AI data centers are making power demand harder to predict, not only larger, a report finds.
Capgemini Research Institute said nearly 80% of utilities expect more extreme and volatile demand patterns, while more than three-quarters say they struggle to forecast future needs accurately.
According to the findings, this uncertainty is changing the data center capacity conversation. Capgemini also discovered 67% of electricity executives refer to “phantom” data-center load requests, with around 19% of data-center power requests, on average, never materializing.
For Paul Cook, global director of technology and innovation at Black & White Engineering, the issue now is “what happens after capacity is secured,” and whether owner-operators have enough visibility to use it properly.
In a comment shared with TechInformed, Cook said businesses are asking more detailed questions about how assets will perform, “how much capacity they can use safely” and how much freedom they will have to adapt as customer requirements change.
The illusion of secured capacity
The pressure is already visible in the market. CBRE said North American inventory across the four largest data center markets rose 33% year over year in the first quarter, while Latin American inventory across its four largest markets rose 41.3%.
Even with that new supply, CBRE said vacancy in the top four U.S. markets fell to an all-time low, including 0.3% in Northern Virginia and 1% in Atlanta.
Cook argued that market supply and vacancy data do not show whether new capacity will translate into usable AI compute once high-density racks are live.
He said: “securing capacity is no longer only a development question.” Owner-operators also need to know how a facility will behave in operation, “how much of that capacity can be used safely” and where constraints are likely to appear across load, cooling, controls and system performance.
High-density racks rewrite facility rules
ASHRAE’s AI data center framework says AI and high-performance computing are changing facility assumptions because rack densities have escalated from about 120 kW to several hundred kilowatts, with megawatt-class racks expected in the near term. ASHRAE also says power and cooling can no longer be treated as separate domains because decisions in one directly affect the other.
Cook adds to this: “land, power and water still dominate data centre development, but they do not show how much usable capacity a facility really has once it is running.” Without that visibility, he said, operators can run too cautiously, leave capacity unused or make decisions later than they should.
When physical constraints hit model performance
AI workloads make this visibility more important because facility constraints can increasingly affect the compute conditions behind model serving. A facility may have contracted power, but user experience depends on whether GPUs can run reliably, whether cooling can sustain dense workloads and whether systems can expose constraints before they affect performance.
The constraints can show up in user-facing AI performance metrics. Nvidia’s LLM benchmarking documentation says time to first token generally includes request queuing time, prefill time and network latency.
The same documentation says longer prompts usually increase time to first token because the model must process the input sequence and create a key-value cache before generation begins.
Nvidia’s metrics make clear that, at the serving layer, model responsiveness is shaped by infrastructure factors such as queuing, latency and GPU resource saturation. When GPU resources saturate, Nvidia says system throughput can fall. In AI data centers, ASHRAE’s framework makes clear that power and cooling choices also shape whether high-density workloads can be supported reliably.
Tracking telemetry and operational data
The operational evidence is already available in cooling and GPU telemetry guidance. ASHRAE recommends liquid cooling, including direct-to-chip systems and rear-door heat exchangers, to support 50-100+ kW racks, while Nvidia says advanced GPU profiling metrics from Data Center GPU Manager can provide visibility into SM utilization, memory bandwidth, tensor core activity and compute pipeline behavior.
This kind of telemetry gives operators and buyers a more practical vocabulary for discussing usable capacity, including sustained GPU availability, latency, cooling headroom and operational constraints.
Cook said owner-operators also need to protect their ability to access, structure and use their own operational data. “If data is difficult to extract, poorly structured or locked into a proprietary environment,” he said, “it can limit future decisions and make it harder to compare performance, bring in new tools or respond to changes in workload.”
His broader point is that operational intelligence has to influence design and delivery, not sit with the operations team after handover. As demand for new space and power continues, Cook said the next test is how capacity is understood, controlled and used once it is built.