Repercep · the inference cloud for every model class
Run any model. For less.
LLMs, VLMs, vision, diffusion, VLAs, world models — served on one cloud, priced off a cost base the others can't reach. We get more work out of each GPU-second, on GPU-seconds that cost less.
Repercep Cloud is in private beta. The runtime underneath it is measured, published, and running today.
Every inference cloud is an LLM company with a container service bolted on the side.
The model zoo outgrew the meter.
The per-token API, the catalog, the autoscaler, the routing — all of it is shaped around a stateless autoregressive text decoder. Bring anything else and you fall off the platform.
A meter that only fits text
Your detector, encoder, diffusion model or policy doesn't emit tokens. There is no per-call price for it — so there's no easy way to buy it.
You pay for idle
Renting a whole GPU-hour to run a 40 ms model is the default answer everywhere. The gap between what you use and what you're billed is the product nobody built.
Many candidates at once
Agents and planners score N hypotheses per decision — long shared prefix, short divergent tails. A serving stack built for one answer per request leaves most of the GPU idle.
One vendor, one price
Every major platform prices a single silicon vendor. When that vendor's price is the only price, there is no arbitrage left for you to capture.
We started at the hard end of that tail — stateful, deadline-bound, multi-modal serving — and grew into LLMs. Not the other way around.
We're not choosing to be cheap. We're able to be.
The cost of serving a model is GPU-seconds per unit of work, times the price of a GPU-second, divided by utilization. We own levers on all three — and none of them is a thinner margin.
LLMs, VLMs, vision, diffusion, VLAs, world models
Any class, one platform. Serverless per token or per call; dedicated per GPU-minute with scale-to-zero.
Fewer GPU-seconds per unit of work
Shared-prefix candidate batching, topology-gated adaptive caching, similarity-scheduled denoising. Measured on both silicon vendors — 2–10× on the vision, diffusion and robotics workloads. The size depends on the workload, and we quote it per workload. On plain LLM tokens we claim no efficiency edge at all; there the advantage is silicon.
Cheaper GPU-seconds
We run production inference on silicon the other platforms don't price at all. On dense LLM serving it matched or beat the incumbent part on output throughput at every workload shape we tested — at 1.7–2.3× lower cost per million output tokens. Across the vision, diffusion and robotics regimes, 26–39% lower cost per unit of work. Both already net of the performance difference. On MoE models the comparison inverts, and we say so below rather than letting you find it.
State-resident session packing
Long-lived sessions stay hot on-device, and density is the denominator of what each one costs. Large-memory silicon turns a hard capacity ceiling into headroom — measured 6× more resident sessions on the same engine code.
Parity-gated, so moving is safe
Cheaper silicon only counts if your outputs don't change. We gate on numerical parity against the reference implementation and publish the result — on one model, bit-identical on both vendors.
What we don't claim. PagedAttention, continuous batching, speculative decoding, prefix caching, FP8 — table stakes. We ride the same open engines everyone else does, and we'd rather say so than let you find out. The levers above are the ones that are ours, and every one of them has a published number and a stated caveat.
One engine, three ways to buy it.
The same code serves all three. That isn't a convenience — it's the reason we can follow the cheapest capable silicon instead of being locked to one vendor's price.
Managed inference, any model class
Serverless per token or per call, dedicated per GPU-minute with scale-to-zero. You bring a model — any class — and we serve it on the cheapest silicon that can run it correctly.
The same engine, inside your walls
Self-host and BYOC for teams with data-sovereignty, on-prem fleet, air-gapped or defense constraints — the segment a hosted-only platform structurally cannot serve.
Hybrid edge ↔ cloud for physical AI
Your processor runs the tight loop; our cloud runs the heavy reasoning. Link decides what runs where, under what deadline, and degrades gracefully to edge-only. Built with robotics and automation partners — we don't build edge silicon.
The cheapest claim is the easiest one to make. Here's ours, measured.
We built this engine on the hardest workloads in the market — six models, three serving regimes, four silicon targets, one unmodified codebase. Every number below is wall-clock from our own harness, with the raw result line archived and the caveat stated next to it. That discipline is why the cost claim is worth anything.
Decoding N candidates as one batched pass over the shared prefix instead of a loop. 5–10× on robot action decode — both vendors, at exact numerical parity. The same technique on LLM parallel sampling measured 1.5×, and we don't count it: every serving engine already ships it, and we could not deliver it to callers who don't already ask for it. We built that, measured it, and took it back out of the plan. Same technique, very different sizes — we publish both, and the one that didn't work.
A multiple-× speedup on diffusion serving, reproduced across sessions, driver stacks and months on both vendors — and opt-in, because it trades a quality dimension we measure and disclose rather than default silently.
Wall-clock · single GPU · state carryover required where the model is stateful · warm vs. cold disclosed
Request the benchmark reportOn dense LLM serving, a single MI300X matched or beat a single H100 on output throughput at every shape we tested — 1.02–1.38×, at 1.7–2.3× lower cost per million output tokens at observed rental prices. Scope, stated rather than buried: dense bf16 models, one box per vendor, and the AMD side ran an older engine build — so those ratios are a lower bound, not a best case. Across the vision, diffusion and robotics regimes, 26–39% lower cost per unit of work.
We went looking for the counter-example to our own strongest claim, on the model family most likely to produce one — and found it. On a DeepSeek-class MoE model the comparison inverts: MI300X reached only 0.53–0.74× of H100, at matched engine versions, failing our pre-registered threshold at every shape. So the silicon claim above is scoped to dense models, by measurement rather than by caution. We publish the boundary because a claim whose edges are unknown isn't worth much.
Moving silicon only counts if your outputs don't move. On a 7B model we reproduce the reference implementation bit-identically — zero maximum error — on both vendors. That gate is what makes the cheaper option usable.
Cost per unit of work · resident sessions/GPU · numerical parity vs. reference implementation
Request the benchmark reportAll numbers measured by us on our own harnesses, single GPU, end-to-end wall time, with raw result lines archived. Where a comparison is not apples-to-apples, the caveat is stated inline rather than in a footnote — including the ones that don't flatter us. Speedups apply to the workload classes named, not to every model. Full methodology, the complete measured set, and the open items available on request.
More work per GPU-second. On GPU-seconds that cost less. And we can prove your outputs didn't change.
Every model class, one cloud. Priced off a cost base the others can't reach.
Work with us
Paying too much to serve a model? Send us the workload.
LLM, VLM, detector, diffusion, policy, world model — tell us what you run and what it costs today, and we'll tell you honestly whether we can beat it.