On November 11, 2026, insurance broker Howden published research accompanying its report “Insuring the data centre supercycle” that reduces the data center power problem to a single ratio. Since 2000, United States data center power demand has risen by about 19 percent a year. Over the same period, the electricity grid has expanded capacity by about 2 percent a year. Demand has grown roughly ten times faster than the wires built to carry it, with delays, lost revenue, and a rush to behind-the-meter power as the consequences.
The ratio is worth pausing on, because it sounds abstract until you compound it. Twenty-five years of 19 percent annual growth multiplies demand by roughly 77 times. Twenty-five years of 2 percent annual growth multiplies grid capacity by about 1.6 times. The ratio of those two multipliers is close to 47 — which is, within rounding, the 50-fold increase Howden reports in data centers’ share of total United States grid capacity, from 0.1 percent in 2000 to 5 percent in 2025. The figures come from Howden’s analysis of EIA data and 451 Research by S&P Global. This is a gap that has been opening for a quarter of a century.
The scale is corroborated by the federal benchmark. Lawrence Berkeley National Laboratory’s 2024 report for the Department of Energy put United States data center electricity use at 176 terawatt-hours in 2023, about 4.4 percent of national consumption, up from 58 terawatt-hours in 2014 — and projected 325 to 580 terawatt-hours by 2028, or 6.7 to 12 percent of the total. The two measures agree on direction and order of magnitude. What has changed is the speed: the AI buildout compressed more than half of the industry’s construction history into 2020–2025, Howden notes.
One caveat is worth stating before going further, because the ratio invites a misreading. Howden’s 19 percent is a growth rate in power demand; its 2 percent is a growth rate in grid capacity. Demand and capacity are not the same unit, and a load that draws nearly flat around the clock is not the same burden on the system as a load that peaks for a few hours. But the caveat cuts both ways: a flat load still contributes to every system peak there is. Whichever way the load is measured, it has grown an order of magnitude faster than the infrastructure serving it for 25 years, and the gap is widening.
The cost of waiting
The first cost of the gap is time. Howden finds that power constraints are adding 24 to 72 months to build times for the most constrained data center projects — two to six years on top of construction schedules that are themselves measured in years. New generation often takes 24 to 60 months to develop, and the grid connection is frequently the slower leg: as this desk reported last week, connecting a 100-megawatt data center in Northern Virginia takes about seven years, according to figures cited by the International Energy Agency.
The second cost is money, and Howden puts a number on it. Across the data centers it tracks, annual revenues approach $400 million at the top end, implying $800 million to $2.4 billion in lost revenue for the largest, most power-constrained builds — the revenue forgone during the years a finished or nearly finished facility sits waiting for power. The range is an estimate, and it applies to the largest projects rather than the typical one. But the mechanism is not in doubt: a data center earns nothing while it waits, and the waiting is now measured in years rather than months.
The 110-gigawatt answer
Howden’s headline figure for what it would take to hold the line is the one to read twice. For data centers to maintain their 5 percent share of grid capacity at current growth rates, they would need to source up to 58 percent of their power — about 110 gigawatts of capacity — from behind-the-meter sources by 2030. The sentence that follows it in Howden’s own release is the one that matters: the firm says that outcome is “unlikely to be achievable.”
The gap between 110 gigawatts needed and what is actually coming is the story. Howden’s report expects US hyperscale data centres to operate around 10 gigawatts of behind-the-meter generation by 2030, roughly one-quarter of hyperscaler power demand — a real buildout, and an order of magnitude short of the 110 gigawatts needed. Howden’s own conclusion is that data centers will meet their power demands through a combination of increasing grid share, moderated growth at power-constrained sites, and behind-the-meter sources. The share will keep rising, some projects will slow or move, and on-site generation will cover part of the difference.
What behind-the-meter generation actually consists of is mostly gas. Howden lists natural gas, renewables, battery storage, and, over the longer term, nuclear energy, but the gas leg is the one already visible and already large. As this desk documented on November 6, converted jet engines are generating power for xAI’s Colossus 2 campus in Memphis, ProEnergy has contracted 13 aeroderivative generator sets totaling 650 megawatts for Crusoe’s hyperscale pipeline, and Cleanview counted more than 100 gigawatts of data-center-owned power plants announced as of October 21, up from about 90 gigawatts in June. The announced pipeline is not the built pipeline, but operators are becoming power plant owners because the grid will not sell them power fast enough.
The rest of the behind-the-meter menu has limits that gas does not. Renewables and batteries help at the margin, but a load that draws flat around the clock cannot be served by solar alone, and batteries are a duration product, not a generation product. Nuclear is the long-term answer in Howden’s framing — long-term being the operative word, since no new reactor will be online by 2030 to serve this load. And a gas plant built as a bridge to a grid connection with no published arrival date has a way of becoming permanent.
What flexibility can and cannot do
The other lever on the table is flexibility: data centers that cut their draw when the grid is stressed, rather than building their own supply. On November 2, OpenAI launched its Grid Protectors Program, pledging that the compute campuses it builds in the United States will cut electricity use when local grids are under strain. The first commitment is the planned campus in Effingham County, Georgia, where OpenAI has agreed to let Georgia Power trim up to 1,000 megawatts of demand during peak periods — about 31 percent of the campus’s planned 3,200-megawatt load, under a 25-year agreement. On November 5, OpenAI widened the program into a National Community Compact and a policy blueprint for future campuses it owns and operates.
Flexibility is real, and it is the cheapest megawatt in the system. But the scale check matters. One thousand megawatts of flexible demand at a single campus is the largest single-facility demand response commitment in the country, by Georgia Power’s description, and it is less than one percent of the 110 gigawatts Howden says behind-the-meter sources would need to supply by 2030. Flexibility shaves the peak; it does not supply the average hour. The Effingham campus is still in development, with power delivery planned in phases from 2028 to 2032 — the pledge is a commitment for later, not a megawatt the grid can call on today. OpenAI itself says so: its announcement states that flexibility “cannot replace investment in new energy supply and transmission.” Flexibility is a complement to supply, not a substitute for it.
The regulatory backdrop reinforces the point. On October 9, the Federal Energy Regulatory Commission rejected most of PJM’s plan to make new data centers bring their own capacity or face curtailment first in emergencies, throwing out the bring-your-own-new-capacity provisions as reaching into retail policy reserved to the states. The market is moving toward self-generation faster than the regulator is willing to require it — and the October 21 post declaring large data centers “mandated” to make their own electricity carried no accompanying executive order or rule. The bypass is being built by developers making commercial decisions, not by regulators issuing orders, which means its pace is set by economics rather than by policy.
The risks that come with the bypass
There is a second half to Howden’s report that has received less attention than the 19-percent figure, and it is the half that matters most to the people underwriting the buildout. Power is not only the constraint on building data centers; it is also the largest source of loss once they are running. Half of the facilities Howden surveyed had experienced an incident with a significant impact over the previous three years, each costing an estimated $500,000 to $1 million, implying global annual losses of approximately $2 billion. Power and backup failures, primarily involving uninterruptible power supply systems, account for most major outages — Howden puts power failures at 45 percent of major data center disruptions.
That creates a protection gap that the insurance market is only beginning to price. Many outages involve no insured physical damage, so losses may fall outside traditional property and business interruption policies; Howden points to parametric cover linked to service-level agreements as the emerging answer. The gap widens as operators take on more of their own power infrastructure: a campus that generates its own electricity also owns its own generation, fuel, and maintenance risk. The values at stake explain why an insurance broker is the one publishing the arithmetic. Howden puts the insurable value of the largest AI data center campuses at $20 billion to $50 billion per site — figures that exceed what traditional insurers can economically cover at a single location — while global data center capacity is expected to double between 2025 and 2030, adding about 100 gigawatts and roughly $1.2 trillion in real-estate value. The exposure is also physical: Howden found 32 percent of United States operating data center square footage has been affected by severe tornado or hail events over the past decade, concentrated in just 20 locations.
Why the grid grows at two percent
The obvious question — why does the grid not simply grow faster — has an answer in Howden’s own release, and it is structural rather than conspiratorial. Deborah Duss, Howden’s global head of advisory and client solutions for global specialty, notes that generation projects often benefit from faster development including permitting, while grid infrastructure “faces more complex project scopes, lengthy permitting, and multiple stakeholder interests, making it harder to keep pace with growing electricity demand.” A generator is one project on one site; a transmission line crosses counties, states, and landowners, and each one of them has a say. The generation side of the system can move at the speed of a developer with a turbine order. The wires side moves at the speed of a permitting process that was designed for a world in which load grew at 2 percent a year — because, for most of the last century, it did. The data center industry has changed the load growth rate by an order of magnitude. The permitting system has not changed at all.
What to watch
Four things will determine whether the gap narrows or widens from here. The first is the behind-the-meter build rate. Test Howden’s 10-gigawatt expectation against delivery schedules: ProEnergy’s 13 Crusoe units due by summer 2027, J&F Power’s batches through November 2027, and how much of the 100-plus gigawatts announced actually gets built. The second is whether the grid side accelerates. Transmission permitting reform has stalled in Congress for years, and the load now arriving in interconnection queues is its largest test yet.
The third is whether flexibility spreads beyond OpenAI. The Grid Protectors pledge is, so far, one company’s program built on one signed agreement, with the terms for future campuses still to be negotiated. If other hyperscalers adopt comparable commitments with published trigger conditions, flexibility becomes a system resource rather than a press release. The fourth is the share itself. Howden’s 5 percent is a number worth re-measuring every year, because it is the single figure that captures the whole story: if it keeps climbing, the grid is absorbing the load and the strain with it; if it plateaus, the bypass has taken over.
The bottom line
Howden’s research quantifies the problem in a way that makes the options legible: demand growing at 19 percent, wires growing at 2 percent, a 50-fold rise in the share of the system one industry occupies, and a behind-the-meter requirement of 110 gigawatts by 2030 that the report’s own authors call unlikely to be achievable.
The honest version of the arithmetic is that there is no single answer to the gap, because the gap is too large for any single answer. Behind-the-meter generation will cover part of it, at a cost in emissions and operational risk that operators are only beginning to price. Flexibility will shave the peaks, but it will not supply the average hour. The grid will absorb a larger share of the load, and the customers already on it will pay some of the cost of that absorption.
Nineteen percent a year is a growth rate the power system was never built to serve. Two percent a year is a growth rate the power system was built to serve. The next four years will decide which of those two numbers gives way.