Meta’s Iris AI Chip Enters Production in September, Tests Clean

Meta Platforms will start manufacturing its in-house Iris AI chip in September, after a six-week bug-testing cycle that found no major issues, according to an internal memo reviewed by Reuters. Iris is the latest chip in Meta’s four-generation Meta Training and Inference Accelerator (MTIA) project, designed in-house to power AI behind Facebook and Instagram and ease the company’s dependence on Nvidia and AMD. The chip is meant to augment the large volumes of Nvidia and AMD graphics processing units (GPUs) Meta already runs.

The September production start is the first tangible sign of momentum in a custom-silicon program the same memo says has “floundered since its launch more than half a decade ago.” Meta plans to double overall computing power to 14 gigawatts by 2027 and expects to spend as much as $145 billion on AI infrastructure this year, a figure the memo calls a significant portion of Big Tech’s more than $700 billion projected outlay on the technology. Working with Broadcom on design and Taiwan Semiconductor Manufacturing Co (TSMC) on manufacturing, Meta tailored the chip for its own workloads rather than chasing Nvidia’s general-purpose GPU market.

What the Memo Says About Iris

The memo, reviewed by Reuters and reported on July 9, 2026, puts Iris on a clear production clock. Manufacturing begins in September. Testing took only six weeks and found no major issues. Meta declined to comment on the memo.

Iris sits inside a four-generation MTIA roadmap that Meta will design in-house. The chip is tailored to Meta’s own needs and will run the AI behind Facebook and Instagram. Per the memo, the strategic aim is to lower Meta’s massive computing costs and gain more independence from chip suppliers such as Nvidia and AMD.

The chip is meant to augment the GPUs Meta already buys from Nvidia and AMD. Even with Iris on the line, Meta is not slowing its GPU orders.

Six Weeks of Testing, Half a Decade in the Making

The quick test cycle is the headline number. Six weeks of bug testing and no major issues signals “positive momentum,” the memo says, for a program that has spent more than half a decade trying to find its footing. The clean test also lands while Meta is still struggling to absorb the latest generation of commercial GPUs.

Meta launched the Meta Training and Inference Accelerator project in 2023 to build custom silicon for AI ranking and recommendation workloads, the dominant Meta workload before generative AI took off, according to the company’s own blog. The broader in-house silicon effort, the memo notes, has “floundered since its launch more than half a decade ago.” Adopting the latest GPUs at Meta’s scale “has been a heavy lift, and it has cost us time,” the memo adds. That framing, building a hedge against a supplier base that has slowed Meta down, is the subtext of the September production start.

Iris is the first chip in that long build to test cleanly. Whether the program can sustain a six-month cadence is the open question.

The pace matters because Meta’s broader AI bill keeps climbing. The company rolled out its first in-house AI image generator, Muse Image, across Instagram, WhatsApp, and the Meta AI app earlier in 2026, the same year Meta’s overall AI spend is set to hit $145 billion, per a Meta Muse Image rollout and the $145B AI bet.

Four MTIA Generations Land in Two Years

In March 2026, Meta unveiled Iris under its technical name alongside three other AI processors, per the company’s own blog. The full MTIA lineup, MTIA 300, 400, 450, and 500, is set to drop into Meta’s data centers on roughly six-month intervals through 2027.

Source: Meta’s technical breakdown of MTIA 300 through 500. The cadence is faster than the industry’s typical one-to-two year release window, and the design is modular: MTIA 400, 450, and 500 use the same chassis, rack, and network infrastructure, so each new generation drops into the same physical footprint. From MTIA 300 to MTIA 500, high-bandwidth memory bandwidth rises 4.5x and compute FLOPS rises 25x, per the same post.

The pitch to investors, in Mark Zuckerberg’s words, is that the chips will “deliver personal superintelligence to billions of people” across WhatsApp, Instagram, and Threads, per a Broadcom and Meta multi-gigawatt MTIA partnership announcement. The pitch to Meta’s own finance team is the more immediate one: lower cost per inference.

Meta’s 14-Gigawatt Build Carries a $145 Billion Price Tag

The chip is one piece of a much larger capital plan. Meta plans to deploy seven gigawatts of computing infrastructure this year, having added one gigawatt in the first half of the year and forecasting another 5.5 gigawatts by the end of 2026, per the memo. One gigawatt of energy is enough to power about 800,000 homes. The seven-gigawatt 2026 total is the floor: Meta plans to double that to 14 gigawatts in 2027.

• 7 gigawatts of compute planned for 2026
• 14 gigawatts targeted for 2027
• $145 billion in AI infrastructure spend this year
• $700 billion+ in projected Big Tech AI outlays
• 1 gigawatt = enough power for about 800,000 homes

“The company plans to double capacity again next year to reach a total of 14 gigawatts in 2027,” the memo says. The $145 billion 2026 spend is “a significant portion of Big Tech’s more than $700 billion projected outlay on the technology.”

That build is the reason Meta is locking in multi-year supply deals rather than buying spot. To expand computing infrastructure, Meta has secured long-term, multi-year supply agreements, the memo shows, with Samsung Electronics for memory chips, Sandisk for flash storage, and Sumitomo Electric for fiber-optic equipment. The infrastructure is taking shape on the ground, too: Meta has committed CAD $13 billion to its first Canadian AI data center in Sturgeon County, Alberta, with a 932 MW private gas plant built specifically for the site, per a Meta’s $13B Alberta data center with 932 MW gas plant report.

The chip, the gigawatts, and the long-term supply contracts all point at the same strategy: build where the cost curve is, buy where it isn’t. Meta’s March 2026 roadmap of four MTIA chip generations through 2027 is the in-house half of that playbook.

Iris Puts Quiet Pressure on Nvidia and AMD

Meta’s MTIA chips are “specifically designed for our workloads” and “achieve greater compute efficiency than general use chips for our intended purposes, making MTIA much more cost efficient,” per the company’s March blog post. The first generations were tuned for ranking and recommendation, the workload that paid Meta’s bills before generative AI. The 450 and 500 generations pivot to GenAI inference, the workload driving the 14-gigawatt build.

The chip is meant to augment the large quantities of GPUs Meta already buys, the memo says. Iris will not show up in a reduced Nvidia or AMD order this year. It will show up in a slower-rising one a few years out.

You can’t become an AI titan if you are dependent on another company for chips. The hyperscalers and even SpaceX all plan chips because it will be the only way to compete on price for model usage.

That is Mike Gualtieri, a vice president and principal analyst at research firm Forrester, speaking to Reuters. Every inference workload Meta moves onto MTIA is a workload that does not need a new Nvidia or AMD GPU. The Broadcom partnership, an initial 1 GW deployment of MTIA silicon, is the first phase of a multi-gigawatt rollout that runs through 2029. The memo’s own framing, that GPU adoption has been “a heavy lift,” telegraphs the direction.

Chipflation and the Memory Lock-Ins Behind the Build

The chip is one bottleneck. Memory is another. Memory and other chip prices have risen rapidly and substantially enough that “chipflation” has become a macroeconomic concern, Morgan Stanley analysts said, a phrase now used to describe price pressure spreading from data centers into the broader economy. AI-driven demand is the trigger.

• Samsung Electronics: multi-year memory chip supply agreement
• Sandisk: multi-year flash storage supply agreement
• Sumitomo Electric: multi-year fiber-optic equipment supply agreement

The Reuters memo lists three long-term supply deals Meta has signed to keep the 14-gigawatt build on track. “Such long-term agreements have become critical for data center expansion targets amid a memory chip shortage that has prompted companies such as Apple to raise prices,” the report says. Sandisk declined to comment on the memo; Samsung Electronics and Sumitomo Electric did not respond to requests for comment. For Meta, the contracts are the insurance policy: even a clean Iris test does not matter if the high-bandwidth memory that wraps around it is unavailable. Apple’s price hike shows the supply squeeze is already hitting consumer devices, not just data centers.

Morgan Stanley’s “chipflation” warning is a signal these contracts are no longer optional. They are the cost of admission to the next gigawatt.

The Market’s Split Read on Meta’s AI Bet

Shares fell after the Reuters story broke, then recovered after Meta announced developer access to an AI coding model positioned against OpenAI and Anthropic. The stock was trading up 4.6% in late afternoon. Meta declined to comment on the memo.

The split reaction, punished for the spend, rewarded for the model, captures how Wall Street is currently pricing Meta’s AI strategy. Investors are weighing a $145 billion 2026 build against a chip that has yet to ship, a six-week test cycle that has yet to scale, and a multi-year supply chain that has yet to clear the memory crunch. The first MTIA silicon rolls off the line in September. The bill for the strategy runs through 2027.