Meta Platforms has begun manufacturing its third-generation Meta Training and Inference Accelerator, code-named Iris, this September — a milestone that ends Nvidia's monopoly on Meta's training infrastructure and cements Broadcom as the internet giant's silicon partner-of-choice.
The chip
MTIA v3, nicknamed Iris, was designed with Broadcom and is being fabbed on TSMC's 3nm process. Meta says the chip cleared its bug-testing phase in about six weeks without meaningful defects and is engineered to handle both large training runs and low-latency inference for feed ranking, ads and generative features across Facebook and Instagram.
Why now
Meta's AI infrastructure spending is projected to hit $145 billion this year, and the company said in an internal memo that its two-step buildout will bring seven gigawatts of compute online in 2026 and 14 gigawatts by 2027. Iris is one of four planned generations of MTIA silicon aimed at cutting per-token training cost and reducing Meta's dependence on Nvidia and AMD — a strategy that mirrors moves by Google (TPUv6), Amazon (Trainium 3) and Microsoft (Maia).
The broader silicon shakeup
Iris arrives at a moment when hyperscalers are actively rerouting demand around Nvidia — see the DOJ's antitrust probe into Nvidia's Groq deal and Positron's $875M Series C. If Broadcom captures another custom-silicon flagship like this one, its ASIC business is on track to rival Nvidia in cumulative hyperscaler revenue by 2028.
What comes next
Meta plans to deploy Iris across its next-generation AI Superclusters in Louisiana and Utah, with rack-scale deployment beginning in Q4 2026. A fourth-generation MTIA is already in tape-out at Broadcom.
Reporting based on coverage from Reuters, CNBC, Electronics Weekly and Meta engineering communications.