An internal Meta memo reviewed by Reuters this month contains an unusually candid admission. Adopting the latest GPUs at a company as large as Meta, the memo says, "has been a heavy lift, and it has cost us time."
Reuters reported on July 9, 2026, that Meta plans to start manufacturing its own AI chip, code-named Iris, in September. The chip cleared six weeks of testing without a major issue, according to the memo. Meta declined to comment on the report.
Six weeks is fast. For a chip program that has struggled for years, it counts as a breakthrough.
Iris is one of four planned chip generations under Meta's Training and Inference Accelerators program, an in-house effort known internally as MTIA. It arrives as Meta tries to nearly double its total AI computing capacity, from 7 gigawatts this year to 14 gigawatts in 2027. That figure describes Meta's entire data center power footprint, the electricity and infrastructure behind all of its computing, not a number specific to Iris chips alone.
The stakes are not abstract. One gigawatt of power is roughly enough to run 800,000 homes. Fourteen gigawatts, Meta's target for next year, would power more than 11 million.
The Six-Week Test
Iris is not Meta's first attempt at building its own AI chip, and it has not been an easy one.
Meta first showed its MTIA chips publicly in 2023, comparing their performance against older Nvidia hardware rather than the newest chips on the market at the time. The comparison drew criticism. Reuters described the broader in-house chip effort as one that "has floundered since its launch more than half a decade ago."
Iris is meant to change that story. Meta is working with Broadcom to design the chip and with Taiwan Semiconductor Manufacturing Co., the world's largest contract chipmaker, to manufacture it. Testing uncovered no major issues in six weeks, a pace Reuters called a notable milestone for the program.
The chip is built to run the workloads that keep Facebook and Instagram running: ranking and recommendation systems that decide what appears in a user's feed, plus the generative AI features spreading across Meta's apps. It is designed to work alongside the large volumes of GPUs Meta continues to buy from Nvidia and AMD, not replace them.
Meta unveiled Iris under its technical name on March 11, 2026, as part of a four-chip roadmap. One of its siblings, MTIA 300, is already running Meta's ranking and recommendation systems in production. Less than two weeks after the roadmap reveal, Meta became the first customer for Arm's first in-house chip in 35 years, a data center processor designed to run alongside Meta's own accelerators.
Seven Gigawatts, Then Fourteen
Meta's chip ambitions sit inside a much bigger number: gigawatts of computing capacity, not chips alone.
According to the memo, Meta added 1 gigawatt of computing infrastructure in the first half of 2026 and plans to add roughly 5.5 gigawatts more by year's end, bringing 2026 capacity to about 7 gigawatts. The company wants to double that to 14 gigawatts in 2027.
Reaching that number requires more than chip design. The memo shows Meta has signed long-term supply agreements with Samsung Electronics for memory chips, Sandisk for flash storage, and Sumitomo Electric for fiber-optic equipment. Sandisk declined to comment; Samsung and Sumitomo did not respond to Reuters. The agreements come as memory and chip prices climb fast enough that Morgan Stanley analysts have described the trend as "chipflation."
Meta expects to spend as much as $145 billion on AI infrastructure this year alone.
That is one piece of a much larger bill. Reuters reported that Big Tech as a whole is on pace to spend more than $700 billion chasing AI compute in 2026.
Broadcom's Common Thread
The design partner behind Iris is not exclusive to Meta.
Broadcom has spent the past year becoming the preferred contractor for hyperscalers that want custom chips without building an internal chip-design team from scratch. It is Meta's design partner on Iris. It is also Google's design partner on its Tensor Processing Units, under an agreement that runs through 2031. And on June 24, 2026, just over two weeks before the Iris memo surfaced, OpenAI and Broadcom unveiled the Jalapeno chip, OpenAI's first custom chip, which reportedly went from initial design to a manufacturable blueprint in nine months.
Three of the five major custom silicon programs now running at frontier labs and hyperscalers, Meta, Google, and OpenAI, route through the same design partner. Amazon and Microsoft work with a different one, Marvell, on Trainium and Maia. Every chip in the group, regardless of who designs it, is manufactured by TSMC.
| Company | Custom Chip | Design Partner | Foundry | Status (mid-2026) |
|---|---|---|---|---|
| Meta | Iris (1 of 4 planned MTIA chips) | Broadcom | TSMC | Production starts September 2026 |
| OpenAI | Jalapeno | Broadcom | TSMC | Unveiled June 2026; deployment targeted by year-end |
| TPU (Ironwood, 7th generation) | Broadcom | TSMC | Generally available since late 2025 | |
| Amazon | Trainium 3 | Marvell | TSMC | Generally available since December 2025 |
| Microsoft | Maia 200 | Marvell | TSMC | Running internally since January 2026; not yet broadly available |
For Broadcom, the arrangement pays well. Analysts at Mizuho estimate the company will collect $21 billion in AI-related revenue tied to Anthropic's use of Google TPU capacity in 2026 alone.
That figure is expected to roughly double, to $42 billion, in 2027. For hyperscalers on the other side of these deals, Broadcom offers a shortcut into custom silicon, the same shortcut Meta is now taking with Iris.
The Six-Month Race
Meta's pace is what stands out most. Most chipmakers take a year or longer between major product releases. Meta wants a new MTIA generation roughly every six months through 2027, compressing what is normally a 12- to 24-month cycle into half that window.
Not Everyone Is Convinced
Not every reaction to the Iris news was enthusiastic, and not every analyst thinks it changes much for Nvidia.
Daniel Newman, chief executive of the technology research firm Futurum Group, responded to the Reuters report on X.
"It isn't replacing AMD and NVDA with in-house," Newman wrote. "It is augmenting to meet ambitious capacity requirements and demand expectations."
Nvidia's own stock reflected similar skepticism. Shares dipped only slightly when the memo report broke, then climbed again over the following day, a sign that investors are not pricing in a meaningful threat to Nvidia's core business anytime soon.
That reading lines up with how messy Meta's own stock reaction was on July 9. Shares fell after the memo report circulated, then recovered only after Meta separately announced developer access to a new AI coding model aimed at OpenAI and Anthropic. By late afternoon, shares were trading up 4.6%, but the chip news and the stock gain happened the same day for two different reasons.
Wall Street's patience with Meta's AI spending is not unlimited. In April, Meta's Q1 2026 capex surge triggered a JPMorgan downgrade, and the stock lost roughly 10% of its value in two trading sessions on concerns that returns were not keeping pace with spending. Iris does not resolve that argument. It adds another line to a budget analysts are already scrutinizing.
Meta's own history with in-house chips adds a further layer of caution. The Information reported in February 2026 that Meta had scrapped a more ambitious training chip, code-named Olympus, after its software and supporting equipment proved less stable than Nvidia's and its 2-nanometer design raised the risk of manufacturing problems. Olympus was meant to replace Nvidia's role in Meta's training clusters entirely. Instead, Meta signed a multiyear deal worth several billion dollars to lease Google's TPUs, on top of what it already spends with Nvidia and AMD.
Microsoft offers a live example of how these programs can slip. Its Maia 200 chip is still trying to land Anthropic as an external customer, and it has been running internally in Arizona and Iowa data centers since around January 2026. Delays pushed mass production from an original 2025 target into 2026, and Maia 200 still has not reached general availability for Azure customers broadly.
The Bottom Line
Strip away the roadmap language and the picture is simple. Meta is still buying tens of billions of dollars in Nvidia and AMD chips this year, and Iris will not change that when production starts in September. What changes is that Meta will, for the first time in years, have real production volume of a chip built for its own workloads, tested in six weeks, and designed with a partner that is doing the same job for two of Meta's biggest AI competitors.
The pace is the real signal. Meta wants a new MTIA generation every six months through 2027, about twice the speed of a typical chip program, and it is not racing alone. Google, Amazon, Microsoft, and OpenAI are all running their own versions of the same sprint.
For a company still explaining an April stock swoon and a capex number Wall Street keeps double-checking, September is the month Meta finds out whether six clean weeks of testing hold up at real scale. Reuters got the memo first. Meta has to prove it right, in public, starting this fall.
Meta's plan to depend less on outside chipmakers currently depends on one.
Sources
- Exclusive: Meta to Put AI Chip Into Production in September as It Looks to Double Computing Capacity, Memo Shows — Reuters, via Euronext (July 9, 2026)
- Meta to Put AI Chip Into Production in September as It Looks to Double Computing Capacity, Memo Shows — Global Banking & Finance Review, citing Reuters (July 9, 2026)
- Meta to Put AI Chip Into Production in September, Report — CNBC, citing Reuters (July 9, 2026)
- Meta to Start Manufacturing Its Own AI Chip in September as It Chases 14 Gigawatts of Computing Power — Crypto Briefing (July 9, 2026)
- Meta to Manufacture Its Own AI Chip 'Iris' as AI Infrastructure Ambitions Grow — ITP.net (July 10, 2026)
- Meta to Begin Manufacturing In-House 'Iris' AI Chip in September — MLQ News (July 9, 2026)
- Meta to Start Production of Iris AI Chip in September 2026 — Yahoo Finance / Quartz (July 9, 2026)
- Meta's Custom AI Chip Timeline Just Got More Interesting — TechRepublic (July 9, 2026)
- Daniel Newman Says Meta Isn't Replacing Nvidia or AMD With In-House AI Chips: 'It Is Augmenting…' — Benzinga (July 10, 2026)
- OpenAI and Broadcom Unveil LLM-Optimized Inference Chip — OpenAI (June 24, 2026)
- Broadcom and OpenAI Unveil Custom-Built Jalapeno Inference Processor — Tom's Hardware (June 2026)
- OpenAI and Broadcom Unveil Jalapeno Inference Chip — Let's Data Science (June 2026)
- Broadcom Agrees to Expanded Chip Deals With Google, Anthropic — CNBC (April 6, 2026)
- Broadcom to Supply Anthropic With 3.5 Gigawatts of Google TPU Capacity From 2027 — Tom's Hardware (April 2026)
- META Stock Tumbles Post Q1 Earnings; JPMorgan Downgrades on Heavy AI Investment Plans — TipRanks (April 2026)
- Meta Stock Drops 10% on $145B AI Capex; JPMorgan Downgrades to Neutral — Let's Data Science (April 2026)
- Meta Scraps Advanced AI Training Chip After Design Roadblocks — Dataconomy, citing The Information (February 27, 2026)
- Arm Is Releasing the First In-House Chip in Its 35-Year History — TechCrunch (March 24, 2026)
- Arm AGI CPU Analysis: Meta Buys Arm's First Chip — Let's Data Science (March 2026)
- Anthropic Eyes Microsoft Maia Chips Amid Compute Crunch — Winbuzzer (May 22, 2026)
- Microsoft Maia 200: Anthropic in Talks to Become First External Customer — Let's Data Science (May 2026)