The most important number in AI right now isn't a benchmark score or a parameter count. It's a gigawatt.

Frontier AI has quietly become an energy business. Training the next generation of models needs power at a scale that used to belong to cities and aluminum smelters — and every major lab is now racing to secure it. The announcements read like utility-company filings: multi-gigawatt campuses, gas-turbine fleets, transmission deals, nuclear partnerships. That's not a side quest. Power is the binding constraint on who gets to build the next model, and the labs know it.

Here's who's building what, where the electrons come from, and why energy — not GPUs — is now the bottleneck.

The scale of the problem#

Put the gigawatt in perspective: 1 GW of computing capacity costs roughly $50 billion to build and draws about as much power as 750,000 U.S. homes. And the totals keep climbing. Lawrence Berkeley National Laboratory estimated in June 2026 that U.S. data centers could use 9.5 to 15 percent of the country's electricity by the end of the decade, up from about 4.5 percent today — and could make up more than 40 percent of all demand growth over the next five years. The lab's report put it bluntly: successive hardware generations get more efficient, but the growth of compute demand "more than offsets these efficiency gains."

Each new flagship training run now draws power in the hundreds of megawatts, with single training clusters pushing toward a gigawatt. That's not an anomaly — it's the plan, for every lab.

Meta: Prometheus and Hyperion#

Meta has the most cinematic plan. In July 2025, CEO Mark Zuckerberg announced the company would invest "hundreds of billions of dollars" in compute, revealing two named superclusters for its Superintelligence Labs effort.

The first, Prometheus, is a roughly 1-gigawatt cluster coming online in 2026, reportedly a network of facilities around New Albany, Ohio. Zuckerberg cited a SemiAnalysis report suggesting Meta was on track to be the first lab to bring a 1GW+ supercluster online — a milestone everyone else was chasing at the same time.

The second, Hyperion, is the big one: a Richland Parish, Louisiana facility designed to scale to 5 gigawatts over several years, building on Meta's earlier $10 billion Louisiana data center project. Zuckerberg described these "titan clusters" as each covering "a significant part of the footprint of Manhattan." He also said Meta is building "multiple more titan clusters" beyond these two — framing compute-per-researcher as the lab's recruiting edge.

Where the power comes from: Louisiana has natural gas and cheap land; Meta has leaned on utility Entergy in the region and, like everyone, is pairing grid power with on-site generation.

OpenAI and Stargate: the $500 billion blueprint#

OpenAI's infrastructure play is Stargate, the joint venture with SoftBank, Oracle, and MGX announced with an initial $100 billion commitment and plans to reach $500 billion over four years, including roughly 20 hyperscale data centers across the U.S.

The flagship site is in Abilene, Texas, where Oracle's campus is operating at around 1.2 GW. Stargate sites have since been announced in New Mexico, Wisconsin, Ohio, and — in October 2025 — a 1 GW+ campus in Saline Township, Michigan, developed with Related Digital. With that addition, the planned Stargate footprint topped 8 GW of total capacity and more than $450 billion in investment.

Not everything has gone smoothly. The planned expansion of the Abilene site from 1.2 GW to 2.0 GW was reportedly scrapped in early 2026 after financing negotiations between Oracle and OpenAI stalled — a reminder that even with committed partners, multi-gigawatt builds carry real counterparty risk. Meta has since been reported to be in talks to take up some of the abandoned capacity.

Where the power comes from: Texas wind and gas, local utility partnerships, and developer-funded grid upgrades — in Michigan, the developers agreed to fund all grid upgrades rather than pass costs to taxpayers, a concession that's becoming standard for getting these projects approved.

Stargate has also gone international. Stargate UAE, unveiled with G42, Oracle, NVIDIA, SoftBank, and Cisco, plans a 1-gigawatt AI cluster in Abu Dhabi inside a 5-gigawatt UAE–U.S. AI Campus, with the first 200 MW phase expected in 2026. Power there will draw from nuclear, solar, and natural gas.

xAI: turbines first, permits later#

Elon Musk's xAI took the most aggressive route to power: don't wait for the grid — build your own power plant.

The original Colossus supercomputer in South Memphis was famously built in 122 days with 100,000 NVIDIA H100 GPUs, later doubled to more than 200,000. But the grid couldn't keep up, so xAI turned to on-site natural-gas turbines. Colossus 2, in Southaven, Mississippi just across the state line, scaled that approach up dramatically: Mississippi regulators granted an air permit for 41 natural-gas turbines capable of generating about 1.2 gigawatts for the site.

The approach has been controversial. The NAACP filed a Clean Air Act lawsuit in April 2026 alleging xAI operated dozens of turbines without federal permits in a historically Black neighborhood in South Memphis — the first major federal test of how the AI industry's compute appetite collides with 1970s-era clean-air permitting. SemiAnalysis has projected Colossus 2 scaling toward 1.1 GW of total capacity, with a joint venture between xAI and Solaris Energy Infrastructure funding the turbine buildout — some $112 million in CapEx in Q2 2025 alone.

Where the power comes from: Mobile gas turbines now, the TVA grid eventually. xAI's playbook is speed over permission — paying for its own substations and generating its own baseload rather than waiting in the grid interconnection queue.

Google and Anthropic: the TPU axis#

Google plays the infrastructure race differently: it owns the silicon. Its seventh-generation Ironwood TPU is the first designed explicitly for the inference era — 4.6 petaFLOPS of FP8 compute per chip, 192 GB of HBM3e, and clusters of 9,216 chips delivering 42.5 exaflops.

The most consequential commercial deal of the year pairs Google's hardware with Anthropic's demand: access to up to one million TPUs, bringing over a gigawatt of AI capacity online starting in 2026 under a multiyear deal reportedly worth tens of billions of dollars. Anthropic deliberately keeps a multi-vendor posture — mixing TPUs, AWS's Trainium, and NVIDIA GPUs — but the TPU deal gives it price leverage and reduces its dependence on scarce NVIDIA allocation.

Then came the escalation: in April 2026, Anthropic signed a second deal with Google and Broadcom for up to 5 gigawatts of next-generation TPU capacity coming online from 2027, alongside reports of Google investing up to $40 billion in Anthropic at a $350 billion valuation. Amazon, not to be left out, has reportedly committed up to $25 billion of its own — with Project Rainier scaling toward a million-plus Trainium chips and around 5 GW of Trainium capacity in view. Anthropic now sits atop roughly 10 gigawatts of reserved compute across hyperscalers: the clearest sign that frontier labs are procuring capacity like utilities, not software companies.

Why energy is the binding constraint#

GPUs are still scarce and expensive, but three things make power the harder ceiling:

  1. Grid interconnection is slow. Getting a new gigawatt-scale load connected to the U.S. grid can take years of studies, permits, and substation construction. That's why labs are buying turbines (xAI), building their own substations, or funding utility upgrades directly (Stargate Michigan).
  1. The money is staggering. At roughly $50 billion per gigawatt, a 5-gigawatt cluster is a quarter-trillion-dollar build. This is why Oracle's balance sheet, not its technology, became the constraint at Abilene — and why Meta's "we have the capital from our business" is a genuine competitive moat.
  1. The politics are getting real. FERC has opened a probe into whether grid upgrade costs are being fairly allocated. Community opposition is rising around noise, water use, and electricity rates. Electricity prices are climbing in several regions, and state regulators — from Michigan to Mississippi — are now central players in the AI race. Speed now depends as much on regulators and turbine vendors as on chip designers.

Takeaway#

The frontier labs spent 2023 and 2024 racing for GPU allocation. In 2026, they're racing for megawatts — and the buildout has become the strategy: Meta's titan clusters, OpenAI's Stargate, xAI's turbine farms, Google's TPU empire. The lab that wins the next model cycle will be the one that locked down power years in advance.

Watch the next bottleneck to form after energy: transformers, cooling, and the gas-turbine order books. The gigawatt race is just the opening round.