Overview
Meta Platforms is preparing to begin production of a custom artificial intelligence processor in September, an internal plan shows. The processor, known within the company by the code name "Iris," is a component of a four-generation effort called Meta Training and Inference Accelerators - MTIA - that Meta has been developing in-house to support the AI that underpins its social media services.
Development and partners
The company's memo indicates that testing of the Iris chip required six weeks and did not surface major problems. Meta tailored the design specifically to its operational needs and enlisted Broadcom to assist with design work while contracting Taiwan Semiconductor Manufacturing Co to handle fabrication. The memo notes that those bug-testing results and the intended production start in September had not previously been made public.
The custom silicon is intended to supplement, rather than fully replace, large volumes of graphics processing units that Meta purchases from Nvidia and Advanced Micro Devices for AI workloads. The memo also acknowledged that integrating the latest GPUs at a company of Meta's scale "has been a heavy lift, and it has cost us time."
Product cadence and specifications
Meta revealed Iris by its technical name in March alongside three other AI processors it is developing. According to the memo, the company plans to roll out a new chip approximately every six months through 2027. That cadence is notably faster than the one-year-or-longer intervals more typical in the AI chip market.
Compute build-out and investment
The memo states that Meta intends to deploy seven gigawatts of computing capacity this year and to double that figure to 14 gigawatts in 2027. To support this expansion, the company expects to spend as much as $145 billion on AI infrastructure in the current year. The memo frames that expenditure as a significant portion of more than $700 billion projected to be spent across major technology firms on AI infrastructure.
Supply agreements and procurement
To secure components needed for rapid data center growth, Meta has entered long-term, multi-year supply agreements with several vendors. The memo lists agreements with Samsung Electronics for memory chips, Sandisk for flash storage and Sumitomo Electric for fiber-optic equipment. The memo indicates these long-term arrangements are intended to insulate Meta's build-out from the strains seen in the memory market.
Sandisk declined to comment on the agreement. Samsung Electronics and Sumitomo Electric did not respond to requests for comment, according to the memo.
Market context and pricing pressure
The memo highlights that components such as memory and AI chips have seen surging demand as companies race to scale data centers to meet AI's large computing needs. That heightened demand has pushed up prices for memory and other chips quickly and substantially. The memo cites the term "chipflation" as a macroeconomic concern, quoting Morgan Stanley analysts on the trend.
Implications and positioning
Meta's strategy is to reduce dependence on external chip suppliers by building specialized processors tailored to its internal workloads while continuing to use industry-standard GPUs. The file of in-house chips and multi-year supply contracts are presented in the memo as steps toward controlling costs and ensuring component availability during a period of rapid capacity expansion.
Key points
- Meta plans to begin producing its Iris AI chip in September and to double data center computing to 14 gigawatts by 2027.
- The Iris design passed six weeks of testing without major issues; Broadcom is assisting design and TSMC will manufacture the chip.
- Meta expects to spend up to $145 billion on AI infrastructure this year and has secured long-term supply agreements for memory, flash and fiber-optic components.
Risks and uncertainties
- Supply chain constraints - Long-term agreements aim to mitigate shortages for memory and other components, which could otherwise delay data center expansion. Affected sectors include semiconductors and data center infrastructure.
- Rising component prices - Accelerating demand for memory and AI chips has driven price increases, a phenomenon described as "chipflation," creating cost pressure for companies expanding compute capacity. This impacts technology capital spending and hardware vendors.
- Integration challenges - The memo notes that adopting the newest GPUs at a company of Meta's scale "has been a heavy lift, and it has cost us time," signaling execution risks in scaling and integrating diverse hardware into large data center operations. Operational and cloud infrastructure sectors are implicated.