Meta Iris AI Chip

Meta just told us where the real AI bottleneck is. It isn't the model. It's the silicon.

Reuters reports Meta will put its custom "Iris" chip into production this September, part of a plan to reach 14 gigawatts of computing power by 2027 — roughly double its 2026 target of 7 GW. TSMC fabricates it. Broadcom designed it. The goal is blunt: less dependence on Nvidia.

The number that reframes the story is $145 billion. That's Meta's projected AI infrastructure spend this year alone — a meaningful slice of Big Tech's $700B-plus in total outlay. When a company spends at that scale, owning the chip stops being a hardware project and becomes a margin strategy. Every watt you don't rent from someone else's supply chain is a watt you control.

There's a cadence signal here too. Meta plans a new chip roughly every six months through 2027, against an industry norm of a year or more. That pace tells you they intend to iterate their way to independence, not buy their way there.

For enterprise leaders, the takeaway isn't "build your own chip." It's that compute is becoming the constraint that decides who can deploy AI at scale and at what cost. The organizations that win the next phase won't just have the best models. They'll have secured the capacity to run them.

Are you factoring compute availability and cost into your AI roadmap yet, or still treating it as someone else's problem?