Meta will begin mass production of its self-designed artificial intelligence (AI) chip, "Iris," starting this September, accelerating its drive toward AI infrastructure independence. Meta is expected to double its AI computing infrastructure next year and to speed up its entry into the cloud sector.
According to internal documents obtained by Reuters on the 9th, Meta will produce Iris as the first mass-production product of its own AI chip project, the Meta Training and Inference Accelerator (MTIA). Through the in-house chip, co-designed with Broadcom and manufactured by TSMC, Meta aims to reduce its dependence on Nvidia and AMD.
According to the internal documents, Iris showed no major problems during six weeks of validation testing. Reuters reported that this shows Meta's own AI chip development, which had struggled for years, has now gotten on track in earnest. The internal documents also stated that the process of deploying the latest GPUs on a large scale had been "a considerable burden and took a lot of time."
Meta plans to release a new AI chip approximately every six months through 2027. This is a faster pace than the existing industry practice of releasing new products on cycles of a year or more. Mike Gualtieri, vice president at Forrester Research, said, "You can't become an AI powerhouse by relying on other companies for chips," adding, "Hyperscalers and even SpaceX are all planning to produce chips, because it is the only way to secure price competitiveness in the competition over model usage fees."
Meta Jumps Into Cloud… Zuckerberg: "Renting Out Computing Externally Could Make Sense"
Meta is also aggressively expanding its AI computing infrastructure. According to the internal documents, Meta plans to expand its AI computing infrastructure to 7 gigawatts (GW) this year and to double that again to 14GW next year. Meta added 1GW in the first half of this year and plans to secure an additional 5.5GW by year-end. One gigawatt is enough to supply power to about 800,000 households, meaning that once the 14GW infrastructure is secured, the equivalent of power for 11.2 million households will be poured into computing.
The infrastructure expansion is expected to accelerate an expansion in scope in which Meta, once one of the world's largest cloud consumers, will also take on the role of a "supplier" armed with its own infrastructure. CEO Zuckerberg said in an interview with Bloomberg that day, "The offered prices for computing usage are so high that for some it could make more sense to rent it out externally rather than for internal use." Earlier, Bloomberg reported on the 1st that Meta was preparing a new cloud business based on its AI infrastructure.
Amazon Web Services' (AWS) "Bedrock" model is being mentioned as one under consideration. Under this approach, Meta would provide its own AI model, "Muse Spark," as a service on its existing AI infrastructure and collect usage fees from developers.
Meta plans to invest up to $145 billion (about 219 trillion won) in AI infrastructure this year. According to the internal documents obtained by Reuters, the company has signed a memory chip contract with Samsung Electronics to expand its data centers, as well as long-term supply contracts for flash storage with SanDisk and for optical cables with Japan's Sumitomo Electric. Amid intensifying competition to secure supply chains — including Apple raising product prices due to memory shortages — Meta is seen as having laid the groundwork for stable supply.
CEO Zuckerberg dismissed market concerns that Meta might be selling surplus computing capacity resulting from AI overinvestment. He told Bloomberg, "(Renting out computing externally) is not because computing resources are in surplus, and even now we are using all the computing power we have internally."
However, Meta is also known to be considering a "neocloud" business that sells pure computing resources themselves, rather than AI models. On this, CEO Zuckerberg emphasized the possibility of commercial success, saying, "Even if we end up not using all the computing resources internally, there is enough demand to sell it sufficiently through long-term contracts like AWS, Azure, and Google Compute."
He added, "It is very interesting in that short-term contracts are signed at a high premium," saying he is also interested in SpaceX's strategy of renting out the xAI data center in Memphis, Tennessee, to Anthropic and Google.
