Supercomputers/Analysis/Build pipeline
Analysis 11
The build pipeline: planned AI campus capacity
67 of the 103 planned or under-construction systems in this dataset state a power capacity, and together they state 55 GW. O'Leary Ventures alone accounts for 14% of it, and USA for 51%. Every figure here is what an operator or developer says it intends to build, not what has been delivered, and it should be read as a ceiling of intent rather than a forecast.
A stated capacity is a claim, not a build. These figures come from press releases, developer pages and filings, and they are not like-for-like: some are IT load, some are total facility or grid power, and some do not say. Of the pipeline, 10 GW is stated as facility power and 43 GW has no stated scope, and 2.0 GW is stated as IT load. The scope is on every row of the table below, and the figures are added together in the headline only because the alternative is to publish no total at all. Where a developer and an independent tracker disagree, the developer's number is used and the disagreement is in the basis column.
Phase 1 seed dataset, compiled by hand. These rows were built from public operator, laboratory and vendor sources. A mechanical second-reader pass has since fetched every cited source: 1092 of 1153 systems have a readable citation that names them, and 40 are genuinely weakly sourced. Every claim carries its source and a confidence tier. Treat anything below verified as a lead, not a citation.
- Under construction
- Planned
Table view: Stated capacity of the build pipeline by operator
| Operator | Systems | Under construction | Planned | Total | Share of pipeline |
|---|---|---|---|---|---|
| O'Leary Ventures | 1 | 0 MW | 7.5 GW | 7.5 GW | 14% |
| Meta Platforms | 3 | 7.0 GW | 0 MW | 7.0 GW | 13% |
| Scala Data Centers | 1 | 0 MW | 4.8 GW | 4.8 GW | 9% |
| Oracle | 4 | 4.3 GW | 0 MW | 4.3 GW | 8% |
| Tallgrass Energy | 1 | 2.7 GW | 0 MW | 2.7 GW | 5% |
| HUMAIN | 2 | 500 MW | 2.0 GW | 2.5 GW | 5% |
| Argyll Data Development | 1 | 0 MW | 2.0 GW | 2.0 GW | 4% |
| IREN | 1 | 2.0 GW | 0 MW | 2.0 GW | 4% |
| Poolside | 1 | 0 MW | 2.0 GW | 2.0 GW | 4% |
| Nebius | 3 | 1.8 GW | 0 MW | 1.8 GW | 3% |
| Brookfield | 1 | 0 MW | 1.8 GW | 1.8 GW | 3% |
| DataVolt | 1 | 1.5 GW | 0 MW | 1.5 GW | 3% |
| Fluidstack | 4 | 1.5 GW | 0 MW | 1.5 GW | 3% |
| Campus AI | 1 | 1.4 GW | 0 MW | 1.4 GW | 3% |
| SB Energy | 1 | 1.2 GW | 0 MW | 1.2 GW | 2% |
| Alphabet | 1 | 1.0 GW | 0 MW | 1.0 GW | 2% |
| G42 | 1 | 1.0 GW | 0 MW | 1.0 GW | 2% |
| Hut 8 | 1 | 1.0 GW | 0 MW | 1.0 GW | 2% |
| Crusoe | 1 | 900 MW | 0 MW | 900 MW | 2% |
| SoftBank Corp. | 2 | 300 MW | 400 MW | 700 MW | 1% |
| Applied Digital | 2 | 630 MW | 0 MW | 630 MW | 1% |
| CoreWeave | 3 | 320 MW | 300 MW | 620 MW | 1% |
| Mistral AI | 1 | 600 MW | 0 MW | 600 MW | 1% |
| NEXTDC | 1 | 0 MW | 550 MW | 550 MW | 1% |
| Racks Central | 1 | 510 MW | 0 MW | 510 MW | 1% |
| Firmus Technologies | 2 | 90 MW | 360 MW | 450 MW | 1% |
| Anthropic | 1 | 0 MW | 401 MW | 401 MW | 1% |
| Microsoft | 2 | 320 MW | 0 MW | 320 MW | 1% |
| Bell Canada | 1 | 300 MW | 0 MW | 300 MW | 1% |
| Cipher Mining | 1 | 300 MW | 0 MW | 300 MW | 1% |
| Cerebras | 2 | 165 MW | 100 MW | 265 MW | 0% |
| LG Uplus | 1 | 0 MW | 200 MW | 200 MW | 0% |
| Omnia Data Centers | 1 | 200 MW | 0 MW | 200 MW | 0% |
| Schwarz Digits | 1 | 200 MW | 0 MW | 200 MW | 0% |
| TELUS | 1 | 150 MW | 0 MW | 150 MW | 0% |
| Lambda | 2 | 123 MW | 0 MW | 123 MW | 0% |
| Reliance Intelligence | 1 | 120 MW | 0 MW | 120 MW | 0% |
| Khazna Data Centers | 1 | 100 MW | 0 MW | 100 MW | 0% |
| SK Telecom | 1 | 100 MW | 0 MW | 100 MW | 0% |
| Synectics Technologies | 1 | 0 MW | 100 MW | 100 MW | 0% |
| Yotta Data Services | 1 | 60 MW | 0 MW | 60 MW | 0% |
| RIKEN | 1 | 0 MW | 40 MW | 40 MW | 0% |
| National Polytechnic Institute (IPN) | 1 | 0 MW | 20 MW | 20 MW | 0% |
| CEA | 1 | 0 MW | 15 MW | 15 MW | 0% |
| Groq | 2 | 7 MW | 5 MW | 12 MW | 0% |
| Hypertec Cloud | 1 | 0 MW | 3 MW | 3 MW | 0% |
| Austrian Scientific Computing | 1 | 0 MW | 2 MW | 2 MW | 0% |
How this was calculated
A system is in the pipeline if its status is planned or under construction and the table of capacity claims holds at least one stated power figure for it. A claim is recorded only when a megawatt or gigawatt number is stated outright in a page the system already cites. Nothing is derived from GPU counts, and nothing is estimated: a system whose sources give a GPU count but no power figure has no row, and is listed below by name.
Where a system states several figures, one row is kept per stage (first phase, full build-out, contracted, operating) and the headline is the first that exists in the order full build-out, contracted, first phase. Bounds and ranges ("more than 1 GW", "700 to 900 MW") are recorded at the stated number with the qualifier in the basis column.
The operator is the organisation this dataset records as operating the system. For campuses built by a developer for a tenant, that is the tenant, following the Stargate Abilene convention, so the operator totals are by who will run the machines, not who pours the concrete.
Every claim in the pipeline
| System | Operator | Country | Status | Capacity | Scope | Stage | Stated window | Basis |
|---|---|---|---|---|---|---|---|---|
| Wonder Valley AI Campus | O'Leary Ventures | Canada | planned | 7.5 GW | unspecified | full buildout | - | 7.5 GW at full build-out per Site Selection and Gensler; a later report cites 9 GW source |
| Meta Hyperion | Meta Platforms | USA | under construction | 5.0 GW | unspecified | full buildout | 2030 (initial operational target) | Up to 5 GW announced by Zuckerberg July 2025; no date for full buildout source |
| Scala AI City | Scala Data Centers | Brazil | planned | 4.8 GW | unspecified | full buildout | - | Expansion potential up to 4.75 GW across 1,730+ acres, a design ceiling source |
| Project Jade (Cheyenne) | Tallgrass Energy | USA | under construction | 2.7 GW | unspecified | full buildout | - | 2.7 GW at buildout per the June 2026 report, up from the 1.8 GW first announced; potential to scale to 10 GW source |
| HUMAIN Project Halo | HUMAIN | Saudi Arabia | planned | 2.0 GW | unspecified | full buildout | - | Planned 2 GW supercluster source |
| IREN Sweetwater Hub | IREN | USA | under construction | 2.0 GW | unspecified | full buildout | - | 2,000 MW across Sweetwater 1 (1,400 MW) and Sweetwater 2 (600 MW), stated as the hub total source |
| Killellan AI Growth Zone | Argyll Data Development | United Kingdom | planned | 2.0 GW | unspecified | full buildout | - | Scaling beyond 2 GW at full build-out source |
| Poolside Project Horizon | Poolside | USA | planned | 2.0 GW | unspecified | full buildout | - | 2 GW campus in eight 250 MW phases source |
| Paducah AI Campus | Brookfield | USA | planned | 1.8 GW | facility | full buildout | - | Campus supports up to 1.8 GW of utility capacity; over 1.2 GW of compute capacity per the same page source |
| DataVolt Oxagon AI Campus | DataVolt | Saudi Arabia | under construction | 1.5 GW | unspecified | full buildout | - | Planned 1.5 GW Oxagon campus per the DataVolt announcement source |
| Campus AI Fouju | Campus AI | France | under construction | 1.4 GW | unspecified | full buildout | - | Planned 1.4 GW total for the Fouju campus per La Voix de France and Middle East AI News source |
| Stargate Shackelford County (Frontier) | Oracle | USA | under construction | 1.4 GW | facility | full buildout | - | Vantage says 1.4 GW campus, ten buildings, first building H2 2026; Epoch AI says 2.0 GW source |
| Nebius Independence AI Factory | Nebius | USA | under construction | 1.2 GW | unspecified | full buildout | - | Campus could reach up to 1.2 GW; first gigawatt-scale Nebius AI factory source |
| Stargate Milam County | SB Energy | USA | under construction | 1.2 GW | facility | full buildout | - | 1.2 GW site design capacity, SB Energy builds; Epoch AI also lists 1.2 GW source |
| Beacon Point (Hut 8) | Hut 8 | USA | under construction | 1.0 GW | facility | full buildout | - | 1,000 MW of utility capacity under an AEP Texas interconnection agreement source |
| Google India AI Hub (Vizag) | Alphabet | India | under construction | 1.0 GW | unspecified | full buildout | - | Google calls it a 1 gigawatt-scale AI ecosystem; Tribune reports about one gigawatt across three campuses; a qualitative target source |
| Meta El Paso (Texas) | Meta Platforms | USA | under construction | 1.0 GW | unspecified | full buildout | - | Meta: 1 gigawatt of compute capacity at the El Paso campus, operations to begin in 2028 source |
| Meta Lebanon (Indiana) | Meta Platforms | USA | under construction | 1.0 GW | unspecified | full buildout | - | Meta: 1GW campus in Lebanon, Indiana; a design capacity, no operational date stated source |
| Stargate Michigan (Saline Township) | Oracle | USA | under construction | 1.0 GW | unspecified | full buildout | - | Related Digital: more than 1 GW planned capacity, three buildings; Epoch AI says 1.4 GW source |
| Stargate Port Washington (Lighthouse) | Oracle | USA | under construction | 1.0 GW | unspecified | full buildout | - | Vantage: close to 1 GW of AI capacity, four data centers; Epoch AI says 1.3 GW source |
| Stargate UAE | G42 | United Arab Emirates | under construction | 1.0 GW | unspecified | full buildout | 2026 (first 200 MW) | One-gigawatt AI cluster in Abu Dhabi announced by G42 May 2025; capacity target source |
| Crusoe Abilene Campus (Microsoft) | Crusoe | USA | under construction | 900 MW | facility | full buildout | Mid-2027 (first building energized) | 900 MW campus, two buildings, first energized mid-2027; combined Abilene 2.1 GW source |
| Stargate Project Jupiter (New Mexico) | Oracle | USA | under construction | 900 MW | facility | full buildout | - | El Paso Matters: standalone microgrid of 700 to 900 MW (upper end); Epoch AI says 1.75 GW IT source |
| Lake Mariner (Fluidstack) | Fluidstack | USA | under construction | 750 MW | facility | full buildout | - | Up to 750 MW gross HPC leasing capacity at the campus, in the Q4 2025 results release source |
| Mistral Borlange AI Data Centre | Mistral AI | Sweden | under construction | 600 MW | unspecified | full buildout | - | Campus potential to scale to 600 MW per EcoDataCenter source |
| NEXTDC S7 Sydney AI Campus | NEXTDC | Australia | planned | 550 MW | it load | full buildout | - | NEXTDC lists S7 Sydney as 550 MW+ of IT capacity, in planning source |
| Racks Central RCJM AI Campus | Racks Central | Malaysia | under construction | 510 MW | unspecified | full buildout | - | Total campus capacity of 510 MW across four facilities source |
| HUMAIN AI Factory Phase 1 | HUMAIN | Saudi Arabia | under construction | 500 MW | unspecified | full buildout | - | Nvidia states a projected AI factory capacity of up to 500 MW over the next five years source |
| Delta Forge 1 (Rapides Parish) | Applied Digital | USA | under construction | 430 MW | facility | full buildout | - | Delta Forge 1 is described as a 430 MW AI factory campus (utility power) source |
| Justified Data (Anthropic) | Anthropic | USA | planned | 401 MW | it load | contracted | - | 20-year Anthropic lease for 401 MW of critical IT load, full load by early 2028 source |
| SoftBank Sakai AI Data Center | SoftBank Corp. | Japan | planned | 400 MW | unspecified | full buildout | - | Future expansion to over 400 MW per the SoftBank announcement source |
| Firmus Batam AI Factory | Firmus Technologies | Indonesia | planned | 360 MW | unspecified | full buildout | - | Firmus: 360 MW AI factory campus for up to 170,000 accelerators; announced 28 June 2026 source |
| Nebius Lappeenranta AI Factory | Nebius | Finland | under construction | 310 MW | unspecified | full buildout | - | 310 MW AI factory, first capacity for customers in 2027 source |
| Barber Lake (Fluidstack) | Fluidstack | USA | under construction | 300 MW | facility | full buildout | - | Barber Lake site 300 MW approved power capacity per Cipher's business update source |
| Bell AI Fabric Saskatchewan | Bell Canada | Canada | under construction | 300 MW | unspecified | full buildout | - | Bell states a 300 MW data centre, coming online in stages from the first half of 2027; scope not specified source |
| Cipher Black Pearl (AWS) | Cipher Mining | USA | under construction | 300 MW | facility | full buildout | - | About 300 MW of gross capacity leased to Amazon Data Services for 15 years source |
| CoreWeave Lancaster AI Data Center | CoreWeave | USA | planned | 300 MW | unspecified | full buildout | - | Expandable to 300 MW at full build-out per the CoreWeave announcement source |
| Nebius Birmingham AI Factory | Nebius | USA | under construction | 300 MW | facility | full buildout | - | Local coverage of the permit application gives 300 MW of power consumption served by Alabama Power; not a Nebius page source |
| SoftBank Tomakomai AI Data Center | SoftBank Corp. | Japan | under construction | 300 MW | unspecified | full buildout | - | SoftBank: power capacity expected to exceed 300 MW source |
| CoreWeave Kenilworth (NEST) | CoreWeave | USA | under construction | 250 MW | unspecified | full buildout | - | Described as a 250 MW AI data center; whether IT or facility load is not stated source |
| River Bend (Fluidstack) | Fluidstack | USA | under construction | 245 MW | it load | contracted | - | 15-year Fluidstack lease for 245 MW of critical IT capacity source |
| Microsoft Azure AI datacenter (Narvik, Norway) | Microsoft | Norway | under construction | 230 MW | facility | full buildout | 2026 (staged deployment begins) | Nscale: more than 30,000 Rubin GPUs on a 230 MW campus; matches site substation_mw source |
| LG Paju AI Data Center | LG Uplus | South Korea | planned | 200 MW | unspecified | full buildout | - | LG Uplus CTO: Paju AI data center capacity of 200 MW, up to 120,000 GPUs, completion in 2027 source |
| Omnia Pecem (TikTok) | Omnia Data Centers | Brazil | under construction | 200 MW | it load | first phase | - | First phase about 200 MW productive IT load inside a roughly 300 MW power envelope (300 MW is the consumption figure) source |
| Polaris Forge 2 (Harwood) | Applied Digital | USA | under construction | 200 MW | it load | contracted | - | 15-year take-or-pay lease for 200 MW of critical IT load, funded by 2.15 billion dollar notes source |
| STACKIT Lubbenau AI Data Center | Schwarz Digits | Germany | under construction | 200 MW | facility | first phase | - | Connected load of about 200 MW for the Lubbenau data center, first phase of three modules due by end of 2027 source |
| Abernathy (Fluidstack) | Fluidstack | USA | under construction | 168 MW | it load | first phase | - | 168 MW critical IT load for delivery in second half of 2026; 240 MW gross capacity source |
| Cerebras Mikkeli AI Data Centre | Cerebras | Finland | under construction | 165 MW | it load | full buildout | - | 165 MW of contracted IT capacity at full scale source |
| TELUS BC Sovereign AI Cluster | TELUS | Canada | under construction | 150 MW | unspecified | full buildout | - | TELUS: cluster will scale to over 60,000 GPUs and 150 MW by 2032 source |
| Reliance Intelligence Jamnagar AI Compute Platform | Reliance Intelligence | India | under construction | 120 MW | unspecified | first phase | - | First 120 MW phase targeted for commissioning by end of 2026 source |
| Cerebras Guyana AI Data Center | Cerebras | Guyana | planned | 100 MW | facility | full buildout | - | MoU for an AI data center of up to 100 MW at Wales, Guyana source |
| Khazna Ajman AI Data Center | Khazna Data Centers | United Arab Emirates | under construction | 100 MW | unspecified | full buildout | - | 100 MW AI-optimized data center, 20 halls of 5 MW each source |
| Lambda Kansas City AI Factory | Lambda | USA | under construction | 100 MW | unspecified | full buildout | - | Scalable to more than 100 MW source |
| SK Ulsan AI Data Center | SK Telecom | South Korea | under construction | 100 MW | unspecified | full buildout | - | 100 MW facility at full build-out by February 2029 with about 60,000 GPUs source |
| USIO | Synectics Technologies | Uganda | planned | 100 MW | facility | full buildout | - | Aeonian campus 100 MW total (USIO plus six 15 MW modules); 100 MW pre-transmission allocation source |
| Firmus Southgate Launceston | Firmus Technologies | Australia | under construction | 90 MW | unspecified | first phase | - | W.Media reports Stage 1 at 90 MW (44 MW then 90 MW) before a planned 300 MW Stage 2; a target relayed from Firmus source |
| Microsoft Azure AI supercomputer (Loughton, UK) | Microsoft | United Kingdom | under construction | 90 MW | facility | full buildout | Q1 2027 (initial GB300 delivery) | Nscale: 50 MW scaling to 90 MW at Loughton campus; matches site substation_mw source |
| Muskogee (CoreWeave) | CoreWeave | USA | under construction | 70 MW | it load | full buildout | - | Core Scientific: 70 MW of critical IT load for GPUs within a 100 MW total facility (30 MW ancillary), expected operational in 2026 source |
| Yotta D2 Blackwell Ultra Cluster | Yotta Data Services | India | under construction | 60 MW | facility | first phase | - | Supercluster deployed at Yotta's 60 MW D2 data center in Greater Noida source |
| FugakuNEXT | RIKEN | Japan | planned | 40 MW | unspecified | full buildout | - | RIKEN: aims to deliver up to a hundredfold application gain within the same approximately 40 MW power constraint as Fugaku; a design constraint source |
| Lambda Chicago AI Factory | Lambda | USA | under construction | 23 MW | unspecified | full buildout | - | Single-tenant 23 MW data centre ready for service in 2026 source |
| Coatlicue | National Polytechnic Institute (IPN) | Mexico | planned | 20 MW | unspecified | full buildout | - | IPN article says the site decision must consider supply of the "approximately 20 megawatts" the machine requires; scope not stated source |
| Alice Recoque | CEA | France | planned | 15 MW | unspecified | full buildout | 2027 | GENCI states under 15 MW (upper bound) for the exascale system source |
| GroqCloud Kamloops Inference Cluster | Groq | Canada | under construction | 7 MW | unspecified | first phase | - | 7 MW Groq facility is the first Bell AI Fabric site source |
| GroqCloud Sydney Inference Cluster | Groq | Australia | planned | 5 MW | unspecified | full buildout | - | 4.5 MW Groq facility in Equinix Sydney source |
| Mila Sovereign AI Research Hub | Hypertec Cloud | Canada | planned | 3 MW | unspecified | full buildout | - | Mila: up to 3 MW of secure AI infrastructure capacity at the hub, a design capacity source |
| AI:AT | Austrian Scientific Computing | Austria | planned | 2 MW | unspecified | full buildout | - | EuroHPC tender document: a total power capacity of up to 2 MW will be made available for the complete system, a design allowance source |
By sector, country and window
| Sector | Systems | Stated capacity |
|---|---|---|
| Neoclouds | 37 | 35 GW |
| Big tech | 17 | 14 GW |
| AI labs | 3 | 3.0 GW |
| Industry | 6 | 3.0 GW |
| Research HPC | 3 | 62 MW |
| Government labs | 1 | 15 MW |
| Country | Systems | Stated capacity |
|---|---|---|
| USA | 29 | 28 GW |
| Canada | 5 | 8.0 GW |
| Brazil | 2 | 5.0 GW |
| Saudi Arabia | 3 | 4.0 GW |
| United Kingdom | 2 | 2.1 GW |
| France | 2 | 1.4 GW |
| India | 3 | 1.2 GW |
| United Arab Emirates | 2 | 1.1 GW |
| Japan | 3 | 740 MW |
| Australia | 3 | 645 MW |
| Sweden | 1 | 600 MW |
| Malaysia | 1 | 510 MW |
| Finland | 2 | 475 MW |
| Indonesia | 1 | 360 MW |
| South Korea | 2 | 300 MW |
| Norway | 1 | 230 MW |
| Germany | 1 | 200 MW |
| Guyana | 1 | 100 MW |
| Uganda | 1 | 100 MW |
| Mexico | 1 | 20 MW |
| Austria | 1 | 2 MW |
| Operator's stated first-power year | Systems | Stated capacity |
|---|---|---|
| 2026 | 2 | 1.2 GW |
| 2027 | 3 | 1.0 GW |
| 2030 | 1 | 5.0 GW |
| No stated window | 61 | 48 GW |
What is missing
36 planned or under-construction systems state no power figure in any source this dataset cites, so they are not in the totals above: DAEDALUS, Doudna, Horizon, Discovery (ORNL), Lux (ORNL), Solstice, Equinox, Blue Lion, KISTI-6, Google TPU (Ironwood) capacity for Anthropic, BMS AI Factory, El Khawarizmi Supercomputer, JD Cloud 100,000-GPU Cluster, Hyundai Motor Group Blackwell AI Factory, Samsung AI Megafactory, Hon Hai Kaohsiung Super Computing Center, HammerHAI, Mimer AI Factory, IT4LIA, Project Ceiba, Korea National AI Computing Center, UK Next National Supercomputer, Gaia AI Factory, LUMI-AI, HUMAIN Qualcomm AI 100 Cluster, Herder, Firmus Southgate Melbourne, Pangea 5, Lynx (Arizona), MeluXina-AI, KarolAIna, SLAIF, Discoverer++, Levente, MareNostrum 5 AI, China Datang Ulanqab Computing Center. Most are national or sovereign machines announced by GPU count or contract value, which is itself a finding: the operators that publish megawatts are the ones building at a scale where the grid connection, not the silicon, is the constraint.
The read
The pipeline is a story about very few organisations. O'Leary Ventures states 7.5 GW, more than the next operator, and the top five operators state 48% of the whole. That concentration is not only an artefact of the sample: these are the campuses that the electricity and switchgear industries are actually being asked to deliver, and 24 of them individually state more capacity than every operating system that states a figure, combined.
For scale, the 12 systems already operating that state a power figure total about 612 MW. The pipeline is 90 times that, and 59% of it is recorded as under construction. Yet 87% of it has no stated first-power year at all, which is what an announcement looks like before a utility interconnection agreement exists.
Two cautions on reading the totals. First, capacity is not compute: a megawatt of 2028 silicon is worth several times a megawatt of 2023 silicon, and no page on this site will convert one into the other without a published basis. Second, announcements are not delivered, and several of the largest figures have already moved between sources. The dated basis on every row is there so the claim can be checked against what actually energises.