Meta's custom Iris AI chip enters production at TSMC this September

Meta’s custom AI chip is real, and it enters production in September.

Iris arrives

Per Reuters, citing an internal memo dated July 10, 2026, Meta will start manufacturing its in-house AI accelerator codenamed Iris this September. The chip was designed together with Broadcom and will be fabricated by TSMC, as part of Meta’s MTIA program, which spans four planned chip generations. Testing wrapped in about six weeks with no major issues found.

A six-week test cycle with no major issues is a notably clean result for custom silicon, and it clears the way for the September production start.

The scale is the story

The chip itself matters less than what it is feeding. Meta is targeting 7 gigawatts of computing capacity in 2026, and plans to double that to 14 gigawatts in 2027.

At that scale, the economics of buying merchant GPUs for every rack become punishing. Custom silicon is how Meta gets to those numbers without paying the Nvidia premium on every rack it deploys. Even a modest per-chip saving, multiplied across gigawatts of capacity, changes the cost curve of the entire AI program.

The unanimous hyperscaler playbook

Meta is not running this play alone. Google has TPUs. Amazon has Trainium. OpenAI is working with Broadcom. Anthropic is in talks with Samsung. Every company operating AI infrastructure at hyperscale has reached the same conclusion: own your chips, own your margins.

That does not necessarily mean Nvidia loses. The open question, and the one worth debating, is whether custom accelerators actually dethrone the GPU incumbent or simply trim the bill at the margins while Nvidia keeps the frontier training market. Custom chips tend to excel at well-understood internal workloads, while cutting-edge research still gravitates to the most flexible hardware available.

Either way, September marks the moment Meta’s answer stops being a memo and starts being silicon.