Vetted H100 and H200 nodes in our Baku datacenter. Need 256 H100s for a week? Done. Eight H200s for a year? Sure. Spin down whenever — you're never locked into a long-term contract.
Reserve nodes at market price, or hold capacity only while it stays under your limit. Bring your keys, your data stays put — there's never an egress fee.
# buy 16 nodes (128 H100s) for 5 days at market price $ bc nodes create -n 16 --zone baku-1 -d 5d # hold 2 nodes (16 H200s) only while price stays below $2.10/gpu·hr $ bc nodes create rl-0 rl-1 --auto --zone baku-1 -p 16.80 # sell back what you don't need — no lock-in $ bc nodes sell rl-1 -d now
Spot prices float with demand on the cluster. Reserve for a fixed term to lock a rate.
| GPU | Node | VRAM / GPU | Interconnect | From (gpu·hr) |
|---|---|---|---|---|
| H100 SXM | 8× | 80 GB | 3.2 Tb/s IB | $1.85 |
| H200 SXM | 8× | 141 GB | 3.2 Tb/s IB | $2.40 |
| A100 SXM | 8× | 80 GB | 1.6 Tb/s IB | $1.10 |
| L40S | 8× | 48 GB | PCIe | $0.65 |
VMs spin up in ~5 min · bare-metal & managed Slurm on request · failed nodes refunded and repacked.
Burst to hundreds of interconnected GPUs for a single training run, then hand them back. No annual commitment to amortize.
Hold dedicated nodes at a reserved rate for production inference, with headroom to scale during peaks.
Per-hour billing that fits grant cycles. Built for teams at ADA, the academies, and regional research groups.
Baku sits on the fiber routes linking Frankfurt, Dubai and the Caspian corridor — low-latency to three regions at once, on power that's among the cheapest in the neighborhood.
Send your workload, GPU count and timeframe — we'll come back with availability and a price.