Dell forecasts AI will make up 75 percent of data centre demand by 2030

Power is the binding constraint on AI data centres. A rack of AI accelerators can draw several times the power of a rack of conventional servers, so capacity is now described in megawatts and gigawatts rather than square metres. Utility interconnection agreements, transformer lead times and substation build-outs typically take years, far longer than the time needed to order servers. That mismatch explains why a hardware vendor is talking about grid capacity on an earnings call. When power is scarce, the sites that already have it become the most valuable asset in the chain.

Training and inference load a facility differently. Training runs a small number of enormous jobs that keep every accelerator busy for weeks and demand very fast networking between them. Inference serves many independent requests, is measured in tokens produced, and scales with user demand rather than model size. Dell's split of an 87-fold rise in inference tokens against a fivefold rise in training compute reflects the expectation that inference grows with adoption. Inference capacity can be distributed closer to users, while training clusters tend to concentrate where power is cheapest.

Server consolidation is a power play, not just a cost saving. Each processor generation delivers more cores and better performance per watt, so one modern machine can absorb the work of several older ones. Retiring a dozen old servers for one new one returns the electrical and cooling headroom those machines used. In a constrained site, that recovered power is what makes room for accelerators. This is why vendors pair AI pitches with refresh pitches: the two compete for the same watts.

Backlog and bookings describe future load, not current load. A backlog is the value of orders received but not yet delivered, and bookings are orders taken within a period. For data centre planners, a large backlog signals hardware that will arrive and need power and cooling in the coming quarters. Read alongside a power forecast, it is a rough leading indicator of where capacity pressure will appear next. It also shows why supply chains for transformers and chillers are watched as closely as chip supply.

Neoclouds and sovereign deployments change where capacity is built. Neoclouds are smaller providers that rent accelerator capacity, and sovereign projects are national or regional deployments meant to keep data and compute within a jurisdiction. Dell's estimate that half of the new power demand comes from these buyers plus enterprises implies many mid-sized facilities rather than a handful of giant campuses. That pattern spreads grid impact across more regions and more utilities. It also means more organisations need in-house skills for siting, power and cooling.