If they can make intelligence cheap, why can’t we do the same for compute?

The biggest clouds made frontier compute expensive, opaque, and inaccessible. The same NVIDIA H100 lists on-demand at $6.88 per GPU-hour on AWS, $10.98 on Google Cloud, and $12.29 on Azure — four to ten times what the same chip costs on the open market — and usually behind a quote, a waitlist, or a multi-year commitment.

That gap decides who gets to build. A 2025 review of 6,517 foundation-model papers found that access to compute tracks national funding and, in turn, citations: the labs that can afford the GPUs are the ones whose work gets seen.

So we asked ourselves a simple question: if frontier labs can use capital and scale to make intelligence affordable, why can’t we bring that same ambition to the infrastructure underneath it?

Why can’t researchers, startups, and open-source teams have reliable frontier compute at radically lower prices? We believe they can.

Frontier compute should belong to everyone.

Compute Cheap combines competition, distributed supply, and a leaner operating model to provide frontier GPU capacity. Each one takes a layer of cost out of the GPU-hour:

• Competition. Capacity owners compete for your workload, so you pay a market price for the hardware rather than a list price set by one vendor.
• Distributed supply. We source capacity across datacenters in the United States and European Union instead of building and financing a single one, and we pass the difference on.
• A leaner operating model. We sell the GPU-hour. Every product decision is measured against whether it makes that hour simpler, more useful, or less expensive — which is why there is no platform fee layered on top of the price.

Today that means NVIDIA H100 capacity from $1.38 per GPU-hour interruptible and $1.19 reserved, and H200 from $1.67 interruptible and $1.99 reserved — published on the pricing page and through a public API, not behind a sales call.

The terms are deliberately plain. Reserved capacity is held for you for the whole block at a fixed rate and carries a 99.9% monthly uptime commitment. Interruptible capacity costs less and can be reclaimed, which suits fault-tolerant training and batch jobs. The minimum is 2,000 GPU-hours per request, we confirm availability before any payment, and a 50% deposit reserves the block. The price you were quoted is the price you pay until the block ends.

So far that has meant 30,000+ GPU-hours delivered to 52 teams — researchers, startups, and AI companies. Arcee AI’s infrastructure team put it this way: “We train and merge small language models on interruptible H100s. Preemptions are rare enough that normal checkpointing covers them, and the per-run cost came in well under what we paid before.”

We are not against frontier labs. We are against a future where only frontier labs can afford to train, serve, experiment, and build at scale.

Not promotional prices that disappear. Not cheap capacity that is never available.

Radically lower prices. Reliably delivered.