fcloud is GPU compute you drive from the terminal. One command uploads your code, provisions a GPU, runs it, and prints the output. You pay for the seconds it runs.
Install fcloud
Just paste it in your terminal to get started. It installs the CLI, signs you in through your browser, and sets up your coding agent.
Manual install instructions: docs · Pay per second, no minimums. Scriptable with --json, or from Python with the same package.
What it looks like
$ fcloud run train.py --sku gpu_1x_l4 # upload, provision, run, print epoch 1/50 loss 2.31 ... $ fcloud create --sku gpu_8x_h100 # persistent session — $0 until used s-abc123 $ fcloud exec --on s-abc123 nvidia-smi -L GPU 0: NVIDIA H100 80GB HBM3 ... $ fcloud stop s-abc123 # spend stops, files stay
The pieces
| Piece | Command | What it does |
|---|---|---|
| Session | fcloud create · --on s-… | A persistent /workspace, not a held GPU. Cold and $0 until used; files, env, and packages persist across commands. |
| SKU | --sku gpu_8x_h100 | The hardware: count × model. Routed to a matching host; `fcloud skus` lists the catalog. |
| Run | fcloud exec · fcloud run | Run a command for its output, or upload a script/dir and run it — foreground, exit code and all. |
| Processes | fcloud spawn · wait · logs | Background work you follow by process id. Durable logs outlive the host. |
| Files | fcloud upload · download · ls | Push code and data onto the box; pull results back off — even after the session stops. |
| Volume | fcloud volume · --volume | A named folder mounted into sessions — how data moves between them. |
| Batch | fcloud job · fcloud map | Run-to-completion jobs, and sweeps that fan one command out over N argument bindings. |
| Lifecycle | hot · warm · cold | Cold sessions cost $0 and come back online the moment you use them; fcloud stop halts spend. |
| Python SDK | project.session(sku=…) | The same sessions, programmatically — for harnesses, pipelines, and agents. |
Compute
| SKU | Hardware | GPU memory |
|---|---|---|
| cpu_generic | CPU only | — |
| gpu_1x_l4 | 1 × NVIDIA L4 | 24 GB |
| gpu_1x_l40s | 1 × NVIDIA L40S | 48 GB |
| gpu_1x_h100 | 1 × NVIDIA H100 | 80 GB |
| gpu_8x_h100 | 8 × NVIDIA H100 | 640 GB |
| gpu_8x_b200 | 8 × NVIDIA B200 | 1.5 TB |
A sample — T4, A10G, RTX 6000 Ada, V100, A100, H200, and B300 SKUs are in the full catalog, or run fcloud skus. Rates are dynamic; your live per-second rate is shown when you provision.
Billing
Metered, per second. Add a card, run compute; usage is invoiced weekly and account credit is applied automatically. Cold sessions cost $0 — you're billed only while a host is running. No subscriptions, no minimums, no reserved commitments.
Get a GPU in seconds.
Pay per second. No minimums.