Meta’s “Iris” AI Chip Enters Production in September, Targeting 14 GW by 2027

Meta Platforms plans to begin manufacturing its custom AI chip, code-named "Iris," in September 2026, according to an internal memo reviewed by Reuters and published on July 13. The chip is the latest generation of Meta's in-house accelerator program and is being built by Taiwan Semiconductor Manufacturing Co (TSMC) with design help from Broadcom. The move is aimed at cutting compute costs and reducing Meta's reliance on Nvidia and AMD as it pushes toward 14 gigawatts of total computing capacity by 2027.

What happened

Detail	Fact
Chip code name	Iris (formally part of the MTIA program)
Production start	September 2026
Manufacturer	Taiwan Semiconductor Manufacturing Co (TSMC)
Design partner	Broadcom
Testing duration	6 weeks, no major issues found
2026 compute target	7 gigawatts
2027 compute target	14 gigawatts
2026 infrastructure spend (projected)	Up to US$145 billion
Release cadence	One new chip roughly every 6 months through 2027

Meta revealed the Iris chip under its technical name back in March 2026 alongside three other AI processors. The chip is the fourth generation of what Meta calls the Meta Training and Inference Accelerators (MTIA) program, a line of custom silicon designed to run the AI that powers Facebook and Instagram.

Iris is not meant to replace Nvidia GPUs entirely. According to the memo, it is designed to work alongside the large volumes of GPUs Meta already buys from Nvidia and AMD. The internal note also acknowledged that integrating the latest GPU generations at Meta’s scale “has been a heavy lift, and it has cost us time.”

How does Meta’s compute build-out break down?

Meta added 1 gigawatt of computing infrastructure in the first half of 2026. It forecasts a further 2.5 gigawatts before the end of the year, reaching 7 gigawatts total for 2026. The plan then doubles that figure to 14 gigawatts in 2027. For context, one gigawatt is enough electricity to power roughly 800,000 homes.

To secure the components needed for this expansion, Meta has signed long-term, multi-year supply agreements with Samsung Electronics for memory chips, Sandisk for flash storage, and Sumitomo Electric for fiber-optic equipment. Those deals matter because a current memory chip shortage has already pushed companies including Apple to raise product prices.

Why it matters

Meta’s chip ambitions sit inside a much bigger industry spending wave. Big Tech is collectively projected to spend more than US$700 billion on AI infrastructure, and Meta’s share alone could hit US$145 billion this year. That level of spending creates pressure on every supplier in the chain and keeps driving what Morgan Stanley analysts have started calling “chipflation.”

For Meta specifically, owning the chip stack means margins improve every time Iris handles a workload that would otherwise run on a purchased Nvidia or AMD GPU. Mike Gualtieri, a vice president and principal analyst at Forrester, put it plainly: “You can’t become an AI titan if you are dependent on another company for chips.”

The six-week testing window is notable. Custom silicon programs at large companies have historically taken years to stabilise. Meta’s MTIA effort had a rough start after launching more than five years ago, so clean test results and a September production date represent a real shift in momentum.

The plan to release a new chip roughly every six months through 2027 is also aggressive. Most AI chip programs run on annual or longer cycles. If Meta can hold that pace, it shortens the window between when a capability is needed and when in-house silicon can deliver it, without waiting on a supplier’s roadmap.

Our take

The Iris story is interesting for what it signals beyond Meta. Custom silicon is no longer just a Google or Apple play. As AI inference (running a trained model to generate outputs) becomes the dominant workload at scale, the economics of renting GPU cycles from Nvidia grow harder to justify. Meta is large enough to absorb the design and manufacturing cost; most companies are not.

For businesses using Meta’s ad platform or AI-powered tools, the practical effect is indirect but real. Lower compute costs for Meta could translate into more headroom for features like Meta Ads automation tools and AI-driven audience targeting, since the underlying inference gets cheaper to run. Whether those savings pass through to advertisers is a separate question.

If you are watching the broader AI infrastructure story, this is a useful data point on how fast custom chip programs can move when a company is willing to spend. Keep an eye on our AI news coverage for updates on Iris milestones and how rival programs at Google and Microsoft compare.

What to do about it

1. If you run Meta ad campaigns, note that Meta’s infrastructure investment is likely to expand its AI targeting and creative tools over the next 12 to 18 months. Review your campaign structure now so you are ready to test new formats as they appear.
2. If you are evaluating AI inference costs for your own products, track the cadence of Meta’s MTIA releases. Chip releases every six months, versus the usual annual cycle, may signal faster price drops in the broader GPU rental market.
3. If you depend on Nvidia GPU availability for any workload, watch how Meta’s supply agreements affect secondary market pricing. Large long-term deals by hyperscalers historically tighten supply for smaller buyers.

The clearest takeaway: custom silicon is now a competitive moat, and companies that can build it are using it to escape supplier pricing power at scale.

Source: Bing News · Meta AI