Summary：Meta to begin production of Iris AI chips to double computing capacity **Introduction** Meta Platf

referrerpolicy="no-referrer"
style="max-width:100%;height:auto;display:block;margin:0 auto;">

Meta to begin production of Iris AI chips to double computing capacity

**Introduction**
Meta Platforms announced that it will start manufacturing its next‑generation AI accelerator, code‑named Iris, in September 2026. The chip is the latest addition to the Meta Training and Inference Accelerator (MTIA) family and is designed to lift the company’s internal compute power by roughly two‑fold. By moving Iris from silicon design to volume production, Meta aims to reduce reliance on third‑party GPUs while supporting the growing demand for large‑scale language models, recommendation systems, and immersive reality workloads.

**Key Developments**
Iris employs a modular chiplet architecture that lets Meta mix and match compute, memory, and I/O dies to tailor performance for specific AI tasks. The first silicon samples, fabricated on a 4‑nm process node, demonstrated a 1.8× improvement in training throughput and a 2.1× gain in inference latency compared with the prior MTIA generation. Meta’s internal roadmap shows that initial wafer starts will occur at its new fab partner in Arizona, with ramp‑up targeting 10 million units per year by 2028. The company also disclosed that Iris will integrate a proprietary interconnect fabric, enabling seamless scaling across its data‑center pods without the bottlenecks seen in traditional GPU clusters.

**Industry Analysis**
The move signals a broader shift among hyperscalers toward custom silicon as a way to control cost, power, and performance. Analysts note that Meta’s investment in Iris could pressure incumbent GPU suppliers to accelerate their own roadmap revisions, especially as AI training budgets continue to climb past $150 billion annually. Moreover, the chiplet approach aligns with industry trends championed by AMD and Intel, suggesting that Meta’s design may become a reference point for future AI accelerators. Critics, however, caution that yield risks at 4‑nm and the complexity of software stack migration could temper early adoption rates.

**Future Outlook**
If production ramps as planned, Iris is expected to power Meta’s next wave of gener