Meta Touts the Cost-Saving Benefits of Latest In-House AI Chips

Meta Platforms Inc. plans to begin deploying a new in-house artificial intelligence chip in data centers during the first half of next year, a move it says will save money and energy when running AI models.

The company, which first announced plans to develop homegrown AI chips in 2023, is testing the third generation of the line, called MTIA 450, or Arke. The next one — dubbed 500, or Astrid — will complete design work in about a month and go into data centers at the end of 2027. Meta expects to use that one even more widely.

“Each one takes on a little bit more risk technologically, and gets us better performance,” Yee Jiun Song, Meta’s vice president of engineering, said in an interview. That includes better performance per watt of energy and per dollar spent, he said.

Meta, the owner of Facebook, Instagram and WhatsApp, is developing its own chips as part of a massive push to develop AI infrastructure. The company is working with Broadcom Inc. on the designs and Taiwan Semiconductor Manufacturing Co. on manufacturing — a project aimed at reducing its reliance on Nvidia Corp.’s industry-leading processors.

When measured by energy use — a key benchmark for the data center industry — the company has committed to more than a gigawatt’s worth of the chips over a 12-month period. After that, “we expect it to accelerate,” said Song, who oversees Meta’s custom silicon program. That forecast assumes there’s no crash in the AI markets or in demand, he said.

Meta Superintelligence Labs, the company’s artificial intelligence division, is helping fine-tune the chips by offering insight into coming AI models and the requirements for running them — a stage known as inference. The end result is that the chips can perform more efficiently than “whatever Nvidia is currently shipping” when running AI models, Song said, “just because we’re doing a lot of the engineering ourselves.”

Twelve of the new model chips arrived at Meta from TSMC on Sept. 1 and performance is within 2% or 3% of what the company’s simulations had indicated. On the first day, the team was able to use the processors to operate Meta models, as well as ones from DeepSeek and Alibaba Group Holding Ltd.

The tests indicate there’s no design snags with the semiconductors. But there’s still a lot of trialing and tuning to go — a few months of work — while production ramps up in the factory. All four generations of the chips largely rely on high-bandwidth memory and are not aimed at the ultrafast inference market — a category where AI models are expected to respond extremely quickly. “These are the workhorse chips that we’re going to use for general-purpose inference,” Song said.

Read More: Meta Preparing to Deploy Four New Homegrown Chips to Handle AI

The company had previously planned to build a chip code-named Olympus that was aimed at both the training stage of AI models and the inference phase. It would have arrived in 2028 or 2029, but Meta canceled that project to focus on inference — in part because of cost considerations, Song said. “When you start to build up gigawatts and gigawatts of capacity, you really care about cost,” he said. If the chip that handles both training and inference is potentially 30% more expensive, that suddenly seems “completely unacceptable,” Song said.

After work is complete on Astrid, Meta will focus on improving speed and throughput — the amount of AI work handled — with future models. Fiber-optic technologies also should let the chip team improve 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.”