Meta to deploy fourth custom AI chip by September, cutting Nvidia reliance
Meta’s in-house AI chip will double computing power by 2027
Meta is accelerating its custom AI chip timeline, planning to deploy its fourth-generation Meta Training and Inference Accelerator by September 2026. Designed in collaboration with Broadcom and manufactured by TSMC, the chip has performed well in early tests, allowing Meta to stay ahead of schedule. This push is part of Meta's strategy to reduce AI infrastructure costs and lessen dependence on Nvidia and AMD—just months after a $100 billion partnership with AMD. Meta aims to release a new chip every six months, a faster cadence than most major chipmakers.
Beyond chips, Meta plans to double its computing capacity from 7 gigawatts in 2026 to 14 gigawatts in 2027 through a mix of owned data centers, neocloud partnerships, and hyperscaler agreements. The company is also reportedly considering selling excess capacity to external AI developers like Anthropic and OpenAI, following SpaceX's model of generating billions in spare capacity revenue. For users, cheaper AI infrastructure could mean faster image generation, smarter assistants, and better recommendations across Facebook, Instagram, and WhatsApp without direct fees. However, it also raises privacy concerns as Meta pushes AI deeper into its smart glasses and social apps.
- Meta's fourth custom AI chip, co-designed with Broadcom, debuts by September 2026 with TSMC manufacturing.
- Meta targets doubling computing capacity from 7 to 14 gigawatts by 2027, rivaling OpenAI's Stargate project.
- Internal discussions explore selling spare cloud capacity to rivals like Anthropic and OpenAI, potentially generating billions.
Why It Matters
Meta's custom chips could lower AI operating costs, enabling more free features while reducing Big Tech's Nvidia dependency.