For years, Meta has been one of Nvidia's biggest customers. That's changing. Starting September, the company will begin mass-producing Iris, its first in-house AI accelerator — designed with Broadcom and manufactured at TSMC facilities.
Details emerged from an internal memo obtained by Reuters. Iris completed its testing phase in just six weeks with no major issues — an unusually fast turnaround for a chip of this complexity. The project sits within Meta's MTIA program (Meta Training and Inference Accelerators), a four-generation custom silicon roadmap that has been quietly underway for years.
The scale of Meta's ambitions is striking. The company wants 7 gigawatts of computing capacity online by end of 2026 and plans to double that to 14 GW by 2027. AI infrastructure spending could hit $145 billion this year alone. Iris is designed to handle the recommendation and ranking algorithms powering Instagram and Facebook feeds, along with generative AI tasks across Meta's growing product lineup.
Amazon built Trainium, Google has its TPUs, Microsoft developed Maia. Meta was late to custom silicon — but the six-week test cycle suggests it's moving fast. Having your own chip means less dependency on a single supplier, better optimization for specific workloads, and more control over how you scale.
How Iris performs against Nvidia's latest accelerators in real production loads is the open question. Those numbers won't arrive until 2027 quarterly results. For now, a clean six-week debut is a solid starting point.
