Dense accelerator fabrics that train sovereign models and serve inference inside the nation's own borders. The owner holds the weights; the cluster never leaves the jurisdiction.
Training throughput is bounded by the fabric that binds accelerators into one coherent pool, so the network is designed first.
GPUs are wired into a rail-optimized, non-blocking topology so all-reduce and all-to-all collectives run at line rate. Fabric bisection is provisioned to match the accelerator generation rather than reused from an older CPU network.
Each accelerator's high-speed NIC is bound to a dedicated switch rail, keeping collective traffic on symmetric paths. Placement is topology-aware so a training job lands on adjacent rails instead of crossing the spine unnecessarily.
Inter-node tensor traffic uses RDMA with congestion control tuned for bursty collective patterns. Loss-sensitive flows are isolated from control and storage traffic on separate virtual lanes.
The fabric is partitioned into failure domains sized to a training pod, so a switch or link fault degrades one pod rather than the whole hall. Domain boundaries map to power and cooling zones for correlated-failure isolation.
Distributed training is all-or-nothing, so the scheduler treats a multi-node job as a single indivisible unit.
Multi-node jobs are placed atomically: either every rank gets an accelerator on the right rail or nothing starts. The placement engine reads live fabric topology so ranks land with minimal hop count.
Accelerators can be sliced into isolated instances for smaller inference tenants and recombined for large training runs. Partition boundaries enforce memory and compute isolation between workloads on the same card.
Long runs checkpoint on a cadence tied to the mean time between fabric faults, so a failed rank resumes from recent state. Lower-priority inference is preempted and requeued when a sovereign training job needs the pool.
Owner workloads run in hard-isolated queues that cannot be co-scheduled with untrusted tenants on the same node. Queue policy is set by the operator institution, not by the platform.
The latest accelerators concentrate heat past the point air alone can carry, so the hall is engineered around liquid.
Cold plates carry heat directly off the accelerator and CPU packages, letting racks run at densities air cannot sustain. Coolant distribution is redundant so a single loop fault does not throttle a pod.
Racks are provisioned for sustained accelerator draw under full training load, not nameplate averages. Feeds are dual-corded to independent paths so a distribution fault does not drop a training pod.
Per-card power, temperature, and clock telemetry stream to the control plane, which caps clocks before thermal limits force an unplanned drop. Operators see the same telemetry the scheduler acts on.
Sovereign AI means the model, its data, and the keys that protect them stay under the operator's control by construction.
Model checkpoints are written only to in-jurisdiction storage and never replicate to external regions. Export of weights requires an explicit, logged action by the owning institution.
Cluster control messages and artifact signatures use ML-DSA-65 (FIPS 204) so job provenance survives a future quantum adversary. Signed checkpoints let an owner prove which run produced a given model.
Encryption keys for datasets and checkpoints are held by the operator, not the platform, so no external party can read training data at rest. Key custody is designed for hardware-backed roots under the owner's control.
Every job submission, placement, and checkpoint is recorded on a hash-chained ledger, giving an auditable history of what ran and on what data. The chain makes silent retroactive edits detectable.