Repercep · the physical-AI operating system

The compute substrate for world models.

Repercep is the optimized lowering layer for world models — the compiler, kernels, and runtime between physical-AI workloads and heterogeneous silicon. Vendor-neutral by design.

For teams serving world models in robotics, autonomy, simulation, and interactive media.

World models are a new kind of compute. The stack that serves them should be too.
01The workload

World models don't compute like LLMs.

An autoregressive decoder emitting one token at a time is the wrong mental model. A world model is a different machine, and it stresses the serving stack in different places.

Encoder + predictor

Two coupled networks per step — perception feeding prediction — not one decoder with a KV-cache.

Large 3D latent state

Spatiotemporal latents that persist across steps and dwarf a token cache — state is the workload.

Many futures at once

Planners score dozens of candidate action sequences per control step — hypothesis batches, not user batches.

Hard deadlines

Closed perception–action loops. The answer is worth nothing after the control deadline passes.

Serving stacks built for autoregressive LLMs leave 2–4× on the table on these workloads. Repercep is the operating system that closes that gap.

02The stack

Repercep owns the seam everything passes through.

From application to silicon, performance is won or lost in the lowering layers — graph capture, optimization, tiling and scheduling, memory, codegen, dispatch. Repercep builds these on an MLIR substrate with one vendor-neutral seam: write the model once, lower it optimally to any silicon.

Your workload

World models, robot policies, AV sim, generative worlds

Captured from PyTorch, JAX, or custom graphs — training and inference.

Compiler

Lowering, deadline-aware

Lowers world-model graphs — encoder, predictor, planner loop — into fused schedules shaped by the control deadline, not by throughput alone.

Kernels

Per-target, tuned

Attention and diffusion kernels tuned for each target — CUDA, ROCm, and CPU paths — dispatched from one engine.

Runtime

State-resident execution

Latent caches, adaptive step-skipping, deterministic dispatch — the state stays hot on-device across the control loop.

Silicon

Backend Protocol — a neutral target, not a rewrite

NVIDIA, AMD, or the custom accelerator that ships next.

Runs today on NVIDIA H100 AMD MI300X Intel CPU · AMX — vendor-neutral by design.

Transformers ride the same rails. The lowering machinery that accelerates world models applies to attention everywhere — so the same stack serves LLMs and agents faster too.

03Products

One stack, four products — in sequence.

Runtime lands first. Everything else compounds on it: serving generates the telemetry, synthesis turns telemetry into speed, and owned models prove the whole loop in production.

Repercep Runtime Now · design partners

World-model-native serving

The engine behind our Cosmos results. A lab adopts it in days and serves on NVIDIA, AMD, or CPU from one runtime.

Repercep Studio Next

Evaluation & CI

Wires world-model evaluation into your pipeline, so every kernel change and model update is regression-checked before it reaches the loop.

Repercep Forge In development

Kernel & compiler synthesis

Generates near-expert world-model kernels from a spec — and, next, full compiler backends from a ChipSpec. The work others do by hand, automated.

Repercep Worlds In development

Owned world models

Repercep's own world models, trained and served on Runtime, licensed into robotics, gaming, and media — hosted or self-hosted on your edge hardware.

04Proof

Already proven — on NVIDIA's own model, on NVIDIA's own silicon.

A Repercep runtime already serves NVIDIA's Cosmos world model multiple-× faster than NVIDIA's published serving stack on their own H100 — and brings Cosmos to AMD, where no public run existed before.

Measured · NVIDIA Cosmos world model
NVIDIA H100

Multiple-× faster than NVIDIA's own published Cosmos serving — on NVIDIA's own silicon.

AMD MI300X

The first public Cosmos run on AMD silicon, to our knowledge — silicon neutrality in practice, not in principle.

Same model · single GPU · end-to-end wall time · Repercep runtime

Request the benchmark report
Single GPUevery claim reproduced on one GPU — no cluster asterisks
Days, not quartersthe runtime is built to drop into a lab's serving stack in days
Beyond H100 memoryMI300X headroom runs world models that don't fit on an H100 at all

System-vs-system comparison against NVIDIA's published reference configuration: same model, same frame count, single GPU, end-to-end wall time. Measured numbers and full methodology available on request.

Fast enough to beat NVIDIA on NVIDIA. Flexible enough to run on whatever ships next.

The category is forming now. We're building its operating system.

Work with us

Serving a world model? We want your workload.

Robotics, autonomy, simulation, video — if a model of the world sits inside your loop, we can make it faster.