Meta will begin production of the latest generation of its in-house AI chips in September, according to an internal memo, part of a push to slow the flow of capital toward Nvidia and AMD as GPU supply tightens. At least one chip cleared its testing phase in about six weeks, an unusually fast turn for custom silicon. The chips fall under Meta's Meta Training and Inference Accelerator program, or MTIA, which Meta has been shipping in some form since 2023.
The economics are the story. Meta said in April it expects capital expenditures of $125 billion to $145 billion this year, most of it going to AI infrastructure. It plans to deploy 7 gigawatts of compute in 2026 and double that in 2027. At that scale, even a modest per-chip cost advantage over merchant GPUs compounds into billions.
Broadcom is the design partner on the new MTIA chips, with Taiwan Semiconductor Manufacturing Company handling fabrication. Meta is sourcing RAM from Samsung, storage from Sandisk, and fiber-optic equipment from Sumitomo Electric. The company detailed four new chips in the MTIA family in March, with some deploying this year and others next year, and it is building them as modular chiplets so individual blocks can be swapped as AI workloads shift.
Key facts
- 01Meta begins production of its latest MTIA AI chips in September, with at least one chip clearing testing in about six weeks.
- 02Broadcom is the design partner; TSMC will fabricate the silicon, with RAM from Samsung, storage from Sandisk, and fiber optics from Sumitomo Electric.
- 03Meta expects $125B to $145B in capital expenditures in 2026, most of it AI-related.
- 04The company plans to deploy 7 gigawatts of compute this year and double that next year.
- 05Four new MTIA chips were detailed in March, some deploying this year and next under a modular chiplet design.
The modular bet is deliberate. Fixed-function accelerators tend to age badly against the pace of model architecture changes, and Meta is trying to design around that risk.
“Each MTIA generation builds on the last, using modular chiplets, incorporating the latest AI workload insights and hardware technologies, and deploying on a shorter cadence”— Meta, Company statement, March
MTIA is aimed at three workloads: training the ranking and recommendation models that power Facebook and Instagram feeds, broader AI training, and inference for Meta's consumer applications, including its Muse Spark series of models. Ranking and recommendation is the most immediate prize because Meta runs those systems continuously at planetary scale, and every point of efficiency shows up in gross margin.
Meta is not abandoning merchant silicon. The company still expects to spend heavily with Nvidia and has a multibillion-dollar deal with AMD for Instinct GPUs, a separate multibillion-dollar arrangement with Amazon to use AWS's homegrown CPUs for AI workloads, and a deal signed last year with ARM to secure compute for its recommendation systems. The MTIA program is additive supply, not a replacement.
The pattern extends across the industry. OpenAI last month unveiled an inference processor it is building with Broadcom, and Anthropic is reportedly weighing a chip development effort with Samsung. Amazon and Google already ship their own training and inference silicon. Every hyperscaler large enough to run its own fabs' worth of chips is now trying to do so, and Broadcom is emerging as the common design partner across several of those efforts.
That concentration is worth watching. If Broadcom is designing custom accelerators for Meta, OpenAI, and a growing list of others, the company becomes a structural chokepoint in AI infrastructure alongside TSMC and Nvidia. It also changes Broadcom's revenue mix in ways that its guidance has not fully caught up with.
The execution risk on MTIA is real. Custom silicon programs at hyperscalers have a mixed track record — Google's TPU line took multiple generations before it was competitive with contemporary Nvidia parts on general workloads, and several other in-house efforts have quietly stalled. A six-week testing pass is a good sign, but production yields, software maturity, and workload coverage are where these programs succeed or fail.
For Meta, the calculation is narrower than most. It does not need MTIA to beat Nvidia across every workload. It needs the chip to be good enough at ranking, recommendation, and Muse Spark inference to displace a meaningful share of the GPU orders it would otherwise place. On a $125 billion to $145 billion capex line, even a 10% swing is a $12 billion to $14 billion decision.
The broader implication is that Nvidia's addressable market at the hyperscaler tier is now visibly capped. Meta, Google, Amazon, and increasingly OpenAI are all designing around the H-series and B-series with silicon they own, tuned to workloads they know. Nvidia keeps the frontier training market and the long tail of enterprise and neocloud demand, but the biggest single buyers are quietly building the exit. September's MTIA ramp is a milestone in that arc, not the end of it.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
