A single datacentre campus reaches 20 GW

Where things stand (16 Oct 2026)

The scale gap. The largest AI datacentre running today draws about 1 GW of IT load. Epoch AI's tracker (updated 12 Oct 2026) puts SpaceXAI's Colossus 2 first at about 946 MW. Next come Anthropic-Amazon New Carlisle (about 910 MW) and Microsoft Fairwater Atlanta (about 636 MW). The 94 sites Epoch covers add up to only 14.1 GW of IT load (https://epoch.ai/data/ai-data-centers). To resolve this question, one campus would need about 20 times today's record.

The announced pipeline. No announced campus reaches 20 GW of IT load:

• PORTS-Pike, Ohio (SB Energy / OpenAI / Nvidia): about 8 GW of IT load, the largest IT-denominated project announced. OpenAI says the first 800 MW arrive in 2028 and the "six-year buildout" runs "through 2032" (https://openai.com/index/openai-joins-ports-pike-project/, 17 Aug 2026).
• Meta Hyperion, Louisiana: planned at 5 GW, completing around 2030 (Meta, 13 Jul 2026).
• Fermi America's Project Matador, Texas: pursuing 17 GW of generation, not IT load, with completion "in the 2030s" (Business Insider, 29 Sep 2026).
• Other Utah and Canadian proposals: the Delta Gigasite (10 GW of generation), Stratos (9 GW), Project Beehive (9.6 GW of nuclear) and Wonder Valley (7.5 GW, with construction now expected around late 2028).
• SpaceX: it is aiming for 10–20 GW of power by the end of 2027 (DCD, Aug 2026). That target is spread over many sites (Memphis, Southaven and others), and separate sites cannot be added together under the question's rules.

Industry direction. Meta's engineering blog (12 Oct 2026) says the industry is moving toward training spread across sites: "a single cluster now outgrows any single region." That reduces the pressure to build one giant campus.

Trend and reference class

I downloaded Epoch's timeline file (https://epoch.ai/data/data_centers/data_center_timelines.csv) and took the largest IT load at any single site each quarter:

• 237 MW in Q3 2024
• 946 MW in Q2 2026
• 2,263 MW projected for Q2 2028 (Fairwater Wisconsin)

A straight-line fit on a log scale over 2024–28 gives about 1.9× growth a year. Carried forward naively, that reaches about 25 GW by the end of 2031. Epoch's 2025 paper (https://arxiv.org/abs/2504.16026) likewise projects a 9 GW leading system by mid-2030 if power keeps doubling yearly, but it expects training to spread across sites instead.

The forward data point to slower growth. Projects already under construction imply only about 1.5× a year for 2026–28. Announced plans top out at 5–8 GW of IT load around 2030–32. Together these suggest the largest single campus will be about 5–10 GW at the end of 2031. Reaching 20 GW of IT load (about 23–26 GW at the meter) would require:

• a project nobody has announced yet;
• building at 5 GW or more a year at one site;
• gas supply of several billion cubic feet a day;
• roughly $1 trillion of hardware in one place.

The verification bar

Statements of capacity, contracted power and "GW campus" figures are all excluded. YES needs one of two things:

• a measured IT demand of at least 20 GW published by the operator; or
• a utility filing of achieved demand, combined with a measured campus PUE (power usage effectiveness) for the same period.

Operators almost always report capacity, not measured load. Utilities report whole-facility load and cannot see power generated on site. Even if the campus were built, I put the chance of a qualifying publication within the window at about 30–45%.

Estimate

• Chance a 20 GW IT campus physically exists by the end of 2031: about 2%. This is mostly the AI hyper-growth scenario (about 20% likely, with about an 8% chance of such a campus within it); the base case is about 0.3%.
• Times the verification chance: this gives about 0.7%.
• Small allowance for lenient resolution: this brings the total to about 1%.

Quarterly hazards are about 0.02% through 2028, since a campus this size is physically impossible by then. They rise to 0.04%, 0.08% and 0.11% per quarter in 2029, 2030 and 2031. Accumulated, the cumulative probability is 0.18% at the end of 2028 and about 1.1% at the end of 2031.

Strongest case against

The yearly doubling of the largest AI supercomputer has held through 2026, and the money and demand are accelerating: SpaceX wants 10 GW or more in 2027, Microsoft has contracted 10 GW so far this year, and projects keep being upsized (Hyperion went from 2 GW to 5 GW, Fermi from 11 GW to 17 GW). If doubling continued at a single site, 20 GW would be near the trend line by late 2031. I discount this for four reasons:

• announced timelines run well below the trend;
• training is moving across sites;
• big sites take 4–6 years or more to build out;
• the verification bar is strict.

Even so, this is why I put the end-2031 figure at about 1% rather than near zero.

Question. By the last day of each quarter, has a single AI datacentre campus reached at least 20 GW of operational IT load, verified by the operator or a utility filing?

Definitions.

• Single campus: one geographically contiguous datacentre development, including adjoining parcels separated only by roads, railways, waterways, or utility rights-of-way, identified as one campus by its operator. Multiple buildings, tenants, and construction phases within that campus may be combined. Separate sites linked by a network, common ownership, power supplier, or project name may not be combined.
• AI datacentre: a campus whose operator identifies AI training or inference as an operating use, not merely a possible future use. The threshold concerns the campus's total IT load; it need not all be attributable to AI. Any operator anywhere in the world is eligible.
• Operational IT load: actual electrical power consumed by operating servers, accelerators, storage, and networking equipment. Cooling, building services, electrical-distribution losses, and other non-IT loads are excluded. Installed or commissioned capacity, contracted power, and equipment nameplate ratings alone are not operational load.
• 20 GW: at least 20,000 MW of simultaneous IT demand. A reported measured operating peak or average over a stated period qualifies; no minimum duration is required. Non-coincident building peaks may not be added. Electricity supplied by the grid and onsite sources counts, without double-counting.

Resolves YES if a qualifying source identifies the campus and verifies an actual operating IT load of at least 20 GW. An explicit numerical statement of achieved IT demand is sufficient without publication of raw meter readings. Alternatively, the threshold may be calculated from published measurements: IT electricity consumption divided by the corresponding elapsed time, or total facility consumption divided by a measured campus-specific power usage effectiveness (PUE), using the same period and facility boundary for both quantities. All necessary inputs must come from qualifying sources. The Observation period convention applies.

Resolves NO / Exclusions.

• No qualifying verification has been published by the quarter-end.
• Announcements, permits, interconnection requests, supply agreements, forecasts, leased capacity, and completed electrical infrastructure do not establish achieved load.
• Total facility demand, generating capacity, or an unspecified “GW campus” figure does not establish IT demand. Design PUE, fleet-average PUE, and assumed industry efficiencies may not be used to infer it.
• Electrical load-bank tests and commissioning tests without operating IT equipment do not count.
• Portfolio totals and geographically separate sites do not count. A qualifying range must have a lower bound of at least 20 GW; “approximately 20 GW” without a verified value or lower bound at least at the threshold is insufficient.

Resolving source. Only (1) a dated operator-authored publication on the operator's official domain or an operator-authored public regulatory filing, or (2) a publicly accessible filing authored by the serving electric utility in an electricity regulator's or grid operator's official docket, qualifies. Operator includes the campus infrastructure operator or the computing operator with responsibility for the reported load. Utility filings must report achieved demand, not only planned service. Public confirmation (A2) applies, including its dispute rule for third-party confirmation; these source restrictions are narrower than A2's general list. Press reporting, executive interviews, and independent estimates cannot resolve the question. The Resolution window and Resolution date (A3) conventions apply. Relevant source formats include operator campus disclosures such as Meta's Richland Parish page and campus measurement tables such as Google's PUE reports; neither a planned campus capacity nor a PUE table alone verifies the required load.