Iris in the Machine and Meta’s 14-Gigawatt Bet

Meta is building its own chip.

That sentence sounds small, and it isn’t. According to an internal memo reported by Reuters, Meta began production of a custom AI chip called Iris in September 2026, built alongside Broadcom and TSMC. Mark Zuckerberg confirmed the production start. The goal behind it is blunt: depend a little less on the GPUs Meta currently buys in enormous quantities from Nvidia and AMD.

If you build or use AI agents and none of this sounds like your problem, stick with me. Chip decisions made in 2026 quietly shape what your assistant can do, how fast it responds, and what it costs, in 2027 and beyond.

What Iris actually is

Iris is described as a data center AI accelerator, not a general-purpose CPU. The distinction matters more than it sounds.

A general-purpose processor is a generalist. It runs your operating system, your browser, your spreadsheet, and it’s decent at all of it. An accelerator is a specialist. It does one narrow category of math extremely well, and that category happens to be exactly the math that AI models need, over and over, billions of times a second.

Iris is meant to handle two jobs. The first is training, which is the expensive process of teaching a model by pushing enormous amounts of data through it until the patterns stick. The second is inference, which is what happens every single time you actually ask a model a question. Inference is the quiet cost center of modern AI. Training happens in bursts. Inference happens forever, every message, every user, every day.

One detail worth holding onto: Iris is meant to supplement Meta’s GPU purchases, not replace them. This is not a divorce from Nvidia. It’s Meta building a second kitchen so it isn’t entirely dependent on someone else’s stove.

The gigawatt number, translated

Meta plans to deploy seven gigawatts of computing infrastructure this year. It added one gigawatt in the first half and forecast another 2.5 to follow. And it intends to double that whole footprint again next year, reaching 14 gigawatts in 2027.

Gigawatts are a strange way to measure computers, and that’s the point. We used to count machines. Now the honest unit is electricity, because power is the actual constraint. To give that number a shape: 14 gigawatts is roughly enough electricity for over 11 million homes, pointed entirely at running AI.

When people say AI infrastructure has become an energy story rather than a software story, this is what they mean. You cannot spin up a gigawatt from a web console. It involves substations, land, cooling, permits, and years of planning.

Why a company would make its own silicon

Designing a chip is slow, costly, and full of ways to fail. Companies do it anyway for a few reasons that are easy to understand once laid out.

• Price. Buying millions of chips from a single supplier means paying that supplier’s margin on every unit. At Meta’s scale, shaving that margin is worth the engineering effort.
• Supply. If everyone in the industry is queuing for the same hardware, your growth plans belong to someone else’s factory schedule.
• Fit. A chip designed for your own models can skip the flexibility a general-purpose part needs and spend those transistors on the work you actually do.
• Negotiating position. A credible in-house option changes how every conversation with an outside vendor goes.

The tradeoff is that a custom chip is only as good as the software around it. Nvidia’s real moat has never been purely the hardware. It’s the years of tooling, libraries, and developer habit stacked on top. Meta is building with Broadcom and TSMC, two of the most capable partners available, which tells you it takes the hardware side seriously. The software side is where these projects usually get hard.

What this means if you just use AI agents

You will never see Iris. You’ll see its effects, indirectly, in three places.

Cost is the first. Inference is the recurring bill for every agent that runs on a schedule, watches an inbox, or answers customers around the clock. Cheaper inference is what turns “neat demo” into “always-on tool.”

Capability is the second. More compute means bigger models, longer memory, and agents that can take more steps before losing the thread.

Concentration is the third, and it deserves a raised eyebrow. When the companies that operate AI services also design the chips and own the power contracts, the number of organizations that can meaningfully compete gets smaller. That’s a reasonable thing for the rest of us to watch.

Meta has told us the plan and the number. September 2026 for production, 14 gigawatts in 2027. The interesting part is no longer whether the ambition exists. It’s whether a first-generation chip can carry that much weight.

🕒 Published: September 11, 2026