Why compute needs a price
Most AI teams do not know what an hour of H100 time costs them. They know what the invoice says at the end of the month. They do not know what the idle time cost, what the queue cost, or what it cost to be on the wrong instance type for three weeks because switching was painful.
We think an hour of compute should have a clear price, a clear owner and a clear record of what it did. That is the whole idea behind RACK.
The problem with opaque hours
On the large clouds, GPU hours arrive bundled with everything else. The list price is one number, the negotiated price is another, the committed use discount is a third, and the effective price after idle time is a fourth that nobody calculates. Capacity appears and disappears by region without notice. A team that needs 8 GPUs for 6 weeks spends the first week finding them.
The result is that compute is planned like weather. Teams over reserve to be safe, then let hardware sit at 40 % utilization. Or they under reserve, and a training run waits two days in a queue.
What a clear price looks like
A clear price is one number per GPU hour, per plan, published on a page, updated on a schedule. Ours are on the pricing page. They change monthly at most, and the change is announced in the capacity report before it takes effect.
A clear owner means the hours you reserve are yours. The scheduler holds them for your team. Nobody else runs on them, and if you leave them idle, the telemetry says so and you can release them or fill them with batch work.
A clear record means that every hour on the rack is accounted for. The monthly capacity report shows how many hours the cluster had, how many were reserved, how many ran on demand, how many ran as batch, and how many sat idle. Those numbers come from the same telemetry feed you can read from the API.
Why we started with one rack
One rack, 8 GPUs, one region. That is small on purpose. It is enough to run a real training job, enough to serve a real inference endpoint, and small enough that we can put every GPU's telemetry on a public page and stand behind it.
The second cluster gets listed when the first one is sold out and the second one has passed 72 hours of soak testing. Not before. The regions page tracks what is next.
What this means for you
You get a price you can put in a budget, capacity you can plan a project around, and a record you can show to whoever signs off on the spend. That is what compute with a price looks like. If that is what your team needs, request access.