Fourteen gigawatts. That’s the number Meta is chasing by 2027, and it’s roughly enough electricity to power over 11 million homes. Except none of those homes get lit. All of it goes into data centers running AI.
Meta’s Iris chip entered production in September 2026, built with Broadcom and TSMC. It’s a data center AI accelerator, not a general-purpose CPU, designed to handle training and inference for the workloads Meta currently pays Nvidia and AMD to run. Mark Zuckerberg confirmed the production timeline. The company deployed seven gigawatts this year, adding one gigawatt in the first half and forecasting another 2.5 after that. Now it wants to double the whole footprint.
What Meta is actually buying with silicon
I review AI tools for a living, which means I spend most of my time watching companies promise things their infrastructure can’t support. Iris is the opposite problem. This is a company building infrastructure so far ahead of its product story that the gap is the story.
Custom silicon makes sense for exactly one reason: you know your workload cold. Meta does. Recommendation engines and generative AI are the two named targets, and recommendation is the older, less glamorous half. Meta has been running ranking models at planetary scale for over a decade. It knows the shape of those computations better than any chip vendor guessing at a general market. A chip tuned to that shape should beat a general-purpose accelerator on cost per inference. That’s not a moonshot. That’s engineering.
The generative AI half is where things get less certain. Meta is committing to a hardware roadmap that has to be locked years before the models running on it exist. Transformer architectures have been stable enough that this is a reasonable bet. But it is a bet, and the whole point of custom silicon is that it’s less flexible than the thing you replaced.
Why this matters if you use AI tools
Here’s my read on the practical consequence for anyone building on or buying AI products.
- Cost pressure on inference. If Meta successfully moves a meaningful chunk of its inference off purchased GPUs, it takes demand out of a market where everyone else is competing for the same supply. That’s good for smaller shops paying rack rate.
- Meta’s models get cheaper to serve. Cheaper serving means more aggressive free tiers, longer context windows, higher rate limits. Whatever Meta ships next, the unit economics behind it improve.
- Vendor lock-in gets weirder. A company running its own silicon has different incentives around open weights than one paying Nvidia margins. Cheaper inference could mean more openness, or it could mean Meta finally has a moat worth defending.
What I’m not going to tell you is that this ends Nvidia. Every hyperscaler has tried building its own accelerator, and the usual outcome is a chip that handles the boring high-volume work while the interesting experimental work stays on GPUs. Meta doubling capacity to 14GW means it needs a lot of chips from a lot of places. Iris supplements. It doesn’t replace, at least not on this timeline.
The number that should bother you
Going from seven gigawatts to fourteen in a year is not a chip problem. It’s a substations, transformers, transmission lines, and cooling water problem. The chip is the easy part. Meta can order more wafers. It cannot order more grid capacity and expect it next quarter.
This is the part of the AI buildout that tech coverage keeps skipping. The constraint moved. It used to be silicon supply. Now it’s electrons and the physical infrastructure that moves them. Any company announcing gigawatt targets is making a claim about utility relationships and permitting timelines, not just fab allocation.
I have no data on how Meta plans to source that power, so I won’t pretend to. But when a target requires doubling energy consumption in twelve months, the interesting question isn’t whether the chip works. It’s whether the plug exists.
My honest take
Iris is a solid, unsexy piece of vertical integration from a company with a genuine reason to build it. Meta has the workload volume to justify custom silicon and the balance sheet to eat the R&D. That combination is rare and it’s why this is more credible than most in-house chip announcements.
The 14GW target is the part I’d treat as aspirational until proven. Doubling infrastructure in a year is the kind of goal that gets quietly restated. Watch what Meta actually deploys, not what it announces.
And watch what shows up in the products. Fourteen gigawatts of compute has to justify itself eventually. So far the pitch is that Meta will use it to recommend things to you and generate things for you, which describes what Meta already does. The chip is real. The reason for that much of it is still a promissory note.
🕒 Published: