AMD's MI355X accelerator undercuts Nvidia's B300 on the cost of running Kimi K3, the Chinese-developed AI model, according to a report from Startup Fortune surfaced in Google News' Nvidia feed.
That single comparison is the whole of the claim, and it is worth being precise about what it covers. The measure at issue is inference cost — the price of actually running a trained model to answer queries — rather than the cost of training a model from scratch. Inference is the recurring bill: every prompt a user types runs through hardware someone is paying for, which makes cost-per-query the number that determines whether an AI product has workable economics.
The report does not, in the material available here, specify the size of the gap, the workload conditions behind it, or who ran the comparison. Those details matter enormously in chip benchmarking, where results shift with batch sizes, precision settings, software stacks and how efficiently a model has been tuned for a given piece of silicon. Readers should treat the headline finding as a claim awaiting fuller documentation.
Still, the framing itself is the news. Nvidia has been the default choice for AI compute, and its pricing power rests on the assumption that no rival is close enough to matter. A credible cost advantage for AMD on a real, widely discussed model chips away at that assumption.
Why it matters: if AMD can genuinely serve popular models more cheaply than Nvidia, the cost of running AI — and eventually the price of AI products — has more room to fall than a one-supplier market allows.