23 Flagship AI Data Centers — An Overview

On July 14, 2025, Zuckerberg posted a brief message on Threads announcing that Meta’s 1 GW-class AI supercomputing campus “Prometheus” in New Albany, Ohio would come online in 2026, while an even more audacious campus — “Hyperion” — was breaking ground in Louisiana with a target of 5 GW and a footprint large enough to “cover a large part of Manhattan”[1].

Behind that post lies a trend quietly taking shape but steeply accelerating: the world’s largest AI data centers, each targeting gigawatt-scale power consumption, are redefining what “massive industrial facility” means.

This is a snapshot of who is moving mountains. The data comes primarily from Epoch AI’s Frontier Data Centers database, published in November 2025 — cross-referenced from satellite imagery, power permit filings, and equipment orders to track the world’s largest AI data centers, with power draw, estimated compute (expressed as H100-equivalent GPU count), and capex as the three key figures for each site[3].

How This List Is Counted

Epoch uses a unified model to estimate capex: $44B per GW of server power ($30B for IT hardware, $14B for construction, power infrastructure, cooling, and land), applied consistently across generations[4]. For compute, if chip counts are reliably disclosed they are used directly; otherwise the figure is back-calculated: compute = (chip FLOP/s/W × total power) / (server overhead × IT overhead × peak PUE).

At the time of the database’s launch (November 2025), the 13 initially tracked sites combined for roughly 2.5 million H100-equivalent GPUs — approximately 15% of all H100-equivalent GPUs shipped globally since the H100’s launch[3].

The figures I use below come from the latest snapshot as of May 31, 2026 (pulled from Epoch’s public CSV)[5]. Compared with the database’s initial release six months ago, the power figures for leading sites have nearly doubled across the board — the pace of change in this space is probably faster than you’d expect.

The 13 Operational Sites (Ranked by Current Power Draw)

The 13 sites combined: ~5.7 GW, ~4.3 million H100-equivalent GPUs, ~$155B in capex. The * marker on tenant assignments reflects Epoch’s designation of those attributions as speculative, carrying some uncertainty.

A few of the most striking changes compared with the November 2025 initial snapshot: Microsoft Fairwater Atlanta went from 433 MW to 859 MW within six months, displacing Meta Prometheus as the world’s #1 operational AI data center; xAI Colossus 2 doubled to 661 MW; Google Pryor (North) surged from a 65 MW node to 454 MW — a 7× increase — making it the single site with the largest incremental gain over the period.

What follows is a site-by-site breakdown of all 13.

Microsoft Fairwater Atlanta

The Atlanta site, located in Fayetteville, Georgia, began small-scale operations in October 2025 and was officially announced by Microsoft on November 12; it spans 85 acres and over one million square feet, and is the second node in the Fairwater family. By May 2026, at 859 MW, it had displaced Meta Prometheus to become the world’s #1 operational AI data center[6]. Power is supplied by Georgia Power (a Southern Company subsidiary). The building design carries over the two-story architecture pioneered at Fairwater Wisconsin, with further density improvements — packed with NVIDIA GB200 NVL72 racks and a closed-loop liquid cooling system that draws near-zero water; UPS systems and diesel generators have been eliminated in favor of grid power backed by batteries.

The most important architectural innovation isn’t at the site level but in the “AI Superfactory” model — linked to Fairwater Wisconsin via a dedicated AI WAN, treating the compute across two states as a single distributed supercomputer for training[7]. On the claim that it is “built for OpenAI workloads”: Epoch’s label is “Likely OpenAI*” — Microsoft has said the facility “will initially be used to train OpenAI models,” but given the renegotiated Microsoft-OpenAI agreement in 2025, this attribution carries uncertainty.

Meta Prometheus

Prometheus is built adjacent to Meta’s existing New Albany Licking County campus (1500 Beech Road) as its newest flagship building; Zuckerberg announced on a brief Threads post on July 14, 2025 that it would “come online in 2026 at 1 GW scale,” framing it as “the world’s first 1 GW+ AI data center”[1]. Epoch’s May 2026 tracking shows 820 MW currently live, having come online in Q1 2026 and held the global #1 position until Atlanta’s 859 MW pushed it to #2.

Power is the central story of Prometheus: local utility AEP Ohio handles high-voltage grid interconnection; Meta subsidiary Sidecat separately won Ohio state approval to build a 200 MW behind-the-meter gas project for direct supply. Beyond gas and grid, Meta also signed a 20-year nuclear PPA with Constellation Energy (June 2025), and in January 2026 finalized a combined 6.6 GW of new and renewed nuclear agreements with Vistra, Oklo, and TerraPower, with Vistra’s two operating Ohio nuclear plants directly supplying over 2.1 GW[8].

xAI Colossus 1 + 2

Colossus 1 was converted from a decommissioned Electrolux factory in South Memphis, bringing 100,000 H100s online in just 122 days in 2024 — Jensen Huang noted that scale would normally take four years — and then expanding to 200,000 within another 92 days[10]. In March 2025, xAI broke ground on Colossus 2 in Southaven, Mississippi (the Memphis metro’s cross-state side), which SemiAnalysis described as having “accomplished in six months what Oracle/Crusoe/OpenAI took 15 months to do.” In late December 2025, Musk confirmed on X that a third Colossus building had been integrated, leading SemiAnalysis to project a total scale of 555,000 GPUs, 2 GW, and $18B in capex, with a target of bringing 1.1 GW of on-site gas turbines fully online by Q2 2027[11].

Power is the project’s most controversial aspect: Memphis Light, Gas and Water (MLGW) can supply only around 150 MW, so xAI erected gas turbines on-site via a JV with publicly traded Solaris Energy Infrastructure (Solaris holding 50.1%, xAI 49.9%) — most of the original 35 turbines at Colossus 1 were unpermitted, and following a warning from SELC (Southern Environmental Law Center), the number was reduced to 15 with permits obtained retroactively. In April 2026, NAACP joined SELC and Earthjustice in suing xAI, alleging that 27 unpermitted gas turbines at the Colossus 2 site emit over 1,700 tons of NOx annually, potentially making it the largest single industrial nitrogen oxide source in Greater Memphis, with pollution falling disproportionately on South Memphis’s Black community[12].

OpenAI Stargate Abilene

OpenAI has no physical infrastructure of its own, instead distributing its training compute across Microsoft, Oracle, and Crusoe. Stargate Abilene was the first to come online: designed and built by Crusoe Energy Systems, running on Oracle Cloud, serving OpenAI. On September 30, 2025, Crusoe officially announced the first two buildings were live, running GB200 clusters; the full plan calls for eight buildings targeting 1 GW or more in total[13]. Epoch’s May 2026 tracking shows 590 MW currently live (roughly 295 MW at the start of the year) — doubling in six months.

This site is frequently misread by outsiders — Stargate is often described as “OpenAI’s data center,” but land, equipment, and grid interconnection are in fact Oracle’s domain, with OpenAI serving as its largest (and currently only) tenant; Crusoe acts as general contractor and operator[14].

Google Pryor (North)

The MidAmerica Industrial Park in Mayes County, Pryor, Oklahoma is Google’s second-largest U.S. data center, with cumulative investment of roughly $4.4 billion. The leap in scale was directly triggered by the August 13, 2025 joint announcement by Alphabet CFO Ruth Porat and Governor Stitt of a $9 billion Oklahoma expansion (Pryor expansion plus a new Stillwater site) — which corresponds precisely to Epoch’s May 2026 snapshot showing Pryor jumping from 65 MW to 454 MW, nearly a 7× increase[15].

Power follows a hybrid architecture of PSO transmission and distribution, GRDA cooperative capacity, and self-procured or purchased clean energy: Google signed a 372 MW Mayes County solar agreement with Leeward Renewable Energy (online by end of 2025), locked in a 600 MW long-term PPA with NextEra, and from late 2025 NextEra also incorporated natural gas into the generation mix supporting Pryor and Stillwater. The campus runs Google’s in-house TPUs rather than NVIDIA GPUs — Google is the only player on the Frontier list not using NVIDIA.

Google New Albany

Google broke ground on its first $600M data center in New Albany, east of Columbus, in November 2019, and the site has snowballed ever since, now constituting Ohio’s second-largest AI compute cluster. In June 2024, Alphabet announced $1.7 billion in additional investment across its New Albany, Columbus, and Lancaster Ohio campuses; another $1 billion expansion and a 618-acre land acquisition followed in 2025[16]. SemiAnalysis estimates that New Albany’s current live compute already exceeds 1 GW, with roughly 500 MW dedicated to AI — Epoch’s 407 MW figure captures only the AI portion.

The most interesting aspect is co-location: Google’s New Albany data center and Meta’s flagship Prometheus (1500 Beech Road) are literally across the street from each other in the same New Albany Business Park, sharing AEP Ohio transmission infrastructure and Columbus municipal water. One of the rare U.S. AI nodes where two hyperscalers have simultaneously placed their bets.

Amazon Madison Mega Site (Canton, MS)

Mississippi Governor Tate Reeves convened a special legislative session in January 2024 to secure AWS’s $10 billion investment at the Madison County Mega Site — a “mega-site” originally reserved for an automotive assembly plant outside the town of Canton — with the largest incentive package in the state’s history: a 10-year corporate income tax exemption, a 3% construction rebate, rolling tax abatements for up to 30 years, a $44M grant, and $215M in local infrastructure[17].

Entergy Mississippi assumed $300M in grid upgrades through its “Superpower Mississippi” initiative, and committed that starting in 2027 reclaimed water would be the sole cooling source (roughly 83 million gallons recycled per year). The Canton site hosts Project Rainier — one of the Trainium2 clusters used to train Anthropic’s Claude across multiple states (nearly 500,000 Trainium2 chips nationwide, with a target of exceeding one million by end of 2026)[18]. In April 2026, AWS announced an additional $11 billion Madison expansion, bringing Mississippi’s total investment commitment to $25 billion.

Google Omaha

In 2022 Google announced the purchase of roughly 260 acres at the intersection of Blair High Road and State Street, northwest of Omaha, for its second Nebraska campus; in 2023 it expanded northward by another 187 acres, bringing total land to nearly 460 acres, with plans for four data center buildings totaling over 2.2 million square feet. The first building came online in 2024 with subsequent buildings under simultaneous construction[19]. Power is supplied by Omaha Public Power District (OPPD), with a dedicated OPPD switching station on campus; Epoch’s May 2026 revision raised the figure from 189 MW to 284 MW, reflecting new buildings coming online in early 2026.

Within Google’s Midwest footprint, Omaha joins Papillion, Council Bluffs, the newly planned Lincoln campus (2.2 million sq ft, developed with Tract), and Cedar Rapids to form a Midwestern compute belt spanning the MISO and SPP grids.

Microsoft Goodyear (PHX-70)

Microsoft’s campus in Goodyear, Arizona, carries the designation PHX-70, situated in the western suburbs of the Phoenix metro area on 279 acres, supplied by Arizona Public Service (APS) with additional renewable power through a Sun Streams 2 solar PPA[20]. The campus is being built in five phases: Phase I/II (143 MW) came online between 2024 and 2025 using direct evaporative cooling; Phase III/IV is planned for 2026; Phase V is slated for 2027, where the new phases shift to a “zero-water” air-cooling design to accommodate AI chip thermal density and address Phoenix’s water constraints. The facility is widely understood to be one of Azure’s core nodes in U.S. West Region 3, providing compute for OpenAI training workloads.

Alibaba Zhangbei

Alibaba Cloud’s flagship compute cluster in Zhangbei County, Zhangjiakou, Hebei, with its first data center commissioned in 2016, currently estimated by Epoch at roughly 200 MW. The core logic behind the Zhangbei location is its natural cold source — latitude 41°N, mean annual temperature 2.6°C — enabling extensive fresh-air free cooling, with a PUE around 1.23, roughly 35% more efficient than conventional mechanical cooling.

The site sits within the Beijing-Zhangjiakou Renewable Energy Demonstration Zone, with wind and solar power from the Bashang plateau feeding the campus directly; it provided the “fully green” cloud compute infrastructure for the 2022 Beijing Winter Olympics[21]. Under China’s “East Data, West Compute” (东数西算) initiative, Zhangjiakou is designated as a Beijing-Tianjin-Hebei hub node, and the Zhangbei cluster serves as the northern anchor of Alibaba’s AI compute (handling Qwen training workloads) in its north-south scheduling architecture. This is currently the only non-U.S. operational site on the Frontier list.

Meta Temple

Meta’s campus in Temple, Texas (Bell County, 2310 Eberhardt Road), general-contracted by JE Dunn, spans 386 acres with three buildings totaling over 750,000 square feet. The project was originally planned to come online by the end of September 2026, but actually went live roughly nine months early, energizing on December 9, 2025[22]. The grid interconnection is through Heart of Texas Electric Cooperative rather than Oncor, cleverly bypassing ERCOT main-grid expansion bottlenecks — a rare “co-op bypass” site selection move for Meta in Texas. In Meta’s AI campus sequence, Temple is a “mid-weight standard site” — far smaller than Louisiana’s Hyperion or Ohio’s Prometheus, and closer in style to Meta’s existing H-type campus template.

Google Council Bluffs East

Council Bluffs, Iowa, is the most senior site in Google’s entire Midwest AI compute footprint — Google has operated data centers in Council Bluffs since 2007, and it now hosts five Google campuses in the area. The “Council Bluffs East” site tracked by Epoch is the newest AI-dedicated expansion among them, at 190 MW. Power is provided by MidAmerican Energy, the vast majority coming from MidAmerican’s large-scale Iowa wind fleet — which is also the core reason Google has long treated Iowa as a strategic AI training location: cheap, reliable, and clean power.

The 10 Under Construction or Planned

This is the list that will actually determine the compute landscape over the next two to three years:

Meta Hyperion

On October 21, 2025, Meta signed a joint venture agreement with funds managed by Blue Owl Capital worth approximately $27 billion — the largest private credit transaction in history. An SPV arranged by Morgan Stanley issued $27B in A+-rated bonds plus $2.5B in equity, anchored by PIMCO ($18B) and BlackRock ($3B)[23]. In the JV structure Blue Owl holds 80% and Meta only 20%, plus a one-time distribution of roughly $3B flowing back to Meta — Meta effectively moved the largest capital expenditure it has ever made “off balance sheet”, with subsequent use of the full facility via an initial 4-year, renewable operating lease. This is a milestone in hyperscaler financial engineering.

The campus is located in Holly Ridge, Richland Parish, Louisiana, and has now expanded to a 2,250-acre primary site plus an adjacent 1,400-acre newly acquired parcel, leading Epoch to estimate “a footprint of roughly 11 km², equivalent to one-fifth of Manhattan Island”[24]. Construction is led by a Mortenson, DPR, and Turner joint venture, targeting full operation by summer 2028 with associated civil works extending to 2030; ultimate compute scale: 5 GW.

On power, Meta’s partnership with Entergy has grown from the initially committed 3 plants to 10 gas-fired plants announced in March 2026 totaling 7.5 GW — adding more than 30% to Louisiana’s existing grid capacity — supplemented by up to 2.5 GW of renewables and battery storage[25]. The practical implication: for a single AI data center, an entire state’s power system is being rebuilt.

Microsoft Fairwater Wisconsin

The Fairwater brand name was first formally introduced in a Microsoft blog post authored by Scott Guthrie’s team on September 18, 2025. The campus sits in Mount Pleasant, Wisconsin, on 315 acres with three large server buildings totaling 1.2 million square feet, with an initial committed investment of $3.3 billion and an early-2026 go-live date; layered with an additional $4 billion, Wisconsin’s total investment exceeds $7.3 billion. In January 2026, Mount Pleasant approved 15 additional new buildings[26].

Epoch AI’s November 2025 projection: once the fourth building is complete, total power will reach 3.3 GW by end of 2027 — exceeding Los Angeles’s average 2023 consumption (roughly 2.4 GW), equivalent to the full output of 3–4 large nuclear reactors. Total capex is estimated to exceed $100 billion[27]. Power is provided by local utility WE Energies, supplemented by a new 250 MW solar project being built in Portage County to support Microsoft’s clean energy procurement.

Fairwater Wisconsin is also the birthplace of the two-story server room design — stacking GB200 NVL72 racks vertically to reduce inter-GPU interconnect latency — a design that Fairwater Atlanta inherited wholesale. In Microsoft’s overall “AI Superfactory” blueprint, Wisconsin is the first node, Atlanta is the second, with more Fairwater sites to follow, all networked via a dedicated AI WAN to train as one supercomputer spanning multiple states.

Anthropic-Amazon New Carlisle

The Project Rainier campus in New Carlisle, Indiana, is AWS’s flagship AI training base, with an $11 billion investment spanning 1,200 acres. It began partial power-on in January 2026 and was fully operational by October, at a final power scale of 2.2 GW. The campus is dedicated exclusively to Anthropic, currently deploying approximately 500,000 AWS in-house Trainium 2 chips networked via UltraServer architecture for training the Claude model family[18].

On April 20, 2026, Anthropic and Amazon announced an expanded partnership: Amazon committing up to $25 billion in additional investment, Anthropic committing to purchase over $100 billion in compute from AWS over the next ten years, with new “up to 5 GW” exclusive Trainium capacity, of which nearly 1 GW of Trainium 2/3 will be online by end of 2026. This is one of the largest single-tenant dedicated AI clusters ever built anywhere in the world, and AWS’s definitive answer to the OpenAI-Microsoft model.

Amazon Ridgeland

On April 9, 2026, Amazon announced $12 billion in additional Mississippi investment — $11 billion to expand the AWS campus in Ridgeland (also in Madison County) and another $1 billion for a new sister facility in Clinton[28]. Ridgeland and the already-operational Canton (Madison Mega Site) are effectively expansion phases within the same county rather than independent sites — the separate town names distinguish locations, but both are extensions of the same Anthropic-dedicated compute belt. Power is supplied by Entergy Mississippi, with Amazon bearing all incremental substation and transmission upgrade costs.

Water cooling is active only roughly 30 days per year (about 9% of the time), with air cooling handling the rest. Including the Warren County and Clinton projects, Amazon’s total Mississippi commitment stands at $25 billion across four campuses — the largest single capital expenditure in state history.

OpenAI Stargate Abu Dhabi

The first overseas Stargate node, jointly announced on May 22, 2025 by OpenAI, Oracle, SoftBank, NVIDIA, Cisco, and G42, sited in Abu Dhabi. Within the broader “5 GW US-UAE AI Campus” framework, the initial target is 1 GW, with the first 200 MW being built and operated by Khazna Data Centres (a G42 subsidiary), using NVIDIA Grace Blackwell GB300 systems[29]. A site visit reported by The National in October 2025 showed the campus taking shape: over 5,000 workers on-site, more than 100,000 cubic meters of concrete poured, steel structural weight equivalent to 1.5 Eiffel Towers, with the first 200 MW phase targeting delivery in Q3 2026 and the full 1 GW cluster to be completed within three years[30].

On the capital side, UAE sovereign fund MGX invested $7 billion in the Stargate parent project, matching Oracle, and recently joined Silver Lake in a stake in Khazna. The biggest uncertainty is export controls — G42 historically had equity or technology ties with ByteDance and Huawei; while the U.S. authorized the export of compute equivalent to roughly 35,000 NVIDIA GB300 chips to G42 in late 2025, this came with strict compliance conditions including on-site bans on Chinese equipment and personnel background checks, with further authorization tranches currently on hold.

The significance of this deal goes beyond adding one more data center — it represents the first time a non-U.S. government has engaged at a national level in an AI compute alliance.

OpenAI Stargate Shackelford

One of the “five new Stargate sites” announced by OpenAI / Oracle / SoftBank on September 23, 2025, located in Shackelford County, Texas, developed by Vantage Data Centers, leased by Oracle, and subleased to OpenAI[14]. The campus is code-named “Frontier”, spanning 1,200 acres, with plans for 10 single-story data center buildings totaling roughly 3.7 million square feet, IT capacity of 1.4 GW with rack densities exceeding 250 kW, using a hybrid air and liquid cooling architecture, with the first building planned to light up in the second half of 2026.

The most contentious aspect is the power strategy: the campus is entirely off the ERCOT public grid, with Voltagrid deploying 210 industrial gas generator sets forming a 700 MW microgrid (197 primary + 13 standby), expandable to 1.4 GW — rather than wait in Texas’s grid interconnection queue, the answer was simply to build their own power plant. Shackelford is the most aggressive “off-grid bypass” example in the OpenAI ecosystem.

Crusoe Goodnight

A gigawatt-class AI campus developed by Crusoe and Hoffman in Armstrong County, in the Texas Panhandle, with Google Cloud confirmed as the anchor customer (not DeepMind-exclusive). Total planned capacity is 1 GW+; the first four buildings connect to the grid, while buildings five and six use a “behind-the-meter” approach, directly powered by 933 MW of on-site natural gas generation — marking the first instance of Google backing an on-site gas plant for a single AI campus.

Annual carbon emissions are estimated at 4.5 million tons, drawing strong opposition from environmental groups[31]. An adjacent phase has signed a net metering agreement with Serena Energy’s Goodnight Wind farm (265.5 MW, commissioned in 2024), with Google locking in 265 MW of that wind capacity — though the company emphasizes it has “not signed a power purchase agreement with the gas plant” to avoid carbon accounting controversy. The project equity transferred from Crusoe to Google in 2025 — marking hyperscalers’ formal entry into the integrated “own-the-power-plant + AI-training” model.

CoreWeave Helios

The Helios campus is located roughly four miles northwest of Afton in Dickens County, Texas, on 160 acres — originally a bitcoin mining facility built by Argo Blockchain in 2022, acquired by Galaxy Digital in December of that year and converted into an AI data center. CoreWeave signed a 15-year long-term lease with Galaxy, by 2025 had locked in the full initial 800 MW capacity and exercised an option for an additional 260 MW; in January 2026 ERCOT approved another 830 MW expansion, more than doubling the campus total to 1.6 GW+, with a $3.5 billion Phase 2 expansion budget[32].

On “Microsoft as the ultimate end customer”: Microsoft is indeed the largest downstream customer in CoreWeave’s overall contract structure (its 2024 IPO prospectus disclosed a revenue concentration exceeding 60%), leading the industry to conclude that Helios compute most likely ultimately serves OpenAI/Microsoft training workloads — though the contract-level characterization can only be stated as “CoreWeave proprietary AI capacity.” The “bitcoin mine → AI training campus” asset conversion path will only become more common through this cycle.

Google Cedar Rapids

In late January 2025, the Cedar Rapids city council approved two hyperscale projects at Alliant Energy’s Big Cedar Industrial Center in a single session — and by late May Google officially confirmed its initial $576 million investment, planning “at least four large buildings” with the first phase targeting a 2026 go-live[33]. Google subsequently increased its total Iowa commitment to $7 billion, making Cedar Rapids alongside the existing Council Bluffs site the dual anchor of its Midwest footprint.

QTS Cedar Rapids

QTS (owned by Blackstone) is building a multi-tenant colocation campus on a 612-acre parcel in southwest Cedar Rapids, within the same Big Cedar Industrial Park as Google, jointly announced in February 2025 by the city, Alliant Energy, and QTS, with a minimum committed investment of $750 million and projected long-term capex of $10 billion — the largest single economic development investment in Iowa and Cedar Rapids history[33]. Seven single-story data center buildings are planned using water-free cooling; tenants have not been officially named, but given QTS’s parcel’s proximity to Google’s approved $576M site in the same Big Cedar park, the industry broadly expects the anchor tenant to be Microsoft or Meta.

Cross-Cutting Observations

Viewing all 23 sites together, several cross-cutting patterns are worth noting.

1. The world’s largest AI compute is concentrated in fewer than thirty locations. Epoch estimates that the 13 operational sites combined are equivalent to roughly 15% of all H100-equivalent GPUs shipped globally. This is a highly concentrated industry — similar to semiconductor manufacturing, with extreme winner-take-most dynamics at the top.

2. The “AI lab × infrastructure provider” axis defines five distinct models:

3. Four geographic clusters:

• Texas corridor (Abilene, Temple, Shackelford, Afton, Goodnight, and cross-state Memphis-Southaven): cheap power, light permitting, wind+gas mix — OpenAI and xAI’s primary battleground
• Ohio / Iowa / Nebraska Midwest corridor (New Albany, Council Bluffs, Omaha, Cedar Rapids): Google and Meta’s traditional home turf, connected to PJM and MISO grids
• Mississippi + Louisiana + Indiana “Southern corridor” (Madison/Ridgeland, Holly Ridge, New Carlisle): Anthropic-AWS and Meta Hyperion both chose this belt — Entergy is the biggest winner of this cycle
• International: Stargate UAE is the first truly non-U.S. sovereign flagship node; Alibaba Zhangbei is currently the only China site tracked by Epoch

4. Construction speed records keep being broken. Epoch’s buildout-speed analysis shows that a 1 GW-class data center typically takes 1.0–3.6 years from groundbreaking to structural completion; yet xAI Colossus in Memphis compressed Phase 1 to 122 days / 100K H100s, with Jensen Huang noting equivalent scale normally takes four years. Colossus 2 targets 1 GW in 12 months — approaching the physical lower bound of what is possible[34]. Meta Prometheus was announced in 2025 and came online in 2026 — an 18-month execution pace is unheard-of in large-scale infrastructure.

5. “A single data center consumes as much power as a mid-sized city” is now a fact. Microsoft Fairwater Wisconsin at full capacity (3.3 GW) exceeds Los Angeles’s peak load; Meta Hyperion at full capacity (5 GW) approaches Singapore’s peak demand; Anthropic-AWS New Carlisle (2.2 GW) equals the total power consumption of a medium-sized city. For one AI data center, Entergy is building 10 gas-fired plants in Louisiana, adding 7.5 GW of generation capacity — the first time a hyperscaler has effectively rewritten an entire state’s power planning.

6. On-site power generation and ERCOT grid bypasses have become standard equipment. xAI (Solaris JV gas turbines), Crusoe Goodnight (933 MW gas generation), Stargate Shackelford (Voltagrid 700 MW microgrid), Microsoft Goodyear (Sun Streams 2 solar PPA), Meta Sidecat (200 MW behind-the-meter gas) — these are no longer exceptions; they are the default configuration for any new gigawatt-class AI campus. I went into detail on the supply-side race in an earlier piece on data center power supply; this list is the “demand side” — they make for useful paired reading.

7. Financial engineering has entered the picture. Meta Hyperion’s $27 billion Blue Owl JV is a milestone in hyperscalers moving massive capex off-balance-sheet — 80/20 ownership structure, $2.5B equity plus $27B in A+-rated bonds, anchored by PIMCO and BlackRock. When capex grows large enough to consume all free cash flow, the SPV + private credit + long-term lease-back playbook will inevitably be replicated.

Lines Worth Digging Into Next

• How xAI Colossus compressed its build cycle to the four-month scale, and where its organizational model diverges from traditional Silicon Valley project management
• OpenAI Stargate’s capital structure: how equity and cash flows are divided among OpenAI, Oracle, SoftBank, G42, and MGX
• The AI data center landscape within China (Epoch currently covers mainly the U.S. plus Alibaba Zhangbei as the sole China site) — ByteDance, Tencent, Baidu, and Huawei’s respective flagship compute nodes
• The “bitcoin mine → AI training campus” asset conversion path, and the new intermediary layer represented by Galaxy/CoreWeave
• Whether Meta Hyperion’s SPV / Blue Owl JV structure is becoming the next standard template for hyperscaler financial engineering

References

• Mark Zuckerberg. Threads post on Prometheus + Hyperion, Jul 14 2025. threads.com/@zuck/post/DMF6uUgx9f9; TechCrunch coverage: techcrunch.com/2025/07/14/zuckerberg-5gw-data-center
• Meta. Richland Parish Data Center. datacenters.atmeta.com/richland-parish-data-center
• Epoch AI. Introducing the Frontier Data Centers Hub, Nov 2025. epoch.ai/blog/introducing-the-frontier-data-centers-hub
• Epoch AI. Frontier Data Centers — Documentation / Methodology. epoch.ai/data/data-centers-documentation/methodology
• Epoch AI. Frontier Data Centers Hub (dataset + CSV download, snapshot 2026-05-31). epoch.ai/data/data-centers
• Atlanta Journal-Constitution. Microsoft’s newest AI superfactory opens at sprawling Fayetteville campus, Nov 2025. ajc.com/business/2025/11/microsoft-ai-superfactory-fayetteville
• Microsoft. From Wisconsin to Atlanta: Microsoft Connects Datacenters to Build Its First AI Superfactory, Nov 12 2025; Infinite scale: the architecture behind the Azure AI Superfactory. news.microsoft.com/…/from-wisconsin-to-atlanta-fairwater-superfactory · blogs.microsoft.com/…/infinite-scale-azure-ai-superfactory
• CNBC. Meta signs nuclear energy deals to power Prometheus AI supercluster, Jan 9 2026. cnbc.com/2026/01/09/meta-nuclear-deals-prometheus
• Data Center Frontier / City of New Albany. Data Centers Are Booming in Ohio’s Digital Heartland. datacenterfrontier.com/…/data-centers-ohio-digital-heartland
• ServeTheHome. Inside the 100,000 NVIDIA GPU xAI Colossus Cluster Supermicro Helped Build for Elon Musk, Oct 2024. servethehome.com/inside-xai-colossus-cluster
• SemiAnalysis. xAI’s Colossus 2: First Gigawatt Datacenter, Dec 2025. newsletter.semianalysis.com/p/xais-colossus-2-first-gigawatt-datacenter
• Earthjustice. xAI Illegal Gas Power Plant for Data Center Colossus (NAACP, SELC, Earthjustice case), Apr 2026. earthjustice.org/case/xai-illegal-gas-power-plant-colossus
• Crusoe. Crusoe announces flagship Abilene data center is live, Sep 30 2025. crusoe.ai/resources/newsroom/abilene-data-center-is-live
• OpenAI / Oracle / SoftBank. Expanding Stargate with five new AI data center sites, Sep 23 2025. openai.com/index/five-new-stargate-sites
• News On 6. Google announces $9B Oklahoma investment (Pryor + Stillwater), Aug 13 2025. newson6.com/story/…/google-9-billion-oklahoma-investment
• DataCenter Dynamics. Google plans $1bn New Albany data center expansion, acquires 618 acres. datacenterdynamics.com/…/google-1bn-new-albany-expansion-618-acres
• Mississippi Today. Amazon $12B Mississippi data center expansion announcement, Apr 9 2026. mississippitoday.org/2026/04/09/data-centers-amazon-investments
• AWS / Amazon. Project Rainier — AI compute cluster for Anthropic, Oct 29 2025; Anthropic. Anthropic and Amazon expand collaboration for up to 5 gigawatts, Apr 20 2026. aboutamazon.com/news/aws/project-rainier · anthropic.com/news/anthropic-amazon-compute
• Baxtel. Google Omaha NE data center. baxtel.com/data-center/google-omaha-ne
• Blackridge Research. Microsoft Goodyear Data Center Arizona — Project Profile (PHX-70). blackridgeresearch.com/project-profiles/microsoft-goodyear-az
• Epoch AI Research. Twitter/X — Zhangbei campus satellite analysis. x.com/EpochAIResearch/status/1997013150072557759
• Datacenter.fyi. Meta Temple TX public record. datacenter.fyi/public-record/f7dc40d2
• Meta. Meta Announces Joint Venture with Funds Managed by Blue Owl Capital to Develop Hyperion Data Center, Oct 21 2025. about.fb.com/news/2025/10/meta-blue-owl-hyperion
• Epoch AI. The largest AI data center campuses will soon be a fifth the size of Manhattan, Nov 19 2025. epoch.ai/data-insights/data-center-sizes
• Fortune. Meta-Entergy: 10 gas power plants to power Hyperion in Louisiana, Mar 27 2026. fortune.com/2026/03/27/meta-hyperion-10-gas-power-plants-louisiana
• Microsoft. Made in Wisconsin: The world’s most powerful AI datacenter, Sep 18 2025. blogs.microsoft.com/on-the-issues/…/made-in-wisconsin-fairwater
• Epoch AI. Microsoft’s Fairwater datacenter will use more power than Los Angeles. epoch.ai/data-insights/fairwater-power-usage
• Amazon. $25 Billion Mississippi data centers announcement, Apr 9 2026. aboutamazon.com/news/…/amazon-25-billion-mississippi-data-centers
• G42. Global Tech Alliance Launches Stargate UAE, May 22 2025. g42.ai/resources/news/global-tech-alliance-launches-stargate-uae
• The National. G42 unit Khazna targets full build-out of 1 GW Stargate UAE project in three years, Oct 16 2025. thenationalnews.com/…/khazna-1gw-stargate-uae-three-years
• Latitude Media. How a Crusoe-Google campus became Texas co-location test case. latitudemedia.com/…/crusoe-google-campus-texas-co-location
• DataCenter Dynamics. CoreWeave leases another 260MW capacity from Galaxy in Texas (Helios). datacenterdynamics.com/…/coreweave-260mw-galaxy-texas
• Data Center Frontier. Cedar Rapids, Iowa gets approval for two major data center investments by Google and QTS. datacenterfrontier.com/…/cedar-rapids-google-qts
• Epoch AI. Build times for gigawatt-scale data centers, Nov 10 2025. epoch.ai/data-insights/data-centers-buildout-speeds