Not true. It's just abstraction

NVDA Chips --> NVDA PTX assembly --> NVDA CUDA --> other companies to use this.

Even if you use PTX, that still bind you to NVDA hardware. DeepSeek is just getting some percentage of efficiency in their case.