Meta's in-house Iris AI chip clears six-week test, enters production in September

Meta's first in-house AI accelerator chip, code-named Iris, has completed roughly six weeks of testing with no major issues found, according to an internal memo reviewed by Reuters. Production begins in September 2026, marking Meta's most concrete step yet toward reducing its dependence on Nvidia and AMD for the silicon powering its AI systems.

The chip's arrival lands at a pivotal moment for the company. Meta has raised its 2026 capital expenditure forecast to a range of $125 billion to $145 billion, with first-quarter capex alone reaching $19.8 billion — spending levels that make custom silicon an increasingly obvious way to control costs at scale rather than paying premium prices for third-party GPUs indefinitely.

Broadcom designed it, TSMC builds it

Broadcom assisted with the Iris chip's design and architecture, while Taiwan Semiconductor Manufacturing Company will handle fabrication — the same division-of-labor arrangement several hyperscalers now use to develop custom accelerators without building in-house fabs. Meta intends to ship new chip generations roughly every six months through 2027, a cadence notably faster than the industry-standard annual refresh cycle for AI accelerators.

The capacity math driving the urgency

Meta plans to bring approximately 7 gigawatts of computing capacity online in 2026, then double that to 14 gigawatts in 2027. That scale of build-out is difficult to sustain on GPU supply alone, given how constrained Nvidia and AMD allocations have become industry-wide — custom silicon gives Meta a second supply line that isn't subject to the same allocation queues as its competitors.

Crucially, Meta has framed Iris as a complement to third-party GPUs rather than a replacement. CEO Mark Zuckerberg described the quarter as having "strong momentum" across the company's platforms, tying the infrastructure buildout directly to broader AI ambitions rather than framing it as a cost-cutting measure in isolation.

The bigger picture: Meta Compute

The Iris chip news follows Meta's early-July announcement of Meta Compute, a new cloud business that will sell surplus AI computing capacity to outside customers on two layers: a hosted model service comparable to Amazon Bedrock, offering API access to Meta's Muse Spark model family, and raw GPU infrastructure-as-a-service that competes directly with CoreWeave and Nebius on price and availability. That announcement sent CoreWeave down 14% and Nebius down 17% in a single trading session, with the broader semiconductor sector — including Micron, AMD, Intel, Samsung, and SK Hynix — also feeling the selloff pressure.

Custom silicon like Iris is what makes the Meta Compute business model viable at scale: chips built and depreciated for Meta's own workloads generate a lower marginal cost per unit of compute than renting equivalent GPU capacity from Nvidia, which matters directly if Meta intends to underprice CoreWeave and Nebius as external cloud competitors rather than just support internal AI training. As reported by The Motley Fool, citing Reuters' review of the internal memo.