Is Data Center Power Forecasting Overstated?

The staggering numbers from utility companies about data center power demand may be misleading. Not misleading in a rounding-error kind of way, but in a manner that quietly reshapes how billions of dollars are allocated across the grid, real estate markets, and clean energy infrastructure.

When a single party controls the forecast, the forecast serves that party. That's not a conspiracy theory; it's just how institutional incentives work.

The Power Forecasting Problem Nobody Talks About

Data centers are among the most energy-intensive facilities ever built at scale. A hyperscale facility can consume anywhere from 100 MW to over 1 GW — enough electricity to power a mid-sized American city. Microsoft, Amazon, Google, and Meta are collectively spending hundreds of billions of dollars on new capacity through 2030. Every one of those projects requires a power interconnection study, a load forecast, and a utility agreement.

Here's the structural problem: the utility company is typically the entity producing or heavily influencing those power consumption forecasts, often working from their own load modeling assumptions, their own grid constraints, and — critically — their own financial interests. Utilities are regulated businesses. More load means more infrastructure investment. More infrastructure investment means a larger rate base. A larger rate base means higher allowable profits.

When the entity doing the forecasting also profits from higher forecast numbers, independent verification isn't optional — it's essential.

This isn't unique to data centers, but the distortion is particularly consequential because of how fast the sector is growing and how long the planning horizons are. A forecast that overstates demand by 20% doesn't just waste paper — it triggers transmission line upgrades, substation builds, and long-term power purchase agreements that ratepayers eventually fund.

What Utility-Driven Forecasts Actually Capture — and What They Miss

Utility load forecasting is a legitimate discipline with real methodology behind it. Companies use historical consumption patterns, economic growth models, and regional demographic data to project future demand. For residential and commercial customers, these models work reasonably well over medium-term horizons.

Data centers break the model.

Unlike a shopping mall or an office building, a data center's actual power draw is radically elastic. Server utilization rates fluctuate. Workloads shift between facilities. A data center that signs a 50 MW interconnection agreement may actually run at 30 MW for years before scaling up — or it may never reach contracted capacity at all. The interconnection request, which is what utilities often use as the basis for their load projections, represents a ceiling, not a prediction.

Forecasting peak contracted capacity as expected load is like projecting highway congestion based on the number of driver's licenses issued.

The implications cascade. Grid operators plan transmission infrastructure around projected peak demand. Utilities file rate cases citing load growth. Investors in adjacent energy infrastructure — battery storage projects, solar farms, gas peakers — make capital allocation decisions based on these numbers. When the underlying forecast is structurally biased toward the high end, every downstream decision inherits that error.

The Real Drivers of Data Center Energy Consumption

Understanding what actually determines a data center's power draw requires getting into the facility itself, not just reading the interconnection application.

Server Density and Compute Efficiency

The single biggest variable in data center energy consumption isn't square footage — it's what's happening inside the racks. A facility built around first-generation AI training chips runs at a fundamentally different power density than one optimized for inference workloads or traditional enterprise applications. NVIDIA's H100 GPU draws 700 watts per chip. A rack dense with H100s can hit 100 kW. Standard enterprise servers might average 5–10 kW per rack.

Efficiency gains compound this complexity. Moore's Law may be slowing for raw compute, but performance-per-watt has continued improving. A workload that required 10 MW of compute in 2018 might require 4 MW today on modern silicon. Forecasts built on older efficiency assumptions bake in energy demand that evaporates as operators refresh hardware.

Cooling: The Hidden Energy Budget

Cooling systems typically account for 30–40% of a data center's total power consumption, measured by a metric called Power Usage Effectiveness (PUE). A facility with a PUE of 2.0 uses one watt of overhead for every watt of compute — essentially doubling the energy footprint. Modern hyperscale facilities often achieve PUEs below 1.2, with some liquid-cooled designs approaching 1.03.

The shift from air cooling to direct liquid cooling — accelerated by the thermal demands of AI hardware — is already compressing facility-level energy consumption in ways that older forecasting models haven't absorbed. A utility projecting load growth based on air-cooled PUE assumptions for a liquid-cooled AI cluster will systematically overstate demand.

Why Current Forecasting Methods Fall Short

The most widely used forecasting approaches in utility planning rely on some combination of bottom-up load building (aggregating individual customer projections) and top-down econometric modeling. Both have significant blind spots when applied to data centers.

Bottom-up approaches depend on information that data center operators often can't or won't provide in detail. Facility utilization rates, workload migration plans, hardware refresh cycles — these are competitively sensitive. Operators aren't going to hand a utility company a detailed capacity roadmap that might end up in a public rate case filing.

Top-down models have a different failure mode: they treat data centers as a monolithic sector rather than a heterogeneous collection of facilities with wildly different operational profiles. A co-location facility serving hundreds of small enterprise tenants looks nothing like a hyperscale campus running LLM training. Averaging across them produces a number that accurately describes neither.

The result is forecasts that are precise in presentation and unreliable in substance — which is arguably worse than a wide confidence interval, because it creates false certainty.

Regulatory frameworks compound the problem. In many states, utility resource plans are filed years in advance and are difficult to revise. Once an inflated load forecast enters the regulatory record, it takes on a life of its own — driving infrastructure decisions long after the underlying assumptions have become obsolete.

Building a More Honest Forecast

Fixing data center power forecasting isn't primarily a technology problem. The tools exist. What's missing is the institutional structure to use them correctly.

Bring Multiple Parties to the Table

The most consequential change would be requiring independent demand forecasting — analysis conducted by parties without a financial stake in the outcome — as a standard component of major data center interconnection reviews. Grid operators like PJM and MISO have begun incorporating more granular load analysis into their interconnection queues, but the process remains heavily dependent on information self-reported by utilities and developers.

Third-party energy analysts, academic modeling groups, and independent system operators can provide meaningful cross-checks. Several European grid operators have moved toward multi-stakeholder forecasting processes precisely because single-party projections proved consistently unreliable for large industrial loads.

Leverage Operational Data, Not Just Contracted Capacity

The gap between contracted capacity and actual consumption is measurable and consistent enough to model. Research from Lawrence Berkeley National Laboratory has documented that data centers routinely operate well below their interconnection capacity, particularly in their first years of operation. Building that utilization curve into standard forecasting methodology — rather than treating contracted capacity as expected demand — would immediately improve accuracy.

Technology helps here too. Advanced metering infrastructure, real-time consumption monitoring, and machine learning models trained on operational rather than contracted load profiles can produce probabilistic forecasts with explicit confidence intervals. That kind of transparency is more useful to grid planners than a single-point projection with false precision.

Align Incentives with Accuracy

Ultimately, the structural issue is incentive alignment. As long as utility revenues are tied to infrastructure investment driven by load forecasts that utilities themselves produce, the pressure toward optimistic demand projections doesn't disappear — it just gets institutionalized.

Some regulators are experimenting with performance-based ratemaking structures that reward utilities for forecast accuracy rather than infrastructure spending. That's a slow reform. But for investors in solar, storage, and transmission infrastructure betting on data center load growth, the near-term implication is clear: scrutinize the forecast provenance before committing capital.

The data center boom is real. The power demands are substantial and growing. But real doesn't mean accurately measured, and substantial doesn't mean the numbers on utility filings are right. For developers, investors, and grid planners, the most expensive mistake isn't missing the trend — it's building infrastructure around a forecast that was structurally designed to be wrong.

The facilities that will anchor the next decade of digital infrastructure deserve forecasts that match their complexity. So far, the methodology hasn't caught up to the moment.

Explore the InfraSale Marketplace for insights and solutions.