Dedicated GPU inference on capacity reserved for you alone.
Reserved AMD or Nvidia capacity for high-volume, low-latency inference on any open model, isolated and GDPR compliant, operated by our engineers.
AMD & Airbus backed · GDPR compliant · VPC-level isolation · US, EU, Southeast Asia
Tell us what you want to run
Five accelerators. All of them in stock.
Pick the hardware, or tell us the model and we will pick it. Everything below is capacity we can put an endpoint on today.
| Accelerator | Memory | Suited to | Availability |
|---|---|---|---|
| Nvidia B300 | 288 GB HBM3e | The largest open models. GLM 5.3 on four. | In stock |
| Nvidia B200 | 180 GB HBM3e | DeepSeek V4 Flash, Qwen 3.8, high-throughput agents. | In stock |
| Nvidia H100 | 80 GB HBM2e | Smaller models and your own fine-tunes. | In stock |
| Nvidia RTX Pro 6000 | 96 GB GDDR7 | 70B-class models and vision workloads. | In stock |
| AMD MI325X | 256 GB HBM3 | GLM 5.3 on four, Kimi K3 on eight to twelve. Our best price per token. | In stock |
Capacity you reserve, run by people who tune it
Renting bare GPUs leaves you to make them fast. A managed node is the difference between hardware you own and hardware that performs.
Five accelerators available now. No waitlist and no procurement cycle, in the US, the EU or Southeast Asia.
Our inference team handles quantisation, batching and ongoing optimisation. Included, not billed separately.
Your own context window and the full concurrency of the hardware you reserved, rather than a shared pool.
We measure your workload on the configuration we would give you, then hold that performance level.
The same model on AMD MI325X, at a third of the hardware bill
Everything above holds whichever accelerator you pick. But if cost drives the decision rather than an existing CUDA dependency, the AMD path is worth knowing.
What 4 × MI325X delivers
Measured on an 8K-token prefill with 1K tokens out, single stream and under concurrency. FP8 is what we serve by default.
| Quantisation | Single stream | 8 conc. / stream | 8 conc. total | 16 conc. total |
|---|---|---|---|---|
| FP8 | 112 tok/s | 70 tok/s | 896 tok/s | 1,034 tok/s |
| Q4 | 143 tok/s | 92 tok/s | 1,176 tok/s | 1,482 tok/s |
Both 4 × B300 (1,152 GB HBM3e) and 4 × MI325X (1,024 GB HBM3) run GLM 5.3 at 753B. The B300 delivers about 11.5% more throughput at over three times the price per GPU-hour. Past two concurrent requests, two MI325X instances match B300 per-stream speed.
We are an AMD partner, and a year of quantisation, batching and kernel work on Instinct is where the gap comes from. It is included, not billed separately. We serve at FP8 and publish what is quantised and how.
Hard CUDA dependency? Stay on Nvidia. We have it in stock, and it costs what Nvidia costs.
For a major telco deploying Featherless in phases, early production results show monthly inference costs for the sustained workload in scope falling 95%.
Read the cost breakdownYour region, and what we can sign
Capacity in the US, the EU and Southeast Asia. You choose the region, multi-region works, and your data stays in it.
| Control | Status |
|---|---|
| GDPR | Compliant |
| Zero data retention | Prompts and completions are never stored |
| Isolation | VPC-level; weights and traffic are not shared with any other customer |
| Data processing agreement | On request |
| Regions | US, EU, Southeast Asia — you choose, multi-region works, and your data stays there |
| Anything else you have to meet | Tell us in the form and we will say plainly what we can support for your setup today |
Three steps, no call in any of them
Send the model and the volume
The model, your volume in tokens per month or peak requests per second, and your latency target. We come back with a sized configuration and a price, usually inside a working day.
We prove it before you sign
We measure your workload on the configuration we would sell you, then stand it up and hold that performance. Send a benchmark script and we will run yours.
The same endpoint, your own GPUs
The same OpenAI-compatible API you already call, now running on isolated capacity reserved for you.
Frequently asked questions
What's the difference between dedicated and serverless inference?
Serverless runs on a shared GPU pool billed per use. Dedicated reserves GPUs for you alone — consistent latency and throughput, no noisy neighbours, up to 1M context against 256k on shared capacity. Which one fits comes down to your load. Sustained volume: dedicated is more predictable and usually cheaper at scale. Light or occasional: per-token on the Developer plan is genuinely cheaper, and we will tell you so. Genuinely bursty: stay per-token, or rent GPUs by the hour — reserved time is paid whether you use it or not. The cost calculator will show you which side of the line you are on.
How many GPUs does my workload need?
Send us three things — the model, your volume in tokens per month or peak requests per second, and your latency target — and we will come back with a sized configuration and a price. No call required.
How is it priced, and what's the commitment?
Per GPU, per month — not per token. What you pay depends on the GPU memory your model needs and the throughput you want on top, so send us the model and your volume and we will price it, usually inside a working day. Included: the GPUs, the serving stack, and our inference team tuning it, plus SLA terms in the contract. Billed monthly, not a year up front. Twelve months is the term we aim for, and it is negotiable.
Which models can we run?
Any of 40,000+ open models — DeepSeek, GLM, Kimi, Qwen, Llama, gpt-oss, MiniMax — plus your own. Public or private Hugging Face repos, including your fine-tunes, redeployed on each update. And you can swap the model on your capacity as better ones ship: we re-benchmark and redeploy. You commit to capacity, not to a model.
AMD or Nvidia — which should we pick?
We offer both: Nvidia B300, B200, RTX Pro 6000 and H100, and AMD MI325X. AMD is usually the better economics — a 4 × MI325X node runs GLM 5.3 at a third of the hardware bill of 4 × B300, for about 11.5% less throughput. If you have a hard CUDA dependency, stay on Nvidia; we have it in stock and it costs what Nvidia costs. Send us the model, the volume and your latency target, and we come back with a configuration.
Where does it run, and what compliance can you meet?
US, EU or Southeast Asia — you choose, and multi-region works. We are GDPR compliant, with zero data retention, VPC-level isolation and a DPA on request. If you have a specific framework to meet, tell us in the form and we will say plainly what we can support for your setup today, rather than after you have signed something.
My traffic is spiky. Am I paying for idle time?
On reserved capacity, yes — it is yours whether or not you use it, which is exactly what makes throughput and latency predictable. If your load is genuinely bursty, per-token or hourly GPU rental is usually the better economics, and we will say so.
Is our data used for training?
Never. Prompts and outputs are not used to train models, and on dedicated capacity they are not shared with any other customer.
What do you guarantee?
We benchmark your workload before you commit and hold that performance level on reserved capacity, with SLA terms in the contract. Performance guarantees are dedicated-only; we do not offer them on shared capacity.
Tell us what you want to run
Pick the hardware and the GPU count. We come back with a configuration within one business day, no call required.
