The metal underneath
Owning the serving stack is what lets us publish served precision, promise capacity and pin a request to a jurisdiction. Here is what it actually runs on.
| Region | Site | Accelerators | Power & cooling | Status |
|---|---|---|---|---|
| us-east | Ashburn, VA | H200 · GB200 (partial) | N+1 UPS, 2N utility feeds | Live |
| us-west | Hillsboro, OR | H200 · MI300X | N+1 UPS, hydro-backed grid | Live |
| eu-west | Frankfurt, DE | H200 | N+1 UPS, 100% renewable contract | Live |
| eu-north | Stockholm, SE | H200 | Free-cooling, renewable grid | Beta |
| ap-south | Mumbai, IN | H200 (planned) | TBA | Coming soon |
| me-central | Abu Dhabi, AE | TBA | TBA | Coming soon |
Serving stack
Continuous batching
vLLM-class runtime with paged attention and prefix caching, tuned per model family rather than one global config.
NVLink & RDMA
Intra-node NVLink and RDMA between nodes for tensor-parallel serving of 400B+ MoE models.
Prefix cache locality
Requests with a shared system prompt land on the same replica, which is where most of the long-context savings come from.
Active/active regions
Each live model runs in at least two regions. A site loss drains to the sibling; the gateway fails over to third parties if both are gone.
No prompt retention
Prompts and completions are held in memory for the request, then dropped. Logging is metadata-only unless you opt in.
Probed like everyone else
The same admission probes we run against third-party providers run against our fleet, on the same cadence, with the score published.
We grade ourselves in public
Our fleet is registered in the gateway as the provider llmcloud-oss and appears in the provider directory, sorted by the same trust score as everybody else. If it slips below the admission bar it loses traffic exactly like a third party would.
Do you own the data centers?+
We own and operate the accelerators, network fabric and serving stack; the facilities are colocation sites with contracted power, space and physical security.
What about carbon?+
EU and Nordic sites run on contracted renewable supply. We report energy per million tokens per region in the attestation log.
Can I audit it?+
Enterprise customers get the evidence bundle: probe results, capacity reports and site attestations, dated and signed.