Meta has agreed a multiyear deal to buy millions of Nvidia chips worth billions of dollars, as the AI data centre boom turns into a knife fight.
The renewal and expansion come as Nvidia tries to stay on top while rivals such as AMD push harder, and Big Tech customers, including Meta, crank out in-house silicon.
Creative Strategies chief executive and principal analyst Ben Bajarin estimated the deal, announced yesterday, would be worth billions of dollars.
Meta chief Mark Zuckerberg last month said the Facebook and WhatsApp owner would nearly double AI infrastructure spending to as much as $135bn this year.
Nvidia has been the main winner of three years of AI infrastructure splurging, but its customers are getting serious about cutting dependency on its kit.
Google, Amazon and Microsoft have all unveiled new in-house chips in recent months, while OpenAI has co-developed a chip with Broadcom and struck a significant deal with AMD.
Meta has invested in developing several AI chips in-house, with Zuckerberg highlighting lower computing costs from processors tuned to its “unique workloads”.
That chip push has hit technical snags and rollout delays, according to a person familiar with the matter, so Meta is still writing cheques to Santa Clara.
Meta has now committed to Nvidia’s next-generation “Vera Rubin” chips, and it is the first Big Tech outfit to say it will buy standalone “central processing units” from Nvidia to run its AI models.
Nvidia is selling CPUs separately, rather than only as part of an integrated bundle with its graphics processing units, a shift in how it tries to own the whole rack.
Graphics processing units handle the brute-force parallel work needed to train the biggest models, but the industry is tilting towards “inference”, meaning running them at scale.
Creative Strategies chief executive and principal analyst Ben Bajarin said “The question of why Meta are deploying Nvidia’s CPUs at scale is the most interesting thing in this announcement. We were in the ‘training’ era, and now we are moving more to the ‘inference era’, which demands a completely different approach.”
Start-ups are pitching inference-specialised chips as an alternative to Nvidia, and in December Nvidia hired talent from inference chip company Groq as it tried to shore up its position.
Nvidia shares dipped late last year after The Information reported Meta was talking to Google about buying Tensor Processing Units, although no such deal has been announced.
The deal lands just before Nvidia’s earnings next week, with investors getting twitchy about whether this spending spree can keep going without cracking credit nerves.
