Meta plans to start putting a new in-house AI chip into its data centers in the first half of next year, betting that custom silicon will make running AI models cheaper as it builds out gigawatts of new capacity.
The chip, MTIA 450, code-named Arke, is the third generation of Meta’s custom accelerator line and is now in testing, Bloomberg reported. The next one, MTIA 500 or Astrid, was set to finish design work in about a month and enter data centers at the end of 2027, and Meta expects to use it more widely. Meta designs the chips with Broadcom and has them made by TSMC.
“Each one takes on a little bit more risk technologically, and gets us better performance,” Yee Jiun Song, Meta’s vice president of engineering, told Bloomberg, as quoted by Quartz. He said that covers performance per watt and per dollar, and that the chips run AI models more efficiently than Nvidia’s current offerings. Meta has committed to more than a gigawatt’s worth of the chips in any 12-month period, and Song said he expects that pace to accelerate.
Inference first, on cost grounds
Twelve Arke units arrived from TSMC on Sept. 1, with performance within 2% to 3% of Meta’s simulations, and on the first day the team ran Meta’s own models plus ones from DeepSeek and Alibaba. The chips target general-purpose inference rather than ultrafast response times. Meta scrapped a chip code-named Olympus, which would have handled both training and inference, because of cost. “When you start to build up gigawatts and gigawatts of capacity, you really care about cost,” Song said, adding that a dual-purpose chip might cost 30% more, which he called “completely unacceptable” at that scale.
Separately, Meta planned to start manufacturing another MTIA chip, code-named Iris, in September as part of a push to raise overall computing power to 14 gigawatts next year, according to an internal memo reviewed by Reuters in July. The custom chips supplement, rather than replace, the GPUs Meta buys from Nvidia and AMD.
Where the capacity goes
The largest site is in Richland Parish, Louisiana. Meta said in July it would expand the campus to 5 gigawatts of compute capacity, up from more than 2 gigawatts, with investment rising past $50 billion, according to Meta’s announcement. The site will house Hyperion, Meta’s largest AI training cluster. The AI Data Center Index tracks 20 Meta facilities in five countries with 15.8 gigawatts of capacity, half of them operational.
WebProNews is an iEntry Publication