Report: Expecting bigger AI workloads, Meta tests out in-house training chip

Like the spaghetti dish you finally got around to making, AI chips might be better when they’re homemade.

Two anonymous sources told Reuters last week that Meta is testing out an in-house chip for training AI models—another investment from a company that’s betting we’ll all find a reason to use artificial intelligence.

“If you know that a workload is here to stay, then it really opens the door for incurring some of the large costs and some of the engineering efforts required to build a chip for that workload,” Benjamin Lee, professor in the department of computer and information science at the University of Pennsylvania, told IT Brew.

Interest in GenAI from the enterprise market dipped in 2024, according to a recent Deloitte survey—in Q4, 59% of global C-suite and executive leaders and 46% of board members had high or very high interest in the technology. In Q1, those enthusiasm numbers were 74% and 62%, respectively.

“We don’t know what the killer app is for generative AI. We don’t know what will suddenly trigger broad, widespread adoption of this workload, but they want to be ready,” Lee said of Meta.

Spending spree. In January 2025, Meta CEO Mark Zuckerberg announced plans to invest up to $65 billion as the company expands GenAI efforts, including a planned Manhattan-sized data center and a reported standalone app.

Just a year before, in 2024, Zuckerberg took to Instagram to share that the company was training its large language model “Llama 3”—an effort requiring some percentage of the 350,000 Nvidia Tensor Core GPUs for the year (as well as almost as many “other GPUs”).

More GPUs, more problems. “As we increase the number of GPUs in a job, the likelihood of an interruption due to a hardware failure also increases,” Meta’s engineers wrote in a June 2024 blog post.

GPUs also use resources to constantly move data into and out of a chip, especially as models expand, Lee told us.

One Reuters source revealed that Meta’s new training chips are designed for AI-specific tasks, and that the “dedicated accelerator,” part of the company’s ​​Meta Training and Inference Accelerator (MTIA) series, is more power efficient than the integrated GPUs generally used for AI workloads.

“They want the accelerator, and I think they want an alternative to GPUs,” Lee said.

Meta recently partnered with Taiwan-based chip manufacturer TSMC, and the test deployment began after Meta successfully completed a “tape-out” of the chip—“a significant marker of success in silicon development work that involves sending an initial design through a chip factory.”

That kind of effort typically costs tens of millions of dollars and takes up to six months, Reuters said in its March 11 story.

DIY. The need for thousands of AI chips has led Big Tech vendors like Google, Microsoft, and Amazon to develop more in-house hardware options.

Meta “may be able to reduce cost because they incur this one time fixed cost of designing a chip and building up an engineering team and expertise, but then they have these smaller marginal costs as they get more and more chips fabricated for them by TSMC,” Lee said.