Meta Iris Chip Enters Production in September to Cut Nvidia Reliance

The Meta Iris chip will enter production in September, according to an internal memo reported by Reuters. Iris is the fourth generation of Meta’s in-house MTIA accelerator programme. It marks the company’s most serious attempt yet to loosen Nvidia’s grip on its AI infrastructure.

Broadcom helped design the chip. TSMC will manufacture it. Testing took just six weeks and surfaced no major issues, the memo said. That is a fast validation cycle for custom silicon, and it suggests the design is conservative by intent.

What the Meta Iris chip is for

Iris will not replace the vast fleets of Nvidia and AMD GPUs Meta already runs. It is designed to augment them. Custom chips handle predictable, high-volume workloads such as recommendation models more cheaply than general-purpose GPUs. Every task shifted to Iris trims Meta’s dependence on outside suppliers and its power bill per query.

The scale of the ambition sits in the numbers. Meta targets 7 gigawatts of compute capacity by the end of 2026. It plans to double that to 14 gigawatts in 2027, as CNBC reports. The company expects to spend up to $145 billion on AI infrastructure this year alone. At that scale, even a small cost advantage per chip compounds into billions.

The strategic logic mirrors moves across the industry. Google has its TPUs. Amazon has Trainium. Microsoft has been pushing its own MAI models to reduce reliance on OpenAI, a shift we covered in our report on Microsoft’s MAI models. Vertical integration has become the default defensive posture of every hyperscaler.

For Nvidia, the threat is gradual rather than sudden. Its GPUs remain the only practical option for frontier training runs. But inference, the everyday work of serving AI to users, is exactly where custom chips bite first. Iris is aimed at that soft underbelly.

September will show whether Meta’s six clean weeks of testing survive contact with mass production.

Fact basis: Reuters reporting on an internal Meta memo, July 9, corroborated by CNBC and US News coverage of the MTIA roadmap.