The server is live: a single NVIDIA H200 (141 GB HBM3e), rented one tenant at a time and billed for the GPU-hours you actually use. This first note covers what it is, how a booking works, and the numbers from the first week.
What you get
- A dedicated machine — no noisy neighbours sharing the card.
- SSH access for the length of your slot.
- Setup, monitoring and security handled on our side.
Booking a slot
Tell us a date and a rough number of hours; we confirm within one business day. Checking the card is a one-liner:
nvidia-smi --query-gpu=name,memory.total --format=csvFirst numbers
| Workload | Throughput | Notes |
|---|---|---|
| Fine-tuning (7B) | good | fits comfortably in VRAM |
| Batch inference | high | memory-bound, not compute |
More detail — and reproducible configs — in the next post. Questions? Get in touch.