Meta's MTIA 450 (Arke) chips enter testing, aimed at cutting AI costs

Meta's chip push did not start from scratch with Arke. The company is now testing its third-generation MTIA 450 processor, while the next one, MTIA 500 or Astrid, is expected to complete design work in about a month. Meta engineering vice president Yee Jiun Song framed the progression bluntly: each generation "takes on a little bit more risk technologically and gets us better performance," pointing to better performance per watt and per dollar. That language explains why the work is being done inside Meta rather than bought entirely from outside suppliers.

The early signal Meta offers is measured. Twelve Arke chips delivered by TSMC on Sept. 1 performed within 2% to 3% of Meta's simulations and have already run Meta models alongside models from DeepSeek and Alibaba. At the same time, Meta has narrowed its ambitions: it canceled the planned Olympus processor, which was meant to handle both AI training and inference, partly because of cost concerns, and is now prioritizing inference. Song described the resulting parts as "the workhorse chips that we're going to use for general-purpose inference."

The stake is whether custom silicon can materially reduce Meta's AI cost curve. If MTIA deployment lowers power consumption and inference expense at scale, Meta could protect margins while expanding AI across advertising, recommendations, assistants and other products. The risk appears to be execution, since competitive chip design demands enormous capital, long development cycles and tight coordination with Broadcom and TSMC. Meta has committed to deploying more than a gigawatt of the chips over a 12-month period, with deployment expected to accelerate if AI demand remains strong.