[Jlagreen]

Simple, because Nvidia can charge twice for the same GPU that way.

Building a cloud data center is cost first and then you earn money by rent.

But if you seel the GPU to a CSP then the cloud provider has the cost and needs to rent while you made immediately money on the GPU.

But what about getting paid continously? Well simple, the end customer is then cloud but the cloud customer. So develop a SW for your GPU which then has a license to it. So the end cloud customer will pay rent to the cloud provider and license fee to the GPU maker. That's Nvidia's business model for now.

Eventually, the demand for GPU sales will drop but by that time, Nvidia will have a large install base of end customers using their enterprise SW. Then Nvidia will start becoming a competitior to CSPs and will easily do so because they can deploy data centers at cost and give better pricing. But today, Nvidia needs the CSPs to spread their SW. The primary goal is to reach all the end customer via DGX cloud and to create mindshare and moats. CSPs are only a means to an end of Nvidia as they could easily out compete them on AI compute if Nvidia wanted to to.