Key takeaways
- Microsoft, Meta, Amazon and Alphabet together disclosed $830.5 billion of data centre leases signed but not yet commenced as of 30 June 2026.
- Microsoft's share is $329.1 billion, commencing between fiscal 2027 and fiscal 2033. Its own capital spending rose 79.6% in a year, to $115.9 billion.
- PJM tightened how it counts announced data centre load in January 2026, and its summer peak forecast for 2028 fell by 4,414 MW, or 2.6%.
- Great Britain's system operator found more than 300 GW of its old connection queue would not move forward, from a queue of over 700 GW.
- Amazon holds energy contracts covering roughly 270 million megawatt hours, at a weighted average remaining duration of about 15 years.
The announced number and the deliverable number are two different numbers
You've read that data centres are about to overwhelm the electricity grid, and you want to know whether the grid data supports that. Here's the direct answer. The capacity companies announce and the capacity grid operators are prepared to plan around are two different numbers, and during 2026 the gap between them stopped being an argument and became a measurement.
Three jurisdictions tested announced data centre demand against what they were actually prepared to plan around, within nine months of each other. Every one of them cut the number. Great Britain's National Energy System Operator found that more than 300 GW of a queue which had grown past 700 GW would not move forward. PJM, the market that includes Virginia's data centre cluster, started derating requests that lack a firm commitment, and cut its own near-term forecast. The Texas governor paused new data centre development on 3 August 2026, and the US Energy Information Administration responded by cutting its 2027 Texas load growth forecast from 14% to 6%.
None of that means the buildout isn't happening. It means data centre power constraints are rationing the buildout rather than capital, and the filings show the buyers know it.
What $830.5 billion of leases not yet commenced actually is
The most useful number in a hyperscaler's accounts isn't capital expenditure. It's the lease that has been signed and hasn't started. That figure is a contractual commitment to occupy capacity that doesn't exist yet, and it sits off the balance sheet until the site is ready.
As of 30 June 2026, Microsoft disclosed "additional leases, primarily for datacenters, that had not yet commenced of $329.1 billion", commencing between fiscal 2027 and fiscal 2033. Meta reported approximately $278.99 billion, running from the remainder of 2026 through 2036, then signed a further $68 billion of data centre leases in July 2026. Amazon's schedule of commitments shows $137.2 billion of leases not yet commenced, of which $93.5 billion falls after 2030. Alphabet reported $85.2 billion, commencing between 2026 and 2031. The chart above plots the four, and they add to $830.5 billion.
Set that against what these companies actually spent. Microsoft's additions to property and equipment were $115.9 billion in the year to 30 June 2026, against $64.6 billion the year before, a rise of 79.6%. Alphabet spent $80.6 billion on property and equipment in the first half of 2026, against $39.6 billion in the same half of 2025. Meta spent $49.1 billion, against $29.5 billion. The pipeline of signed leases is roughly seven times a single year of Microsoft's capital spending, and it stretches to 2036.
That's the announced side. It's a contractual queue, not a physical one, and the two only converge if the electricity arrives on schedule.
Data centre power constraints show up first in the grid connection queue
A grid connection queue is a waiting list. A developer asks the network for a given number of megawatts at a given site, and gets a place in line. NESO's own account of what went wrong in Great Britain is blunt: "Outdated rules meant that every project, whether it was ready-to-go or not, took up space in the queue." A place in line was not a commitment to build, so the queue total and the delivered total drifted a long way apart.
Then Great Britain measured the drift. NESO said in December 2025 that its old queue held "more than 700 GW of generation and storage projects waiting for grid access", which it put at "around four times what Great Britain requires" for 2030. After applying readiness and alignment tests, it confirmed 283 GW of generation and storage plus 99 GW of transmission connected demand, and stated that "more than 300GW of projects in the old connections queue will not move forward". Jonathan Brearley, chief executive of Ofgem, named the demand side directly, citing "AI-driven technologies and power-hungry data centres" in the same announcement.
So the first honest measurement of data centre power constraints isn't a shortage of electrons. It's the discovery that a large share of the announced pipeline was never going to be built as announced.
PJM asked for proof, and its own forecast fell
PJM ran the same test on the demand side, and published the result. Its 2026 Load Forecast Report, dated January 2026, says near-term forecast years need firm commitments, such as an Electric Service Obligation or a Construction Commitment. Longer-term projects, in PJM's words, "will be considered 'non-firm' and will be derated because of their greater uncertainty".
The effect was immediate. Compared with the 2025 report, PJM's summer peak forecast fell by 2,564 MW for 2026, by 4,414 MW for 2028, and by 1,630 MW for 2031. The 2028 cut is 2.6% of the peak. PJM attributed the near-term reduction to "updates to the electric vehicle forecast, economics and large load adjustments".
Read that carefully, because it cuts both ways. A grid operator that stopped taking announced data centre load at face value produced a smaller forecast. It did not produce a small forecast. PJM still projects summer peak load growth averaging 3.6% a year over ten years, reaching 222,106 MW in 2036, an increase of 65,733 MW. Net energy grows 5.3% a year, an increase of 581,554 GWh over the decade. That single regional increase is equal to about 13.9% of total United States electricity consumption in 2025.
Virginia already has the data centres, so it shows the bill first
Virginia Power is the closest thing to a live experiment, because Dominion's filings describe the concentration of data centres in Loudoun County as a unique challenge that already requires significant investment in transmission and generation. Its 2025 Form 10-K states that "Data centers represent 28% and 26% of Virginia Power's electricity sales for the years ended December 31, 2025 and 2024, respectively", and that PJM "has projected a 5.4% average peak annual load growth over the next ten years for the PJM DOM Zone".
The revenue line separates the customer classes. Virginia Power reports a "high load" class, defined in the filing as customers "with actual or anticipated forecast demand of 25 MW or higher and annual load factor of 75% or higher". In the quarter to 30 June 2026 that class produced $814 million of revenue, against $422 million a year earlier, a rise of 92.9%. Residential revenue over the same quarter rose from $991 million to $1,164 million, or 17.5%. High load customers were 23.8% of Virginia Power's total operating revenue for the quarter.
Then comes the part that matters for anyone reading announced capacity as a forecast. Dominion's 10-K says Virginia Power "will, in accordance with the terms of the 2025 Biennial Review order, begin in January 2027 to require a 14-year contract, collateral over that period and demand minimums for distribution, transmission and generation revenues for high load customers at connection". The filing states these provisions are "designed to minimize both cross-rate class subsidies and stranded costs".
A 14-year contract with collateral and demand minimums is a price on optionality. It converts a queue position from a free option into a funded liability, and a developer holding five speculative sites has to decide which one it actually wants. That's how a regulator shrinks an announced pipeline without refusing anyone.
Amazon books its power risk as a derivative
The clearest statement of the constraint is in Amazon's own accounting policy. Its 10-Q for the quarter ended 30 June 2026 explains that it enters "into energy contracts to secure electricity supply for our existing and future operations, some of which extend 20 years", and then adds the sentence worth reading twice: "We may make or receive net cash payments, rather than take delivery of electricity, when our consumption is less than committed quantities due to operational variability."
Because those payments can settle in cash, the contracts are derivatives, and Amazon marks them to fair value. As of 30 June 2026 the quantities involved were "approximately 270 million megawatt-hours", at a weighted average remaining duration of "approximately 15 years", with most of the volume delivered beyond the next nine years.
Alphabet discloses the same idea in different words. It reported $707.0 billion of expected future fixed or guaranteed commitments as of 30 June 2026, including energy service agreements with "terms ranging from two to 26 years, with obligations through 2054", which "generally include take-or-pay provisions for minimum quantities of energy supply and substantive termination fees". It also carries a maximum exposure of $7.6 billion under backstop agreements that "support counterparty procurement of long-lead time equipment for our future power purchase and energy agreements".
Microsoft states the constraint plainly in its risk factors. "The availability, reliability, and cost of electrical power are critical to the operation and expansion of our datacenters. In many regions, electricity generation, transmission, and distribution infrastructure is experiencing increasing demand and capacity constraints." The same passage names "delays in obtaining power connections" and "requirements imposed by utilities, regulators, or other market participants" as things that could restrict its ability to expand. This is the same cost pressure that runs through AI capex depreciation once the assets are switched on.
The strongest objection: 18 TWh a year is not a national emergency
Here's the best case against reading any of this as a crisis, and it's a good one. Divide Amazon's 270 million megawatt hours by the roughly 15 year weighted average duration and you get about 18 TWh a year of contracted electricity. Total United States electricity consumption was 4,195 billion kWh in 2025. Amazon's contracted volume is therefore around 0.43% of the American grid a year.
The EIA's August 2026 Short-Term Energy Outlook, completed on 6 August 2026, forecasts total US consumption of 4,268 billion kWh in 2026 and 4,391 billion kWh in 2027. That's growth of 4.7% across two years. Inside it, commercial sector sales rise from 1,493.5 to 1,608.6 billion kWh, or 7.7%, while residential sales go from 1,515 to 1,530 billion kWh, a rise of 1.0%. Data centre electricity demand is visible in the commercial line and invisible in the residential one, which is roughly what a concentrated industrial load looks like.
EIA's separate analysis of commercial buildings, published on 19 May 2026, put data centre servers at "an estimated 7% of commercial sector electricity consumption in 2025", growing to a range of 22% to 33% of commercial building electricity use by 2050. Large, and not sudden.
The objection holds at national scale and dissolves at local scale. A grid is not a single pool. Virginia Power's high load class already produces 23.8% of its quarterly revenue, and PJM's ten year energy increase equals 13.9% of current national consumption inside one regional market. National averages are the wrong denominator for a load that the filings measure in blocks of 25 MW or more, and that clusters in a handful of counties.
What this data cannot tell you
Four limitations are worth stating before anyone builds a view on these figures.
The lease figures are dollars, not megawatts. No filing above converts $830.5 billion into gigawatts, and any conversion you see is someone's assumption about price per megawatt, not a disclosure. Treat the dollar figure as a commitment to spend, not as a quantity of power.
Queue reform measures the queue, not demand. NESO's 300 GW removal tells you the old queue was inflated. It doesn't tell you how much of the removed capacity was real projects that failed a paperwork test rather than speculative applications.
PJM's derating is a methodology change, so the forecast cut is partly a measurement artefact, not purely a change in the world. The comparison of 2026 against 2025 mixes both.
And the sample is short. Every firmness test above was applied within the last twelve months, so nobody yet knows the eventual conversion rate from a Gate 2 offer or a Construction Commitment to a live, energised building. Two data points are not a trend, and this is closer to one.
What would change the conclusion
If the firmness tests keep cutting forecasts, the announced pipeline is the number that's wrong, and the operating question moves away from data centre power constraints and toward lease obligations on capacity nobody needs. Amazon's disclosure already names that outcome: net cash payments instead of delivery when consumption falls short of committed quantities. Alphabet's take-or-pay energy agreements run to 2054 on the same logic.
If instead the tests stop biting, and PJM's next forecast revises upward while Dominion's high load class keeps compounding near 90% a year, then the 2026 cuts were a pause in the paperwork and the constraint is physical after all. Dominion's January 2027 requirement is the cleanest single test available, because it puts a 14-year contract and collateral in front of every new high load connection. The share of announced Virginia capacity that signs one is a direct readout of how much of the pipeline was ever real.
The number to watch isn't announced gigawatts, and it isn't capital expenditure either. It's the gap between what developers have signed and what grid operators are willing to count, published in each operator's own load forecast, and reconcilable against the leases in each 10-Q. That reconciliation is also where cloud capex stops being an accounting question and starts being an engineering one, long before it shows up in AI gross margin.