Which company is most likely to challenge Nvidia's ML training market?

Honestly, its Intel, AMD, or bust.

Part of Nvidia's advantage is the ubiquity of their hardware. CUDA became the monster that is because students with Nvidia gaming GPUs grew up using it, and thats never going to happen with AI startups who keep their hardware locked away in datacenters or in $10K PCIe cards you cant even buy as an individual.

Apple is formidable but too closed off. I am not convinced other ARM licensees will make a chip beefy enough for hobbyist training/tuning, but would love to be proven wrong. Cloud chip designers like Amazon and Google will never make desktop products.

Intel, on the other hand, is ubiquitous. OpenVINO performance in Stable Diffusion is already very good on Arc, and the port is basically 2 changed lines. And people forget that Intel has a history of some graphics heavy SoCs and a fab advantage.

ROCM Stable Diffusion performance on desktop cards is... truly awful. But 3rd party frameworks are better, so maybe it wont always be awful?

Probably SiFive or someone similar. People want an open acceleration path for ML, and open ISA adoption will enable that. If not them, then it's up to AMD or Apple to displace the market, and neither of them are likely capable. Nvidia has extremely perverse incentives in the current market, and I don't think there are any easy paths to dethroning them. OpenCL and ROCm exist, but there's not much money in promoting or developing them.

Tenstorrent. (CEO: Jim Keller)