computer.com › Services › GPU compute
GPU compute service

The computer
brings its own GPUs.

H100s and A100s, in our own data centers, ready when the agent asks for them. Train, fine-tune, infer. No procurement, no contracts, no drivers. Just a prompt.

Open the computer → See GPU pricing
Fine-tune Llama-3-8B on my support tickets, deploy as a private API
On it. Reaching for a GPU.
gpu.provision H100 · eu-fr-1 · ready in 47s
storage.read tickets.parquet · 142k rows
hf.finetune Llama-3-8B · 3 epochs running…
I'll deploy it behind a private endpoint when training finishes (~18 min).
GPUs available
H100 · A100 · L40S
Spin-up time
< 60s typical
Billing
Per second min 60s
Regions
12 across 3 continents
gpu lineup

The right chip
for the job at hand.

The computer picks automatically based on the workload. You can override per-prompt or pin defaults for a project.

Inference / light training
L40S
$1.10/hr
  • 48 GB GDDR6
  • Ada Lovelace architecture
  • Great for < 70B inference
  • Stable Diffusion XL / LoRA
Use L40S →
Most popular
H100
$2.95/hr
  • 80 GB HBM3
  • FP8 + Transformer Engine
  • Fine-tune up to 70B
  • NVLink-attached pairs available
Use H100 →
Long-running training
A100
$1.85/hr
  • 80 GB HBM2e
  • Ampere — broad framework support
  • Best $/hr for known workloads
  • 4× and 8× clusters available
Use A100 →

Reserved 1-month and 1-year contracts available with up to 40% discount. Talk to sales →

how it works

Three steps you don't take.

The computer handles provisioning, environment setup, and teardown. You describe outcomes; it manages hardware.

01 / ask

Describe the job

"Fine-tune this model on this dataset." "Run inference on these 10k prompts." The computer figures out what kind of GPU, how many, and for how long.

02 / provision

Agent brings the hardware

Sub-minute provisioning. CUDA, PyTorch, your env vars, your dataset already mounted. No SSH, no driver hell. Just a working machine.

03 / deliver

Results, not VMs

Trained weights pushed to storage. Inference exposed as a private API. Logs and metrics piped back into the chat. GPU released. You pay only for the work.

use cases

What people ask their computer
to do with GPUs.

Fine-tune an LLM

Adapt Llama, Mistral, or any HuggingFace model to your data. The computer handles QLoRA, full FT, or PEFT depending on your dataset and budget.

Fine-tune Mistral-7B on my customer support tickets

Train image models

Stable Diffusion, FLUX, ControlNets. Train a LoRA on a hundred reference images and get a deployable image model in minutes.

Train a LoRA on these product photos for an SDXL pipeline

Batch inference

Process millions of rows through an LLM cheaply. The computer batches, retries, and tracks cost. Results land in storage.

Classify these 2M reviews with Llama-3 — sentiment + topic

Run a research notebook

Attach a Jupyter kernel to an H100 from your prompt. The computer mounts your repo, installs deps, and gives you a working URL.

Run my training notebook on an H100 with persistent storage

Stop renting GPUs.
Get a computer that uses them.

Free to start. Pay only for the seconds you compute. Your AI computer manages the rest.

Open the computer → Read the docs
mockup · services / gpu