The short version: The public conversation about AI data centers has collapsed into a single number — how many gigawatts a campus can secure. But nameplate power is the easy part. The two harder problems, both missing from most buildout headlines, are how violently AI training load swings second to second, and how much of the installed compute never does useful work.
The gigawatt story everyone tells
Read anything about AI in 2026 and the story is identical: demand is exploding and the race is to lock down power and pack racks denser. The numbers are huge. Goldman Sachs Research projects data-center power demand will rise as much as 165% by the end of the decade against a 2023 baseline. Deloitte has forecast that next-generation AI racks could reach 370kW in 2026, with U.S. AI data-center demand climbing from a few gigawatts to well over one hundred by the mid-2030s. Individual accelerators now draw on the order of a kilowatt each, which is why direct-to-chip liquid cooling has gone from exotic to a requirement.
All of that is real, and all of it measures the same thing: how much power a site can pull. What it does not describe is how that power behaves once the GPUs actually start training — and that is where the interesting engineering lives. For a sober view of the demand curve itself, the IEA’s Energy and AI analysis is a better starting point than most vendor decks.
Why AI compute sites don’t behave like traditional halls
A conventional enterprise data hall draws a fairly flat baseload; utilization drifts gently across the day. An AI training cluster does the opposite. Tens of thousands of GPUs execute the same job in lockstep, so they ramp up and drop out together — waiting on a checkpoint, a gradient sync, or a collective-communication step. The result is a coordinated power oscillation.
This is not theoretical. Meta’s Llama 3 work documented power-fluctuation problems on a 24,000-GPU cluster serious enough that engineers reportedly added a dummy-workload switch to keep the draw from collapsing between steps. Research on power stabilization for AI training describes swings of tens to hundreds of megawatts inside seconds — sometimes a fraction of a second. Data Center Dynamics has covered the same phenomenon under the apt label “the AI power swingers”. At the wrong frequency those swings can resonate with grid components, which is why the topic has moved from operator forums into Uptime Institute’s electrical guidance.
Volatility reshapes the thermal problem too. Air-cooled rooms can absorb minutes of disruption before hitting limits; liquid loops can tolerate only seconds. A load that spikes and collapses on a second-by-second basis, cooled by a system with a second-scale safety margin, leaves almost no room for the slow, human-paced response that older facilities were designed around.
The utilization gap inside AI factories
The second blind spot is quieter but just as expensive: a lot of the compute that gets built never runs. Industry estimates put typical GPU utilization at roughly 60–70% against a realistic ceiling in the mid-90s, and some analyses suggest 30–40% of provisioned GPU capacity can sit idle thanks to overprovisioning, scheduling gaps, and data-pipeline bottlenecks. Given that GPUs account for roughly 40% of the power draw in an AI facility, idle silicon is not just stranded capex — it is stranded power and stranded cooling, provisioned and paid for but converting electricity into little more than heat.
The same gap shows up on the inference side: fleets sized for peak query bursts spend much of the day underused, and a model that under-fills its accelerators wastes memory bandwidth long before it saturates its compute. That reframes efficiency. The metric that matters is not how many gigawatts you secured but how much useful work — tokens, training steps, inferences — you extract per watt. A campus running at 60% utilization has effectively overpaid for every part of its stack, from substation to chiller.
What the AI data centers buildout actually measures
Gigawatts, in other words, is a proxy for ambition, not for performance. It says nothing about whether the load is smooth or spiky, and nothing about whether the expensive hardware is busy. Even PUE — the industry’s default efficiency ratio — is a facility-level number: it can look excellent on a site whose GPUs are half-idle and whose power profile is sawtoothing the local grid.
There is also a contractual mismatch hiding in the averages. A site can hold a firm gigawatt-scale power contract while its actual draw swings far below that ceiling for hours at a time — capacity booked, grid reserved, utilization thin. The operators pulling ahead are the ones treating power as a system to be shaped rather than a quantity to be bought: securing capacity years in advance, yes, but also smoothing training load so it doesn’t slam the grid, and scheduling aggressively so the silicon stays fed. Those are operating disciplines, not procurement wins.
What operators should actually watch
For anyone running or specifying AI infrastructure, a few things follow directly from the two gaps above.
- Measure at the timescale the load moves. Fifteen-minute power averages and room-level temperature readings hide exactly the second-scale swings and per-rack hotspots that define AI behavior. Observability has to match the physics.
- Treat load-shaping as first-class. Scheduling, checkpoint staggering, and the power-smoothing features now shipping with the silicon exist specifically to blunt synchronized swings; using them is part of the operating model, not an optional tweak.
- Budget for ride-through, not just capacity. With only seconds of thermal margin, buffering and failover behavior deserve the same design attention as raw megawatts.
- Score useful work, not nameplate. Utilization and work-per-watt tell you whether the buildout is paying off in a way that a capacity figure never will.
None of this shows up in a gigawatt headline. But volatility and utilization are where the money — and the risk — actually sit. The AI data centers that win the next few years will not simply be the biggest; they will be the ones that keep a violently variable load steady and a very expensive fleet of chips genuinely busy.