The new constraint is speed to power
Compute equipment can be ordered faster than major grid assets are planned and built. Campus economics therefore depend on interconnection timing, substation equipment, firm capacity, cooling design, and whether workloads can move in time or geography.
Three numbers that locate the frontier
These are modeled scenarios, not commitments or measured 2030 demand. Outcomes depend heavily on shipments, utilization, chip life, and efficiency.
Measure delivered accelerator-hours, not nameplate megawatts
Useful capacity is the intersection of energized IT, memory and network balance, thermal headroom, software utilization, and job completion. A stranded accelerator and an unenergized building produce the same output.
Training favors tightly synchronized networks; inference can often be more geographically distributed.
Higher rack density shortens links but concentrates power delivery and heat removal.
The headline metric sits on a system
Each layer can become the bottleneck even when the layer before it improves.
Grid and onsite power
Generation, transmission, interconnection, substations, storage, and backup establish electrical capacity.
- Measure
- MW · time to energize
- Failure mode
- Queues and firm supply
Facility
Switchgear, UPS, cooling, water systems, and buildings deliver conditioned power and remove heat.
- Measure
- PUE · W/rack
- Failure mode
- Thermal density
Compute fabric
Accelerators, CPUs, HBM, switches, and optical links act as one machine.
- Measure
- Useful accelerator-hours
- Failure mode
- Memory and network balance
Workload system
Schedulers, checkpointing, data pipelines, and service objectives turn hardware into completed jobs.
- Measure
- Jobs/MWh · tail latency
- Failure mode
- Utilization and reliability
Every watt entering compute exits as heat
Power conversion can improve, but the first law remains: nearly all electrical energy consumed onsite ultimately becomes heat that must leave the equipment and facility.
Compute procurement moves upstream into infrastructure
The binding lead time increasingly sits in interconnection studies, transformers, transmission, construction, cooling equipment, and community permission rather than in server assembly alone.
Flexible workloads
Shift delay-tolerant jobs across hours or regions.
Balanced systems
Buy memory and network capacity that keeps expensive compute occupied.
Thermal co-design
Match rack density, liquid loops, heat rejection, climate, and water constraints.
Firm interconnection
Coordinate grid investment, onsite assets, tariffs, and realistic ramp profiles.
An optimistic view, with conditions
The data center becomes a grid participant
Facilities that expose safe flexibility, stage load growth, and coordinate onsite assets can connect sooner and use infrastructure more productively while bulk supply catches up.
Sources, method, and boundaries
Energy quantities are Berkeley Lab scenarios. Flexibility mechanisms come from its 2026 grid-integration review; no scenario is presented as a forecast certainty.
