Meta plans to begin manufacturing an in-house AI chip in September, according to an internal memo reviewed by Reuters, as part of a push to roughly double its computing power to 14 gigawatts next year.
The data center chip, code-named Iris, passed testing in about six weeks with no major issues, a relatively quick result that signals momentum for a custom-silicon effort that has struggled since it began more than half a decade ago. Iris is the latest in Meta’s four-generation Meta Training and Inference Accelerator, or MTIA, family, and corresponds to the MTIA 400 that Meta unveiled publicly in March. Meta designs the chips itself, works with Broadcom on the design and relies on Taiwan Semiconductor to manufacture them.
The strategic goal is to reduce Meta’s dependence on Nvidia and AMD, whose GPUs it buys in enormous quantities, and to lower the cost of the compute that powers Facebook and Instagram. The memo was candid about the difficulty of relying on outside suppliers, noting that adopting the latest GPUs at Meta’s scale “has been a heavy lift, and it has cost us time.”
Iris is aimed at augmenting rather than replacing those GPUs, and it is built for inference, the work of running AI models, particularly the ranking, recommendation and generative features across Meta’s apps, rather than training the largest models. Meta plans to release a new MTIA chip roughly every six months through 2027, far faster than the industry norm of a year or more, an approach enabled by reusing modular designs.
The chip is one piece of a vast buildout. Meta intends to deploy seven gigawatts of computing infrastructure this year and double that in 2027, underpinning a capital spending budget of as much as $145 billion for 2026, a large share of the more than $700 billion Big Tech is expected to spend on AI. To hit those targets, Meta has locked in long-term supply deals, including memory chips from Samsung, flash storage from Sandisk and fiber-optic equipment from Sumitomo Electric.
Why It Matters
Owning its silicon gives Meta leverage on cost, supply and design. Custom chips tuned for its specific workloads can deliver more performance per dollar than general-purpose GPUs, and a diversified supply chain insulates the company from price swings and shortages at any single vendor. That independence matters more as GPU demand outstrips supply and prices climb.
It also fits a broader industry shift, with Google, Amazon and Microsoft all building their own AI chips to escape reliance on Nvidia, whose dominance rests partly on scarce, expensive hardware. Broadcom, which is helping design Iris and also builds custom chips for Google and Anthropic, is emerging as a key winner of this move toward bespoke silicon.
The Supply Squeeze
The memo underscores how acute the components crunch has become. Meta’s emphasis on multi-year supply agreements reflects a memory shortage severe enough that companies including Apple have raised product prices, and Morgan Stanley analysts have coined the term “chipflation” to describe the macroeconomic effect of rapidly rising chip costs.
Meta executives have said they are worried about securing enough high-bandwidth memory for their roadmap, even as they believe they have locked in what they need. The push also arrives amid pressure to prove the payoff of Meta’s spending, after CEO Mark Zuckerberg recently conceded that some AI bets have not yet delivered. A cheaper, in-house chip that reaches production on schedule would be tangible evidence that the enormous outlay is translating into real efficiency, though the harder test is whether Iris performs at scale in live data centers.
