This resonates strongly with what I wrote recently about China and the West not even looking at the same scoreboard.
In my China observations (check my page for a series of paper on this topic), what stood out wasn’t a singular obsession with “beating the US to AGI”, but how normalised, embedded, and utilitarian AI already is: in drones, robotics, logistics, manufacturing, public services, consumer apps. The metric isn’t “who has the smartest model”, it’s “how much of the real economy and daily life is already AI-augmented.”
The Western narrative frames competition as a frontier-model race. The Chinese narrative — and policy reality — looks much closer to what you describe here: system-level capability, deployment at scale, resilience, and integration into industrial and social infrastructure.
So when we say “China is behind” or “catching up,” we’re implicitly projecting a US-style scoreboard: parameter count, benchmarks, AGI timelines. But if the scoreboard is AI-as-general-purpose infrastructure, not AI-as-mythical intelligence breakthrough, then the relative positions look very different.
In that sense, this piece is right: it’s not a single race. It’s two civilisational strategies optimising for different futures — and misunderstanding that may be the West’s bigger strategic blind spot than any single model gap.
This resonates with the recent announcement that Google will be “white labeling” Gemini to power Apple’s Siri. Apparently OpenAI turned down the deal to “focus on compute” and their own upcoming devices designed by Jony Ive. It seems Apple decided to work on its own coordination solutions rather than jump into a pointless race. OpenAI remains obsessed with having the smartest model.
Thank you for your work, Sangeet. I read and recommend Reshuffle to everyone I know who’s working on anything AI-related — and even if they aren’t! It’s a clear framework for a very complex issue.
I really like this piece. Rarely you would find an article going so deep and so broad and at the same time so frontier in the latest development. Sangeet, congratulations for making such a cohesive theory-building article. Thanks again for making this happen.
As always, great writeup. That strongly resonates.
I am curious though, what fundamental change could America do at this at this point in time? Apart from just somehow trying to reduce the rate of acceleration of the increasing gap. Even if they think they are ahead, then let's call it the increasing rate of reducing the gap.
The best America may be trying for is, finding ways to buy more time.. But how?
Insightful analysis, Sangeet...and I agree with your POV! In a very simplistic way, this is manifested in the distinct way Manus approaches a requirement, as compared to ChatGPT...the former is sharply focused on the 'HOW'...'coordination' and 'getting things done', while the latter is focused on the WHAT...essentially the 'intelligence' layer...
Amazing analysis and layout of this race. This thought process along with China’s goal to make the intelligence layer more indigenous renders the importance of export controls almost irrelevant. Coordination with less capable chips can easily close the gap in adoption and maximizing implementation across all layers.
Great analyzes because China massively invest in DRC (Congo) where 70% of rare earth and critical minerals come from. Those minerals are the backbone of AI infrastructure
There is also an insight as to how UK missed the political might of being the leading colonialist because of its economic model centered on invention rather than coordination, which I think it carried out in its colonies cue cheap labour and also due to protectorate obligations. Still, despite having a lead from their research group on atomic bombs, because they lacked coordination systems, it came from US and thus changed the global locus of control or sphere of influence.
The core flaw in most debates about AI supremacy is the assumption that technology can be durably controlled. History shows the opposite: technology resists centralization and inevitably democratizes.
Even under export controls and geopolitical pressure, China continues to learn, adapt, and advance in AI. Restrictions slow diffusion, but they do not stop it. Knowledge leaks, efficiency improves, alternatives emerge. Control is always temporary.
Argentina’s development of nuclear technology offers a parallel lesson. Despite nuclear know-how being tightly guarded by a few powers, Argentina built indigenous capabilities and integrated into the global nuclear ecosystem. Once a technology exists, it spreads through talent, institutions, and necessity, not permission.
This pattern is not new. The printing press did not remain a tool of kings or the Church. Once introduced, it escaped centralized control and reshaped society through mass access to knowledge. The same happened with the internet and microprocessors: openness beat enclosure.
AI is no different. As intelligence becomes cheaper and more abundant, advantage shifts away from owning the most advanced models and toward who enables adoption, integration, and use at scale.
The real winner of the AI era will not be the country that best controls AI, but the one that democratizes it fastest. Power will belong to those who turn intelligence into a widely accessible capability rather than a guarded asset.
In the long run, containment loses. Diffusion wins.
Martin, you might be familiar with their work, but if not, check out the 2024 Nobel prize winners in economics. They showed that iclusive institutions where many people win beat exploitive ones where a few win at the cost of manu. Or as you said, containment loses, diffusion wins.
This resonates strongly with what I wrote recently about China and the West not even looking at the same scoreboard.
In my China observations (check my page for a series of paper on this topic), what stood out wasn’t a singular obsession with “beating the US to AGI”, but how normalised, embedded, and utilitarian AI already is: in drones, robotics, logistics, manufacturing, public services, consumer apps. The metric isn’t “who has the smartest model”, it’s “how much of the real economy and daily life is already AI-augmented.”
The Western narrative frames competition as a frontier-model race. The Chinese narrative — and policy reality — looks much closer to what you describe here: system-level capability, deployment at scale, resilience, and integration into industrial and social infrastructure.
So when we say “China is behind” or “catching up,” we’re implicitly projecting a US-style scoreboard: parameter count, benchmarks, AGI timelines. But if the scoreboard is AI-as-general-purpose infrastructure, not AI-as-mythical intelligence breakthrough, then the relative positions look very different.
In that sense, this piece is right: it’s not a single race. It’s two civilisational strategies optimising for different futures — and misunderstanding that may be the West’s bigger strategic blind spot than any single model gap.
Great way to boil down the US-China challenge today with the analogy to the US-UK history in the early part of the last century.
This resonates with the recent announcement that Google will be “white labeling” Gemini to power Apple’s Siri. Apparently OpenAI turned down the deal to “focus on compute” and their own upcoming devices designed by Jony Ive. It seems Apple decided to work on its own coordination solutions rather than jump into a pointless race. OpenAI remains obsessed with having the smartest model.
Thank you for your work, Sangeet. I read and recommend Reshuffle to everyone I know who’s working on anything AI-related — and even if they aren’t! It’s a clear framework for a very complex issue.
I really like this piece. Rarely you would find an article going so deep and so broad and at the same time so frontier in the latest development. Sangeet, congratulations for making such a cohesive theory-building article. Thanks again for making this happen.
As always, great writeup. That strongly resonates.
I am curious though, what fundamental change could America do at this at this point in time? Apart from just somehow trying to reduce the rate of acceleration of the increasing gap. Even if they think they are ahead, then let's call it the increasing rate of reducing the gap.
The best America may be trying for is, finding ways to buy more time.. But how?
Insightful analysis, Sangeet...and I agree with your POV! In a very simplistic way, this is manifested in the distinct way Manus approaches a requirement, as compared to ChatGPT...the former is sharply focused on the 'HOW'...'coordination' and 'getting things done', while the latter is focused on the WHAT...essentially the 'intelligence' layer...
Amazing analysis and layout of this race. This thought process along with China’s goal to make the intelligence layer more indigenous renders the importance of export controls almost irrelevant. Coordination with less capable chips can easily close the gap in adoption and maximizing implementation across all layers.
Very well pull, this is something I’ve been thinking much about lately. The us tends to be short sighted and focused here, missing the bigger picture
Great analyzes because China massively invest in DRC (Congo) where 70% of rare earth and critical minerals come from. Those minerals are the backbone of AI infrastructure
There is also an insight as to how UK missed the political might of being the leading colonialist because of its economic model centered on invention rather than coordination, which I think it carried out in its colonies cue cheap labour and also due to protectorate obligations. Still, despite having a lead from their research group on atomic bombs, because they lacked coordination systems, it came from US and thus changed the global locus of control or sphere of influence.
The core flaw in most debates about AI supremacy is the assumption that technology can be durably controlled. History shows the opposite: technology resists centralization and inevitably democratizes.
Even under export controls and geopolitical pressure, China continues to learn, adapt, and advance in AI. Restrictions slow diffusion, but they do not stop it. Knowledge leaks, efficiency improves, alternatives emerge. Control is always temporary.
Argentina’s development of nuclear technology offers a parallel lesson. Despite nuclear know-how being tightly guarded by a few powers, Argentina built indigenous capabilities and integrated into the global nuclear ecosystem. Once a technology exists, it spreads through talent, institutions, and necessity, not permission.
This pattern is not new. The printing press did not remain a tool of kings or the Church. Once introduced, it escaped centralized control and reshaped society through mass access to knowledge. The same happened with the internet and microprocessors: openness beat enclosure.
AI is no different. As intelligence becomes cheaper and more abundant, advantage shifts away from owning the most advanced models and toward who enables adoption, integration, and use at scale.
The real winner of the AI era will not be the country that best controls AI, but the one that democratizes it fastest. Power will belong to those who turn intelligence into a widely accessible capability rather than a guarded asset.
In the long run, containment loses. Diffusion wins.
Martin, you might be familiar with their work, but if not, check out the 2024 Nobel prize winners in economics. They showed that iclusive institutions where many people win beat exploitive ones where a few win at the cost of manu. Or as you said, containment loses, diffusion wins.
I have not thought of it this way… 🤯🤯🤯
https://drive.google.com/file/d/1FCkciku5CM4fOFP9GmXv1IE8u6a2W9p0/view?usp=drivesdk