What is RACK

RACK SYSTEMS is a neocloud for AI teams. We own racks of NVIDIA H100 GPUs, run them in a data center in us-west, and sell the hours on them. You get the whole machine, live telemetry from every GPU, and a price you can plan around.

The first cluster, C1, is live. It is a single rack: 8 H100 SXM GPUs on one NVLink fabric, 2 TB of system memory, 30 TB of local NVMe. It runs training jobs, inference endpoints and batch work. Every GPU reports utilization, memory, temperature and power once a second. That feed is the same feed shown on the site, in the console and in the monthly capacity reports.

What you can do today

During preview, console and API access is by invitation. You can request access and tell us what you run. We reply within two business days. Once you have an API key, you can submit a job with one call, read its telemetry with another, and reserve capacity by the month.

There are three ways to buy hours. On demand hours are billed per minute with no commitment. Reserved hours are held for your team for a month or longer, at a lower rate. Batch hours are preemptible and run when reserved capacity sits idle. The pricing page has the numbers.

What RACK is not

RACK is not a reseller of someone else's cloud. The hardware is ours, the rack is ours, the power contract is ours. When something breaks, the person who fixes it works here.

RACK is not a general purpose cloud either. There is no object storage product, no managed database, no serverless runtime. There are GPUs, a scheduler, a telemetry feed and a small API. If you need more than that around your job, bring it.

Where to go next

Start with the quickstart if you want to see the API. Read why compute needs a price if you want to know why we built this. Read clusters if you want the hardware spec. Labs that need a month or more of dedicated capacity should read reserved capacity.

Questions go to access@racksystems.cloud.