Meta's Iris AI Chip Enters Production in September

Meta will begin manufacturing its own AI accelerator, code-named Iris, as early as September. The chip was designed with Broadcom, will be fabbed by TSMC, and is the centerpiece of a plan to roughly double Meta's compute footprint from 7 gigawatts this year to 14 gigawatts in 2027, as first reported by Reuters.

The goal is not subtle: cut the dependence on Nvidia and AMD before the next wave of capacity gets ordered.

What Iris actually is

Iris sits in Meta's MTIA line (Meta Training and Inference Accelerators), the in-house silicon program that has so far produced inference chips for ranking and recommendation workloads on Facebook and Instagram. Iris is one of four planned chip generations, which tells you this is a roadmap, not an experiment.

One detail from the internal memo stands out: at least one Iris chip cleared its initial bug-testing phase in about six weeks with no major architectural problems. For silicon of this complexity, that's unusually fast. First bring-up is where custom chip programs typically stall for months, and Meta's went smoothly enough that leadership greenlit a September production start.

[!NOTE] Designing "with Broadcom" is the same playbook Google used for early TPUs. The partner handles the physical design heavy lifting while the customer owns the architecture. It compresses years off the schedule at the cost of margin.

Why every hyperscaler is doing this

Meta joins a club that already includes Google (TPU), Amazon (Trainium and Inferentia), and Microsoft (Maia). The economics are straightforward:

Industry coverage has started calling the demand-driven price pressure "chipflation," and that pressure is explicitly part of Meta's motivation. When you're planning to double to 14 gigawatts, a chip that costs 40% less per unit of inference is worth billions annually, even if it only handles a slice of the workload mix.

The catch: internal silicon only pays off if the software stack makes it usable. Nvidia's real moat has always been CUDA and the ecosystem above it. Meta has an advantage here that most companies don't: it controls PyTorch's direction and its own model stack end to end, so it can target Iris without waiting for the rest of the world to port anything.

What it means beyond Meta

Two things worth watching:

1. Nvidia's biggest customers are systematically becoming its competitors for their own internal demand. Nvidia keeps selling everything it makes anyway, but the hyperscaler share of that demand is the part with an expiry date.
2. The 14-gigawatt number is the story. That's the compute Meta thinks it needs by next year, and it's being announced the same week New York froze large data center construction and the industry's financing went into overdrive. The buildout and the backlash are now running side by side.

Production starting in September means Iris silicon lands in Meta data centers in volume through 2027. Whether it dents Nvidia's numbers is a 2028 question. Whether it dents Meta's capex line, we'll see in the earnings calls a lot sooner. Data Center Dynamics has additional reporting on the deployment plans.