Strategy, Workload Archetypes & Economics

The workload sets density, cooling, fabric, redundancy and siting; your rack, measured traffic and uptime target size each. Get the archetype wrong and the fix is moving, a new utility service request, floor rework, or tearing out power or cooling plant under a running hall.

A training cluster is one synchronous supercomputer that moves at its slowest GPU's pace. Design for goodput per megawatt, dense and checkpointable, with cooling and fabric chosen from your rack's specs and measured traffic, before you cut steel.

An inference data center serves many independent requests against a latency SLO. Size it to goodput per dollar and tokens per watt at real batch sizes, not to peak training throughput; designing it as a de-rated training cluster gets nearly every subsystem choice wrong.

Post-training is two workloads under one name. Supervised and offline preference tuning are small, bursty training jobs; rollout-heavy RL spends most of its time generating, which is inference. Size each phase from measured GPU-seconds, or you buy a trainer's fabric for an inference fleet.

Edge inference trades cheap power and consolidation for proximity and local control. It pays when a centralized region fails a latency, data-boundary, WAN-autonomy or input-transport-cost requirement.

Match capacity to the durability of your demand: compare ownership for durable, contracted load and rental for uncertain load, valuing only the exit option the contract actually grants.

The workload archetype narrows every subsystem choice but settles none. Cooling, fabric, storage, redundancy and siting each need measured evidence, frozen in a signed design basis; fill a cell from the label alone and you strand capital or starve the GPUs.

Four numbers decide an AI data center's return: capex per watt, assumed economic earning life, billable utilization and post-deflation price. Get any one wrong and the asset strands. Model book depreciation and debt service separately.