Meta Compute
Last reviewed
Jun 3, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 1,705 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 1,705 words
Add missing citations, update stale details, or suggest a clearer explanation.
Meta Compute is a top-level organization that Meta created in January 2026 to plan, build, and run the gigawatt-scale data center capacity behind its push toward artificial general intelligence and what the company calls "superintelligence."[1][2] Chief executive Mark Zuckerberg announced it on January 12, 2026, in a post on Threads, writing that "Meta is planning to build tens of gigawatts this decade, and hundreds of gigawatts or more over time" and that "how we engineer, invest, and partner to build this infrastructure will become a strategic advantage."[1][3] The new group pulls Meta's hardware, networking, silicon, and site-development work under one umbrella, and it folds in long-range capacity planning and energy procurement that had previously been spread across teams.[2][4]
The move was, in large part, an organizational acknowledgement of something that had already happened. By early 2026 Meta was spending more on compute than almost any company in the world, and it had announced a string of single-site clusters drawing more than a gigawatt each, including Prometheus in Ohio and Hyperion in Louisiana. Meta Compute formalized that buildout as a standing function rather than a series of one-off projects.[2][5]
The clusters that train and serve large AI models are now Meta's single largest cost. The company reported roughly $72 billion in capital expenditures for 2025, the bulk of it data center construction and the chips that go inside, and it has signaled spending on the order of $600 billion in the United States through 2028.[2][4] Numbers that size are hard to run as side projects inside an engineering org built for apps and ads, so Zuckerberg elevated infrastructure to its own initiative reporting near the top of the company.[2][5]
The framing is competitive. Meta's view, stated repeatedly through 2025, is that the lab with the most compute per researcher will move fastest on frontier models, and that securing electrical capacity years in advance is now part of the race rather than a back-office concern.[5][6] Meta Compute exists to lock in that capacity: land, power, chips, and the long-lead-time deals with utilities and suppliers that a gigawatt-scale site needs before a single server is racked.[2][4]
Meta Compute is co-led by two people with very different backgrounds.[1][2] Santosh Janardhan, Meta's head of global infrastructure and a co-head of engineering, runs the technical side: system architecture, the company's in-house silicon effort, the software and tooling stack, and day-to-day data center operations. Janardhan has been at Meta since 2009 and had already been overseeing much of this work before the reorganization.[2][7]
The other lead is Daniel Gross, a more recent and more closely watched hire. Gross had co-founded and run Safe Superintelligence, the lab started by former OpenAI chief scientist Ilya Sutskever; he joined Meta in the summer of 2025 after Meta's attempt to acquire SSI fell through.[2][8] At Meta Compute he heads a new group focused on the longer horizon: forecasting how much computing the company will need before it needs it, managing strategic supplier relationships, and modeling the business case for spending at this scale.[2][4]
A third name comes up alongside them. Dina Powell McCormick, whom Meta named president and vice chair on the same day in January 2026, works with both leads on government and sovereign partnerships that touch infrastructure financing and where new sites get built.[2][9] Meta Compute sits next to, rather than inside, Meta Superintelligence Labs, the research organization led by Alexandr Wang that designs the models the clusters are meant to train.[5][6] In effect, the labs decide what to build and Meta Compute builds the machine it runs on.
Meta Compute oversees a set of campuses that Meta has been announcing since 2024. The flagship multi-gigawatt sites carry internal cluster names borrowed from Greek mythology, while several others are identified by location.
| Site / cluster | Location | Reported capacity | Status |
|---|---|---|---|
| Prometheus | New Albany, Ohio | ~1 GW | Coming online in 2026[5][10] |
| Hyperion (Richland Parish) | Richland Parish, Louisiana | Scaling toward ~5 GW | Under construction, 2025 to 2030[6][11] |
| Lebanon campus | Lebanon, Indiana | ~1 GW | Broke ground February 2026[12][13] |
Prometheus is a roughly one-gigawatt cluster at Meta's New Albany campus northeast of Columbus, Ohio, and is expected to be among the first single-site facilities anywhere to draw more than a gigawatt when it comes online in 2026.[5][10] Hyperion, in Richland Parish in northeast Louisiana, is the larger of the two. Meta plans about 4 million square feet of buildings on a tract of roughly 2,250 acres and has said the site will scale toward about 5 gigawatts of capacity over several years, with construction running from 2025 through 2030.[6][11] In October 2025 Meta set up a roughly $27 billion joint venture with Blue Owl Capital to finance the Hyperion development, taking part of the cost off its own balance sheet.[14]
The newest of the three broke ground in February 2026. On February 11, Meta started construction on a 1-gigawatt campus in Lebanon, Indiana, inside the state's LEAP innovation district, an investment it put at more than $10 billion.[12][13] The Indiana site uses a closed-loop liquid cooling system that recirculates the same water, and it sits alongside Meta's existing Indiana facility as part of the company's growing U.S. footprint.[13] Meta has been candid that it is moving faster than conventional construction allows; through 2025 it described putting servers under temporary tent-like structures to bring capacity online while permanent buildings were still going up.[6]
The hard constraint on all of this is electricity, and it is where Meta Compute's planning role is most visible. A single 5-gigawatt site like Hyperion would draw more power than many American cities, and the grid in rural Louisiana was not built for that kind of load.[6][15]
Meta's answer has leaned heavily on new natural gas generation, mostly through deals with the utility Entergy. Three gas-fired plants tied to Hyperion, totaling roughly 2.3 gigawatts, won approval from state regulators in 2025.[15][16] Then on March 27, 2026, Meta and Entergy announced a much larger commitment: seven more gas plants adding about 5.2 gigawatts, which would bring the campus to ten plants and roughly 7.5 gigawatts of dedicated gas capacity.[15][16] Meta agreed to pay for the construction and financing of all ten plants, at an estimated cost of nearly $11 billion, and to help fund up to 2.5 gigawatts of renewable energy and battery storage on top of that.[15][16] The seven new plants still need their own sign-off from the Louisiana Public Service Commission before construction can begin.[16]
That much new fossil generation for one customer has drawn scrutiny. The plants would represent a large share of new capacity on the regional grid, and critics in Louisiana have questioned the climate impact and who ultimately bears the cost and risk if the data center's demand changes.[15] Meta, for its part, has paired the gas deals with renewable and transmission commitments and argues that it is funding the new supply rather than crowding out existing ratepayers.[16] I find this the genuinely unresolved part of the story: the compute math is straightforward, but the question of whether you should build out gigawatts of gas to chase frontier AI is not, and it is being decided one utility filing at a time.
Meta Compute is the infrastructure half of a two-part bet the company made through 2025 and 2026. The research half is Meta Superintelligence Labs, assembled after an expensive talent push that brought in Wang, Gross, and others; the infrastructure half is the organization that has to keep those researchers supplied with compute.[5][6] The clusters Meta Compute runs are also where the company's Llama family of models and its in-house MTIA accelerators get exercised at scale, tying the org back into Meta's existing AI hardware and model work rather than standing apart from it.[5][7]
Whether the spending pays off is an open question that even Meta's own statements treat cautiously. The company has acknowledged it could be overbuilding relative to near-term demand, while arguing that the bigger risk is having too little compute if a frontier breakthrough arrives.[2][5] Meta Compute is the structure built to make that gamble, at the scale of gigawatts and hundreds of billions of dollars, with a single set of people accountable for it.