Meta Quietly Starts Making Its Own AI Chips—Iris Rollout

They began building chips in secret. Not a dramatic launch. No flashy keynote. Just engineering and production lining up for September.

Internal notes seen by Reuters say Meta will start manufacturing a bespoke AI chip, codenamed Iris, next month. Iris is part of a fourth-generation internal accelerator program called MTIA — Meta Training and Inference Accelerators — and it was designed to shoulder a growing share of the company’s AI workloads across Facebook and Instagram.

Testing moved fast. Six weeks. No major faults flagged. For a project that has weathered years of setbacks, that kind of momentum feels like a small victory and a sign the architecture is finally coherent.

Meta didn't build Iris alone. The design came in-house, but the company is tapping Broadcom for engineering collaboration and TSMC to handle fabrication. The idea is simple: custom silicon tuned to Meta’s models and data flows, so the company can lower the staggering cost and supply risk of running giant GPU fleets.

Iris won’t replace NVIDIA or AMD GPUs. Think of it as complements to the thousands of graphics processors Meta already buys. The chip's role is to absorb specific training and inference tasks at scale, freeing up GPUs for the workloads where they excel.

Ambition here is not modest. Meta plans to bring online roughly 7 gigawatts of compute this year and scale to 14 gigawatts by 2027. To get there, the company has committed to long-term supply deals: Samsung for memory chips, SanDisk for flash storage, and Sumitomo Electric for fiber optics, among others. Those contracts are a hedge against global memory shortages that have squeezed tech supply chains and driven what analysts call "chipflation."

Meta also signaled a faster product cadence than the industry: Iris was unveiled alongside three other AI processors in March, and the company intends to roll out new chips about every six months through 2027, while others typically refresh hardware on a yearly or slower timetable.

Meta expects to spend up to $145 billion on AI infrastructure as it chases scale and cost efficiency.

The move marks a broader shift in how hyperscalers approach hardware: owning more of the stack, from design through supply, to control performance and price. It’s an expensive bet. But when operating at the scale of billions of users and models that demand ever more compute, custom silicon starts to look less like an experiment and more like a necessity.

Will Iris change the economics of AI for Meta? Time will tell. For now, the company is quietly building the tools it thinks will keep its platforms fast, sustainable, and competitive as models grow larger and demand explodes.