Rent the GPU. Host the model.
Ship the endpoint.
Inferix runs the whole path from bare silicon to an OpenAI-compatible URL — scheduler, model hosting, object storage, and a GPU marketplace, on one account and one bill. No glue code. No second vendor.
Four vendors' worth of infrastructure, on one bill.
Most teams stitch together a GPU host, a model registry, an object store, and an inference vendor. Inferix is all four, wired together, so your execution stays linked to your artifacts.
One account, zero glue
Repos, endpoints, storage, and rented GPUs share one API key and one invoice — no S3 account, no separate inference bill.
Your model, OpenAI-compatible
Push a repo, get a /v1/chat/completions URL. The SDK you already use works unchanged — swap the base URL.
Open & self-hostable
Open data model, GGUF-native, no lock-in. Run it on our fleet, or on your own hardware. Zero vendor lock-in by design.
Everything the workload needs, nothing you have to wire.
Developers buy specifics, not outcomes. Here's exactly what each layer is, what it replaces, and what it lets you do.
Rent H100 / B200 / A100 / RTX 5090 by the second. KubeRay dispatch, GPU attestation before a job runs, 18-second cold starts.
Push a repo, get an OpenAI-compatible endpoint. GGUF-native, Git-backed, versioned, autoscaled to zero when idle.
Datasets and weights sit next to the GPUs that read them. No egress tax, no separate S3 account to provision and secure.
Rent idle capacity from verified providers, or list your own hardware and earn on usage. Proof-of-Render settles every job.
Repo to endpoint in three commands.
No Dockerfiles to hand-write, no cluster to stand up. Push, attach, ship.
Model files, revisions, and metadata — versioned and shareable, Git-backed.
Pick a GPU; KubeRay schedules it; the runtime builds and verifies before it runs.
An OpenAI-compatible URL goes live, autoscaled, billed by the second.
$ inferix push ./llama-ft ⠿ uploading weights … done 41.2 GB $ inferix deploy --gpu h100 ⠿ scheduling h100-3 … ready ⠿ building runtime … done ⠿ verifying GPU … attested ✓ endpoint live https://api.inferix.co/v1/…/llama-ft cold start 18s · $1.89/GPU-hr
If you can call OpenAI, you can call your own model.
Same request shape, same SDK, same response. Point the base URL at Inferix and your code doesn't change.
curl https://api.inferix.co/v1/chat/completions \
-H "Authorization: Bearer $INFERIX_KEY" \
-d '{
"model": "llama-3.3-70b",
"messages": [{"role":"user",
"content":"ship it"}]
}'{
"id": "cmpl_9f2a…",
"model": "llama-3.3-70b",
"choices": [{
"message": {
"role": "assistant",
"content": "Shipped."
}
}],
"usage": { "total_tokens": 11 }
}Priced in public. Pay for seconds, not seats.
No "contact sales" to see a number. Rent by the second; hosting autoscales to zero when idle.
| GPU | Form | Memory | On-demand | Available now |
|---|---|---|---|---|
| NVIDIA H100 | SXM5 | 80 GB | $1.89/hr | 38% |
| NVIDIA B200 | SXM | 192 GB | $4.20/hr | 22% |
| NVIDIA A100 | PCIe | 80 GB | $1.10/hr | 61% |
| RTX 5090 | consumer | 32 GB | $0.44/hr | 74% |
“We moved three inference workloads off a hosting vendor and onto Inferix in an afternoon — same SDK, half the bill, and we finally see which GPU ran what.”
Does this work with the OpenAI SDK and my stack? +
openai SDKs (Python, Node, Go), LangChain, LlamaIndex, and anything that speaks /v1/chat/completions or /v1/embeddings work unchanged. Set base_url=https://api.inferix.co/v1 and your key.Which GPUs can I rent, and how is it billed? +
Can I self-host or keep data in a specific region? +
Is it actively maintained, and what support do I get? +
Docs, an API changelog, and a live status page; paid tiers add priority support and SSO/audit controls for teams.