Nvidia inks order for 16,000 AI GPUs worth $500 million — Indian data center company seeks to own 32,000 Nvidia H100 and GH200 GPUs by 2025
India is an essential market for Nvidia now that U.S. export restrictions block off China.
According to a report from Reuters, Yotta, a data center and server company based in India, is set to buy 16,000 Nvidia GPUs worth $500 million by March 2025. This deal is coming soon after U.S. sanctions banned the export of many of Nvidia's top-end GPUs to China, and it seems India is at least partially filling the void where Chinese demand used to be.
Reuters talked to Yotta CEO Sunil Gupta, who said the company would order 16,000 H100 and GH200 GPUs by March 2025. These GPUs are estimated to be worth around $500 million altogether, though it's unclear how many Yotta is ordering. In 2022, the price of one H100 GPU was over $30,000, but rough math implies $500 million would be enough to buy several thousand H100 chips in addition to GH200, which is presumably significantly more expensive. Yotta may be getting a better deal from Nvidia, and perhaps the order doesn't include many GH200s.
Before this latest purchase, the company also bought another 16,000 H100 GPUs last year, and they're due for delivery this July. This deal was worth about $500 million, as Gupta claimed its order for H100 and GH200 would bring the total cost of Yotta's AI GPU purchases to $1 billion. The fact that Yotta can purchase higher-end hardware without significantly increasing its budget implies that perhaps H100 isn't as expensive as it used to be. This may be the case, considering that H200 has since succeeded H100.
The report links this Yotta-Nvidia deal to the latest U.S. sanctions on China, which have banned the export of many high-end GPUs to the country. Nvidia wasn't allowed to sell its H100 GPUs in China, so it made a slower version of H800 to comply with the sanctions. Months later, however, H800 was banned from China, meaning Nvidia had to create more variants of its GPUs that it could sell in China.
Although the U.S. government seems to be finally easing up on increasing the strictness of its sanctions, it's uncertain whether more rules will be put in place, and uncertainty isn't conducive to business. This may be why India is a more appealing place to strike AI GPU deals because even though it's not an ally of the U.S., India is unlikely to become the next target for sanctions soon.
Stay On the Cutting Edge: Get the Tom's Hardware Newsletter
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Matthew Connatser is a freelancing writer for Tom's Hardware US. He writes articles about CPUs, GPUs, SSDs, and computers in general.
-
Neilbob
31k for an AI doohicky vs 2k for a graphics thingamabob. I think the priorities are quite clear at this point.hotaru251 said:still waiting for Jensen to change name since they arent a graphics company anymore -
peachpuff
Until the ai bubble bursts or someone figures out how to do this cheaply.Neilbob said:31k for an AI doohicky vs 2k for a graphics thingamabob. I think the priorities are quite clear at this point. -
Tom Sunday I wish that AMD with the help from TSMC and AMDs other required contracted production services gotten that job!Reply -
kjfatl The smart players will write AI software that is GPU agnostic. In 5 years it is likely that Intel, AMD, Tesla, Google and Microsoft have solutions of their own twice as fast as what we have now at 20% of the current cost.Reply
The chances of NVIDIA maintaining a commanding lead in AI 5 years from now are about as likely as the US federal government balancing it's budget for a single year.
It will also be interesting to see what happens with Intel's A818 process and AI hardware.
Competition is great! -
endocine kjfatl said:The chances of NVIDIA maintaining a commanding lead in AI 5 years from now are about as likely as the US federal government balancing it's budget for a single year.
Couldn't agree more, as Pat Gelsinger said, the entire industry is out to dethrone nvidia, and nvidia isn't the only player now, they are really good, but not the only player. Pat isn't dumb, but has said some stupid things, this statement however is an astute observation if not an obvious one. $60K "compute" devices aren't sustainable for nvidia, it would be a miracle if nvidia were to keep going at its present run rate.