Four regions, Tier III+ facilities, and a fabric built for collective operations rather than web traffic.
| Region | Facility | Typical latency | GPU classes | Status |
|---|---|---|---|---|
| us-east-1 | US East (Ashburn, VA) | 12 ms | H200, H100 SXM, H100 PCIe, A100 80GB, A100 40GB, L40S, RTX A6000, RTX 4090, RTX 3090, V100 | Operational |
| us-west-2 | US West (Hillsboro, OR) | 38 ms | H100 SXM, A100 80GB, L40S | Operational |
| eu-west-1 | EU West (Dublin, IE) | 84 ms | H200, H100 SXM, H100 PCIe, A100 80GB, A100 40GB, L40S, RTX A6000, RTX 4090, RTX 3090 | Operational |
| ap-south-1 | AP South (Mumbai, IN) | 146 ms | H100 PCIe, A100 80GB, RTX 4090 | Degraded |
N+1 UPS with diesel backup rated for 72 hours at full load. Two independent utility feeds per facility. Annual PUE across the fleet sits at 1.18.
Direct-to-chip liquid cooling on all Hopper racks, rear-door heat exchangers elsewhere. Sustained 700 W per card without thermal throttling.
NDR InfiniBand at 400 Gbps per port, eight ports per SXM node, in a rail-optimised fat tree. Non-blocking within a pod of 512 GPUs.
All-NVMe parallel filesystem, 2 PB usable per region, delivering 180 GB/s aggregate read to a single pod.
Biometric plus badge at every threshold, mantrap entry, 90-day CCTV retention, escorted access only for third-party engineers.
Cards are burned in for 72 hours before entering the pool and retired at five years or on the first uncorrectable ECC event.
Instances are scheduled onto hosts by a bin-packing scheduler that prefers to fill a partially used node before waking a cold one. That keeps power draw down and, more usefully for you, means multi-GPU requests are far more likely to land on a single node with NVLink between the cards.
When a region runs short, requests queue rather than failing. You are never billed while queued, and you can cancel at any point. The catalogue marks a card as limited when median queue time in the last hour exceeded two minutes.
Host maintenance is announced 14 days ahead on the status page and never interrupts a running instance — hosts are drained by refusing new placements until existing work finishes. Emergency firmware patches are the exception and are announced as they happen.