The owner of Facebook has agreed to buy $60bn (£44.5bn) of artificial intelligence chips from the US semiconductor company Advanced Micro Devices despite fears over the vast sums being spent on the AI industry.
Meta, which also owns Instagram and WhatsApp, has clinched the five-year deal in which it will also buy 10% of the chip company.
AMD signed a similar pact with OpenAI last year, which was hailed as a vote of confidence in its chips and software, significantly boosting its stock price.
A recent series of chip supply agreements underscores the AI industry’s appetite for processors. Meta has separately struck a deal with AMD’s larger rival Nvidia to buy millions of AI chips.
AMD would supply 6GW worth of chips to Meta, starting with 1GW of the company’s forthcoming MI450 hardware in the second half of this year, said AMD’s chief executive, Lisa Su.
In addition to AMD’s flagship graphics chips (GPUs), Meta also plans to buy central processors (CPUs), including a variant that will be customised for the social media platform’s needs.
The custom CPU would be tuned to deliver powerful performance while keeping energy consumption as low as possible, Su said. The deal will include two generations of AMD’s CPUs.
“So no question Mark is very, very ambitious in what he wants to accomplish, and we want to use every aspect of our technology to really help Meta to accomplish that,” Su said, referring to Meta’s chief executive, Mark Zuckerberg. “Meta is making a big bet on AMD.”
Meta contributed to the MI450 design, which is optimised for a computing process known as inference – when a chatbot such as OpenAI’s ChatGPT responds to a user’s queries.
Industry analysts expect the market for inference hardware to dwarf the size of the market for the equipment needed to train large AI models.
Meta planned to continue to buy chips from other vendors and develop its in-house processors at the same time, said Santosh Janardhan, the company’s infrastructure head. Meta has been in discussions with Google about using the company’s tensor processors (TPUs) for AI work, Reuters reported.
The scale at which Meta was building datacentres and infrastructure required multiple chip vendors and approaches, Janardhan said.
“All of the chipmakers end up having sort of a seat at the table,” Janardhan added.
