100% concur. But there is a missing link - a gaping hole! - in China's hardware. And without it, the Edge cannot be properly 'tamed' by Chinese AI. China needs a Mac Mini — a consumer‑scale, unified‑memory, $799 desktop that runs DeepSeek V4‑Flash (and other Chinese LLMs though - after Qwen - DS is closest to being 'Mini Mac' ready) offline, with one click. Until that exists, the Hardware Edge will belong to Apple. Of the Chinese LLMs, only Qwen3.6-35B-A3B is compatible with Mac Mini (along with Nemotron 3 and Gemma 4.) Of course - by definition - no closed weight models are compatible with the Mac Mini.
Hi Grace, thank you for laying out the contrasting mindset between the U.S. Frontier strategy vs the DeepSeek V4 Inference focusing on the economic dimension of the Chinese AI ecosystem. In the long run, the Chinese will reap harvest with a much shorter payback period, and use such gain to re-invest more into her home-grown Frontier model and propel the ongoing inference strength into a credible and defensible MOAT.Hi Grace, thank you for laying out the contrasting mindset between the U.S. Frontier strategy vs the DeepSeek V4 Inference focusing on the economic dimension of the Chinese AI ecosystem. In the long run, the Chinese will reap the harvest with a much shorter payback period, and use such gains to re-invest more into their home-grown Frontier model and propel the ongoing inference strength into a credible and defensible MOAT.
These are extremely helpful analyses with rich insights, thanks.
As someone who is congenitally predisposed to support export controls for national security reasons, I agree that this accelerated (and, as you point out, somewhat coordinated) shift from training to inference and deploying intelligence at scale, reliably, and cheaply does, in fact, bolster some of Jensen's arguments about building on a US stack.
If the US is competing on capability, fair to say this reinforces the argument that China is competing on the economics of intelligence.
What matters more: individual AI model performance, or "good enough" AI ubiquity at inference across society? The jury is still out, but we're beginning to see why despite claims to the contrary, it's no longer a simple answer. The competition is being redefined as we speak.
100% concur. But there is a missing link - a gaping hole! - in China's hardware. And without it, the Edge cannot be properly 'tamed' by Chinese AI. China needs a Mac Mini — a consumer‑scale, unified‑memory, $799 desktop that runs DeepSeek V4‑Flash (and other Chinese LLMs though - after Qwen - DS is closest to being 'Mini Mac' ready) offline, with one click. Until that exists, the Hardware Edge will belong to Apple. Of the Chinese LLMs, only Qwen3.6-35B-A3B is compatible with Mac Mini (along with Nemotron 3 and Gemma 4.) Of course - by definition - no closed weight models are compatible with the Mac Mini.
...if you want to work offline.
Hi Grace, thank you for laying out the contrasting mindset between the U.S. Frontier strategy vs the DeepSeek V4 Inference focusing on the economic dimension of the Chinese AI ecosystem. In the long run, the Chinese will reap harvest with a much shorter payback period, and use such gain to re-invest more into her home-grown Frontier model and propel the ongoing inference strength into a credible and defensible MOAT.Hi Grace, thank you for laying out the contrasting mindset between the U.S. Frontier strategy vs the DeepSeek V4 Inference focusing on the economic dimension of the Chinese AI ecosystem. In the long run, the Chinese will reap the harvest with a much shorter payback period, and use such gains to re-invest more into their home-grown Frontier model and propel the ongoing inference strength into a credible and defensible MOAT.
These are extremely helpful analyses with rich insights, thanks.
As someone who is congenitally predisposed to support export controls for national security reasons, I agree that this accelerated (and, as you point out, somewhat coordinated) shift from training to inference and deploying intelligence at scale, reliably, and cheaply does, in fact, bolster some of Jensen's arguments about building on a US stack.
If the US is competing on capability, fair to say this reinforces the argument that China is competing on the economics of intelligence.
What matters more: individual AI model performance, or "good enough" AI ubiquity at inference across society? The jury is still out, but we're beginning to see why despite claims to the contrary, it's no longer a simple answer. The competition is being redefined as we speak.