Any sort of continuous video processing, especially low-latency.

Imagine you were tracking items on video at a self-service checkout. Sure, you could compress the video down to 15 Mbps or so and send it to the cloud. But now, a store with 20 self-checkouts needs 300 Mbps of upload bandwidth. That's one more problem making it harder for Wal-Mart to buy and roll out your product.

Also, if you know you need an NVIDIA L4 dedicated to you 24/7 for a year, a g6.xlarge will cost $7,000/year on-demand or $4,300/year reserved [1] while you can buy the card for $2,500.

Of course for many other use cases the cloud is a fine choice. If you only need a fraction of a GPU, or you only need a monster GPU a tiny fraction of the time, or you need an enormous LLM that demands water cooling and tolerates latency easily, the cloud can be a fine choice.

[1] https://instances.vantage.sh/aws/ec2/g6.xlarge?currency=USD&...