Nvidia Steps Back From Cloud Effort to Compete with AWS
The Takeaway
• Nvidia is primarily using its GPU cloud service for its own research and development
• The change suggests Nvidia didn’t attract many external customers to the service
• The chip designer committed $13 billion to rent back its own chips for the service
Nvidia is stepping back from its nascent cloud computing business, which had put it in quasi-competition with Amazon Web Services. It has lessened its efforts to attract businesses to the cloud service, dubbed DGX Cloud, said a person with direct knowledge of the situation.
Instead, Nvidia plans to primarily use the service, which consists of servers powered by its AI chips, for its own researchers, this person said. Employees use such servers for everything from designing new chips to developing AI models that can run on the chips.
Nvidia has committed to spend $13 billion to rent back its own AI chips from bigger cloud providers that had purchased them, including AWS. Nvidia has then turned around and rented out some of those chips to companies such as Amgen and ServiceNow that want to develop their own AI applications. Nvidia believed the cloud business could one day help it generate $150 billion in revenue.
The change to DGX Cloud, which hasn’t been previously reported, suggests Nvidia found limited demand for the service. Nvidia hasn’t stopped renting chips out to some new customers: it recently said quantum and AI startup SandboxAQ would use the service. (Nvidia has invested in SandboxAQ.)
Some AI developers have balked at Nvidia’s prices for DGX Cloud servers, which are typically more expensive than those sold by traditional providers, according to someone who has helped multiple companies negotiate cloud deals.
In setting up DGX Cloud, CEO Jensen Huang moved beyond just selling chips and other data center equipment. He envisioned banks, pharmaceutical companies and other large businesses coming to Nvidia to rent servers with its AI chips, similar to how they rent such servers from Amazon and Microsoft.
“We’re going to democratize the access of this infrastructure…really make this technology and this capability quite accessible,” Huang said on an earnings call in early 2023, speaking about DGX Cloud. He hoped the service would ease a shortage of Nvidia chip servers for business customers at the time.
Huang’s competitive move ruffled feathers at large cloud providers—especially at AWS—that are Nvidia’s biggest customers.
The deals Nvidia made to rent its own chips for DGX Cloud served a dual purpose: making sure large enterprises could access Nvidia’s graphics processing units “without having to wait,” Huang said at the time, and boosting younger cloud providers such as CoreWeave and Lambda—which also provided AI chip servers to Nvidia for DGX Cloud.
Nvidia wanted to do deals with those younger providers to increase competition in the cloud market, partly at the expense of AWS, Google Cloud and Microsoft Azure. Those bigger clouds account for half of Nvidia’s revenue, increasing the risk that that chipmaker could end up relying too much on them. AWS, Google and Microsoft have also developed their own AI chips to lessen their reliance on Nvidia’s, posing a threat to Nvidia’s dominance.
By August 2023, Huang told investors that DGX Cloud was “an enormous success.” In late 2024, CFO Colette Kress said the company’s software business, a category she had said included DGX Cloud, was generating revenue at an annualized rate of $2 billion.
Now, though, Nvidia uses most of the DGX Cloud server capacity for its own internal efforts, according to the person with direct knowledge of the service. And in its most recent quarterly report, Nvidia stopped stating that its cloud spending commitments are partly for DGX Cloud, despite doing so in previous quarters. The disclosure change implied Nvidia was no longer prioritizing the service for external customers.
In an interview, Alexis Black Bjorlin, who runs DGX Cloud, disputed that Nvidia’s strategy had shifted. “Our internal researchers need significant computing, so do all of our customers,” she said. “So our strategy is still the same.”
DGX Cloud customers typically sign short-term contracts of less than a year, then find GPU servers to rent from a larger cloud provider, according to someone with direct knowledge.
For Nvidia, one benefit of renting GPU servers from cloud providers for its researchers is it won’t have to spend the money and time to set up data centers on its own.
Nvidia’s Other Cloud Play
Nvidia still wants some business customers to come to Nvidia to rent GPUs. Earlier this summer, Nvidia launched another kind of cloud service, DGX Cloud Lepton. Though similar to the original DGX Cloud, Lepton is a “marketplace” in which companies access GPUs from a network of cloud providers.
Cloud providers choose to list their available GPU server capacity on the marketplace, rather than Nvidia signing multi-year contracts to rent that capacity to customers the way it does with DGX Cloud.
The marketplace didn’t get off to a fast start. Executives at several of the cloud firms that are listed in the Nvidia Lepton marketplace said they viewed the new service as competitive to their own because it gives the chip designer direct relationships with their customers, The Information previously reported.
Representatives of two of the nearly two dozen cloud providers listed in the Lepton marketplace said they participated in the service because Nvidia asked them to be a part of a marketing push for it, not because they wanted to be a part of it.