Meta to put its ‘Iris’ AI chip into production in September

TL;DR:

  • An internal memo seen by Reuters shows Meta will begin manufacturing its custom “Iris” AI chip in September as it aims to reach 14 gigawatts of computing by 2027.
  • The in-house silicon, built with Broadcom and TSMC, is designed to cut Meta’s dependence on Nvidia and AMD.
  • Meta expects to spend up to $145bn (about £110bn) on AI infrastructure this year.

Meta will start producing a custom AI chip in September, according to an internal memo reviewed by Reuters, as part of a plan to roughly double its computing power over the next year. The processor, code-named “Iris”, is the latest in Meta’s in-house accelerator line and is intended to loosen the company’s reliance on Nvidia and AMD.

Breaking the Nvidia habit

Iris passed testing in just six weeks with no major issues — brisk progress for an effort that had struggled since launching more than five years ago. Meta designed the chip with Broadcom and will have it manufactured by TSMC, aiming to lower the huge cost of the graphics processors it buys to run AI across Facebook and Instagram. “You can’t become an AI titan if you are dependent on another company for chips,” Forrester analyst Mike Gualtieri told Reuters, noting that hyperscalers increasingly see custom silicon as the only way to compete on price.

The scale is striking. Meta plans seven gigawatts of computing this year and 14 by 2027 — one gigawatt powers roughly 800,000 homes — backed by AI infrastructure spending of up to $145bn (about £110bn), part of Big Tech’s projected $700bn-plus outlay. To secure supply it has locked in long-term deals with Samsung for memory, Sandisk for storage and Sumitomo for fibre-optics, a response to the same memory shortage now driving up chip prices worldwide. The trend mirrors DeepSeek’s move to build its own chip.

Looking forward

Meta intends to release a new chip roughly every six months through 2027, an unusually fast cadence. For the wider market, its push is another sign that the biggest AI spenders want independence from Nvidia — a shift that could reshape a supply chain currently built around a single dominant vendor. For UK firms renting AI compute, cheaper in-house silicon at the hyperscalers may eventually filter through to lower cloud prices, though not before the current capacity crunch eases.

Share this article