TLDR version:

• For much of what subsciption/metered frontier models are used for, a workflow using templates and self hosted frozen-version open weight models works better.
• Small low power devices like a Mac mini, NVIDIA Thor and GB10 appliances are not super cheap nor have the raw inference power of a more expensiven fat PC with RTX 5000 or RTX 6000 ADA Graphical card(s) in it, but already lower the self-hosting barrier.
• Crypto PoW moved from CPU to GPU to ASICs, and with the Canaddian company Taalas having shown smaller language models can already run on ASICs, it is likely that faster than GPU cheaper than GB10 appliances inference will become available for self hosting relatively soon.
• A company like Alibaba with one foot in top open models (Qwen3) and one foot in low margin cheap hardware is in a position where it could start selling plug and play model-in-a-box systems, either GB10 style or Taalas style at a real low price point.
• If this pans out, subsciption/metered frontier centered labs like Antropic and OpenAI, and to a lesser extent xAI may be largely overvalued for their IPO valuation.