Chamath @ 00:36:30Right
techai
Within the next several years (by around 2030), Nvidia will expand beyond selling GPUs/servers into offering its own large-scale cloud or data-center compute service that competes directly with major hyperscalers such as AWS, Google Cloud Platform, and Microsoft Azure (i.e., customers will be able to rent Nvidia-operated GPU compute capacity as an alternative to those hyperscalers).
I think Nvidia this is to build on Sachs's point is going to get pulled into competing directly with the hyperscalers. So if you were just selling chips, you probably wouldn't. But Sachs is right. Like these are these big, bulky actual machines. Then all of a sudden you're like, well, why don't I just create my own physical plant and just stack these things and create racks and racks of these machines and go head to head with AWS instead of selling to them?... I think it's it's likely that Nvidia goes on a full frontal assault against GCP and Amazon and Microsoft.
Explanation
Public information since 2023 shows that Nvidia did move beyond just selling GPUs/servers into offering its own large-scale cloud-style compute service for AI, rented directly to customers:
- Nvidia launched DGX Cloud in 2023 as its first cloud service, explicitly designed to rent out GPU servers directly to large enterprises (e.g., SAP, Genentech) for AI development, rather than only selling chips or boxes. (theinformation.com)
- DGX Cloud is marketed by Nvidia as a fully managed AI supercomputing / AI‑training‑as‑a‑service platform that provides enterprises with scalable access to Nvidia DGX infrastructure and NVIDIA AI Enterprise software over the cloud. (nvidia.com)
- The service is hosted on infrastructure in data centers of major cloud providers (AWS, Microsoft Azure, Google Cloud, Oracle), but it is sold and branded as Nvidia’s own DGX Cloud service; AWS and Nvidia explicitly describe AWS as hosting NVIDIA DGX Cloud, not merely reselling hardware. (techpowerup.com)
- Reporting characterizes this as Nvidia “entering cloud computing” and “muscling into the GPU cloud market”, noting that DGX Cloud competes with established hyperscalers for AI training workloads and has even “rankled” AWS because it gives customers a Nvidia‑run alternative to simply renting GPUs from AWS itself. (theinformation.com)
Caveat: Chamath’s colorful language about Nvidia building its own physical plants isn’t literally how this has played out—Nvidia relies on partner data centers rather than operating a standalone hyperscaler network. But the core prediction that Nvidia would get pulled into directly offering cloud AI compute that competes with the big hyperscalers has clearly come true, and it was in fact already true (via DGX Cloud) by the time of the May 2024 episode, with Nvidia doubling down on that strategy afterward.