We started in 2023 with one rented RTX 4090 in a friend's garage and the belief that renting a GPU should not mean a sales call. It has not gotten much fancier since, on purpose.
Four people, three time zones, one shared pager rotation. The support inbox is read by whoever is awake.
Buys cards, racks them, fixes the network. Wrote the billing meter so we could stop lying about hours.
Makes instances come up fast and stay up. The boot-time work was his. He has strong opinions about image layers.
Built the scheduler that packs models onto cards. Answers the hard support questions about tokens and latency.
Talks to customers, writes the docs and the changelog, and answers the overnight support mail.
No sales team. If you email us, you reach the people who run the
hardware, not a forwarding address.We answer support within a day, usually faster. Overnight issues get
the on-call person, who is one of the four of us.We buy cards on the used market when the prices are sane, which means
our prices move. The pricing page is current.We do not keep a private roadmap. The changelog is the roadmap, in
reverse.Three racks of cards we own, in three colocation halls. Nothing virtual, nothing shared out of a third-party pool: when you rent an H100 from us, it is a card in one of these boxes with your name on it in the sheet by the door.
Power and cooling, which is most of it.Replacing cards that die. It happens more than the marketing material
of bigger clouds would suggest.Colocation in Frankfurt, Ashburn, and Singapore.The remainder is ours, and it is not a lot. That is how a four person
company keeps prices like $0.19 an hour.| 2023 | First 4090 online. Support was one email address and a phone that never rang. |
| 2024 | Second region. Started model hosting. Per-second billing shipped in October. |
| 2025 | A100s, then the first H100. Third region. Boot time work cut instance provisioning in half. |
| 2026 | Still four people. Roughly 300 models hosted, and the third year of a pricing page that moves when power prices do. |