Meta AI chips begin production in September, four MTIA designs

Meta will start production of its latest AI-specific chips in September. Reuters reported, citing an internal memo, that at least one of the new designs completed testing in about six weeks.

How will Meta build these chips?

Meta will work with Broadcom on chip design and rely on Taiwan Semiconductor Manufacturing Company, TSMC, for fabrication, while sourcing RAM from Samsung, storage from Sandisk, and fiber-optic equipment from Sumitomo Electric. The memo says Meta is assembling a supply chain that pairs Broadcom-led design work with TSMC manufacturing and third-party memory and networking parts.

Testing has moved fast for at least one device, the memo notes: one chip "sailed through its testing phase in about six weeks." That speed underpins Meta’s plan to shift some AI workload away from externally purchased GPUs toward its own silicon.

What are the MTIA chips for and how are they designed?

The chips are four designs under Meta Training and Inference Accelerator, or MTIA, intended for training ranking and recommendation models, broader AI workloads, and inference for its applications; some MTIA chips are already in deployment and others will roll out this year or next. Meta described the MTIA approach as modular, using chiplets so each generation can "builds on the last, using modular chiplets, incorporating the latest AI workload insights and hardware technologies, and deploying on a shorter cadence," the company wrote when it detailed the program in March.

Meta began producing its own AI chips in 2023 and first outlined the four MTIA designs in March. The modular chiplet strategy is meant to let Meta iterate on designs faster as AI workloads evolve between design and production windows.

Why does this matter?

Meta hopes the MTIA chips will reduce how much it must buy from GPU suppliers like Nvidia and AMD, even as it continues to spend heavily with them. The company told investors in April it expects capital expenditures between $125 billion and $145 billion this year, much of it directed to AI efforts. The internal memo cited by Reuters says Meta plans to deploy 7 gigawatts of compute this year and to double that next year, signaling very large capacity needs that make owning some custom silicon financially attractive.

Other major AI players are also moving the same direction. OpenAI unveiled an inference processor it is building with Broadcom, Anthropic is reported to be considering chips with Samsung, and both Amazon and Google develop their own chips for training and inference. Those moves show Meta’s initiative sits inside a broader industry push to cut dependence on external GPU vendors and control more of the compute stack.

What to watch

The immediate milestones are clear: TSMC beginning manufacture in September and broader deployment of MTIA chips this year. Two concrete signals to track are whether Meta meets its compute rollout targets — 7 gigawatts this year and a plan to double that next year — and whether the MTIA chips materially reduce purchases of Nvidia or AMD GPUs in upcoming procurement disclosures.