Meta's Iris Chip Starts Production This Month

Meta's first mass-produced AI chip, code-named Iris, cleared testing in six weeks and enters production this month — the clearest sign yet that Meta wants to buy fewer Nvidia and AMD chips per unit of AI compute, not zero.

An internal memo reviewed by Reuters, first reported on July 9, showed Meta telling staff that Iris, its custom AI accelerator, had cleared its bug-testing phase without turning up significant problems and was on track to enter production in September, according to CNBC. The same memo laid out something bigger than one chip: a plan to nearly double Meta's total AI computing capacity, from 7 gigawatts online this year to 14 gigawatts by 2027.

What Iris actually is

Iris is not Meta's first attempt at custom silicon — it is the latest generation of the Meta Training and Inference Accelerator program, which the company has run since 2023. What makes this rollout notable is the partner list behind it: Broadcom is serving as Meta's design partner, and Taiwan Semiconductor Manufacturing Co. has been tapped to handle fabrication, according to DataCenterDynamics. That is the same Broadcom-TSMC pairing that Google and OpenAI have each leaned on for their own custom accelerators, which makes Iris less a solo bet than Meta joining a design pattern the rest of the industry has already converged on.

Meta's own March announcement of its MTIA roadmap, published on its Newsroom, described four new chip generations arriving within two years — MTIA 300 already in production for ranking and recommendations, with MTIA 400, 450 and 500 aimed primarily at generative-AI inference. Meta says it can now ship a new chip generation every six months or less, versus the industry's typical one-to-two-year cycle, by reusing modular designs that drop into existing rack infrastructure rather than requiring new data-center layouts each time.

The compute math behind the chip

The 14-gigawatt target is the more consequential number in the memo, and it is worth being precise about what it measures: Meta's total data-center power footprint across every chip vendor, not capacity attributable to Iris alone. AndroidHeadlines reported that Meta has deployed roughly 1 gigawatt so far this year, with the memo detailing plans to add another 5.5 gigawatts in the second half of 2026 alone — a pace that would make Iris one input into a much larger buildout still dominated by GPUs bought from Nvidia and AMD.

That distinction matters because it shapes what Iris is actually for. The chip is designed to supplement Meta's GPU purchases, not replace them — a strategy aimed at cutting the average cost per unit of inference compute rather than exiting the Nvidia relationship entirely. Meta deploys hundreds of thousands of MTIA chips today across ranking and recommendation workloads inside Facebook and Instagram, and the newer generations are built inference-first: rather than repurposing a chip designed for the heaviest workload (large-scale model training) down to cheaper jobs, Meta is optimizing MTIA 450 and 500 for inference first and letting them handle training only as a secondary use.

Who this changes things for

For Meta, custom silicon is a direct lever on the AI capital-expenditure numbers that have unsettled investors all year — every MTIA chip that replaces a marginal Nvidia GPU purchase is a chip Meta didn't have to buy from a supplier charging a premium during a global compute shortage. The company's stated goal of creating "personal superintelligence for all" requires that ever-larger compute budgets grow slower than user engagement and ad revenue, and in-house chips are the only lever available for that math besides simply spending less.

For Broadcom, the design win cements its position as the industry's preferred silicon partner for hyperscalers building custom AI accelerators, alongside similar work the company has done for Google's TPU line — a role that has made Broadcom's AI revenue one of the most closely watched line items on Wall Street this year, distinct from Nvidia's own approach of open-sourcing robotics models to widen its own hardware's reach instead.

For Nvidia and AMD, the near-term risk is muted precisely because Meta has been explicit that Iris supplements rather than replaces GPU purchases — but the long-term signal is less comfortable: every hyperscaler that proves it can design, test and ship its own accelerator on a six-month cadence is a hyperscaler with more leverage the next time it negotiates GPU pricing, a dynamic playing out in parallel with AMD's own decade-long push to break Nvidia's software moat.

The bigger pattern

Meta joins Google, Amazon, OpenAI and Microsoft in treating custom AI silicon as core infrastructure rather than a side experiment, and the timing is not a coincidence — every major AI lab is running into the same wall of GPU scarcity and pricing power that Nvidia has enjoyed since 2023. What distinguishes Meta's approach is the speed it claims: a six-month chip cadence, if it holds past this first Iris generation, would let Meta iterate on silicon roughly as fast as it iterates on its own AI models, closing a gap between hardware and software development cycles that has defined the entire generative-AI boom. Whether Iris delivers the cost savings Meta is banking on will only be visible once the 14-gigawatt buildout is finished and someone can point to an actual bill.

Frequently Asked Questions

What is Meta's Iris AI chip?

Iris is the code name for Meta's newest custom AI accelerator, part of its Meta Training and Inference Accelerator (MTIA) program. It cleared testing in about six weeks and was set to enter production in September 2026, according to an internal memo reviewed by Reuters.

Who is helping Meta build the Iris chip?

Broadcom is serving as Meta's design partner for Iris, while Taiwan Semiconductor Manufacturing Co. (TSMC) is handling fabrication — the same pairing several other hyperscalers have used for their own custom AI silicon.

Will Iris replace Meta's Nvidia and AMD chip purchases?

No. Meta has described Iris as supplementing its GPU purchases from Nvidia and AMD, not replacing them. The goal is to reduce the average cost per unit of AI compute as Meta scales its total data-center capacity.

How much AI computing capacity is Meta planning to add?

Meta's internal memo outlined a plan to nearly double its total AI computing capacity, from about 7 gigawatts online in 2026 to 14 gigawatts by 2027, with roughly 5.5 gigawatts planned for the second half of 2026 alone. That figure covers Meta's entire data-center footprint, not chips attributable to Iris specifically.

Editor's note — sources: CNBC, DataCenterDynamics, AndroidHeadlines, Meta Newsroom.