Meta to produce custom Iris AI chip starting September; targets 14 gigawatts by 2027

Meta plans to begin manufacturing its custom Iris AI chip in September, according to an internal company memo reviewed by Reuters. The chip completed bug testing in six weeks with no major technical issues and represents a major step toward vertically integrated AI infrastructure. Broadcom serves as the design partner while TSMC handles fabrication. Iris is one of four generations in Meta's MTIA (Meta Training and Inference Accelerators) lineup, with new chip variants planned roughly every six months through 2027.

The Iris production milestone is part of Meta's ambitious infrastructure buildout: the company plans to deploy 7 gigawatts of computing capacity in 2026 (adding 5.5 GW by year-end after 1 GW in H1) and double that to 14 gigawatts in 2027. Meta expects to spend between $125 billion and $145 billion on AI infrastructure in 2026 alone—a significant slice of Big Tech's projected $700 billion collective AI spending. To secure components amid global shortages, Meta has locked in multi-year supply agreements with Samsung Electronics for memory, SanDisk for flash storage, and Sumitomo Electric for fiber-optic equipment.

Iris is designed to augment, not replace, NVIDIA and AMD GPUs that Meta continues to purchase at scale. The custom silicon allows Meta to deploy inference workloads on its own hardware, reducing scheduling bottlenecks and improving cost efficiency as it scales inference-heavy AI applications. The chip's design aims to accelerate inference for ranking, recommendation, and generative AI tasks across Facebook, Instagram, WhatsApp, and Meta AI products.

For infrastructure builders: Meta's ability to execute a custom silicon roadmap at this cadence—doubling GW every 12 months while maintaining quality (bug-free in 6 weeks)—suggests that hyperscalers can now credibly move inference workloads off merchant GPUs. The long-term message: competitive advantages in AI ops increasingly require custom silicon control, and cloud tenants should expect accelerators to commoditize further as every major player builds in-house.