I'm not looking to enter the ecosystem, I'm already deep in it and want to fix the AMD problem so that I can build big projects around it and undercut everyone who's using Nvidia.

You can purchase H100 and A100 PCIe cards over the counter. They're great for compiling CUDA code, testing code before you launch a multi-node job into a cluster, and for running evaluations.

AMD has nothing of the sort, and it's hurting them.

I cannot blow 250K on an SMCI server, nor do I have the electricity setup for it. I can blow 20K on a PCIe GPU and start contributing to the ecosystem, or maybe prove out an idea on one GPU before trying to raise millions from a VC to build a more cost-effective datacenter that actually works.