Key Points
- Meta will begin manufacturing its in-house Iris AI chip in September 2026.
- Broadcom is the design partner; TSMC will fabricate the chip.
- Iris is part of Meta's MTIA program, extended with Broadcom through 2029.
- Meta aims to cut compute costs and reduce reliance on Nvidia and AMD.
Meta will start manufacturing its custom AI chip, code-named Iris, in September, moving the company's silicon program from roadmap to production line. According to a company memo, the processor cleared its bug-testing phase in roughly six weeks without turning up any significant problems.
Iris is the newest entry in Meta's Meta Training and Inference Accelerators (MTIA) program, the in-house effort Meta framed this year as a four-generation chip roadmap. Broadcom is the design partner and Taiwan Semiconductor Manufacturing Co. will fabricate the chip; the two companies have extended their partnership through 2029 to cover multiple MTIA generations. It is the same design-partner model Broadcom ran with OpenAI on its Jalapeño processor.
The chip is one piece of a much larger buildout. Meta has committed to an initial deployment of more than one gigawatt of computing capacity, part of a plan for seven gigawatts online in 2026 scaling to 14 gigawatts by 2027. Owning the design is how Meta intends to make that math survivable — cutting compute spending and leaning less on outside suppliers even as big tech spends $700 billion on AI this year. Some of that capacity is already sited: Meta is putting a C$13 billion data center in Alberta, its first in Canada.
The catch, and the reason this is a hedge rather than a coup, is that Iris supplements Nvidia and AMD rather than replacing them. Meta is still buying GPUs by the warehouse; the in-house chips take on specific inference and ranking workloads where a purpose-built part is cheaper than a general one. That is the real prize here — not escaping Nvidia, but blunting its pricing power on the workloads Meta runs most.
Source: Yahoo Finance, Meta.
Comments