AI Development Lab

Most AI teams lose more time to cluster scheduling than to training. A researcher waits for an allocation, gets four hours, and spends the first of them re-establishing an environment somebody else has since changed. The work that genuinely needs eight accelerators is a small fraction of the work that queues for them.

A GB10 machine changes that arithmetic because of its memory rather than its throughput. 128GB of coherent unified memory means the model, the optimiser state and the batch sit in one address space with no partitioning and no host-to-device copy to plan around. Up to 1 PFLOP of FP4 is useful, but the memory is the reason the work can happen at a desk at all.

One thing to settle before ordering rather than after: GB10 is Arm. The CUDA stack and the containers are the same, but anything in the pipeline that ships as an x86 binary needs checking, and some internal tooling and older vendor libraries will not have an Arm build. Where that check fails, the E403 is the answer in the same room: a single Xeon, up to 2TB of DDR5 and a full-height GPU, at 117mm tall.

The cluster then does what it is good at, which is long production runs submitted by people who have already validated their code. Utilisation goes up because the queue is no longer full of exploratory work that never needed rack-scale hardware.