Your tenants share one fabric and must never see each other on it. Genesis Grid enforces that in the switch and in the NIC: InfiniBand partition keys on the training fabric, per-tenant segmentation on Ethernet, and a policy surface your tenants configure themselves. The performance clause in every tenant contract you sign rests on this layer.
What the fabric enforces is membership, not fairness: partition keys separate tenants, they do not reserve bandwidth. So the figure you sell is the one measured on your own fabric under simulated contention, and where your topology forces two tenants through one congestion domain, we tell you which tenant classes can carry a latency commitment and which cannot.
Ethernet or InfiniBand, existing switches, existing NICs. There is no forklift and no dependency on a single silicon vendor. Where a capability is genuinely missing, you hear it before the contract rather than during the rollout.
Multi-tenancy is enforced inside a compute cluster. If you operate several sites you operate several clusters, each with its own fabric and its own failure domain, and a problem in one never travels to another.
Long-term tenants and on-demand tenants share the fabric without seeing each other, so you keep the committed base that finances the build and still capture the higher average hourly price the short-term market pays.
The envelope depends on the fabric already in your racks, so we benchmark it under simulated multi-tenant contention before anyone signs. What enters your tenants' service level is the measured figure, never the datasheet figure.
RDMA bandwidth per accelerator on the training fabric
per-node tenant traffic carried in user space with VPP
to benchmark your fabric under multi-tenant contention
With partition keys enforced on the fabric. Each tenant gets its own partition, and a node outside it cannot address the traffic. That mechanism surprised NVIDIA's own engineers when we showed it to them. On the Ethernet side the equivalent is per-tenant segmentation with default-deny east-west.
.svg){kind=small}
Yes, inside a cluster, against the figure we measured on your fabric under contention and committed per deployment. Partition keys give you separation, not a bandwidth reservation, so the number comes from the measurement rather than the mechanism. What you cannot promise is one fabric spanning sites: separate sites are separate clusters, each with its own fabric.
You do. You keep the facility, the physical links and the carrier contracts, and you operate the fabric; Genesis carries escalation when the fault is in the stack. What that takes in headcount is set out on the data centres page.
Not by accident. There is no cross-site scheduling in the stack, so a workload never migrates across a border to fill a gap in demand. What leaves a site is what your tenant explicitly sends.
