Meta has announced it will begin manufacturing its custom-designed artificial intelligence chip, code-named "Iris," in September 2026. The chip, part of Meta's Meta Training and Inference Accelerator (MTIA) program, passed its bug-testing phase in just six weeks with no major issues found, according to internal memos reviewed by Reuters.
The Iris chip is designed specifically for AI inference workloads that power Meta's family of applications — Facebook, Instagram, WhatsApp, and Threads — which collectively serve over three billion users worldwide. By tailoring the hardware architecture to its specific AI models, Meta expects to achieve significantly better performance-per-watt compared to general-purpose AI accelerators like Nvidia's H100 or B200 series. This efficiency gain translates directly into lower data center operating costs, a critical advantage as Meta projects 2026 capital expenditures exceeding $60 billion, largely driven by AI infrastructure investments.
Meta's custom silicon strategy also serves as a hedge against supply chain concentration in the AI chip market. Nvidia currently commands an estimated 80% of the AI accelerator market, giving it substantial pricing power and allocation leverage. By designing its own chips, Meta joins a growing club of tech giants — Google with its TPU line, Amazon with Trainium and Inferentia, and Microsoft with its partnership-based Maia chip — seeking to reduce dependency on a single vendor. Broadcom has partnered with Meta on the Iris chip's design and manufacturing integration, building on an earlier partnership agreement signed in early 2026.
The production timeline puts Meta on track to deploy Iris chips in its data centers by late 2026 or early 2027. Initial production volumes will be modest, focused on validating manufacturing yields and performance characteristics, before ramping to full production in 2027. The chips will initially handle inference workloads — the stage where trained AI models process user requests in real time — before potentially expanding into training workloads in future generations.
This development reflects a fundamental transformation in the semiconductor industry. As AI models grow larger and more computationally demanding, the one-size-fits-all approach of general-purpose processors becomes increasingly inefficient. Custom silicon allows hyperscale companies to eliminate unnecessary circuits, optimize memory hierarchies for their specific algorithms, and control their hardware roadmaps independently of the chip industry's traditional cadence. For Meta, which has doubled its AI compute capacity every year since 2022, Iris represents both a cost-saving measure and a strategic necessity.