Meta Puts Its Iris AI Chip Into Production in September to Double Compute to 14 Gigawatts
An internal memo reveals Meta's custom MTIA accelerator, designed with Broadcom and fabbed by TSMC, enters production in September. The chip cleared bug testing in six weeks and anchors a plan to double datacenter compute from 7 to 14 gigawatts.
Meta will start manufacturing its in-house AI chip, code-named Iris, in September, according to an internal memo reported on July 9. The accelerator was designed with Broadcom, will be fabricated by TSMC, and is the sharpest move yet in Meta’s campaign to cut its dependence on Nvidia and AMD GPUs.
The numbers around it are enormous. Meta intends to deploy 7 gigawatts of compute in 2026 and double that to 14 gigawatts next year. Total capital expenditure for the year is projected at $125 billion to $145 billion, most of it aimed at AI infrastructure. Iris exists to make those gigawatts cheaper: every recommendation query Facebook and Instagram serve on custom silicon is a query that doesn’t pay Nvidia’s margin.
Iris comes out of Meta’s MTIA program — Meta Training and Inference Accelerator — which has been shipping inference chips since 2023 but never at this scale or ambition. The supply chain reads like a who’s-who of the industry: Broadcom on physical design, TSMC on fabrication, Samsung supplying RAM, Sandisk on storage, Sumitomo Electric providing fiber-optic gear. According to the memo, at least one chip cleared its initial bug-testing phase in roughly six weeks without major architectural problems — unusually fast for silicon of this complexity.
The target workloads are the ones Meta knows best: ranking and recommendation algorithms across Facebook and Instagram, plus training and inference for broader AI workloads. Meta says it takes “a modular approach to designing these chips,” an acknowledgment that AI requirements will shift faster than a silicon production cycle.
Nobody should read this as Nvidia losing a customer. Meta explicitly expects to remain a major buyer of both Nvidia and AMD GPUs — frontier training still happens on merchant silicon. What changes is the mix. Inference at Meta’s scale is a cost-of-goods problem, and custom accelerators tuned to known workloads are how hyperscalers attack it. Google proved the model with TPUs; Amazon followed with Trainium and Inferentia. Meta joining production-scale custom silicon means every top-five AI spender now has an in-house alternative for at least part of its fleet.
Broadcom is the quieter winner. The same week this memo leaked, Apple committed over $30 billion to Broadcom for US chipmaking, and Anthropic already counts Broadcom as a compute partner. Designing custom accelerators for companies that would rather not pay GPU prices has become one of the best businesses in semiconductors.
September is close. If Iris ships on schedule and holds up in production, expect the 2027 conversation about AI economics to be less about GPU allocation and more about who owns their own silicon.