Meta Platforms will begin mass production in September of AI chips it designed in-house. Alongside large-scale data center investments, the company is also accelerating moves toward AI Semiconductor self-sufficiency.
On the 9th (local time), Reuters reported this, citing an internal Meta memo. In March, Meta unveiled four in-house AI chips, including the MTIA 400, known by the codename "Iris." The MTIA 400 is the core chip of the fourth-generation Meta Training and Inference Accelerator (MTIA) project under development, with Broadcom handling design and Taiwan's TSMC manufacturing it.
It is a strategy to reduce dependence on Nvidia and AMD and cut AI computing expense. Chip testing was completed in six weeks, and no critical defects were reported. Meta is also sharply shortening the semiconductor development cycle. The internal memo said securing the latest GPUs was "very difficult and time-consuming," and noted the company plans to reduce the product launch cycle, which typically took more than a year, to around six months to roll out new products through 2027 in order to reduce reliance on external supply chains.
This push for semiconductor self-sufficiency is tied to aggressive AI infrastructure investments. Meta plans to expand computing infrastructure to a total of 7 gigawatts (GW) this year. It built 1 GW in the first half and aims to add 5.5 GW by year-end to meet this year's target. Next year, it plans to add another 7 GW, increasing total computing capacity to 14 GW.
Focusing on AI infrastructure, Meta set capital expenditures (CAPEX) this year at up to $145 billion (about 220 trillion won). That is about 20% of the $700 billion full-year CAPEX forecast for Big Tech overall.
To support large-scale data center expansion, it also secured key component supply chains in advance. Meta said it has signed long-term supply agreements with Samsung Electronics for memory semiconductors, SanDisk for flash storage, and Sumitomo Electric for the optical fiber equipment institutional sector. The move is aimed at preparing for a memory shortage as AI investment expands.
Morgan Stanley recently labeled the surge in memory prices as "chipflation," pointing to it as a new macroeconomic variable. Mike Gualtieri, a vice president at the research firm Forrester, said, "You cannot become an AI giant while relying on other companies for chips."
Meta is moving to strengthen not only hardware but also AI software competitiveness. That day, it released the AI coding model "Muse Spark 1.1" to developers as a paid API to take on OpenAI and Anthropic. The model was opened for the first time in a public preview, after previously being available only to select partners. It is also the first case of offering an in-house AI model as a paid service, departing from the previous open-source-centered "Llama" strategy.
Meta is also set to launch the image-generation model "Muse Image" this week, rapidly expanding its AI service portfolio and moving in earnest into AI platform competition against OpenAI and Anthropic.