You’re on your phone, thumb moving down Instagram Reels. A video about sourdough starter. Then one about a cat that hates Mondays. Then a short clip explaining a news story you hadn’t heard about yet. It feels like nothing. Your thumb barely registers the effort.
Somewhere in a building the size of several football fields, that scroll just triggered thousands of tiny calculations. What do you like? What will hold you for four more seconds? What should come next? Multiply that by billions of people scrolling at the same time, and you start to understand why Meta decided it needed to build its own computer chip.
That chip is called Iris. Production started in September 2026, and Meta wants it running across 14 gigawatts of computing capacity by 2027.
What a chip actually is, in plain terms
Think of a chip as a very specialized worker. A general-purpose processor is like a capable assistant who can do a bit of everything: answer emails, do some math, organize a calendar. Useful, but not fast at any one thing.
An AI accelerator is different. It’s a worker who does exactly one type of task, millions of times a second, and does nothing else. Iris falls into this category. Reuters reported from an internal memo that Iris is a data center AI accelerator, not a general-purpose processor. It was built for two jobs specifically:
- Recommendation engines — the systems deciding what shows up in your feed
- Generative AI — the systems producing text, images, and responses when you ask an AI assistant something
Meta made Iris with Broadcom and TSMC. It handles both training (teaching a model) and inference (the model actually answering you). Before this, Meta bought that hardware from Nvidia and AMD.
Why build instead of buy
Buying chips from Nvidia works fine. It’s also expensive, and you’re standing in the same line as every other company that wants them.
There’s a second reason that matters more for how AI agents behave. When you design your own chip, you design it around your own work. Meta knows exactly what its recommendation systems need to do because Meta built them. A chip made for those specific patterns can be faster and cheaper to run than a general one doing the same job.
It’s the difference between a rented van and a truck built for your exact cargo. Both move things. One wastes less.
The 14 gigawatt number, made human
Gigawatts measure electrical power, which is an odd way to talk about computers until you realize that power is the real limit now. Chips need electricity. Electricity needs generation and cooling and physical space.
Here’s how Meta’s buildout has gone: seven gigawatts of computing infrastructure planned for deployment this year. One gigawatt added in the first half, with another 2.5 forecast after that. Then the plan is to double the whole footprint again, hitting 14 gigawatts in 2027.
One analysis put 14 gigawatts at roughly enough power for over 11 million homes, dedicated entirely to running AI. That’s not a metaphor. That’s a real amount of electricity going toward machines that answer questions and pick videos.
What this means if you use AI agents
I write about AI agents for people who don’t build them, so let me connect this to something you can feel.
Every time an AI agent does something for you — drafts a reply, summarizes a document, plans a trip — a data center somewhere runs the calculation. The speed of that response, the cost of the service, and how much the company can afford to give you for free all trace back to hardware.
Cheaper compute tends to show up in three places:
- Response speed. Faster chips mean less waiting for an agent to finish thinking.
- Price. Companies that own their hardware have more room on what they charge.
- Ambition. When running a model gets cheaper, companies let agents attempt bigger, slower, more complicated tasks.
That third one is the interesting one. A lot of what AI agents currently can’t do isn’t a mystery of intelligence. It’s a matter of cost. An agent that could handle a forty-step task will happily do so once forty steps stop being expensive.
The honest caveats
Production starting is not the same as production working at scale. Chips are difficult to manufacture, and doubling a computing footprint in a year involves power grids, construction schedules, and supply chains that don’t always cooperate.
The size of the bet is the story here regardless. Meta looked at its own scroll-and-answer machine, calculated what the next few years would demand, and decided the hardware underneath needed to be its own. Your thumb, moving down a feed, is what all of it points at.
🕒 Published: