Meta is on track to begin producing the latest generation of its own AI accelerators in September 2026, according to an internal memo cited by Reuters. The chips fall under the Meta Training and Inference Accelerator (MTIA) program, are co-designed with Broadcom, and will be manufactured in TSMC fabs. The goal is straightforward: lower the GPU bill during an unprecedented component shortage.
Key takeaways
- Production of the newest MTIA chips is planned for September 2026, per an internal memo cited by Reuters.
- The chips are designed with Broadcom and made by TSMC, with RAM from Samsung, storage from SanDisk, and fiber-optic gear from Sumitomo Electric.
- At least one chip cleared its testing phase in roughly six weeks.
- Meta projects capital expenditures of $125–145 billion in 2026, much of it on AI.
- The company plans to deploy 7 gigawatts of compute this year and double that next year.
Modular chiplets instead of one monolith
Meta detailed four new MTIA-family chips back in March 2026. Some are already deploying, others will enter use this year or next. The company favors a modular approach — building processors from chiplets on the assumption that its needs will shift faster than the silicon can be produced.
"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 wrote when it introduced the chips.
The use cases are practical rather than prestige-driven. MTIA are meant to train models for ranking and recommendation algorithms, handle broader AI workloads, and run inference for Meta apps. These are areas where the company generates enormous, repeatable traffic — and where custom silicon pays off fastest. Meta has produced its own AI chips since 2023, so the September start is a next step, not a debut.
Not a break from Nvidia, but diversification
Reuters notes the in-house chips should help curb GPU purchases from Nvidia and AMD, yet Meta still plans to spend heavily with both. This is risk-spreading, not a rupture. Alongside MTIA, Meta has a deal with ARM for recommendation-system compute, a multibillion-dollar contract for AMD Instinct GPUs, and an agreement with Amazon for millions of its homegrown CPUs for AI work.
Meta is not alone here. OpenAI unveiled an inference processor built with Broadcom last month, and Anthropic is reportedly discussing its own chip with Samsung. Amazon and Google have developed custom training and inference silicon for years. The common thread is clear: every large player is trying to stem the flow of capital to Nvidia and regain control over cost per token.
The scale of that pressure shows in Meta's own figures. The projected $125–145 billion in 2026 capex is largely consumed by the AI infrastructure needed to train and deploy the Muse Spark model series. Stated plans for 7 gigawatts of compute this year and a doubling next year suggest custom silicon is an economic necessity rather than an experiment. Meta declined to comment.
Why it matters
A September production start signals that vertical integration in AI is no longer the sole domain of Google and Amazon. When an ad-and-social company decides to design and order its own silicon, it means the cost of compute for recommendation and inference has grown large enough that a GPU vendor’s margin becomes a real drag on the balance sheet. Modular chiplets matter strategically here — they let Meta shorten deployment cycles and tune hardware to shifting workloads rather than freezing an architecture for years. At the same time, the diversified supply chain (TSMC, Samsung, SanDisk, Sumitomo, ARM, AMD, Amazon) shows that even the largest firms cannot break free of outside partners — they can only spread the risk. For the market, it confirms that the war over inference cost is now being fought at the silicon layer, not just in the models.
What's next?
- Production of the newest MTIA chips is planned for September 2026 — the first production deployments will reveal real performance against Nvidia and AMD GPUs.
- Meta says it will deploy 7 GW of compute in 2026 and double that in 2027 — the pace will show how quickly custom silicon replaces external purchases.
- Future MTIA generations are meant to ship on a shorter cadence thanks to modular chiplets, translating into faster hardware turnover in data centers.
