Meta's Custom AI Chip MTIA Enters Production in September — Trimming the 'Nvidia Bill'

Meta is on track to begin mass-producing the latest version of its self-designed AI accelerator chip in September, Reuters reported on July 9, 2026, citing an internal memo. The chip is the newest generation of the MTIA (Meta Training and Inference Accelerator) program Meta has run since 2023, and it is central to the company's push to lower the cost of the GPUs it buys from Nvidia and AMD. Per the memo, at least one chip cleared its testing phase in roughly six weeks.

Who Does What — Reshaping the Supply Chain

This chip isn't something Meta builds alone. Meta works with Broadcom on the chip design, but the actual manufacturing is handled by Taiwan's TSMC. On top of that, memory (RAM) comes from Samsung, storage from SanDisk, and fiber-optic equipment from Sumitomo Electric. In other words, behind a single AI chip sits a multinational semiconductor supply chain spanning the US, Taiwan, Korea, and Japan.

Why Build Its Own Chip — the 'Nvidia Bill' and Modular Design

The reason Meta keeps investing in custom silicon is simple. As AI training and inference demand explodes, the GPU bill it pays Nvidia and AMD has ballooned to a level that's hard to sustain. In March, Meta detailed four new chips developed under the MTIA program; some are already deployed or will be this year or next. The company says it designs these chips using a modular chiplet approach — precisely because AI workloads shift so fast that it needs flexibility for how requirements will have changed by the time chips reach production.

At the time, Meta wrote that "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." The intended uses are training its ranking and recommendation models, broader AI workloads, and inference aimed at its applications.

The Compute Arms Race — Every Big Tech Firm Is Building Silicon

This isn't a Meta-only move. Trying to stem the flow of capital to Nvidia via in-house chips is an industry-wide trend. Last month OpenAI unveiled its first custom inference processor built with Broadcom, and Anthropic is reported to be discussing developing its own chips with Samsung. Amazon (Trainium) and Google (TPU) have designed their own AI training and inference chips for years. A wave of AI-chip startups has sprung up to meet skyrocketing demand.

Behind all of it sits Meta's enormous compute-expansion plan. Meta is pouring tens of billions into data center and power deals worldwide to train and deploy its Muse Spark series of AI models. The company plans to deploy 7GW of compute this year and double that next year, with 2026 capital expenditure projected at $125–145 billion. Shaving even a little off that spending curve makes custom silicon a necessity, not a luxury.

What to Watch Next

Three things are worth watching. First, whether September production stays on schedule and how much it lowers Meta's GPU procurement share from Q4 onward. Second, Broadcom's rise as an "anti-Nvidia design house" as it designs custom chips for Meta and OpenAI at the same time. Third, whether custom chips can expand beyond recommendation and inference into frontier model training. The moment that boundary is crossed will mark the real start of any "de-Nvidia" shift.