It’s 7:12 in the morning. You’re still half awake, thumb on the screen, and Instagram already knows to show you a video of someone restoring a rusted cast iron pan. You didn’t search for it. You never told anyone you find that satisfying. Somewhere in a building you’ll never visit, a rack of silicon made a guess about your mood before you finished making coffee.
That guess costs electricity. A lot of it. And Meta just decided it would rather build the hardware doing the guessing than keep buying it from someone else.
What Meta actually announced
Meta’s own AI chip, called Iris, began production in September 2026. Reuters reported the plan from an internal memo, and Mark Zuckerberg confirmed the timing. The chip is being manufactured with Broadcom and TSMC, and it’s meant to support Meta’s goal of doubling its computing capacity to 14 gigawatts by 2027.
Iris isn’t a general-purpose processor like the one in your laptop. It’s a data center AI accelerator, which is a narrower and more interesting thing. It exists to do two jobs: training, which is teaching a model patterns from enormous piles of data, and inference, which is the model actually answering something in the moment. Right now Meta buys chips for both of those tasks from Nvidia and AMD. Iris is Meta’s attempt to do it in-house.
Gigawatts, translated
Fourteen gigawatts is a number that slides right off the brain, so let’s anchor it. That’s enough power for over 11 million homes, dedicated entirely to running AI. Not heating them, not lighting them. Just math.
The climb is steep and already underway. Meta planned to deploy seven gigawatts of computing infrastructure this year. It added one gigawatt in the first half of the year, then forecast another 2.5 gigawatts to close the gap. Then the plan is to double that whole footprint again to hit 14 gigawatts in 2027.
If you’ve ever wondered why AI companies keep showing up in conversations about power grids and water usage, this is why. The interesting constraint on AI right now isn’t clever ideas. It’s electricity and the physical space to put chips in.
Why build your own chip when you can buy one
This is the part I find genuinely clarifying, because it says something about how AI systems get built.
A chip you buy off the shelf has to be good at everything, because it’s sold to thousands of customers doing thousands of different things. A chip you design yourself only has to be good at what you do. Meta’s memo describes Iris as designed for Meta’s specific workloads, which means recommendation engines and generative AI.
Recommendation engines are the quiet giant here. They’re the systems deciding which Reel, which post, which ad you see next, billions of times a day. That’s a different shape of problem than generating a paragraph of text. If most of your electricity bill goes to one repeated task, designing hardware tuned for exactly that task starts to look less like vanity and more like arithmetic.
There’s a strategic layer too, and it’s not subtle. Every dollar Meta spends on Nvidia and AMD chips is a dollar of its AI future sitting in someone else’s supply chain. Owning the silicon means owning the schedule.
What this means if you use AI agents
Most readers here care about AI agents, the assistants that book things, summarize things, and take multi-step actions on your behalf. Iris isn’t an agent and won’t feel like one. But it sits under them, and that matters in three practical ways.
- Cost shapes availability. Every message you send an AI agent is an inference request running on hardware someone pays for. Cheaper inference is the difference between an agent you use occasionally and one that runs quietly in the background all day.
- Speed shapes usefulness. Agents that take multiple steps to finish a task multiply latency. Hardware tuned for a company’s specific models is one of the main levers for cutting that lag.
- Capacity shapes ambition. Doubling compute doesn’t just mean more users served. It means bigger models and more steps per task become affordable to try.
The honest caveats
Production starting is not the same as production working at scale. Custom chips have a long history of arriving late, underperforming their targets, or being quietly reassigned to less glamorous duties. Meta hitting 14 gigawatts in 2027 depends on power agreements, construction timelines, and manufacturing yields that no memo can guarantee.
What we can say is the direction of travel. The companies building AI have decided the winning move is to own the physical layer, not rent it. Iris is Meta placing that bet with a date attached.
So the next time an app reads your mood at 7 in the morning, picture the power plant. That’s the real story of AI in 2026: not the magic on your screen, but the enormous, unglamorous machinery humming behind it.
🕒 Published: