The token,
recomputed in light.
What if a token cost almost nothing?
Not less. Almost nothing.
Every product decision in AI today — what to ship, what to charge, what to throttle, what to refuse — is downstream of one number: the cost of an inference. Lower that number by a percent and the industry tunes. Lower it by a factor large enough to change the decimal place, and the industry rewrites itself.
We built the appliance that does the second one.
It runs the models you already deploy, on memory you already understand, in a rack you already own. What it changes is the floor — the absolute minimum cost and latency of an inference call. We dropped the floor through the basement.
The question is no longer whether photonic compute works. It is what you build when it does.
What it is.
A complete inference appliance. Inside it: a wafer of programmable optics, conventional high-bandwidth memory matched to the capacity of a flagship GPU, and an optical pipeline that moves weights between them at a bandwidth conventional interconnects cannot approach.
The compute happens in light. Not optical interconnect bolted to an electronic chip — actual computation, performed by photons traveling through silicon. Millions of optical elements fire in picoseconds, in parallel, across many wavelengths at once.
It plugs into a standard rack. It draws standard power. The only thing unusual about deploying it is what comes out of it.
What changes.
Place the appliance beside a flagship GPU. Give them the same memory. Give them the same batch. Ask both to serve the same workload.
The GPU produces tokens at the rate its memory bus allows. The appliance produces tokens at the rate light allows.
Throughput does not improve. It moves by orders of magnitude. Per-user latency does not shrink. It changes units — from milliseconds to microseconds. Energy per token does not drop. It collapses to a fraction so small that a year of inference on the appliance fits inside a single day of inference on the GPU.
The hardware premium pays itself back in days, not years. After that, every token the appliance produces is priced at a level no GPU operator can match without losing money on the transaction.
Full benchmark methodology and head-to-head numbers are shared with qualified partners under NDA.
Why the physics wins.
Two facts about light do the work. Neither is a clever optimization. Both are properties of the medium.
A photon does more work per joule than an electron.
It carries a multiply through a waveguide and emerges still useful — capable of driving the next operation before it is finally absorbed. An electron in a transistor does one job and dissipates as heat. The energy gap between those two facts is the energy gap on your power bill.
A waveguide carries many computations at once.
Different wavelengths of light occupy the same fiber without interfering. Each one is an independent stream of work. Multiplying the bandwidth feeding the compute substrate does not require a wider pipe — it requires more colors of light in the same pipe.
The appliance combines both. Energy per operation falls. Bandwidth feeding compute climbs. The window that batches users into a queue on a GPU becomes the window that serves them all in parallel here.
There is no queue.
Where it lands.
Voice agents that interrupt the way humans do.
When per-user latency falls below the threshold of human perception, the conversational loop closes. The product stops feeling like a chatbot and starts feeling like a presence in the room.
Agentic systems where every tool call compounds.
A ten-step agent on a millisecond accelerator spends ten milliseconds waiting for itself. On this appliance, the same agent finishes before the user has lifted their finger from the keyboard.
Code completion that lands as you type.
Not after the keystroke. With it. The latency budget that defined a decade of IDE design assumptions is gone.
Routers, classifiers, ranking layers — the silent infrastructure of every AI product.
A single appliance saturates a tier that today consumes a row of GPUs and the cooling loop to match.
Inference-as-a-service at a price the market cannot currently set.
When the cost of a token falls by a factor large enough to change the unit economics, business models that were uneconomic become obvious.
The economics.
A trillion tokens of inference per year is a real workload at a real customer. On a GPU cluster, it is a substantial annual energy bill. On this appliance, it is a rounding error.
The hardware costs more than the GPU it replaces. It earns the difference back in roughly the time it takes to ship the box from the factory to the rack. Everything after that is a different cost curve than anyone else in the market is operating under.
This is the conversation the appliance starts: not whether photonics is faster, but what an inference business looks like when the marginal cost of a token approaches zero.
Built once. Scaled by light.
The compute wafer is built once. Performance scales by widening the pipe that feeds it — adding wavelengths, multiplying effective throughput, on the same silicon, in the same power envelope, with no forklift.
Customers buy the appliance once. The performance curve runs underneath them for years.
This is not a faster accelerator.
This is the first time intelligence has been computed in a medium other than the electron.