Nvidia does have their gaming cloud. They also have excellent researchers, and engineers improving the performance of published research, including consulting work.
I suppose for the self-serve cloud market, that's set to blow up. For all those people who want to train their own models, LoRA adapt, or explore possibilities.
Jensen has talked about creating a cloud including their customers selling back compute time on DGX nodes.
Lots of potential reasons Nvidia doesn't want to be selling a public cloud. For one, they want to sell shovels to the gold rush, not pan for themselves.
Two, they'd be directly competing with their major datacenter customers. They already do this in other markets, but there's a balancing act with how much you can push the big cloud players and how much they'll spend trying to replace you. Nvidia is already struggling to keep them in line (controlling customers is in their DNA) and every new market segment they enter makes antitrust regulators wary.
Having an investment also doesn't tie them to the fate of any particular company. If it goes bankrupt tomorrow, Nvidia doesn't have to sit there and fire the employees/unwind the company.
Lots of others too, like not wanting to get involved with future cloud legislation.
Nvidia dominates the market largely because they're the best and people choose to buy from them, rather than primarily from shady OEM tactics like Microsoft did. People could buy Cerberus, or AMD, or Imagination, or a dozen other vendors, they just don't.
Google famously made the choice to design TPUs rather than buy Nvidia. They still have Nvidia in GCloud due to customer pressure. I've been involved with at least one GPU design to avoid buying from Nvidia.
Nvidia holds grudges for a long, long time though.
This would also resolve a branding problem for Lambda, people always confusing it w AWS’ similarly named product..