Meta will begin manufacturing the newest generation of its in-house AI chip in September, according to TechCrunch, which cited a Reuters report on an internal company memo. The chips belong to Meta’s Training and Inference Accelerator program, known as MTIA, and represent the company’s most concrete move yet to reduce how much it spends on Nvidia and AMD processors.
At least one of the new designs passed testing in roughly six weeks, according to the internal memo Reuters cited. Broadcom is Meta’s partner on chip design, the same role it plays for Google and, since last month, OpenAI on their own custom silicon efforts. Production will run through Taiwan Semiconductor Manufacturing Company, the foundry that also builds Nvidia’s and Apple’s chips. Component sourcing extends further: Samsung is supplying memory, Sandisk is supplying storage, and Sumitomo Electric is supplying the fiber-optic gear for the buildout, TechCrunch reported.
Meta first detailed four new MTIA designs in March, built around a modular chiplet strategy meant to let each generation absorb new AI workload demands without a ground-up redesign. Some of those four are already deployed in limited form. The rest are expected to ship later this year or in 2027. The approach lets Meta iterate on silicon closer to the pace at which its model architectures change, rather than locking in a chip specification years ahead of deployment.
The chips serve three jobs at Meta: training the ranking and recommendation systems behind its ad business and feeds, running broader AI research workloads, and handling inference for consumer products including the Muse Spark model family. Meta has built its own AI silicon since 2023. September is when the newest chiplet-based generation moves from validated design into actual manufacturing output.
This is a hedge, not a substitute. Meta still intends to keep buying heavily from Nvidia and AMD as MTIA volume ramps, TechCrunch reported. The company guided to $125 billion to $145 billion in capital spending this year, with most of the increase tied to AI infrastructure. On top of that, Meta carries a large AMD agreement for Instinct GPUs, a separate arrangement to run AI workloads on Amazon’s own CPUs, and an ARM deal signed last year to support recommendation-system compute. Custom silicon lowers the cost of each unit of capacity. It does not eliminate the need for capacity at the scale Meta is buying.
The real stakes are Meta’s AI compute bill, not a chip rivalry for its own sake. Meta plans to bring 7 gigawatts of AI compute online this year and roughly double that capacity in 2027, according to the memo cited by Reuters, a target that turns small percentage gains on training and inference hardware into meaningful savings once multiplied across that footprint. OpenAI’s Broadcom-designed inference processor and Anthropic’s reported talks with Samsung over custom chips point to the same logic. Every frontier lab now treats an in-house silicon program as leverage against Nvidia’s pricing power, even while remaining one of Nvidia’s largest customers.
For teams budgeting AI infrastructure spend into 2027, Meta’s September production start is an early marker of how much training and inference capacity could shift away from merchant GPUs, and how soon that shift shows up in cloud and inference pricing.
TechCrunch (Ram Iyer), citing a Reuters report on an internal Meta memo, published July 9, 2026.
