What Meta is building and when
Meta first laid out its custom AI silicon plans in 2023 and is now preparing to place a new internally developed chip in its data centers in the first half of next year. The company is currently testing the third generation of the family, called MTIA 450 and code-named Arke. The follow-on, MTIA 500, known as Astrid, is about a month away from design completion and is planned to start appearing in Meta's data centers at the close of 2027, with even wider deployment to follow.
How the chips were tested and who is involved
A dozen of the new chips reached Meta from Taiwan Semiconductor Manufacturing Co. on Sept. 1, and their performance landed within 2% to 3% of internal simulation targets. On day one, Meta used the processors to run its own models as well as models from DeepSeek and Alibaba Group Holding Ltd., and early testing indicated no design issues. There is still a few months of trialing and tuning ahead while the factory scales up.
Meta is building this silicon as part of a big push to expand its AI infrastructure. It is collaborating with Broadcom on chip design and with TSMC on manufacturing - a strategy intended to reduce reliance on Nvidia processors. Meta Superintelligence Labs, the AI division, is feeding insights about upcoming models and inference needs to refine the chips. Yee Jiun Song, who leads Meta's custom silicon as its vice president of engineering, said the tight feedback loop lets the hardware run inference for AI models more efficiently than "whatever Nvidia is currently shipping."
When companies invest in new technology, investors benefit from planning for change. Join Briefs Finance CEO Jaspreet Singh on September 29th for a FREE live investor workshop, How to Profit From A Dollar That's Losing its Value, where he shows how we're spotting investment opportunities as the dollar falls. Save your spot.
Why Meta is doing this and what the chips are for
Song said each generation takes on more technical risk and delivers stronger performance, including better output per watt and per dollar spent. Using energy consumption as the yardstick, a key data center metric, Meta intends to roll out in excess of a gigawatt's capacity of these chips across a 12-month period and then accelerate thereafter, provided AI markets and demand don't slump.
All four generations rely heavily on high-bandwidth memory and are not built for the ultrafast inference segment where responses need to be near instantaneous. As Song put it, "These are the workhorse chips that we're going to use for general-purpose inference."
Meta had previously scoped a chip called Olympus that would handle both training and inference and was expected around 2028 or 2029, but the company canceled it to focus on inference, citing cost. "When you start to build up gigawatts and gigawatts of capacity, you really care about cost," Song said, adding that if a combined chip is potentially 30% pricier, that becomes "completely unacceptable." After Astrid, Meta plans to push for higher speed and throughput in future models, with fiber-optic technologies expected to help lift performance. "We have a very robust road map," Song said. "For the next few years, we expect to continue to build chips that are competitive with what our vendors are building for us."
What this means for your portfolio
Meta, which owns Facebook, Instagram and WhatsApp, is shifting from mainly buying third-party processors to building and deploying its own at scale. That could reshape its spending profile and operating efficiency as the chips roll into production and deployment timelines extend through 2027 and beyond. The bigger picture for investors: if Meta's in-house silicon consistently delivers better performance per watt and per dollar for inference, it could lower its long-term AI operating costs while dialing in performance to its own models.
Steady saving and thoughtful choices help your portfolio weather technological shifts. Our CEO Jaspreet Singh is hosting a FREE live investor workshop, How to Profit From A Dollar That's Losing its Value, on September 29th. Sign up free to join him live.
