MEASURE
Measure everything
The numbers on this site come from the test suite, not from a slide. If a deploy regresses p99 routing overhead past 235µs, it does not ship, and the published figure changes only when the suite says it changed.
The whole company is one serving path
Nyx is a bootstrapped inference provider. We win on serving efficiency
per GPU-hour, not on fleet size: open-weight models behind one
OpenAI-compatible endpoint at https://nyxprovider.com/v1,
priced under every listing.
Why Nyx exists
Nyx started from a bet about waste. Most inference capacity sits idle or runs far from its ceiling, and customers pay for that idleness in the price per token. A small fleet held near saturation beats a warehouse of GPUs waiting for traffic, so we built one serving path, a 5.5 MB statically linked binary in front of tuned open-model runtimes, and spent our time on microseconds instead of procurement.
The second bet was that measured numbers age better than marketing. Routing overhead on this site is quoted at 71µs p50 and 235µs p99 because that is what the latency harness recorded on the build serving you now, and the harness gates every deploy. Prices are read from /v1/models at request time, the same endpoint your billing runs on, so the site cannot drift from the invoice.
The wafer became the mascot of that idea. A wafer is one design stepped across a disc until the pattern runs out of edge, and it is worth something only when each die yields. That is the standard we hold the serving path to.
2.1B+
tokens served monthly
71µs
p50 routing overhead
5.5 MB
serving binary, statically linked
24/7
engineer on call
How we operate
Three rules decide most arguments here. Each one is checkable from the outside, which is the point.
REFUSE
Refuse before degrading
At capacity the endpoint returns 429 immediately instead of queueing you into a slow tail. Your client backs off on honest information, and latency stays flat for every request already in flight.
PRICE
Price under every listing
Every model is priced under its cheapest listed price, published live at /v1/models. The margin comes from serving efficiency per GPU-hour, not from your bill.
Every die the same design
The color on a wafer is structural. The surface acts as a diffraction grating, so the iridescence appears only when the light moves across it, and a still wafer reads as plain gold. There is one pattern on the disc, repeated exactly, and the interesting part is how it behaves under changing light.
The disc on the home page works the same way. It is a live WebGL shader, ray-traced in your browser, and the color shifts because a virtual light is orbiting it. A photograph would have been easier to ship. It would also have been a picture of someone else's wafer.
Work on the serving path
Nyx is seven people, remote across four time zones, and the entire product is one serving path, so new engineers ship to production in their first week. We hire people who profile before they guess. If that is you, write to us about something you made fast and how you proved it.