Meta Is Quietly Becoming One of the World's Biggest Chipmakers

Meta plans to begin production of its new AI chip, Iris, in September, aiming to reduce reliance on Nvidia GPUs. The company targets 14 gigawatts of AI computing capacity by 2027 and has raised capital expenditure guidance to $125-145 billion for 2026. Meta joins Google, Amazon, and Microsoft in developing custom AI chips to control costs and infrastructure.

Meta is preparing to put its own AI chip into production in September, and the important part is not the chip alone. It is the bill Meta is trying to stop Nvidia from writing forever. According to an internal memo reviewed by Reuters, Meta plans to begin production in September of a new AI chip codenamed Iris, built with Broadcom and manufactured by Taiwan Semiconductor Manufacturing Co. Reuters reported on July 9 that the chip cleared its bug-testing phase in about six weeks and turned up no significant issues. For a custom silicon project at this scale, that is the result you want. Iris isn't a one-off. It is the latest entry in Meta's MTIA program, short for Meta Training and Inference Accelerator, and Reuters said it is one of four chip generations the company has planned. Meta wants to ship a new version roughly every six months through 2027. That is an aggressive cadence for custom silicon, and it tells you Mark Zuckerberg isn't treating this as a side project anymore. The stakes are the computing capacity behind it. Meta is targeting 7 gigawatts of AI computing capacity in 2026 and plans to double that to 14 gigawatts in 2027, according to the memo cited by Reuters. Reaching that number with Nvidia GPUs alone would mean writing an even bigger check to Jensen Huang's company than Meta already does. Meta already spends enormously. The company raised its 2026 capital expenditure guidance to between 125 billion and 145 billion dollars, nearly double the 72 billion dollars it spent in 2025. Zuckerberg has said some of that jump comes from rising memory prices, not only GPU demand, but the larger point still holds. AI infrastructure is becoming the largest strategic cost in the company. Frankly, that math is the whole story here. Nvidia GPUs remain the fastest way to train and run frontier models, but they are also expensive, scarce and controlled by a supplier with enormous pricing power. Every dollar Meta can shift onto its own hardware is a dollar it keeps out of Nvidia's margin. The memo frames Iris as a supplement to the Nvidia and AMD chips Meta already buys, not a replacement. Read that as careful phrasing. The direction is obvious. Meta Is Joining a Crowded Field Google has been building its TPUs for over a decade. Amazon has Trainium. Microsoft has Maia. Meta is not early to this fight, but it is arriving with a huge spending plan and a much clearer reason to care than it had during earlier chip efforts. You do not need to believe Meta will beat Nvidia at Nvidia's own game to see why this matters. Meta only needs enough workloads to run better, or cheaper, on its own silicon. The cloud angle makes the chip work more important. Bloomberg reported earlier this month that Meta has been exploring ways to sell excess AI computing capacity to outside customers, and MarketWatch noted that the report helped fuel a sharp jump in Meta's stock. Zuckerberg has also said renting out spare compute is on the table if the company builds more capacity than it needs. That is not the same as an official AWS-style launch, and the distinction matters. But you do not build a serious compute business on hardware whose cost structure you cannot influence. Broadcom is the quiet winner in all of this. Reuters reported that Broadcom worked with Meta on Iris, and the chip company already has similar custom silicon relationships with major cloud operators. Every time a hyperscaler decides to design its own accelerator instead of simply buying off Nvidia's shelf, Broadcom tends to be somewhere in the room. That is why investors have treated Broadcom as one of the cleaner ways to bet on custom AI chips without betting against Nvidia outright. None of this means Nvidia is in trouble. Meta's own memo says Iris supplements, it does not replace. Nvidia's data center revenue is still climbing, and no major AI company is walking away from GPUs soon, because CUDA remains one of the strongest moats in the industry. But revenue and pricing power are different things. If Meta, Google, Amazon and Microsoft can move even a meaningful slice of inference or internal workloads onto in-house chips, Nvidia can keep growing and still face more pressure at the margin. What To Watch Next The next test is not whether Iris exists. Reuters says it passed bug testing. The real test is whether it deploys cleanly after September production begins and whether Meta can keep the six-month rhythm it has set for the next MTIA generations. Chip roadmaps are easy to announce and hard to keep, especially when the customer is also the designer, the buyer and the company trying to run the models on top. You should also watch the 14 gigawatt target for 2027. That number is bigger than a chip story. It points to power contracts, data center buildouts, memory supply and the uncomfortable fact that AI ambition now runs straight into the physical limits of electricity and cooling. A successful Iris chip will not solve all of that. It will give Meta one more part of the stack it can control. That is why this story matters. Meta is still a massive Nvidia customer today, and it will be one tomorrow. But if Iris works, Meta starts looking less like a company buying AI infrastructure from the outside and more like one assembling the whole machine itself. Also read: Robinhood Lets AI Agents Trade Your Crypto Around the Clock Now https://startupfortune.com/robinhood-lets-ai-agents-trade-your-crypto-around-the-clock-now/ • Communities Are Turning Against AI Data Centers and Washington Is Starting to Notice https://startupfortune.com/communities-are-turning-against-ai-data-centers-and-washington-is-starting-to-notice/ • Most Americans now want the government to seize half of AI companies' equity https://startupfortune.com/most-americans-now-want-the-government-to-seize-half-of-ai-companies-equity/