On September 13, according to The Information, Nvidia is quietly "shrinking" its fledgling cloud computing business.
Nvidia has reduced its efforts to attract enterprise customers to its DGX Cloud service, and even plans to keep the service mainly for its own use - such as providing computing power support to AI researchers within the company.
This means that Nvidia's "cloud business", which originally wanted to do a big job, has now stepped on the brakes.
Why the sudden adjustment?
The main reason is that the market does not buy it, and the demand is too weak.
Nvidia has boasted that it will spend $13 billion to lease its AI chips back from large cloud service providers such as Amazon and Microsoft, and then package them into DGX Cloud services to sublease to enterprise customers, and set a revenue target of up to $150 billion.
But the reality is skinny: many AI developers think DGX Cloud is too expensive! Compared with traditional cloud platforms such as Amazon AWS and Microsoft Azure, Nvidia's quotation is much higher, and customers are naturally hesitant. To put it simply, the price is too high, and everyone thinks it is too expensive and is unwilling to pay for it.
Another "unexpected benefit": it eases conflicts with large customers.
Nvidia makes its own cloud services, which is actually a bit of a "fight for itself" - because its main customers (such as AWS, Azure) are cloud giants themselves. It's like selling a knife to someone else and opening your own martial arts gym to grab business, and the relationship must be awkward.
Now Nvidia has taken a step back and no longer vigorously promoted DGX Cloud to grab the market, but has relieved these big customers and eased the tension of the partnership. After all, these cloud service providers contribute nearly half of Nvidia's revenue and cannot be offended.
The focus of the business shifts to internal R&D
According to people familiar with the matter, Nvidia now reserves most of its DGX Cloud's computing power for its own use. In-house engineers and researchers are using these powerful servers for all kinds of core work:
Designing the next generation of AI chips,
Train your own AI model,
Do various cutting-edge technology research and development.
换句话说,原本打算“租给客户赚钱”的云服务,现在更像是英伟达的“内部研发工具”。
However, Alexis Black Bjorlin, head of DGX Cloud, denied in an interview that the company had changed its strategy. "Our R&D team does need a lot of computing power, but so do our customers," she said. "Our strategy hasn't changed, it's still serving ourselves and our customers." ”
But there is one detail that is very telling: in its recent earnings report, Nvidia quietly deleted a sentence - it has mentioned that "part of the cloud spending commitment is for DGX Cloud" in the past few quarters, but this time it was not mentioned. This change is seen as a signal that Nvidia may no longer make providing DGX Cloud services a priority.
Brief summary: The mouth says "nothing has changed", but the action has "turned". Although it has not been officially announced, judging from the resource allocation and financial report wording, Nvidia obviously attaches more importance to using these computing power to get its own technology first, rather than rushing to rely on it to make money.
New service models seek breakthroughs
Although Nvidia has shrunk its DGX Cloud service, it has not completely abandoned the cloud market. This summer, it launched a new way to play: DGX Cloud Lepton.
This is not like the previous "self-operated model", but more like a "GPU version of Taobao" - NVIDIA has built a platform that allows major cloud service providers (such as AWS, Azure, etc.) to "put on the shelves" of idle GPU computing power, and enterprise customers can rent it directly through this platform.
What is the difference from the previous one?
Old model (DGX Cloud): Nvidia rents a large number of GPUs by itself, packages them into services and subleases them to customers, which is equivalent to "opening a store and selling it yourself".
New model (Lepton): Nvidia only does a "platform", and cloud service providers voluntarily put up computing power, which is equivalent to "let others open a store, and I will provide a market".
Sounds beautiful, but the reality is a bit cold.
According to some cloud service provider executives involved in the Lepton platform, the new service is progressing slowly Why? Because they feel that Nvidia is "grabbing business". They are worried that Nvidia will directly connect with their big customers through this platform, and over time, customers will no longer need them as middlemen. It is equivalent to helping Nvidia build a bridge to "jump over me"
Nvidia wants to be a "platform party", selling chips and matching transactions. But cloud service providers don't buy it - for fear of being overheaded, they simply treat it coldly. This "battle for computing power" is ostensibly technical cooperation, but behind it is actually a contest of who controls customers and who controls the right to speak.
Jingtai Observation|Not "giving up", but "focusing"
Many people think that Nvidia's "cloud business has failed", but Jingtai believes that it has just changed its playing style: from "selling services in person" to "selling chips + empowering giants". Its core advantage has never been "operating the cloud", but - creating the world's most sought-after AI chips.
1. For Nvidia (NVDA): Short-term bearish, long-term positive
Short-term: The 150 billion cloud revenue target may be lowered, and the market may be disappointed;
Long-term: Focusing on the main chip business and consolidating cooperative relationships with cloud giants is a more sustainable model. There is no need to panic, the core logic has not changed.
2. For cloud giants (AWS, Microsoft, Google): Less pressure and more stable cooperation
there is one less opponent who is "both a referee and an athlete"; you can purchase Nvidia's top chips with more confidence without worrying about "robbing customers". It is good for cloud service providers' AI expansion strategies.