Meta is preparing to manufacture its own AI chip starting in September, part of an aggressive push to cut its dependence on Nvidia and AMD while massively scaling up its computing infrastructure. An internal memo reviewed by Reuters outlines plans to reach 14 gigawatts of total computing capacity by 2027, double what the company expects to have in place by end of 2026.
The chip, code-named “Iris,” is part of Meta's broader Meta Training and Inference Accelerators (MTIA) program, a four-generation in-house silicon project designed to power the AI behind Facebook and Instagram. Bug testing took just six weeks and turned up no major problems, which is notable given that Meta's custom chip program has had a rocky road since it launched more than five years ago.
Meta is working with Broadcom on design and Taiwan Semiconductor Manufacturing Co (TSMC) on manufacturing. The chip is not meant to replace the graphics processing units (GPUs) Meta buys from Nvidia and AMD, but to work alongside them and reduce the company's overall reliance on outside suppliers.
The scale of what Meta is building is hard to overstate. The company plans to deploy seven gigawatts of computing infrastructure in 2026 alone. To put that in context, one gigawatt is enough to power roughly 800,000 homes. Meta added one gigawatt in the first half of this year and expects to add another 5.5 gigawatts before the year is out. That total then doubles to 14 gigawatts in 2027.
To fund this, Meta expects to spend up to $145 billion on AI infrastructure in 2026, a major slice of the more than $700 billion that Big Tech as a whole is projected to spend on AI this year. The memo also reveals Meta has locked in long-term supply agreements to support this buildout:
- Samsung Electronics for memory chips
- Sandisk for flash storage
- Sumitomo Electric for fiber-optic equipment
Those agreements matter because memory chips are in short supply right now. The shortage has already pushed companies like Apple to raise prices, and analysts at Morgan Stanley have flagged “chipflation” as a growing macroeconomic concern. Locking in multi-year deals gives Meta more predictability in a market where component prices have risen fast.
The broader context here is that every major tech company is racing to secure AI computing resources, and those that control their own silicon have a significant cost advantage. “You can't become an AI titan if you are dependent on another company for chips,” said Mike Gualtieri, a vice president and principal analyst at Forrester. “The hyperscalers and even SpaceX all plan chips because it will be the only way to compete on price for model usage.”
Meta is also planning to release new chips at an unusually fast pace, roughly every six months through 2027. Most companies in this space release AI chips once a year or more. That cadence, if Meta can maintain it, gives the company a way to keep up with the rapid pace of AI model development without waiting on outside vendors.
The memo also acknowledged that keeping up with the latest GPU generations from Nvidia and AMD has been difficult internally. “Adopting the latest GPUs at a firm as large as Meta has been a heavy lift, and it has cost us time,” the document stated. That admission makes the push toward custom silicon look less like an optional upgrade and more like a strategic necessity.
Meta declined to comment on the memo. Samsung Electronics and Sumitomo Electric did not respond to requests for comment. Sandisk also declined to comment.
