This is the pattern behind NVIDIA and SpaceX: own and make the engine underneath, drive it down a curve, and an ecosystem forms on the abundant output. High-power light is the next such ground, and we're built to be the ones who make it. Today we're a vertical builder; the photon platform is where the curve ends, not where it starts.