Research as of October 11, 2026
Summer 2025
Oracle would build five campuses in Texas, New Mexico, Wisconsin and Michigan, then rent the computing power back to OpenAI for about five years starting in 2027.
Where the $300 billion goes (estimates)
The five campuses, drawn to scale as one-story footprints. Each filled block is a building.
OpenAI, SoftBank, Oracle and MGX promise $500 billion over four years: 10 GW, framed as about 20 data centers.
The SoftBank side of Stargate has built almost nothing. OpenAI signs with Oracle for 4.5 GW of new capacity, about $30 billion a year.
The Wall Street Journal reports the contract runs about five years, starting around 2027.
Shackelford County, Texas and Doña Ana County, New Mexico, plus a Midwest site that becomes Port Washington, Wisconsin and Saline Township, Michigan. A 600 MW expansion at Abilene, Texas is still being negotiated.
Blue Owl declines to fund the $10 billion of equity for the Michigan campus. Some sites slip from 2027 to 2028.
OpenAI cuts its 2030 computing target from about $1.4 trillion to about $600 billion.
Financing dragged and OpenAI's needs kept changing. Three weeks later, Crusoe leases the land to Microsoft instead.
Stargate Norway goes to Microsoft and Stargate UK is paused. Reports say OpenAI has effectively abandoned building its own Stargate data centers in favor of leasing.
S&P cuts Oracle's credit rating to BBB- and names OpenAI as a key risk.
Oracle sends a force-majeure notice on Project Jupiter, seeking up to three years of rent deferral if it misses 2028.
As of October 2026 the $300 billion contract still stands, on paper. It is the most closely watched data center deal in the world.
Every large data center campus finished since 2023, under construction, or planned, around the world. Then an estimate of the ones nobody has announced.
Each circle is one data center campus, sized by its power. In the middle are OpenAI's five Oracle sites. The dashed one is the Abilene expansion, dropped in March 2026.
533 campuses in 41 countries: running, under construction, or planned. Each one is a row in the research file with its own source.
The biggest are bigger than the whole OpenAI deal on their own.
Big campuses hide behind shell companies and code names. Meta's Indiana campus was filed as Project Domino by Orla LLC. Ohio approved Project Klondike under a non-disclosure agreement.
Our estimate: about a quarter of the buildout is unreported (15 to 40%). The dashed circles stand in for it.
Outlined: single campuses over 4.5 GW. Four are in Utah. Most are being built by companies few people have heard of, like Joule, Creekstone, Valar Atomics and Pacifico Energy.
Each square is the entire OpenAI and Oracle deal. The amber square is the deal itself; the blue ones are what's being built everywhere else.
1.62 billion sq ft. Laid out flat, about 58 square miles: nearly the land area of Washington, D.C.
547 GW. More than the whole US grid delivers on an average hour today, about 480 GW.
$17.0 trillion: $5.5T for buildings, land, power and cooling, and $11.5T for the servers and chips inside. About 2.4 years of the entire federal budget.
Counting only what is likely to actually get built, the total is still $10.6 trillion, or 35 OpenAI deals. Money is the smallest multiple because the OpenAI contract also pays for five years of electricity, operations and Oracle's margin.
The US holds 283 of the 411 known gigawatts, spread over 362 campuses. Each square shows a state's planned data center power.
16 campuses. On an average day the entire state uses about 4 GW of electricity. The Stratos project alone, 9 GW on 40,000 acres in Box Elder County, would use about 2.3 times all the electricity Utah sold in 2024.
76 campuses, including three of OpenAI's sites. The state grid operator, ERCOT, has 438 GW of requests from large new power users, 89% of them data centers.
Places with few data centers before 2023 now host some of the biggest: Project Jade (10 GW), Monarch (8 GW) and Meta's Hyperion (5 GW).
The 16 largest known campuses by planned power. The dashed amber line is everything OpenAI contracted from Oracle: 4.5 GW across five sites.
One deal was negotiated in public. Every gigawatt and dollar was reported, argued over, and partly walked back. The buildout around it is 57 times larger, mostly unpriced, partly hidden, and paid for in unusual ways.
In the same months OpenAI shrank its 2030 computing plan by more than half and walked away from the Abilene expansion, the four biggest cloud companies planned to raise their spending by another 75%.
OpenAI target: reported Feb 20, 2026. Capex: company filings and 2026 guidance for Microsoft, Alphabet, Amazon and Meta.
229 campuses in the list announced a dollar figure. Together they add up to $1.67 trillion for 156 GW, or $10.7 million per megawatt. That is about what the shell, power and cooling alone cost.
Fill the same building with AI servers and the bill roughly triples. Two thirds of the cost of an AI data center almost never appears in the announcement.
Cost per MW: JLL, Cushman and Wakefield, Bernstein, Epoch AI ($38B per GW all-in), Nvidia's Jensen Huang ($50B to $60B per GW).
The Census Bureau's tally of data center construction reached an $85 billion annual rate in August 2026. But the 66 GW that JLL counts under construction in North America implies $250 to $320 billion of building work a year.
Power plants and substations get booked elsewhere, some campuses are filed as warehouses, and land isn't counted. A member of Congress signed a non-disclosure agreement with Laidley LLC, the shell company behind Meta's Hyperion campus.
US Census C30 private construction spending (2023 $20.0B, 2024 $34.8B, 2025 $49.7B). JLL North America midyear 2026.
Texas has about 94 GW of named data center projects. Its grid operator has 438 GW of requests to connect large new power users, and 89% of those are data centers. Someone is planning far more than anyone has announced.
ERCOT large-load queue, June 2026. Named Texas projects from the research list and trackers.
Sixteen known campuses in Utah add up to about 49 GW. The whole state averages about 4 GW. Four of the world's six largest planned campuses are in Utah, all in rural Millard, Emery and Box Elder counties, and several plan to build their own gas or nuclear power.
Stratos was approved by Box Elder County on May 4, 2026. A referendum was rejected, and an appeal and a civil-rights lawsuit are pending.
Utah electricity sales 2024: 34.7 TWh. Stratos: Deseret News, May 12, 2026.
About 4 GWUtah's average electricity use
49 GWplanned Utah data center campuses
FactSet, Introl, S&P, UBS, Bain, WSJ. Details in the research notes.
Each track is 100% of capex.
Add up everything announced plus the unreported share and you get $17 trillion. Discount the plans that probably won't happen and it is still $10.6 trillion. Count only what top-down spending can support by about 2030 and it is $6.4 trillion. The usual guess is around $2 trillion.
Only 61 of the 533 known campuses are running today, with 25 GW. 213 are under construction. The rest are plans, and the tracker Currence expects 55% of announced US hyperscale capacity never to come online.
Typical guess: the Gemini and Grok answers to the same research prompt ($2.1T and $2.7T). Top-down: JLL, Synergy, Census, capex, McKinsey ($6.7T by 2030).
The deal OpenAI said was too big to build itself is under 2% of the money and under 1% of the floor space now planned.
5 campuses, 12 million sq ft, 4.5 GW, $300 billion
About 711 campuses, 1.6 billion sq ft, 547 GW, $17 trillion
| Measure | Value | Range |
|---|---|---|
| OpenAI deal: campuses | 5 | 4 to 6 |
| OpenAI deal: floor space | 12.05M sq ft | 9.85M to 17.75M |
| OpenAI deal: power | 4.5 GW | 3.9 to 6.5 |
| OpenAI deal: contract | $300B | $150B to $319B |
| OpenAI deal: buildings (est.) | $60B | $36B to $75B |
| OpenAI deal: equipment (est.) | $130B | $117B to $158B |
| Known campuses | 533 | 533 to 593 |
| Known floor space | 1.21B sq ft | 0.94B to 1.48B |
| Known power | 410.5 GW | |
| Known buildings and land | $4.11T | $2.96T to $6.57T |
| Known hardware | $8.62T | $6.32T to $12.07T |
| Share publicly known | 75% | 60 to 85% |
| Unreported campuses | 178 | 94 to 355 |
| Unreported floor space | 404M sq ft | 214M to 808M |
| Unreported power | 136.8 GW | 72 to 274 GW |
| Total buildings and land | $5.47T | $2.16T to $10.95T |
| Total hardware | $11.50T | $4.63T to $20.12T |
| Total | $16.97T | $6.79T to $31.07T |
| Likely to get built | $10.56T | |
| Floor space multiple | 134× | |
| Power multiple | 122× | |
| Money multiple | 57× | 35× (likely built) to 57× |
Every campus in the circle chart is a row in claude/data_centers.csv with its own source link. Full notes: claude/final_answer.mdx and claude/research/.