Summary: Mark Zuckerberg announced Meta Compute on January 12, 2026 as a top-level initiative to build AI infrastructure measured in gigawatts. The plan targets tens of gigawatts this decade and hundreds or more over time, with infrastructure, capacity strategy, and government relations split among three executives. The article argues power and land, not model architecture, now pace the frontier compute race.
“Today we’re establishing a new top-level initiative called Meta Compute.” With that one line in a Threads post on January 12, 2026, Mark Zuckerberg committed Meta to a new kind of organization. The move is not one more round of data center purchasing — it lifts AI infrastructure into a company-wide undertaking that sits on par with the model business. Reuters put it bluntly in its headline: the Meta Compute effort exists to build computing capacity measured in gigawatts.
Zuckerberg supplied the numbers himself: what Meta plans to put in place is “tens of gigawatts this decade, and hundreds of gigawatts or more over time.” For scale, a big nuclear reactor unit outputs about 1 GW, which makes the pledge the equivalent of dozens of power plants’ worth of compute for AI alone. After the string of compute announcements that opened 2026, nothing says it louder: the agenda is now set by power and land rather than by model architecture.
Who Runs Meta Compute: Three Names, One Mandate
TechCrunch reports that three senior executives share command of Meta Compute, and that their assignments line up one-to-one with infrastructure’s three hardest problems.
Santosh Janardhan, running global infrastructure, has been with Meta since 2009; his brief spans silicon, the software stack, developer productivity, technical architecture, and the worldwide datacenter fleet and network — he operates everything Meta already has in production. Daniel Gross came on board in 2025, after co-founding Safe Superintelligence alongside Ilya Sutskever; he carries planning, supplier partnerships, business modeling, and long-term capacity strategy — the desk that decides what gets bought, when, and on what terms. Dina Powell McCormick, newly installed as president and vice chairman, takes government relations and works to finance the facilities and get them built — at this scale, getting data centers built is in the end a matter of politics and permits.
The most revealing slot in the arrangement is capacity strategy going to an SSI co-founder: Meta is not shopping for additional procurement managers — it wants someone who can read the demand curve of frontier models and plan compute years out.
Electricity, Not GPUs, Sets the Pace
The yardstick of this race has changed: where labs once counted GPUs, they now count gigawatts. Axios relays Zuckerberg’s goals of tens of gigawatts inside this decade and hundreds or more in the long run, while TechCrunch cites projections under which US AI electricity demand climbs from 5 GW to 50 GW by 2030. Even the nearer “tens of gigawatts” target, on those numbers, would put one company’s construction pipeline on the same order of magnitude as the AI electricity use projected for the whole country.
Rivals are not standing still either: Microsoft continues to scale up via infrastructure partnerships; Alphabet, for its part, bought data center operator Intersect in December 2025. Every frontier lab is competing for the same grid capacity, the same transformers, and the same land — whoever locks up power first buys their way into the next model generation.
What the Money Was Already Saying
The direction of capital was on record long before the announcement. Meta’s capex guidance for 2025 peaks at $72 billion, and CFO Susan Li has described leading AI infrastructure as “a core advantage.” What Meta Compute adds is permanence: the spending line becomes a standing organization instead of a budget item renegotiated every quarter.
Flows across the industry point the same way. Etched, an AI chip startup, used the same week to announce $500 million in funding aimed squarely at NVIDIA; a few days earlier, xAI had set a record with a $20 billion Series E. Chips, power plants, server halls — the AI capital race opened 2026 at full throttle. Our 2026 opening outlook already flagged compute acquisition as a battlefield splitting away from the model race, and Meta Compute is the most explicit organizational evidence of that split so far.
What Developers and Product Teams Should Take Away
Three effects matter in practice. First, model supply becomes easier to plan around: with capacity now mapped across a ten-year horizon, teams that lean on Llama-family models or Meta’s ecosystem get a steadier pool of training and inference resources. Second, in-house silicon carries more weight: Janardhan’s brief lists silicon explicitly, so Meta’s in-house accelerator push beyond NVIDIA keeps moving forward, and that makes it a live variable for AI frameworks and chip supply chains alike. Third comes power costs: they will eventually surface in API pricing and in deployment limits — in a gigawatt-scale era, an architecture pinned to a single cloud or a single chip vendor is an open operational risk; multi-source compute planning should already sit on the engineering agenda.