What Reuters Reports
According to a Reuters exclusive report, Meta plans to begin mass production of a new generation of its own AI chips in September 2026. The basis of the reporting is an internal memo that was provided to the news agency according to a Reuters secondary publication. Reporters Katie Paul, Max A. Cherney, and Stephen Nellis paint a picture of a corporation that aims to systematically reduce its dependence on Nvidia and AMD graphics chips. Meta itself declined to comment on the details of the memo according to the reports.
Both TechCrunch and The Next Web independently confirmed the core statements: An initial chip of the new generation has reportedly gone through about six weeks of testing without major issues, so nothing stands in the way of the production start in September.
The Timeline Until 2027
According to the information from the memo, Meta aims to double its global computing capacity from around 7 gigawatts in 2026 to about 14 gigawatts in 2027. The new, self-developed chips are expected to take on an increasing share of the training and inference workloads, which have so far predominantly run on purchased Nvidia and AMD graphics processors.
For manufacturing, Meta is reportedly collaborating with chip designer Broadcom; the semiconductors will be produced by the Taiwanese contract manufacturer TSMC. Supplier contracts also exist with Samsung Electronics for memory, with SanDisk for flash storage, and with Sumitomo Electric for fiber optic technology in the data center network.
An internal codename for the new chip circulating in several secondary reports cannot be confirmed through the sources deemed reliable and is therefore intentionally not included here.
Technical Background: Meta’s MTIA Roadmap
The new chips are part of the in-house series “Meta Training and Inference Accelerator” (MTIA), which Meta has been developing since 2023. As Meta outlines in its corporate blog, the company has introduced four additional MTIA generations within two years, internally referred to as 300, 400, 450, and 500. The MTIA 300 generation is already reportedly running productively for ranking and recommendation algorithms, MTIA 400 has completed lab tests, while the generations 450 and 500, optimized for generative image and video applications, are expected to enter mass production only in 2027 according to Meta.
Between generations 300 and 500, the bandwidth of the memory (HBM) increases by a factor of 4.5, and the computing power in floating-point operations increases by a factor of 25, according to the company. Meta describes its chip strategy in the blog with three principles: fast iterative development cycles, an “inference-first” approach, and native integration with industry standards like PyTorch. The goal is to deliver a new chip generation approximately every six months – a significantly faster pace than is typical in the rest of the semiconductor industry, according to several industry media.
Additionally, Meta reported back in March 2026 about the expansion of its own chip manufacturing in the context of billion-dollar additional contracts with AMD and other partners.
Why Own Chips? The Cost Logic Behind the Strategy
The background of the strategy is of an economic nature. According to the memo that was available to Reuters, the transition to the latest Nvidia graphics chips has been cumbersome and time-consuming for a company the size of Meta. Every computing load that runs on self-developed chips instead of purchased Nvidia or AMD graphics processors reduces dependence on the margins of chip manufacturers.
The financial framework behind this is considerable: Meta plans investments of up to 145 billion US dollars in its AI infrastructure for 2026 according to the reports. According to several analysts quoted in the reports, the total AI infrastructure spending of the major tech companies is expected to exceed 700 billion US dollars in 2026 – a magnitude that could not be conclusively verified independently and should be understood as an industry estimate.
Context: An Industry-Wide Trend
Meta is not alone with this strategy. OpenAI, Anthropic, Amazon, and Google are also developing their own AI accelerators to become less dependent on Nvidia’s graphics chips in the long term. For Meta, the investment in its own chips comes at a time when the company reportedly cut around 8,000 jobs in May 2026 according to Yahoo Finance – primarily in the areas of Integrity, Cybersecurity, Content Design, and Reality Labs, while AI infrastructure and foundation model teams were exempt from the cuts. At an internal town hall in early July, Zuckerberg reportedly admitted that the development of AI agents had not accelerated as hoped in recent months. Whether the new chip generation will actually bring the hoped-for cost reductions can only be assessed once mass production begins in September and initial independent evaluations of the performance data are available.
