Meta to Begin Production of In-House AI Chip Iris in September

Meta will begin production of its own artificial intelligence chip, codenamed Iris, in September 2026. It is the first step in a four-year program to build an in-house line of AI chips, aimed at reducing the company's reliance on Nvidia and supporting a doubling of Meta's computing capacity to 14 gigawatts by 2027.

Reuters first reported the news, citing an internal Meta memo, and the report was later confirmed by CNBC and several other business outlets. Iris is an ASIC, an integrated circuit designed for a specific task, in this case optimizing recommendation ranking and content feeds across Meta apps such as Facebook and Instagram.

A Four-Generation MTIA Roadmap

Iris is part of the Meta Training and Inference Accelerator program, or MTIA, which calls for releasing a new generation of the company's own AI processor every six months through 2027. The successive variants are designated MTIA 300, 400, 450 and 500. That pace is meant to help the company keep up with growing demand for computing power to train and run AI models.

Broadcom is responsible for the chip's design, having positioned itself over the past several years as the go-to design partner for tech companies building their own AI chips instead of relying solely on Nvidia. Production is handled by Taiwan's TSMC, using the most advanced lithography processes available on the market.

Why Meta Is Moving Beyond Nvidia

Building in-house inference chips is a strategy already pursued by Google with its TPUs and Amazon with Trainium. For Meta, the goal is to cut per-unit costs at the scale needed to serve content recommendations to billions of users daily, as well as to reduce the risk of depending on a single hardware supplier at a time when demand for Nvidia chips far outstrips supply.

Testing the Iris chip took just six weeks and, according to the memo cited in the reports, revealed no major technical problems, which sped up the decision to move into production. That matters because earlier MTIA generations faced delays and saw limited use, mostly confined to recommendation tasks rather than training large language models.

The Scale of Investment

Meta raised its 2026 capital expenditure forecast to a range of $125-145 billion, with most of it going toward data centers, GPUs and now its own custom chips. The company plans to double its total data center computing capacity from 7 gigawatts in 2026 to 14 gigawatts in 2027.

The market reacted positively to the news. Shares of Broadcom and other companies in the semiconductor supply chain rose after the reports emerged, as investors read the scale of Meta's spending as confirmation that the investment boom around AI infrastructure is still accelerating rather than slowing down.

What It Means for the Market

For TSMC, Meta's decision means another major customer using its most advanced production lines, regardless of whose logo ends up on the finished chip. For Nvidia, it's a sign that its biggest customers are consistently building alternatives to its processors for at least part of their workloads, especially where large-scale inference cost matters more than raw training power.

For companies relying on cloud infrastructure and Meta's ad platforms, the effect will be indirect, through potentially lower operating costs that could translate into more stable service pricing. Direct access to Iris chips outside of Meta is not currently planned, unlike Nvidia's chips, which are widely available through public clouds.

Sources: Reuters via CNBC (cnbc.com), MLQ News (mlq.ai), Tom's Hardware (tomshardware.com).