Data Centres
Data centres, the interconnection queue as the real constraint, co-location and behind-the-meter, and who carries the risk when a fifteen-year grid asset is built for a five-year-old business.
Words the industry uses are marked like this at the sentence that defines them, so you can tell a term you will hear on a desk from a phrase this book happens to be using.
Data centres consumed somewhere around 4 to 5% of American electricity in recent years. Lawrence Berkeley National Laboratory's projections for 2030 span 9.5 to 15.3%.
A range that wide, from a serious institution, over five years, is the most informative thing in this chapter. Every participant in this industry is making twenty-year commitments against a forecast whose credible spread is a factor of one and a half, and the commitments are transmission lines, capacity auctions and power plants.
This chapter therefore concentrates on the parts of the situation that hold whichever forecast turns out to be right.
The queue measures interest
Chapters 17 and 18 gave the figures. Nationally the generation interconnection queue (generators waiting to be connected to the grid) has run to thousands of gigawatts, against a peak US current demand somewhat over 700. ERCOT's large-load queue reached about 474 gigawatts by June 2026, against a peak near 91.
Those numbers describe requests rather than projects. A speculative developer shopping one power campus across six sites appears six times, and a queue position costs a fraction of what a data centre costs. Both ERCOT and the federal regulator reached the same conclusion within about two years and applied the same remedy: stop studying requests one at a time, study them in batches, and require applicants to demonstrate readiness to hold a place.
The constraint for anyone trying to build has stopped being the cost of generation and become the wait for permission to connect, which is a procedural problem rather than an engineering one. Developers are finding sites, constructing shells and even energising plants while they wait for permission to connect.
Counting the build-out
Banks and research houses publish their own estimates of the build-out and revise them often, so each figure below carries the month it was published.
Table 25-1: Estimates of US data centre demand published in 2026
| Source | Published | Measures | Estimate |
|---|---|---|---|
| EPRI | Feb 2026 | Share of electricity, and capacity of the computing equipment alone | 9 to 17% of US electricity by 2030, from 4 to 5% now; 56 to 132 GW of IT capacity, from 35 to 44 GW in 2024 |
| Lawrence Berkeley National Laboratory | Jun 2026 | Electricity used in a year | 649 TWh in 2030, 11.8% of US electricity, within bounds of 521 to 843 TWh |
| BloombergNEF | Jul 2026 | Power demand | 194 GW of demand by 2035, up from 106 GW in its December 2025 outlook |
| Wood Mackenzie | Aug 2026 | Requests for connection, not projects | 1,066 GW requested from US utilities, of which about 28% is likely to be built |
| Goldman Sachs | Nov 2026 | Capacity built | 152 GW of US data centre capacity by 2030, up from 125 GW in its previous forecast, at about $50 billion of capital per gigawatt |
The figures measure different things, which is why they look further apart than they are. Electricity used in a year converts to average power by dividing by the 8,760 hours in a year, so Berkeley’s 649 TWh is an average draw of about 74 GW. Capacity built is larger than average draw, because data centres do not run at full load around the clock and new campuses fill over years: at 60 to 70% utilisation, Goldman’s 152 GW would draw about 90 to 105 GW, at or just above the top of the Berkeley and EPRI ranges. Capacity of the computing equipment alone excludes cooling and the rest of the site, which typically adds 20 to 40%. Banks also tend to work down from announced capital spending, which assumes the money is spent and the sites get power, while the laboratories build up from server shipments and utilisation.
The spending is on the same scale. Alphabet guided to between $195 and $205 billion of capital spending for 2026 and Amazon to about $220 billion, and analysts at Evercore and Bank of America expect AI capital spending to pass a trillion dollars in 2027. The IEA puts cumulative data centre investment at $3.9 trillion between 2026 and 2030.
The most useful figure is the gap between what is requested and what is signed. A tracker published by Latitude Media in November 2026, covering 53 utility territories, counted 537 GW of data centre load reported by utilities, 120 GW under signed contracts and 22 GW actually operating. Georgia Power reported 80.1 GW against 12.8 GW signed. Wood Mackenzie found that only about 12% of national requests held firm service, and in August 2026 Texas paused new large-load approvals while ERCOT’s queue was audited. On Latitude’s count the average project takes 33 months from announcement to first power.
Construction is concentrated in a few markets. CBRE counted a record 7,481 MW under construction in the main North American markets in the first half of 2026, more than 80% of it leased before completion, with vacancy at 1.4%. Atlanta overtook Northern Virginia in construction for the first time, with nearly 2,900 MW underway, while Northern Virginia remains the largest market in operation at about 4,500 MW. EPRI expects data centres to use between 39% and 57% of Virginia’s electricity by 2030, up from about a quarter today. A forecast prepared for PJM by Charles River Associates in November 2026 put data centre demand in PJM at 30 GW by 2030, warned that the market could be about 11.5 GW short of capacity that year, and counted 294 local moratoriums or bans on data centres passed in the first seven months of 2026.
The largest campuses belong to a handful of buyers. Epoch AI’s estimates of operating IT load put the Amazon campus at New Carlisle, Indiana, built for Anthropic, at about 1,040 MW, xAI’s Colossus 2 in Memphis at about 950 MW and OpenAI’s Stargate site at Abilene, Texas, at about 820 MW. The Stargate sites announced in the United States add up to more than 9 GW, most of it due by the end of 2028. Meta’s Hyperion campus in Richland Parish, Louisiana, is to be served by seven new Entergy gas plants totalling 5.2 GW, and SoftBank’s planned 10 GW campus at Piketon, Ohio, by 9.2 GW of its own gas generation.
The hyperscalers’ cash flow pays for only part of it. Morgan Stanley estimated in July 2025 that global data centre spending of about $2.9 trillion between 2025 and 2028 left a financing gap of about $1.5 trillion, and expected private credit to supply about $800 billion of it. JPMorgan put the five-year requirement at $5 trillion or more. The bond market carries a growing share: JPMorgan Asset Management counts $109 billion of investment-grade bonds issued by hyperscalers in 2025 and $194 billion in the first half of 2026 alone, and notes that the first large issue of the second half drew orders of 1.4 times the amount offered, against 3.5 to 4.5 times earlier in the year. Some of the financing has moved off the tenants’ balance sheets, the model being Meta’s Hyperion joint venture, 80% owned by funds managed by Blue Owl with Meta as the tenant. No published source estimates what share of the announced pipeline has its financing arranged.
Behind the grid connection sits a second queue, for gas turbines, which Chapter 26 covers, including the $3.7 billion SoftBank wrote off in November 2026 when it cancelled turbine reservations it no longer needed.
What it means for gas
Most of the new generation built for data centres burns natural gas, and the conversion is arithmetic on the heat rate from Chapter 4. A gigawatt of load running around the clock on a combined-cycle plant at 7,000 Btu per kilowatt-hour burns about 168,000 MMBtu a day, roughly 0.16 billion cubic feet a day (Bcf/d). On simple-cycle turbines or engines on site, at around 10,000 Btu per kilowatt-hour, the same gigawatt takes about 0.23 Bcf/d. Every 10 GW of data centre load served by gas is therefore between about 1.6 and 2.3 Bcf/d. EQT’s agreement to supply up to 665,000 MMBtu a day, about 0.64 Bcf/d, to the 4.4 GW Homer City plant in Pennsylvania fits the combined-cycle end of that range.
The share of the new load that gas serves is the assumption that moves the answer most, and the published estimates spread accordingly.
Table 25-2: Estimates of added US gas demand from data centres
| Source | Published | Estimate |
|---|---|---|
| East Daley Analytics | Feb 2025 | 4.2 to 6.1 Bcf/d by 2030, assuming about 81 GW of new data centres with gas meeting 40% of it |
| Morgan Stanley | Feb 2026 | About 4.8 Bcf/d by 2030 in its base case, and 14 to 15 Bcf/d if gas served all of the new load |
| RBC Capital Markets | May 2026 | About 6.1 Bcf/d by 2030, with a likely range of 6 to 7 |
| BloombergNEF | Sep 2026 | About 15 Bcf/d by 2035, more than double its December 2025 estimate of 6.9 |
| Energy Flux | Nov 2026 | 1.4, 3.8 or 6.4 Bcf/d by 2030 in three scenarios, from a slowdown to an AI race |
The middle of the range, 4 to 7 Bcf/d by 2030, is comparable to the 3.9 Bcf/d the EIA expects US LNG exports to add between 2025 and 2027, and equal to roughly 4 to 6% of the 107.6 Bcf/d the United States produced in 2025. Power plants already burned 35.8 Bcf/d that year. Readers of NatGas 101 will look for where the new demand reaches the pipelines, and the answer is in the projects now contracted: Kinder Morgan’s Permian Link, about 2.8 Bcf/d for an unnamed hyperscaler from 2030, and Energy Transfer’s laterals carrying about 0.9 Bcf/d to three Oracle data centres, two of them in Texas.
The plug in the side of the power station
If connecting through the grid takes years, an obvious alternative presents itself. Build the data centre next to an existing power station and connect it directly, behind the meter, bypassing the transmission system entirely.
The test case was Talen Energy's Susquehanna nuclear station in Pennsylvania, alongside which Amazon Web Services built a campus. When the parties sought to expand the behind-the-meter arrangement from 300 to 480 megawatts, the federal regulator rejected the interconnection agreement, and upheld that rejection on rehearing in April 2025.
The objection was about cost allocation rather than engineering. A large load sitting behind the meter uses the grid for backup when its host plant trips, and benefits from the reliability the transmission system provides, while contributing nothing toward its cost. Every dollar it avoids is a dollar recovered from everybody else.
The commercial answer arrived in June 2025, when the parties restructured the deal as a seventeen-year power purchase agreement worth around 18 billion dollars for up to 1,920 megawatts from the 2.5 gigawatt station. The transaction happened. It happened in front of the meter, through the market, paying network charges.
The regulatory answer followed in December 2025, when the commission directed PJM to write proper rules for co-location, creating defined service options rather than case-by-case adjudication, with further compliance orders on bring-your-own-generation arrangements through 2026.
The co-location fight is the oldest question in regulated utilities. When a large customer can partially exit the shared system, who pays for the shared system? The same argument was had about industrial self-generation and rooftop solar net metering, and it produces the same fault line each time.
Net metering is the one that has run furthest, and the way it resolved is the template. The original design paid exports at the full retail rate, which bundles the energy together with the network and the policy costs, so an exporting customer stopped paying for a system it still leaned on. The answer was to unbundle: pay for the commodity at avoided cost, which is what the utility would otherwise have spent on that energy, and charge for the network separately. California’s net billing tariff of April 2023 cut the average export rate by roughly three quarters, from about 30 cents a kilowatt-hour to about 8, and Hawaii had made the same move in 2015 with Arkansas, Arizona, Indiana, Louisiana, Michigan and Utah following. North Carolina added a grid access charge and a minimum bill so the network is paid for whatever the net energy turns out to be, and every one of them grandfathered existing customers for a term, which is what keeps a repricing from being a confiscation.
So the question has an answer, and it is the one the commission is reaching for in asking PJM to write defined service options rather than settling each case on its own facts. Charge the exiting party for the commodity, bill it for the network directly, and protect what was already built.
The assumption nobody had examined
Every queue, capacity auction and planning study in this book treats a large load as firm: present at full size, at every hour, forever.
A study from Duke University's Nicholas Institute asked what happens if that assumption is relaxed slightly. Examining 22 large balancing authorities, the entities responsible for matching supply and demand across a defined area, covering about 95% of national peak demand, it found that the existing system could absorb around 76 gigawatts of new load, close to 10% of national peak, if that load accepted curtailment for an average of 0.25% of its maximum uptime. At 0.5% the figure rose to about 98 gigawatts, and at 1.0% to roughly 126. By region at the half-per-cent level, PJM could take about 18 gigawatts, MISO 15, and ERCOT and SPP around 10 each.
A quarter of one per cent of a year is about twenty-two hours. The finding is that a very small quantity of flexibility converts into a very large quantity of headroom, because the grid is sized for a peak that occurs rarely and sits underused the rest of the time.
Duke University Nicholas Institute, covering 22 balancing authorities and about 95% of national peak. The finding is contested and its assumptions matter: that curtailment can be called where and when the system needs it, that load genuinely reduces rather than switching to onsite generation, and that the contracting machinery exists. Its contribution is that the firmness of large loads is an assumption rather than a fact.
The result is contested, and reasonably so. It assumes curtailment can be called where and when the system needs it, that the load can genuinely reduce rather than merely shifting onsite generation, and that the contractual and computational machinery exists to arrange it. None of that is free. What the study establishes is that the firmness of large loads is an assumption rather than a fact.
Data centres are also better placed to provide flexibility than most load. AI training runs can be paused and rescheduled in ways a steel furnace cannot, AI inference (day to day AI queries) can be slowed imperceptibly, and Chapter 7 noted that these are the first new loads in decades large enough, concentrated enough and sophisticated enough to negotiate about when they consume.
Whose risk is it
The durable question underneath all of this concerns duration mismatch.
A transmission line is planned, permitted and built over the better part of a decade and depreciates across forty years. A nat gas plant is a thirty-year asset. They are being built for customers whose industry is a few years old, whose demand forecasts have a fifty per cent spread, and whose siting decisions can change with a tax regime or a fibre route.
If the load arrives and stays, the investment is sound and the cost is spread across a larger base, which lowers everybody's bills. If it arrives, then consolidates onto fewer, more efficient sites elsewhere, the asset remains in the rate base and existing customers pay for infrastructure built for a tenant who left.
Which is why the serious regulatory work is happening in large-load tariffs rather than in market design: minimum take obligations, contract terms matched to asset lives, collateral, and exit fees. Those provisions decide who holds the risk, and they are being written now, mostly in state proceedings, at speed, with far less attention than the capacity auctions in Chapter 16 attract. If you move fast, sometimes things break.