Blog
Absract: Along with Thomas Grahame, I published articles in 2001 on data center electrical consumption and raised important policy questions about how the then build out of data centers would impact the electric grid. While the references to data center electrial usage and and size are outdated, many of the policy issues and solutions discussed are even more relevant 25 years later.
Introduction
In 2001, Thomas J. Grahame and I wrote a two-part series on the electricity demands of the first generation of large Internet data centers. Much of that series was descriptive — documenting how utilities were being blindsided by requests for hundreds of megawatts of power with almost no lead time. But buried inside it was something closer to a policy menu: a set of specific mechanisms utilities and regulators were experimenting with to manage large, dense, uncertain new loads without either bankrupting the utility or dumping the cost on everyone else.
Twenty-five years later, regulators are reaching for the same toolkit to manage data centers — at a scale we never anticipated. This piece goes through each tool, traces it from its 2001 origins to its current form, and asks a simple question of each: is it actually working?
Tool 1: Large-Load Tariff Classes
2001 origin. Our article described early efforts by the Seattle City Council and the Washington Utilities and Transportation Commission to develop new tariffs specifically for data centers, separate from standard commercial rate classes. The logic was straightforward: a customer requesting 40-50 watts per square foot creates a fundamentally different planning and cost profile than a typical office tenant at 4-6 watts per square foot, so it made sense to price them differently rather than force them into an existing rate class built around very different assumptions.
Where it stands now. This is the tool that has scaled up most cleanly. Large-load or "high-density" tariff classes for data centers are now common across major utility territories, and the fight has shifted from whether to have a separate class to how aggressively it should be structured — minimum contract terms, demand charges, exit fees, and credit requirements have all gotten sharper as utilities have gained more experience with these large load, AI-scale customers.
Open question. Our 2001 benchmark of 20 watts per square foot as a threshold for triggering special treatment is almost quaint against today's reality of 30 to 110+ kilowatts per rack. The thresholds that made sense to distinguish "genuinely dense load" from ordinary commercial growth in 2001 need a wholesale recalibration for the AI era — a rate class built around square footage assumptions from the server-farm period doesn't map cleanly onto a facility with a small footprint and enormous per-rack draw. This is arguably overdue work most jurisdictions haven't finished.
Tool 2: Upfront Payments and Minimum-Take Contracts
2001 origin. We described a common alternative to new tariff classes: requiring data centers to make upfront payments toward the cost of network upgrades, which would then be refunded through rates over time as the facility reached full occupancy and its actual power demand matched what had been requested. This was explicitly a hedge against the uncertainty we flagged as one of the biggest risks in the whole sector — how many requesters would drop out, how many leased centers would never fill up.
Where it stands now. This mechanism has not just survived, it's gotten more aggressive, and for good reason. Utilities that got burned by speculative dot-com-era requests are now asking AI-era customers for firmer commitments: minimum-take (take-or-pay) contracts, multi-year terms, and in some cases collateral or creditworthiness requirements that go beyond what a typical commercial customer would face. The uncertainty we described in 2001 — real demand tangled up with speculative reservations — hasn't gone away; if anything it's a live issue again given how capital-intensive and front-loaded AI infrastructure investment is, and how sensitive that investment cycle is to shifts in AI business models.
Open question. The dollar amounts involved now are large enough that getting the balance wrong matters more. Terms too soft, and utilities risk another round of stranded investment if AI capital spending cools. Terms too aggressive, and utilities risk pricing out legitimate long-term customers (see Tool 3).
Tool 3: Dedicated and Self-Supplied Generation
2001 origin. Our article flagged distributed generation (DG) as an emerging response to reliability requirements — data centers were already installing onsite backup generation to guarantee the "six nines" uptime and reliability their business models demanded. We noted this was tightly constrained by environmental regulation: New York regulators, for example, only allowed backup generators to run during declared emergencies or outages, not as a routine hedge.
Where it stands now. This has grown from a backup-power afterthought into a primary strategy. "Bring Your Own New Capacity" (BYONC) arrangements, dedicated onsite gas turbines, and behind-the-meter generation deals are now central to how the largest AI data centers are being sited and financed, driven less by backup-power logic and more by a simple fact: building or contracting for new capacity, particularly on-site or nearby, is often faster than waiting years in an interconnection queue.
Open question. The regulatory tension we described in 2001 — environmental permitting limiting when onsite generation can run — has become a bigger fight now that onsite generation is a primary power source rather than an emergency backstop. Permitting new gas capacity at the scale to serve large load data centers raises air-quality and emissions questions that regulators are still working through, and the answer looks different depending on whether the onsite unit is framed as "backup" or "baseload." The use of on-site batteries is becoming a viable and economic option as relability backstop and to address environmental constraints.
Tool 4: Interconnection Queue Reform
2001 origin. This tool essentially didn't exist in our article, and that absence is itself informative. In 2001, we treated transmission siting and generation construction timelines (3-4 years) as the binding constraint, and discussed opposition to new transmission lines in Northern Virginia as the closest analog to today's queue problem. There was no equivalent then to a formal, multi-year interconnection study process as the primary bottleneck — the constraint was mostly about permitting and construction, not about getting in line to connect to an already-crowded grid.
Where it stands now. Interconnection queues have become arguably the single most-discussed bottleneck in the current cycle, with wait times stretching to years in many regions — PJM's interconnection queue stretched to 4-7 years — long enough that it exceeded the construction timeline for the generation or transformer equipment itself. This has pushed FERC and individual utilities toward fast-track interconnection processes and flexible interconnection specifically carved out for large loads, a genuinely new policy category that has no real precedent in the 2001 story.
Open question. This is the area with the least settled track record. Fast-track and flexible interconnection processes for large loads are new enough that it's not yet clear whether they'll meaningfully compress timelines or simply reshuffle who's at the front of the line — and there's a legitimate concern that jumping AI data centers ahead of other pending interconnection requests (industrial facilities, EV charging infrastructure, renewable generation) just moves the queue problem onto someone else.
Tool 5: Demand Transparency and Forecasting
2001 origin. This was the most explicitly prescriptive part of our original article. We argued that utility planners and engineers were poorly positioned to forecast fast-moving technology demand on their own, and proposed that the broadband industry put together a technical committee to map internet growth, estimate current electricity use per unit of internet activity, and project how that ratio would change with future technology — specifically so utilities could plan further ahead than year-to-year requests allowed.
Where it stands now. That specific proposal never materialized in the form we described, but pieces of the underlying function now exist in a more fragmented way: EPRI, national labs, and organizations like the International Energy Agency now publish regular data center and AI electricity demand forecasts, and some jurisdictions are beginning to discuss mandatory long-range load disclosure requirements for very large customers as a condition of interconnection.
Open question. This is the tool with the biggest gap between what we called for and what exists today, and I'd argue it's the most consequential one. The organizations with the best visibility into future data center electricity demand — the hyperscalers, chipmakers, and AI labs themselves — are not the ones doing utility-grade long-range forecasting, and utilities are still largely reacting to requests rather than planning against credible multi-year demand curves supplied by the industry that's driving the growth. Closing this gap would do more to prevent the next round of scramble-and-catch-up than any tariff or interconnection reform, because it addresses the root problem: utilities finding out about demand only when it shows up as a formal request.
A Scorecard
|
Tool |
2001 status |
Today's status |
Verdict |
|
Large-load tariffs |
Emerging, city-by-city |
Widely adopted, but our thresholds outdated |
Landed, needs recalibration |
|
Upfront payments / minimum-take |
Early experimentation |
Standard practice, terms tightening |
Landed, appropriately so |
|
Self-generation / DG |
Backup-only, tightly regulated |
Primary strategy at scale, BYONC encouraged |
Transformed, regulation lagging |
|
Interconnection reform |
Didn't exist as a category |
Most active area of policy reform |
New, unproven |
|
Demand forecasting |
Proposed, never built |
Fragmented, reactive |
Largest unfinished lesson |
Read across the row, the pattern is consistent: the tools aimed at protecting utilities and ratepayers financially (tariffs, upfront payments) have matured reasonably well over 25 years. The tools aimed at anticipating demand before it shows up as a crisis have not. That's not a coincidence — financial protection tools get built by the party bearing the risk (the utility), while forecasting tools require voluntary cooperation from the party that has the information (the tech sector) and no strong incentive to share it early. If this cycle is going to go better than the last one, that's the gap most worth closing.
