gpyou routes training, finetuning and inference across seven providers. Always available. Lowest price. Scheduled to the greenest grid the moment you run.
$ gpyou train ./run.py --gpus 8xH100 --max-price 2.10 scanning 7 providers · 31 regions ...... ok candidates under cap .................. 12 greenest: eu-north (Nebius) 34 g/kWh $1.98/h allocating 8×H100 80GB ................ ok mounting s3://acme-ckpt ............... ok ▶ job 4f21 running · logs streaming step 120/4000 loss 2.114 1.9 it/s $0.42 spent
No dashboards, no consoles. The CLI speaks plain stdout and JSON, so an agent can request, monitor and tear down GPUs the same way it runs tests.
┌──────────────┐ ┌─────────────────────┐ ┌───────────────┐
│ coding agent │──cli──▶│ gpyou │──────▶ │ Runpod │
│ (any shell) │◀─json──│ price · avail · │──────▶ │ Vast.ai │
└──────────────┘ │ grid carbon score │──────▶ │ Nebius │
│ │──────▶ │ CoreWeave │
$ gpyou train │ failover · ckpt │──────▶ │ Lambda │
$ gpyou serve │ one bill · one log │──────▶ │ Scaleway │
$ gpyou cost └─────────────────────┘──────▶ │ Varda │
│ └───────────────┘
▼
┌─────────────────────┐
│ your bucket / vpc │
└─────────────────────┘ Capacity pooled across every provider. If a region sells out mid-job, checkpoints move and the run continues.
Median saving versus a single provider. Spot and reserved inventory are compared on every request, not once a quarter.
Grid intensity of the region we would pick for you right now. Jobs go where the wind is blowing.
Every candidate region is scored on price and grid carbon intensity at the moment of your request. The lowest-carbon region under your price cap wins. Override with --carbon off if you must.
$ gpyou finetune qwen2.5-coder-7b --data ./repo.jsonl --json
{"job":"a91c","gpus":"4xA100-80","region":"eu-north","provider":"nebius",
"price_per_hour":1.62,"grid_gco2_kwh":34,"eta_min":41}
$ gpyou status a91c --wait
epoch 3/3 loss 0.71 eta 00:04 $1.08 spent
$ gpyou serve a91c --name repo-coder
https://repo-coder.gpyou.run ready (cold start 3.1s)
$ gpyou cost --month
training $184.20 provider $173.77 margin $10.43
inference $12.06 4.1M tokens · 38 min audio
carbon 41 kg CO₂e avoided vs. default region Every command has a --json flag, idempotent retries and exit codes an agent can reason about. Long jobs stream logs; short ones return. Secrets stay in your environment.
For the parts of an agent loop that don't need a frontier model. Same routing, same bill, priced per million tokens or per audio minute. gpyou serve <model> for anything not listed.
Pay per GPU-hour, billed by the second. No subscription, no minimum, no egress markup. You see the provider's price and our margin on every line of gpyou cost.
No. gpyou is CLI only, by design. Every state the dashboard would show is one command away and machine-readable with --json.
In your own object storage. gpyou mounts it into the job; nothing persists on provider disks after teardown.
It only chooses among regions under your --max-price. If the greenest region is also the most expensive you allow, you pay that. Set --carbon off to route on price alone.
The job checkpoints, gpyou reroutes to the next best region and resumes. You see one interruption line in the log and nothing on the bill for the gap.
Runpod, Vast.ai, Nebius, CoreWeave, Lambda, Scaleway and Varda today. H100, H200, B200, A100, L40S and RTX 4090/5090 in most regions.