Buy a thousand GPUs and you have a very expensive cluster. You don't have a cloud yet. The difference is nearly all software, and it's the part that never makes the launch photo.
A cloud has problems a private cluster never meets.
Tenants can't see each other's data or traffic. In NVIDIA's design that fence is pushed down into the network card. A BlueField DPU runs the network and security policy outside the host, so a tenant handed a whole bare-metal server still can't touch the rules that keep them in.
Tenants serve themselves. Someone signs up, asks for 64 GPUs with storage and a Kubernetes cluster attached, gets them in minutes and hands them back on Friday. Behind that sits a self-service layer, quotas, metering and billing. Each one is software somebody has to build or buy, and then keep running.
And more and more, tenants want proof the provider can't see in either. That's confidential computing. On the Vera Rubin NVL72, NVIDIA stretches it across the whole rack: one encrypted security domain covering 72 GPUs, 36 CPUs and the NVLink between them.
An enterprise running models for its own staff needs almost none of this. There's one tenant and one set of policies. A scheduler and decent access control will do the job.
That's the real line between a provider and everyone else. It isn't how many GPUs you own. It's whether you're building for strangers.