You decide how much you want to pay
You shouldn't need a spreadsheet to predict your bill
Our pricing will be available as of September, but for now don't hesitate to play around in our sandbox!
Get sandbox access in September
What actually drives the cost
Rather than relying on a spreadsheet, our calculator uses four simple parameters to calculate your GPU usage cost.
Want to know your number?
We're building the GPU cost calculator now. Model your own bill, workload by workload, the moment the prices land in September.
01
Timing
Energy is more expensive at some hours than others, and cleaner at some hours than others. Almost every provider flattens that variance into a single rate and keeps the difference. We'd rather pass it on. Same GPU-hour, three ways to buy it. The only variable is how much control over the timing you hand us.
Priorityrun it now
Full price. It runs the moment you send it.
Flexiblewithin a specific timeframe
Cheaper, because we can place it when the grid is kinder.
Carbon-optimisedwhen the sun is up
Cheapest, and the lowest footprint. Cheaper compute and a lower footprint come out of the same mechanism.
02
Performance or maximum privacy
Run on shared capacity for the best throughput per euro, or on isolated, single-tenant hardware when the data can't sit on a machine with anyone else. The models are the same either way. You decide, per workload, how much isolation it needs, and you pay for that and nothing more.
03
Which model you run
Open weights mean the model is your choice, not ours. A Chinese open model, a European one, whichever fits the job. Bigger models need more GPU-hours, so the choice shows up in the price. And because it stays open weights, you can swap models or leave whenever you want.
04
The carbon it leaves behind
Every choice above moves the number: when you run it, how isolated it is, which model and how large. Automated per-workload CO₂ measurement is in development — soon available, and when it lands the footprint is yours to see and to lower.
What you can expect from our pricing
Priced against real usage from the summer sandbox rather than a guess, then held for at least a quarter so you can plan against it.
Our CEO, Özturk Taspinar, explains why we think compute pricing should work differently here.
Our pricing will be published here as of September. Subscribe to stay up to date.
What we won't do
No egress fees. Your data leaves as freely as it arrived.
No hidden meter. Every line on the invoice maps to something you can point at.
No price that moves while you're not looking. When we publish, we hold it for at least a quarter.
The hardware, honestly
RTX PRO 6000 S. Right-sized for inference, which is most of what actually runs in production. It isn't the biggest number on the market, and we'd rather tell you that than sell you headroom you'll never use. We add H100s when a customer's workload proves the need, not before.
Carbon-displacing by designAudit in progress
Carbon-aware scheduling, our own renewable energy and heat reuse together push the footprint of a compute-hour below what the same work costs elsewhere. We call that carbon-displacing, not carbon-negative, and we won't use the stronger word until an independent audit confirms it.