Meta Moves Its Own AI Chip Into Production

Meta has told staff it will begin mass production of its first in-house artificial intelligence training chip in September, according to an internal memo reported by Reuters. The processor, called Iris, is the clearest sign yet that the company wants to build its own path off Nvidia's hardware rather than keep writing ever larger cheques for it.

The memo lays out an aggressive timeline. Meta plans to run about 7 gigawatts of computing capacity this year and double that to 14 gigawatts in 2027, which would put its data-centre fleet among the largest on the planet. Iris is meant to help carry that load. Testing reportedly took only six weeks and turned up no serious problems, which is quick for a chip of this complexity.

Built with Broadcom, made by TSMC

Meta designed Iris with Broadcom and will have it manufactured by Taiwan Semiconductor Manufacturing Company, the same foundry that prints Nvidia's and Apple's most advanced silicon. The chip belongs to Meta's MTIA line of accelerators, and the company reportedly wants a new generation roughly every six months through 2027. That cadence is unusually fast. Most custom chips move from design to production on a one to two year cycle.

For now Iris supplements Meta's Nvidia and AMD graphics processors rather than replacing them, and that distinction matters. Building a chip that works in a lab is one thing. Getting thousands of them to train frontier models reliably, at scale, is where several would-be Nvidia challengers have stumbled before.

Why bother

The appeal is money and supply. Nvidia's data-centre chips carry famously high margins, and demand has outstripped supply for two years running. Every hyperscaler that can design a workable in-house accelerator claws back some of that margin and insulates itself from the queue. Google has done it with its TPUs and Amazon with its Trainium parts. Now Meta is pushing hard to join them.

The move fits the wider hardware story we have been tracking. Memory maker SK Hynix just staged a record Nasdaq debut on the back of AI demand, and Meta itself has been breaking ground on multi-billion-dollar data centres and even looking to rent out spare compute. Those gigawatt numbers only work if the chips inside the buildings get cheaper, and Iris is Meta's bet on making that happen.

There is reason for caution. Custom silicon has a long history of slipped schedules and quiet cancellations, and a six-week test window is short. If Iris performs, Meta gains real leverage over its single biggest supplier. If it does not, the company has committed to a compute build-out it will still have to fill with Nvidia hardware at Nvidia's prices. September will be the first honest test.