Meta to begin production of new custom AI chips this September
Meta plans to begin production of its custom MTIA AI accelerator chips in September to decrease reliance on third-party GPUs for training ranking and recommendation models. The initiative forms part of a broader infrastructure strategy to scale compute capacity while managing massive capital expenditure on AI hardware.
Key Takeaways
- New custom MTIA chips, including the Iris processor, enter mass production in September 2026 after completing testing in six weeks.
- Meta's infrastructure strategy aims to deploy 7 gigawatts of computing power in 2026, doubling to 14 gigawatts by 2027.
- The company secured multi-year hardware agreements with Broadcom for design, TSMC for manufacturing, and Samsung for high-bandwidth memory (HBM).
- Capital expenditure for 2026 is forecast between $125 billion and $145 billion, driven by AI hardware and data center costs.
Why It Matters
Meta's shift to custom silicon reduces its extreme financial exposure to third-party GPU vendors as it scales its AI-driven feed and recommendation engines. By developing modular, chiplet-based hardware on a shorter six-month cadence, Meta can adapt its infrastructure to evolving model architectures faster than traditional silicon cycles allow. This vertical integration directly supports the deployment of its new Muse Spark model series and underpins a broader strategy to monetize excess compute capacity through its emerging cloud and API business. Watch for Meta's 2027 data center efficiency ratings to signal if this custom silicon approach successfully delivers the projected 35% reduction in operating costs.
Additional Context
The production launch of the Iris chip follows a period of significant strategic pivots for Meta's hardware and software divisions. In April 2026, per Meta’s Q1 earnings report, the company raised its full-year capital expenditure forecast to a range of $125-$145 billion, an increase of $10 billion from prior guidance. This revised outlook, which rattled investors and briefly lowered the stock price, was attributed to rising component prices and a global shortage of flash storage and high-bandwidth memory. To secure its supply chain, Meta has established long-term procurement frameworks, including a multibillion-dollar agreement with Amazon Web Services in April 2026 to deploy tens of millions of Graviton5 processors for CPU-intensive inference tasks. Meta is also diversifying its high-end GPU sources beyond Nvidia. In February 2026, per Forbes and CNET, Meta signed a multi-year deal with AMD to purchase 6 gigawatts of Instinct GPUs, a deal that included warrants for Meta to acquire up to 10% of AMD’s common stock. This dual-sourcing strategy, combined with the internal MTIA program, is part of a broader effort to provide infrastructure for the newly launched Muse Spark 1.1 model. Released by the Alexandria Wang-led Meta Superintelligence Labs in July 2026, Muse Spark 1.1 is Meta’s first frontier model to be offered via a paid developer API, signaling a shift from its traditional focus on open-weights distribution toward a commercial cloud services model similar to OpenAI and Anthropic.
Read full article at techcrunch.com
Get this in your inbox → Subscribe
Enjoy our coverage?
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source