For some context, in the machine learning GPU market, Nvidia makes basically two products:

• Supercomputer chips with HBM memory. This is the A100, H100, B200, along with AMD’s MI325X and such; things you won’t even find as PCIe cards. These cannot game.

• Then there’s repurposed gaming silicon, like the L40, A40 and such. These are 3090/4090/5090 silicon dies with more VRAM hanging off, but otherwise basically identical. These can game.

In a nutshell, everyone wants the former. No one wants the latter; to the point where Nvidia stipulates “if you buy H200s, so must also buy this many L40s” in big business contracts. They can’t even get buyers to take them.

Where I’m going with this is that, contrary to what you hear, Nvidia has a whole bunch of “budget” server GPUs piling up that no one really wants for their still-astronomical asking price. This has been happening for a few years. They’ve got to do something with them, and they’re basically perfect for cloud gaming.

Hence cloud gaming is relatively cheap now, even with price adjustments. It’s cheap because the hardware is in surplus and the service is new. But this is not going to last, and I guarantee you will see major cloud gaming price hikes in the future over the shrinking surplus.

However, I’m saying that it is really cheap now and it’s gonna stay that way in the short term. The adjustments so far are pretty minor.