Holonomy Systemsthe photon-engine platform
We make high-power light manufacturable.
Holonomy builds high-power surface-emitting lasers (SEL) on a wafer platform: a family of photon engines, power and cost on the learning curve.
Energy, defense, space, and fusion are converging on one primitive: high-power light that scales. Its bottleneck was never physics; it's manufacturing. Light sources are still built by hand, one bespoke system at a time. We make them the way the world already makes its hardest things: on a wafer, thousands at a time.
The convergence
Four trajectories, one primitive.
Energy, defense, space, fusion: each hits its limit at the light source.
Clean power that has to cross distance a grid can't reach. Defense whose marginal cost has to fall to zero. Energy beamed across vacuum. Fusion lit by laser. Different fields, one ask: energy projected as light, focused at range, scaling with electricity. Fusion is our entry point, not the whole map.
Why now
The physics is proven. The driver isn't.
AI made energy the binding input, and fusion just proved the frontier is real.
AI's hunger for compute has made energy the binding input of the decade. And the hardest energy frontier just broke open: laser-driven fusion crossed ignition in 2022 and has repeated it, with target gain climbing roughly a thousandfold in fourteen years to above four. But ignition is not a power plant. Those shots fire about once a day, from a stadium-sized laser barely half a percent efficient at the wall. Inertial fusion energy needs a driver that fires ten times a second, at fifteen to eighteen percent efficiency, cheaply. The limit is no longer the physics; it is the driver.
The problem
Why the driver is a manufacturing problem.
The driver's cost is one component: the semiconductor pump diode. Fusion scale needs it far cheaper, at far higher volume, and no less capable; the serial edge-emitter that makes it today cannot ride that curve. The physics is solved. The manufacturing is not.
What an inertial-fusion plant can spend on diodes versus what one plant needs made every year: neither wall is physics. Figures are per peak optical watt.
- Plant CAPEX
- $4 to 6 billion
- Laser driver
- about $1.75 billion, 35% of plant CAPEX
- Diodes
- about $0.5 billion, roughly a third of the driver
- Required diode price
- about $0.01 per watt — that $0.5B has to buy the driver’s full ~50 GW of peak optical power
- One plant’s need
- ~50 GW peak optical power
- World output today
- ~18 million bars a year at 500 W to 1 kW each, i.e. 9 to 18 GW
- Time to supply one plant
- 3 to 6 years of the whole world’s output
The mechanism
Change the topology.
The fix is not a better bar; it is a different shape. A surface-emitting laser emits from the wafer face instead of a cleaved edge, so it is made and tested thousands to a wafer, like a chip. That single change answers all three walls where they bite hardest: wafer-parallel manufacturing puts cost and capacity on the learning curve instead of a labor floor; dropping the facet removes the failure mode, spreads the heat, and lifts the brightness ceiling. Same physics, different topology. That is the whole difference.
The proof
The answer already holds.
The evidence, not lab records.
The building blocks are proven; integration is the bet. The surface-emitter runs uncooled with a litho-set wavelength instead of one that drifts with heat; it has no cleaved facet, so the edge-emitter's signature failure mode is designed out; and it rides the same wafer cost curve consumer 3-D sensing already drove down. We state the open item up front: areal power density at the pulse is the one thing left to validate.
Three building blocks of the surface-emitter pump are already proven on the bench — uncooled with a litho-set wavelength, no cleaved facet so no COD failure mode, and it rides a wafer cost curve already driven down — leaving one open item: areal power density at the pulse.
- Proven · thermal and spectral
- uncooled, with a litho-set wavelength
- Proven · reliability
- no cleaved facet, so no COD failure mode
- Proven · cost
- rides a wafer cost curve already driven down
- One open item
- areal power density at the pulse
The bet
Own the slowest step, own the loop.
Generative Hardware: the bottleneck of the AI era is physical.
AI now designs and simulates hardware faster than anyone can physically validate and build it. That is Generative Hardware: the loop from token to watt, jammed at the one step still physical: making and proving the hardware. Value concentrates at that bottleneck, and it is the step we are built to own. (Thesis, not a proven moat.)
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
Get involved
Where do you come in?
Talk specs.
Wavelength, power, duty cycle: start there.
Tell us what you need the light to do. We'll tell you what's on the wafer.