Eleven million homes. That’s roughly what 14 gigawatts of electricity would power if you handed it all to residential customers. Meta wants that much capacity pointed at AI compute by 2027, and it just started building its own silicon to fill the racks.
The chip is called Iris. Production began in September 2026, built with Broadcom and manufactured by TSMC. Zuckerberg confirmed the start date. It’s a data center AI accelerator, not a general-purpose CPU, designed to handle training and inference for the models Meta currently buys Nvidia and AMD hardware to run. Reuters reported the plan out of an internal memo, which is usually how you learn that a company has been quietly building an escape hatch.
The math on that gigawatt number
Meta planned to deploy seven gigawatts of computing infrastructure this year. It added one gigawatt in the first half and forecast another 2.5 after that. Then the goal is to double the entire footprint again to reach 14 gigawatts in 2027.
Doubling a datacenter footprint in a year is not a software problem. It’s a power problem, a construction problem, a permits-and-transformers problem. Every gigawatt has to be generated, delivered, cooled, and staffed. When a company talks about doubling compute capacity, the chip is the part that gets the press release and the substation is the part that decides whether it actually happens.
The reported spend attached to this effort sits around $145 billion. Wall Street responded the way Wall Street does, with a $743 target on META after a 15% week. That’s a bet on the plan, not on the results, because there are no results yet.
What I can and can’t tell you as a reviewer
My job on this site is to test things and tell you whether they work. Nobody outside Meta can. There are no published benchmarks, no throughput numbers, no perf-per-watt comparisons against an H100 or an MI300. What exists is a production start date, a partner list, and a capacity target.
So treat every confident take about Iris beating Nvidia as fan fiction. Custom accelerators have a long history of looking great on internal slides and mediocre in practice, usually because the hardware ships before the compiler stack is ready. Google took multiple TPU generations to get software that developers could actually use. Amazon is still working on making Trainium feel like a real option rather than a discount. Meta’s advantage is that it doesn’t need to sell Iris to anyone. It only needs Iris to run Meta’s own workloads, which it controls end to end.
That’s the smart part of the design choice. Iris is aimed at recommendation engines and generative AI, which happen to be the two things Meta runs at absurd scale. A chip built for two known workloads under one roof is a far easier engineering target than a chip built to satisfy every customer on the planet. Narrow scope is an advantage here, not a limitation.
Why anyone building on AI should care
You will never touch this chip. You will feel its consequences.
- Inference cost pressure. If Meta can serve its own models on its own silicon, the cost floor for open-weight model hosting drops. Meta has been giving models away to commoditize the layer that competitors sell. Cheaper internal compute makes that strategy sustainable for longer.
- Less Nvidia dependency at the top. Meta is not abandoning Nvidia and AMD tomorrow. Iris takes on work currently bought from them, which changes the negotiating position more than it changes the supply chain. Every hyperscaler with its own accelerator gets better pricing on the ones it still buys.
- Feature velocity in products you use. Recommendation ranking is where Meta makes money. More dedicated capacity there means more aggressive model iteration in the surfaces billions of people already open daily.
- A power ceiling nobody has solved. If 14 gigawatts turns out to be unreachable, the constraint hits everyone’s inference bills, not just Meta’s.
My honest read
The chip is the least interesting part of this story. Every large AI operator is building custom silicon now, and the engineering path is well documented enough that Broadcom and TSMC can get you there if you write the check. What separates the announcements from the outcomes is whether the electricity and the software stack arrive on schedule.
I’d watch three things over the next year. Whether Meta publishes any performance data on Iris, which would signal real confidence. Whether the 2.5-gigawatt addition lands as forecast, which tells you if the infrastructure timeline is credible. And whether Nvidia orders from Meta actually shrink, which is the only proof that Iris is doing production work rather than pilot work.
Until then, this is a serious plan with a serious budget and zero verifiable performance. That’s not a criticism. It’s just where the evidence currently stops.
🕒 Published: