"Meta signs AI chip deal with Google after Nvidia and AMD…in-house chip efforts hit hurdles"

Meta, which has joined the race to develop superintelligence that surpasses humans, has signed a deal to bring in artificial intelligence (AI) chips from Google, following agreements with Nvidia and AMD.

The Information, a U.S. IT-focused outlet, reported on the 26th (local time), citing sources, that Meta has signed a multiyear contract worth several billion dollars to lease Google’s AI chips, Tensor Processing Units (TPUs). In addition to this compute-leasing agreement, Meta is also in talks over a purchase deal to directly procure and install Google TPUs in its own data centers.

Meta previously signed an AI chip supply agreement with Nvidia on the 17th, estimated to be worth tens of billions of dollars. On the 24th, it also entered into a partnership with AMD and decided to receive chip supplies worth $100 billion.

Meta has been rolling out large-scale AI chip deals in quick succession. This is seen as a response to its long-standing plan to develop its own chips running into obstacles.

Meta had pursued a plan to develop cutting-edge in-house AI chips under the name “Meta Training and Inference Accelerator” (MTIA), but halted it. The Information said, citing sources, that Meta has shifted its development direction to a simpler version.

Meta canceled the development of an AI training chip it had been pursuing under the codename “Olympus.” It also scrapped one version under its other chip plan, “Iris.”

The decision was recently communicated to the AI infrastructure division as well. Meta had originally planned to develop the Olympus chip, deploy it in large server clusters, and use it for AI training. However, executives reportedly judged that proceeding with the plan could pose significant risks in the race to develop AI models against OpenAI and Google. This was because the related equipment and software were not as stable as Nvidia’s products, and the chip design was complex, raising expectations of difficulties in mass production.

An official working in Meta’s in-house chip development unit said, “Given the risks of development delays or redesigns, there is substantial internal skepticism about whether we can build a chip that rivals Nvidia’s performance.”

In a statement, a Meta spokesperson said, “We are keeping our commitment to continue investing in a diverse silicon portfolio, including advancing the MTIA portfolio,” adding, “We will share more later this year.” Meta itself expects to spend $135 billion (about 194 trillion won) this year alone on AI infrastructure investment.

Separately, Google plans to step up its TPU leasing business by establishing joint ventures with major investment firms, in addition to Meta.

Park Subin, Hankyung.com reporter waterbean@hankyung.com