Meta plans $125–145B AI capex

Meta just turned the AI spending race into an infrastructure story. On April 29, the company raised its 2026 capital-expenditure forecast to $125 billion to $145 billion, up from $115 billion to $135 billion, while saying the increase comes from higher component prices and extra data-center costs to support future capacity. That is not a normal budget revision. It is a signal that Meta now sees compute, power, and physical buildout as the main bottlenecks to winning in AI. ### Why is this bigger than a normal capex bump? Because this is not just more servers in existing rooms. Meta is telling investors it needs a much larger physical footprint — chips, networking gear, and data centers — to train and run larger models. The company spent $19.84 billion in capital expenditures in the first quarter alone, and its multiyear infrastructure commitments rose by $107 billion. Basically, the AI race is no longer just about model quality. It is about who can secure land, transformers, grid access, and enough equipment to keep scaling. ### What changed this week? The new wrinkle is that the power side of the story is getting harder to ignore. John Perella, a longtime data-center executive, pointed to a Virginia near-miss where multiple facilities dropped off the grid or switched to backup power during a disturbance, then created a dangerous mismatch when the grid expected load that was no longer there. His warning was simple — if AI campuses keep getting bigger, the failure modes get bigger too. ### Was the Virginia event real? Yes — and the official version was even larger than the anecdote now circulating. NERC’s incident review says a transmission-line fault in July 2024 triggered the simultaneous loss of about 1,500 MW of data-center load in Northern Virginia. Operators had to intervene because the grid had been planned for big generation losses, not for a huge block of load vanishing all at once. NERC later told FERC that Virginia saw a 1,500 MW load-loss event in July 2024 and another 1,800 MW event in February 2025. ### Why does losing load cause trouble? It feels backward, but the grid has to balance supply and demand every second. If a giant data center suddenly disappears as a customer, generation is still flowing for a moment, so frequency and voltage can jump. Think of it like a truck dropping its cargo at highway speed — the vehicle does not smoothly settle, it lurches. That is why grid planners are now treating hyperscale data centers less like passive customers and more like system-critical actors that need tighter coordination. ### Why does Meta matter more than the others here? Because Meta is not a cloud landlord selling spare compute. It has to build for its own models, products, and inference demand across apps used by billions of people. That makes its spending more direct and, in some ways, less optional. If Zuckerberg wants faster model training and wider deployment, Meta cannot wait for someone else’s capacity. It has to buy and build. ### So what is the real constraint now? Power delivery — not just chips. The industry spent the last two years talking about GPU shortages. But the catch is that a gigawatt-scale AI campus needs substations, transmission upgrades, backup systems, and operating rules that utilities and regulators are still figuring out. FERC has already been hearing about reliability risks from large-load integration, which means the politics of AI buildout are moving from Silicon Valley to public-utility territory. ### What should investors and everyone else watch? Watch whether Meta’s spending actually translates into energized capacity, not just announced capacity. A company can order chips quickly. It cannot summon grid upgrades on demand. If the next phase of AI depends on campuses that pull power like small cities, then the winners will be the companies that solve the boring stuff — interconnection, load management, and reliability — before the flashy model demos arrive. The bottom line is that Meta’s $125 billion to $145 billion plan is not just a bet on AI models. It is a bet that the physical world — power systems, data centers, and grid coordination — can keep up. That is where the real risk has moved.