The gigawatt datacenter race: what's announced, what's running, and what's holding it back

A gigawatt has become the standard unit of ambition in AI datacenters. OpenAI talks about Stargate in gigawatts. Meta describes Hyperion in gigawatts. Anthropic's compute deals are now priced in gigawatts too. Add up the headline numbers and you get well past 30 GW. Primary sources tell a more modest story. Most of that capacity is a letter of intent, a land plan or a chip-supply contract. Very little of it has a published, dated "this is running" status.

We built a table from company posts and filings to separate the two. The full dataset, with a source link for every row, accompanies this article as a downloadable CSV.

What's actually announced (and what's actually running)

Last verified: October 8, 2026

The chip and cloud contracts sit on top of these sites:

Two things stand out. First, the same gigawatts get counted more than once. OpenAI's 26 GW of chip letters with Nvidia, AMD and Broadcom describe hardware that has to go into Stargate and partner sites. It is not extra capacity. Anthropic's deals with Amazon, Google and Microsoft are cloud capacity that may live in buildings the clouds already announced. Adding every headline number together inflates the total.

Second, operating status is the rarest data point. Crusoe gave the clearest public update: as of June 9, 2026, two of eight buildings at the 1.2 GW Abilene campus were operational. That campus broke ground in June 2024. Microsoft's Fairwater posts give GPU counts and rack power (about 140 kW per rack in Atlanta) but no gigawatt figure. xAI's site lists GPU counts and build times but no power figure. We could not find a primary source for the size or status of "Colossus 2", so it is not in our table.

How much power are we talking about?

The International Energy Agency's Energy and AI report, published April 10, 2025, estimated that data centres used about 415 TWh in 2024, roughly 1.5% of global electricity. The US accounted for 45% of that. The IEA projected about 945 TWh by 2030, "slightly more than the entire electricity consumption of Japan today". It also said data centres would drive nearly half of US electricity demand growth to 2030.

The IEA's April 16, 2026 follow-up, Key Questions on Energy and AI, says data centre electricity use grew 17% in 2025 to about 485 TWh. Use by AI-focused data centres rose 50%. The 2030 projection is about 950 TWh, around 3% of global demand.

For the US specifically, Lawrence Berkeley National Laboratory's December 2024 report for the Department of Energy put data centre use at 176 TWh in 2023 (4.4% of US electricity). It projected 325–580 TWh (6.7%–12%) by 2028. The width of that range is the point. Even government analysts can't say which announcements will be built.

For scale, a 1 GW campus running flat out all year would use about 8.8 TWh. That is an upper bound, since real facilities don't run at 100%. Ten such campuses would equal about 18% of the IEA's 2025 global estimate.

What's real and what's PR

Real: the money is moving. The IEA says capital spending by the largest tech companies exceeded $400 billion in 2025 and is expected to rise about 75% in 2026. Hardware is shipping into these buildings. Nvidia said in August 2026 that Vera Rubin was "in full production" with racks running at several clouds (see our chip news roundup). Amazon says Project Rainier is fully operational, and Anthropic says it uses more than one million Trainium2 chips.

PR-shaped: headline gigawatts without dates. A "5 GW" figure for Hyperion is a multi-year ambition. Meta's own site describes the Louisiana campus as both "5 gigawatts" and "over two gigawatts", and Zuckerberg's July 2025 post put it at 2 GW by 2030. OpenAI's September 2025 post said Stargate was "ahead of schedule" to secure its full 10 GW commitment by the end of 2025. Its October 2025 update reported "over 8 GW" planned, and we found no 2026 update with a new total. "Secured" and "planned" are not "energized".

The constraints that decide what gets built

Grid connections and turbines. The IEA's 2026 report says slow grid connections are pushing developers to onsite gas, and gas turbine orders rose 70% in 2025 amid a supply crunch. It estimates 15–27 GW of onsite gas could power data centres by 2030, mostly in the US. Running that gas reliably needs 30–70% more generation than the load it serves. The new AWS campus in Homer City, Pennsylvania shows where this leads. According to DCD, it will be powered by a 4.5 GW gas plant on the site of a former coal station, with more than 1 GW of that output going back to the grid. Construction is planned in four phases starting in December 2026, the last beginning in January 2032, with completion targeted for June 2036. Google's 3.6 GW agreement with Constellation, also reported October 6, 2026, takes the other route: it locks in existing nuclear and grid capacity.

Transformers and memory. The IEA lists transformers and power electronics as supply-chain pressure points. It also expects the shortage of high-bandwidth memory (HBM) to last through at least the end of 2027. A datacenter without HBM-equipped accelerators is an expensive empty shell.

Water. Operators are designing around it. Microsoft says more than 90% of Fairwater uses closed-loop liquid cooling that is filled once. Meta says Hyperion uses closed-loop and dry cooling for most of the year. That has not ended local disputes over water and onsite generation, which often decide how fast a site can grow.

Financing and circular deals. Much of the capital moves in loops:

• Nvidia may invest up to $100 billion in OpenAI, released as each gigawatt goes live. OpenAI then spends that money on Nvidia systems. At a Morgan Stanley conference on March 4, 2026, Nvidia CEO Jensen Huang said Nvidia's $30 billion investment in OpenAI "might be the last" before OpenAI goes public, and that he didn't foresee Nvidia investing the full $100 billion, according to CNBC and Benzinga. The staged amount in the letter of intent is not guaranteed.
• AMD gave OpenAI warrants for up to 160 million AMD shares that vest as OpenAI deploys AMD GPUs.
• In April 2026, Anthropic announced that Amazon would invest $5 billion in it immediately and up to $20 billion more in the future, on top of the $8 billion Amazon had invested earlier. Anthropic has committed more than $100 billion to AWS technologies over ten years (Anthropic, April 20, 2026).
• Microsoft and Nvidia agreed to invest up to $5 billion and $10 billion in Anthropic, which committed $30 billion to Azure.

None of this is hidden, since it's all in the announcements. But it means part of the "demand" for these campuses is paid for by the suppliers. If model revenue doesn't keep up, the loops can unwind. That is why token pricing isn't the whole story and why OpenAI's latest funding talks matter for its buildout. Our explainers on Anthropic and OpenAI cover how each company pays for compute.

What to watch

• Energization dates, not announcements. Treat a site as real when a utility, developer or filing confirms power is flowing.
• Second-half-2026 milestones. Nvidia's, AMD's and Broadcom's OpenAI deals all promise first deployments in 2H 2026. AMD says OpenAI expects Helios online from Q4 2026. If those dates slip, the gigawatt totals mean less.
• Updated totals. OpenAI hasn't published a Stargate total since October 2025, and Meta hasn't published a Prometheus status. We'll update this table when they do.