“This would keep [US] AI token prices high, giving China an opportunity to export cost-effective inference services, supported by abundant electricity supplies,” Ng said.
This sentence contains many layers, and all of them are somewhat wrong in multiple dimensions, in my opinion.
AI token costs didn’t drop at all since all of this started, they rose significantly because of the increasingly larger models used. Barring an scientific breakthrough, the hardware side is already in the area of physics where further significant improvements are unlikely, as we all saw in the slowdown in performance gains over the last 15 years. And it is in the nature of GenAI that it is computational intensive - all this matrix multiplication is necessary, or else they would have already replaced it.
The most important export China has made in the AI space are open weight models like DeepSeek and GLM 5.3, devaluing tokens coming from Anthropic, OpenAI, Meta, and so on.
the “abundant electricity supplies” that the US are missing are a direct result of the regulatory capture in the US, which emphasizes power generation using carbohydrates to funnel more money into specific companies instead of widespread solar and wind power adoption with battery storage. (The US has more than enough strategic reasons to strive towards being able to build their own solar and wind generation modules, which would actually create jobs and encourage adoption.)
I heard that tokens got cheaper, but the models now use more tokens to accomplish the same tasks.
Also, non-AI companies are widely halting tokenmaxxing, instead telling their work staff to use AI for the use-cases that work, and otherwise go back to human methods.
I also heard that engineers ordered to tokenmaxx (and given incentives to do so) would just prompt ridiculously difficult complex problems, and that would burn through and overrun company token budgets.
This sentence contains many layers, and all of them are somewhat wrong in multiple dimensions, in my opinion.
I heard that tokens got cheaper, but the models now use more tokens to accomplish the same tasks.
Also, non-AI companies are widely halting tokenmaxxing, instead telling their work staff to use AI for the use-cases that work, and otherwise go back to human methods.
I also heard that engineers ordered to tokenmaxx (and given incentives to do so) would just prompt ridiculously difficult complex problems, and that would burn through and overrun company token budgets.
The metric is not the goal.