On January 12, 2026, Mark Zuckerberg announced in a Threads post: “Today we’re establishing a new top-level initiative called Meta Compute.” This is not another data center purchase — it elevates AI infrastructure to a company-level program standing alongside the model business itself. Reuters framed it plainly: Meta will build gigawatt-scale computing capacity under the Meta Compute effort.
Zuckerberg set the scale target himself: Meta plans to build “tens of gigawatts this decade, and hundreds of gigawatts or more over time.” For context, a large nuclear reactor unit generates roughly 1 GW. Meta is effectively declaring it will lay down compute equivalent to dozens of power plants for AI alone. After a 2026 opening streak of compute announcements, this is the loudest signal yet that power and land, not model architecture, now set the agenda.
Three Leaders, One Top-Level Program
TechCrunch reports that Meta Compute is co-led by three executives whose remits map exactly onto the three hard problems of infrastructure.
Santosh Janardhan, head of global infrastructure who joined Meta in 2009, runs technical architecture, the software stack, silicon, developer productivity, and the global datacenter fleet and network — the operator of everything Meta already runs. Daniel Gross, who joined in 2025 and co-founded Safe Superintelligence with Ilya Sutskever, heads long-term capacity strategy, supplier partnerships, planning, and business modeling — the “what to buy, when, and at what price” function. Dina Powell McCormick, recently named president and vice chairman, handles government relations to help finance and deploy the infrastructure — because at this scale, building data centers is ultimately a permitting and politics problem.
Putting an SSI co-founder in charge of capacity strategy is the most telling move of the arrangement: Meta does not want more procurement managers — it wants someone who understands the demand curve of frontier models planning compute years ahead.
Power Is the Real Battlefield
The unit of measurement in this race has shifted from GPU counts to gigawatts. Axios cites Zuckerberg’s targets — tens of gigawatts this decade, hundreds or more over time — while TechCrunch points to estimates that US AI power demand could grow from 5 GW to 50 GW by 2030. In other words, even the near-term “tens of gigawatts” goal puts a single company’s buildout in the same league as the projected AI electricity consumption of the entire United States.
Competitors are moving too: Microsoft keeps expanding through infrastructure partnerships, and Alphabet acquired data center firm Intersect in December 2025. With every frontier lab chasing the same grid capacity, transformers, and land, whoever secures power first secures a ticket to the next model generation.
Capex and Market Context
The money already told this story. Meta’s 2025 capital expenditure guidance ran as high as $72 billion, with CFO Susan Li calling leading AI infrastructure “a core advantage.” Meta Compute turns that spending line into a permanent organization rather than a budget item renegotiated quarter by quarter.
The direction matches capital flows across the industry. In the same week, AI chip startup Etched announced a $500 million round to challenge NVIDIA directly, and days earlier xAI had closed a record $20 billion Series E. From chips to power plants to server halls, the AI capital race heated up across the board at the start of 2026. As we noted in our opening outlook for 2026, compute acquisition has become a battlefield separate from the model race — Meta Compute is its clearest organizational proof so far.
What It Means for Developers and Product Teams
Three practical effects. First, longer visibility into model supply: with capacity planning now on a decade scale, teams building on Llama-family models or the Meta ecosystem can expect a more stable pool of training and inference resources. Second, more leverage for internal silicon: Janardhan’s remit explicitly includes silicon, so Meta’s in-house accelerator track beyond NVIDIA will keep advancing — a live variable for chip supply chains and AI frameworks alike. Third, power costs will gradually show up in API pricing and deployment limits: in the gigawatt era, an architecture tied to a single cloud or single chip vendor is an explicit operational risk, and multi-source compute planning belongs on the engineering agenda now.
Sources
- Mark Zuckerberg says Meta is launching its own AI infrastructure initiative — TechCrunch
- Meta to build gigawatt-scale computing capacity under Meta Compute effort — Reuters
- Meta launches new Meta Compute initiative to build AI infrastructure — Axios
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.