Who is building the biggest AI data center?
Whether you're building or investing in the AI data center market, our report covers everything you need to know: the players, the money, and what's coming next.
Summary
SB Energy, for OpenAI, is building one of the world's biggest planned single-campus AI data centers: the 8-gigawatt PORTS-Pike development in Ohio. But Fermi America has a larger power-generation ambition, while xAI is ahead in one of the clearest published comparisons of GPUs already deployed.
The distinction is important because the biggest announced data center and the biggest operating supercomputer are currently very different things. Gigawatts of proposed electricity generation, planned IT capacity and installed processors cannot be ranked as though they measure the same achievement.
OpenAI's Ohio development has an unusually strong commercial foundation. It combines a 20-year lease, NVIDIA's $1.5 billion investment in SB Energy, credit support for an initial 4.25 gigawatts and an option for another 3.75 gigawatts. The first 800 megawatts are expected in 2028, leaving most of the announced capacity dependent on later construction.
Fermi America's Project Matador has the largest headline figure at up to 17 gigawatts, but that describes potential power generation rather than installed AI computing. Its first binding customer lease covers 222 megawatts, revealing the enormous distance between the proposed energy campus and its contracted data center business.
Meta's Hyperion may be one of the most credible giant campuses already under substantial construction. Its 5-gigawatt target is supported by a development program exceeding $50 billion, an institutional financing partnership with Blue Owl Capital and significant construction spending in Louisiana.
xAI presents the opposite picture: less emphasis on a distant multigigawatt target and more evidence of rapid hardware deployment. Colossus 2 reportedly contained 550,000 NVIDIA GB200 and GB300 GPUs in September 2026, with additional large expansions proposed but not yet confirmed as operational.
Amazon's Project Rainier complicates any claim that xAI owns the world's biggest AI supercomputer. Anthropic reported using more than one million AWS Trainium2 processors, but differences in chip architecture and undisclosed single-site performance make a direct comparison unreliable.
Electricity is becoming as important as access to advanced chips. Fermi, Meta and OpenAI are all tied to massive power-generation or transmission projects, and a campus can secure financing and customers years before it receives enough electricity to operate at its advertised scale.
The financing structures reveal another shift. Meta is using a major infrastructure joint venture, Fermi is relying on long-term customer leases, and SB Energy will own and operate the Ohio facilities for OpenAI. Securing a creditworthy customer has become a major part of turning data center ambition into a financeable project.
Our conclusion is that OpenAI and SB Energy have one of the strongest claims to the biggest planned AI computing campus, Meta is among the most credible builders of a giant campus already under construction, Fermi leads on proposed power-generation scale, and xAI stands out for disclosed NVIDIA GPU deployment. The eventual winner will be the company that brings the most usable computing capacity online, not the one that announces the largest number.
Written by the New Market Pitch teamWe track new markets so founders and investors can move faster. About us
Why are companies suddenly building AI data centers with several gigawatts of power?
AI data centers have entered a completely different size category, with OpenAI, Meta and other major companies now planning campuses that would have seemed extraordinary just a few years ago.
In January 2025, OpenAI, SoftBank and Oracle introduced Stargate, targeting $500 billion of investment and 10 gigawatts of AI infrastructure. Meta subsequently expanded its Hyperion project in Louisiana to 5 gigawatts. Then, in August 2026, OpenAI announced an agreement for approximately 8 gigawatts of IT capacity at the PORTS-Pike campus in Ohio.
For perspective, a 1-gigawatt power requirement is comparable to the output of a large nuclear reactor. A 5-gigawatt data center campus would need electricity on the scale of several major power stations, even though its actual demand would vary as construction progresses and workloads change.
The biggest change is the amount of computing power companies want to concentrate in one location. Training advanced AI models requires enormous clusters of processors connected by high-speed networks. Running those models for millions of customers creates additional demand.
The International Energy Agency estimated that data centers consumed around 415 terawatt-hours of electricity worldwide in 2024, approximately 1.5% of global electricity consumption. Its April 2025 Energy and AI report projected consumption could reach around 945 terawatt-hours by 2030.
Those forecasts help explain why AI infrastructure projects now involve power plants, transmission networks and multibillion-dollar financing arrangements alongside the actual data center buildings.
What makes an AI data center the biggest in the world?
The world's biggest AI data center can mean the campus with the most planned electricity, the most computing capacity or the largest number of processors actually running. Those definitions currently produce different winners.
Consider three projects. Fermi America's Project Matador in Texas has discussed up to 17 gigawatts of integrated power-generation capacity. OpenAI's PORTS-Pike development targets approximately 8 gigawatts of IT capacity. Meanwhile, xAI's Colossus 2 has one of the world's largest publicly reported deployed GPU clusters.
A comparison based only on gigawatts would make Fermi look like the obvious leader. But power generation and IT capacity are different measurements. An on-site power plant can produce electricity for cooling, electrical systems and other infrastructure, while some of its generation capacity may remain unused until additional buildings and customers arrive.
GPU counts answer another question: how much AI hardware has actually been installed? Even then, 100,000 older NVIDIA H100 GPUs cannot be treated as equivalent to 100,000 newer GB300 GPUs.
For the main ranking, we focus on the largest planned AI computing campus. We also examine operational hardware separately, because a completed 500-megawatt supercomputer tells us more about today's available compute than an unbuilt 10-gigawatt project.
In our report covering the AI data center market, we always try to gather several sources and make estimates that make sense, so you can work with solid numbers.
Is Fermi America building the world's biggest AI data center?
Fermi America's Project Matador currently has one of the biggest announced AI infrastructure ambitions, reaching up to 17 gigawatts of power generation, but its first contracted data center covers just 222 megawatts.
The company is developing Project Matador near Amarillo, Texas, on land associated with the Texas Tech University System. Its long-term plans include approximately 11 gigawatts of private power generation, potentially expanding to 17 gigawatts with additional land and infrastructure.
That proposed ceiling is more than three times Meta Hyperion's 5-gigawatt capacity target. Yet comparing those figures directly would exaggerate Fermi's lead, because Hyperion's announcement concerns AI computing capacity while Fermi's larger figure concerns potential energy generation.
On August 10, 2026, Fermi announced a binding lease with TensorWave for a 222-megawatt facility designed to support tens of thousands of AMD Instinct GPUs. Initial delivery is expected in the second half of 2027. TensorWave also secured expansion rights that could take the partnership above 650 megawatts.
According to Fermi, approximately 6 gigawatts of its planned 17 gigawatts had received permits, and more than $1.5 billion had been invested in the campus by the time the TensorWave agreement was announced.
The contrast is striking. The first contracted facility represents roughly 1.3% of the proposed 17-gigawatt generation ceiling, although the figures measure different layers of the project. Even the potential TensorWave expansion would account for less than 4%.
Fermi is developing one of the most ambitious energy-integrated AI campuses ever proposed. Calling Project Matador the world's biggest functioning AI data center would be premature.
If you want more recent data on this point, please see our latest AI data center market report.
Is OpenAI building an even bigger AI data center in Ohio?
OpenAI's PORTS-Pike project in Ohio is currently one of the strongest candidates for the world's largest planned AI computing campus, with approximately 8 gigawatts of targeted IT capacity.
On August 17, 2026, OpenAI announced that it had reached an agreement with SB Energy, NVIDIA and the US Department of Energy for the PORTS-Pike Technology Campus near Piketon, Ohio.
SB Energy will build, own and operate the facility under a 20-year lease to OpenAI. NVIDIA will supply the AI computing infrastructure and has agreed to invest $1.5 billion in SB Energy, while providing credit support for an initial 4.25 gigawatts of IT capacity. The arrangement also includes an option covering another 3.75 gigawatts.
That adds up to an announced 8-gigawatt IT capacity target, including the optional expansion. It is 60% larger than Meta Hyperion's announced 5-gigawatt target.
However, the full campus will be developed over several years. OpenAI expects the first 800 megawatts to become available in 2028, with further phases requiring additional power plants, transmission lines, financing and permits.
OpenAI also says it will begin paying for capacity as completed infrastructure becomes available, rather than paying immediately for the entire planned project.
The Ohio development has a stronger commercial foundation than a speculative proposal with no major tenant. Its 8-gigawatt target is nevertheless a future computing ambition, and the first phase represents only one-tenth of that total.
Theory only gets you so far. In our research about AI data centers, we dig up the real signals from the field and show you the ones worth watching.
Is Meta's Hyperion the biggest AI data center already under construction?
Meta Hyperion is one of the most advanced giant AI campus developments in the United States, with substantial construction already underway and a planned 5 gigawatts of computing capacity.
Meta began constructing its Richland Parish data center in Louisiana in December 2024. In July 2026, the company announced that the project would expand to 5 gigawatts, making Hyperion the largest AI computing facility in Meta's own network.
The expanded development represents more than $50 billion of planned investment in the region. Meta reported that it had already contracted over $1.6 billion with Louisiana businesses, providing a concrete indication of how much local construction and infrastructure activity had begun.
The project also has a major financing partner. In October 2025, Meta established a joint venture with funds managed by Blue Owl Capital for the original development scope. Blue Owl-managed funds received an 80% ownership stake, while Meta retained 20%.
Compared with OpenAI's Ohio project, whose initial capacity is scheduled for 2028, Hyperion has a longer construction history. Compared with Fermi's Project Matador, Meta has already put a major financing structure and a large regional contracting program in place.
We still cannot treat its 5 gigawatts as operational capacity. But among the largest AI projects being physically developed, Hyperion has a particularly strong record of construction and financial commitments.
If you want more recent data on this point, please see our latest AI data center market report.
Which companies are planning the biggest AI data centers right now?
Fermi America, OpenAI's infrastructure partners and Meta dominate the largest announced US AI campus projects, although their published capacity figures measure different things.
The comparison below uses the capacity described by each developer, rather than converting every announcement into an estimated operating IT load.
| AI data center project | Company | Announced capacity | Current interpretation |
|---|---|---|---|
| Project Matador, Texas | Fermi America | Up to 17 GW | Potential power-generation capacity |
| PORTS-Pike, Ohio | SB Energy / OpenAI | ~8 GW IT | Planned AI computing capacity, including an expansion option |
| Hyperion, Louisiana | Meta | 5 GW | Planned AI compute capacity |
| Stargate, multiple US sites | OpenAI / partners | 10 GW target | Distributed infrastructure program |
| Colossus, Tennessee region | xAI | ~2 GW planned | Expanded campus power target |
| Cameron County, Texas | Fluidstack | Up to 1.5 GW | Proposed initial development |
| Vantage Frontier, Texas | Vantage Data Centers | 1.4 GW | Planned critical IT load |
| Stargate Abilene, Texas | Crusoe / Oracle / OpenAI | 1.2 GW | Planned campus power capacity |
In our market report about AI data centers, we don't want to repeat what boring consulting reports say. Instead, we show you the real and latest market signals that matter.
Who has the largest AI GPU cluster actually running today?
xAI's Colossus 2 has the largest deployed GPU cluster in the State of AI Report's selected inventory, with 550,000 NVIDIA GB200 and GB300 GPUs reported in September 2026.
According to the State of AI Report Compute Index, which used an October 1 cutoff, Colossus 2's reported configuration included 110,000 GB200 GPUs and 440,000 GB300 GPUs.
The inventory relies on figures reported by the operator, rather than an independent hardware audit. It also excludes some private computing installations that have never disclosed their chip counts.
Even with those limitations, Colossus 2 represents an extraordinary deployment. xAI began building the original Colossus infrastructure in Memphis during 2024. Within roughly two years, the company had progressed from its first major GPU installation to a reported cluster containing more than half a million Blackwell-generation processors.
Elon Musk also described plans for three additional batches of 220,000 GB300 GPUs. If all three arrive and become operational, Colossus 2 would reach approximately 1.21 million GPUs.
That would more than double its September configuration. However, we should keep the additional 660,000 GPUs separate from the reported deployed count until their installation is confirmed.
For readers asking which company has already assembled the most impressive publicly disclosed NVIDIA GPU cluster, xAI has a stronger claim than the companies whose largest campuses are still years from completion.
If you want more recent data on this point, please see our latest AI data center market report.
Is Amazon's AI infrastructure bigger than xAI's Colossus?
Amazon has deployed AI infrastructure at a comparable industrial scale, but its Trainium processors make a direct chip-count comparison with xAI's NVIDIA clusters misleading.
Amazon's Project Rainier was developed to support Anthropic's AI workloads using AWS Trainium processors. In April 2026, Anthropic reported using more than one million Trainium2 chips through its Amazon infrastructure relationship.
Numerically, that exceeds the 550,000 GPUs reported for Colossus 2. But Trainium2 and NVIDIA's GB200 and GB300 systems differ in architecture, performance and how they are assembled into computing clusters.
Amazon and Anthropic have also announced plans for up to 5 gigawatts of new AI capacity, with nearly 1 gigawatt of Trainium2 and Trainium3 capacity expected by the end of 2026.
Amazon's advantage is its ability to build custom processors and operate them inside AWS's existing infrastructure. Rather than depending entirely on NVIDIA's GPU roadmap, Amazon can design chips around its own cloud systems and Anthropic's computing requirements.
What we cannot establish from public disclosures is whether Project Rainier has surpassed Colossus 2 in usable AI training performance at a single site.
Both companies have built serious computing infrastructure. The reported processor totals alone do not settle which one operates the larger supercomputer.
In our document detailing everything you need to know about AI data center, we like to show you the big picture. We know your time is precious, so we make every page quick to read and packed with as much learning as possible per minute.
Is OpenAI's Stargate bigger than Meta Hyperion?
OpenAI's Stargate is larger than Meta Hyperion as a broader AI infrastructure program, but Stargate's total capacity is spread across multiple sites.
The original Stargate announcement in January 2025 outlined a $500 billion investment ambition and a 10-gigawatt infrastructure target. By September 2025, OpenAI and its partners had identified nearly 7 gigawatts of planned capacity across Abilene and other US locations.
Meta Hyperion, by comparison, is a concentrated campus in Richland Parish with a planned 5-gigawatt computing capacity.
Stargate's scale comes from assembling several enormous facilities into one computing network. The Abilene campus in Texas is being developed with Crusoe and Oracle, while other Stargate-related developments involve additional infrastructure partners.
The newer Ohio agreement makes this comparison more complicated because the PORTS-Pike project alone targets approximately 8 gigawatts of IT capacity. We must avoid adding overlapping Stargate announcements together when determining the total.
In September 2025, OpenAI said its announced infrastructure pipeline was approaching 7 gigawatts, before the later Ohio agreement. Those disclosures cover commitments made at different times and under different project arrangements.
OpenAI may build a larger overall AI infrastructure network, while Meta is developing one of the largest individual campuses. For the question of who is building the biggest single data center, PORTS-Pike is the more relevant OpenAI comparison.
How much of OpenAI's Stargate data center in Texas is already running?
Stargate Abilene is already operating AI computing infrastructure, although most of its planned 1.2-gigawatt capacity still depends on later construction phases.
The project in Abilene, Texas, is being developed by Crusoe using infrastructure associated with Lancium, with Oracle supplying computing capacity for OpenAI workloads.
The full campus is designed around eight large data center buildings. According to research published by Epoch AI in April 2026, approximately 300 megawatts of facility power had become operational, based on the organization's estimates at that time.
That would represent roughly one-quarter of the planned 1.2-gigawatt campus. Epoch AI subsequently updated its estimates as additional Oracle disclosures and construction information became available.
Abilene gives us something that many giant AI campus proposals do not yet offer: an established operating footprint. Its published construction plans can be compared against buildings and computing infrastructure already in use.
The campus also illustrates how long the last stretch of development can take. Moving from hundreds of megawatts to more than a gigawatt requires additional buildings, electrical equipment, cooling systems and GPU installations.
Abilene's operational progress is impressive, even though its eventual 1.2-gigawatt capacity would be smaller than several newer megaprojects.
In our report covering the AI data center market, we turn what is vague and complex into simple, useful visuals. Market maps like this one show you at a glance where every startup stands.
Could Microsoft's AI data centers be bigger than everyone else's?
Microsoft is already among the world's largest AI infrastructure operators, but it has not disclosed enough comparable information to establish Fairwater as the biggest single AI data center.
In September 2025, Microsoft introduced Fairwater in Wisconsin as its largest and most sophisticated AI data center project to date.
The facility was designed around hundreds of thousands of advanced AI chips, high-performance networking and specialized cooling infrastructure. Microsoft also announced additional Fairwater developments and highlighted its broader cloud network of more than 400 data centers across 70 regions.
Microsoft's approach allows AI workloads to run across many facilities rather than depending on a single enormous campus. That can help with geographic coverage, customer demand and access to computing resources.
However, a network of hundreds of data centers cannot be compared directly with one 5-gigawatt campus. Neither does a total processor count establish the electrical capacity or computing performance of one specific location.
We therefore consider Microsoft a major competitor in overall AI infrastructure, while leaving Fairwater out of any definitive single-campus capacity ranking until more comparable data becomes available.
Is electricity becoming the biggest problem for AI data centers?
Electricity is currently one of the biggest barriers to building giant AI data centers, particularly when companies need several gigawatts at one location.
The clearest evidence comes from how developers are changing their plans. Instead of simply requesting a large connection from the local utility, several projects now include dedicated electricity generation and transmission infrastructure.
Fermi's Project Matador plans to combine natural gas generation with potential nuclear power and other energy resources. Its development depends heavily on generation equipment, fuel supply, permits and transmission.
Meta faces similar challenges in Louisiana. In April 2026, the Union of Concerned Scientists reported that Entergy Louisiana was seeking approval for seven additional gas-fired power plants alongside three previously approved facilities. The ten plants could provide approximately 7.5 gigawatts of generation capacity.
Ohio's PORTS-Pike development also involves extensive new electricity infrastructure. Its plans include at least 9.2 gigawatts of new natural-gas generation and major transmission upgrades.
Power development is becoming a decisive competitive advantage. A company with access to GPUs and construction funding can still face years of delay if the electricity supply cannot support its plans.
There is also a practical difference between securing the right to build a power plant and delivering electricity to thousands of running servers. Permits, fuel contracts, equipment delivery and grid upgrades can all affect when capacity becomes available.
For the largest AI campuses, securing reliable electricity increasingly determines how fast construction can turn into actual computing power.
Our research about AI data centers gets updated several times a month, so you always get the latest view of where things stand.
Who is paying for the world's biggest AI data centers?
The biggest AI data centers increasingly depend on infrastructure investors, long-term leases and project financing, because the cost of constructing an entire multigigawatt campus can exceed tens of billions of dollars.
Meta's Hyperion provides a good example. Its October 2025 joint venture with Blue Owl Capital covered approximately $27 billion in expected development costs for the original financed scope. Blue Owl-managed funds received 80% ownership, while Meta retained 20%.
That arrangement allowed Meta to work with an infrastructure investment partner on a major physical asset while continuing to use the data center for its own AI operations.
The company's later announcement of more than $50 billion in Richland Parish investment covers a broader project scope. The figures overlap and should not be added together.
Fermi has taken a different approach, signing long-term leases with infrastructure customers. Its initial TensorWave agreement is expected to generate approximately $6.5 billion in contracted revenue over 15 years.
That works out to about $433 million in average annual contractual revenue, assuming the full amount is earned over the initial term. It does not mean Fermi has already received $6.5 billion or that all the projected revenue is guaranteed regardless of future events.
OpenAI's Ohio campus follows another structure. SB Energy will own and operate the facilities, OpenAI will lease the computing capacity, and NVIDIA is providing investment and credit support.
Across these projects, developers are using long-term customer commitments to support construction financing. The ability to secure those commitments increasingly separates serious projects from ambitious announcements.
Could Fermi's 17-gigawatt AI data center actually reach that size?
Fermi's Project Matador could become an extraordinary AI infrastructure campus, but reaching its full 17-gigawatt power-generation target would require a massive expansion beyond the commitments disclosed so far.
The company has made tangible progress. Its initial land arrangement covers thousands of acres, approximately 6 gigawatts of planned infrastructure had reportedly received permits by August 2026, and TensorWave has signed the first binding customer lease.
Fermi has also discussed nuclear power as part of its longer-term energy strategy, including Westinghouse AP1000 reactors. Nuclear generation could eventually support a campus with enormous electricity requirements, although licensing, financing and construction would add years of development work.
The challenge is the size of the remaining gap. The 222-megawatt initial TensorWave facility is backed by an identified customer, while most of the project's proposed energy-generation capacity still needs further construction and commercial development.
The economics of the first phase also show the scale involved. Its expected $6.5 billion in revenue over 15 years applies to a facility representing only a small part of Fermi's ultimate ambition.
Even reaching the broader potential TensorWave arrangement above 650 megawatts would leave the project far short of 17 gigawatts of generating capacity.
Fermi has established more than a conceptual proposal, but the full campus remains a long-term development scenario. We would need to see multiple additional binding leases, completed power infrastructure and operating data halls before treating its maximum proposed capacity as likely to be delivered.
If you want more recent data on this point, please see our latest AI data center market report.
In our market report about AI data centers, you will always find fresh funding data, taken straight from our own verified funding database.
Which company is expanding its AI computing capacity the fastest?
xAI has demonstrated one of the fastest reported expansions of deployed AI processors, while OpenAI and Meta are pursuing larger infrastructure programs that will take years to complete.
The most striking example is the reported growth of Colossus. xAI began building its original Memphis supercomputer in 2024 and had reported 550,000 Blackwell-generation GPUs in Colossus 2 by September 2026.
Elon Musk subsequently outlined three additional batches of 220,000 GB300 GPUs. Each batch would add computing hardware equal to 40% of the previously reported 550,000-GPU configuration.
If all three were delivered, the installed GPU count would grow by 120%. The final batch was explicitly presented as conditional, so the 1.21-million-GPU target should not be confused with verified deployment.
The pace looks extraordinary beside the longer construction schedules at other major campuses. Hyperion started construction in 2024 and is still working toward its multigigawatt target. OpenAI's Ohio project expects the first 800 megawatts in 2028, with later phases extending through the following years.
The comparison has limits because companies report different measurements. xAI discloses chip counts, Meta emphasizes computing capacity, and other developers focus on electricity and construction milestones.
Still, among companies publishing detailed figures for large GPU clusters, xAI has shown unusually rapid hardware deployment. Whether that makes it the fastest-growing AI computing operator across all architectures cannot be established from public data alone.
Which giant AI data centers are most likely to become fully operational?
Meta Hyperion and OpenAI's PORTS-Pike have particularly strong evidence behind their long-term development plans, while xAI and Stargate Abilene already have large operating computing installations.
We compared the publicly reported construction progress, financing arrangements, customer agreements and operational milestones of the leading projects. The evidence is stronger for some than others, and no ranking can guarantee that an unfinished campus will reach its proposed capacity.
| AI data center | Evidence already available | Biggest remaining challenge |
|---|---|---|
| xAI Colossus 2 | Reported deployment of 550,000 advanced GPUs | Verifying and completing further expansion |
| Stargate Abilene | Operational computing infrastructure and ongoing construction | Reaching its planned 1.2 GW |
| Meta Hyperion | Major construction, institutional financing and local contracts | Completing the 5 GW expansion |
| Amazon Project Rainier | Large deployed Trainium infrastructure | Comparable single-site capacity disclosure |
| Microsoft Fairwater | Specialized AI facility and hardware development | Comparable IT capacity disclosure |
| PORTS-Pike | OpenAI lease, NVIDIA backing and development agreements | Delivering power and the first computing phases |
| Project Matador | Tenant lease, permits and infrastructure investment | Contracting and building the remaining capacity |
| Fluidstack Cameron County | Announced development and initial construction plans | Delivering the proposed 1.5 GW |
In our document detailing everything you need to know about AI data center, we look at what the winners do differently, so you can learn from the best instead of guessing.
Who is really building the biggest AI data center in the world?
SoftBank's SB Energy is building one of the largest planned AI computing campuses for OpenAI, with approximately 8 gigawatts of targeted IT capacity in Ohio. But the answer changes when we look at power-generation ambitions or the amount of hardware already running.
For planned AI computing capacity at a single identified development, PORTS-Pike has the strongest headline among the major projects reviewed. Its 20-year customer agreement with OpenAI, NVIDIA's involvement and 8-gigawatt target make it a particularly serious contender. The full campus will still need years of construction before that capacity exists, and 3.75 gigawatts of the total is covered by an expansion option.
For proposed power-generation capacity, Fermi America's 17-gigawatt Project Matador is larger. Its initial contracted facility is 222 megawatts, however, and its long-term generation target cannot be treated as equivalent to installed AI compute.
Meta Hyperion looks particularly credible among the giant campuses already under substantial construction. It has a 5-gigawatt target, major financing partners and a documented record of physical development.
Meanwhile, xAI's Colossus 2 remains the standout in publicly reported deployed NVIDIA GPU hardware. Its 550,000-GPU figure gives us a much more concrete measure of operating computing infrastructure than the gigawatt targets attached to unfinished projects.
Our final judgment is that OpenAI and SB Energy have one of the strongest claims to the world's biggest planned AI computing campus, while xAI leads the selected disclosed GPU-cluster ranking. Fermi has the largest proposed energy ambition, and Meta is among the most credible builders of a giant campus already well into construction.
The winner will ultimately be determined by how much computing power these companies actually deliver, rather than how many gigawatts they announce.
If you want more recent data on this point, please see our latest AI data center market report.
Our methodology
This analysis examines who is building the world's biggest AI data center using publicly available information through October 2026. We compare announced computing capacity, proposed power-generation infrastructure, deployed processors, construction milestones, financing arrangements and customer commitments. Our goal is to distinguish projects already delivering computing power from much larger campuses that remain under development.
We treat IT capacity, facility power, generation capacity and processor counts as separate measurements. A proposed 17-gigawatt power campus cannot be directly compared with an 8-gigawatt IT development or a running cluster of 550,000 GPUs. We also distinguish binding commercial arrangements from expansion options and long-term targets. Where a developer publishes a maximum planned capacity, that figure describes the proposed project rather than the amount already operational.
For construction and commercial progress, we consider physical development, permitting, infrastructure investment, customer leases, financing support and announced delivery schedules. These factors help assess how credible an expansion may be, but they do not guarantee completion. The analysis gives greater weight to operating infrastructure and documented commercial commitments than to announcements without comparable evidence of execution.
For operational computing comparisons, we use the State of AI Report's reported GPU inventory, AWS and Anthropic's processor disclosures, and independent analysis from Epoch AI. We do not convert NVIDIA GPUs and AWS Trainium chips into a single estimated performance measure because their architectures and workloads differ. Reported chip counts are treated as disclosed configurations, not independently audited inventories, and additional proposed deployments are excluded from current operating totals.
Key sources for project announcements and commercial arrangements include OpenAI's PORTS-Pike announcement, NVIDIA's PORTS-Pike agreement, SB Energy's project disclosure, Fermi America's TensorWave lease announcement, Meta's Louisiana expansion announcement, Meta's Blue Owl Capital joint venture disclosure, and OpenAI's original Stargate announcement.
For operating infrastructure and competing developments, we use the State of AI Report Compute Index, Anthropic's Amazon compute announcement, AWS's Project Rainier update, Crusoe's Abilene expansion disclosure, Epoch AI's Stargate site analysis, Microsoft's Fairwater announcement, Vantage Data Centers' Frontier announcement, and Fluidstack's Cameron County announcement.
For electricity demand and power infrastructure, we reference the International Energy Agency's Energy and AI electricity demand analysis and energy supply assessment, alongside the Union of Concerned Scientists' analysis of proposed Louisiana gas generation. These sources provide context for why power availability, generation development and transmission construction increasingly shape the delivery schedules of large AI campuses.
If you build or invest in the AI data center market, you need to know which business models pay off now and which ones are fading. That's why our report covering it has a whole section on it.
Written by the New Market Pitch teamWe track new markets so founders and investors can move faster. About us
