Great stuff Ryan, and soooo much better than any of the other ideologically tinged "analysis" from the usual suspects and China podcasts.. great working with you on these issues...
Thanks so much for putting in the work on this Ryan! Really impressive stuff.
more thoughts to come as I digest but just getting through the initial part I think it's hard to say containment is no long possible if we're not factoring in the ability not just to design but to manufacture at scale
thanks Jordan! yeah only focusing on fabless design firms misses the big questions around SMIC and CXMT yields, which I try to touch on in relevant sections (CMD+F "unit costs")
there's also the follow-up question around production capacity and horse trading between the S and A-tiers. truthfully I'm saving those for a separate future deep dive
Ryan, exceptional data work and the finest piece of analysis on the Chinese chip sector. But I can add the "System Level".
You are comparing Card vs. Card (Ascend vs. H100). In the ChinArb framework, System B (China) has already accepted defeat at the "Lithography" layer. They are moving the war to the "Cluster" layer.
Here is the Alpha your dataset implies but doesn't explicitly state:
1. The "Supernode" is the New Unit of Compute System A optimizes for the Chip (Nvidia's margin). System B optimizes for the Rack (Huawei's CloudMatrix). When you can't make the transistor smaller (SMIC limit), you make the interconnect faster (optical backplanes) and the cooling cheaper (liquid cooling at scale). China isn't trying to build a Ferrari; they are building a High-Speed Rail network. Individual trains (chips) can be slower, but the system throughput is higher.
2. DeepSeek is a "Cost Discipline" Mechanism You noted the DeepSeek standard. The strategic intent here isn't just "good enough" performance. DeepSeek R1 is designed to commoditize Intelligence down to the price of electricity. It forces every domestic chipmaker to optimize for Inference Cost (Tokens/$), not Training Speed. Why? Because System B plans to embed this AI into everything from BYD cars to $10 toasters.
3. The Missing Variable: "Stranded Energy" Your "Tokens/Joule" metric is brilliant. But the killer equation is: (Tokens/Joule) × (Cost/Joule). System B has massive amounts of "stranded renewable power" in the West (East Data West Compute). They are routing these "good enough" chips to where electricity is effectively free. System A has better chips, but runs on an inflationary grid.
Don't watch the Fab. Watch the PUE (Power Usage Effectiveness) of the Guizhou clusters. That’s where the real "Sovereign AI" is being forged.
Hi Ryan, I've been very interested in reading your essays! I'm a Tech Reporter and former Asia Correspondent for the Swedish newspaper Svenska Dagbladet, spending a fair bit of time trying to cover the geopolitics in tech development. I'd be very interested in chatting to you about the diverging AI paths of China and the US – would you have time soon-ish? Please write me at per.liljas@svd.se Much appreciated!
Great stuff Ryan, and soooo much better than any of the other ideologically tinged "analysis" from the usual suspects and China podcasts.. great working with you on these issues...
Thank you Paul! Always learning from you
Great overview! Looking forward to the next parts!
"is by no means an authoritative declaration - it’s subjective and shouldn’t be taken too seriously" lol just wait for the domestic coverage
wait you mean to say people don't read disclaimers?
Thanks so much for putting in the work on this Ryan! Really impressive stuff.
more thoughts to come as I digest but just getting through the initial part I think it's hard to say containment is no long possible if we're not factoring in the ability not just to design but to manufacture at scale
thanks Jordan! yeah only focusing on fabless design firms misses the big questions around SMIC and CXMT yields, which I try to touch on in relevant sections (CMD+F "unit costs")
there's also the follow-up question around production capacity and horse trading between the S and A-tiers. truthfully I'm saving those for a separate future deep dive
Ryan, exceptional data work and the finest piece of analysis on the Chinese chip sector. But I can add the "System Level".
You are comparing Card vs. Card (Ascend vs. H100). In the ChinArb framework, System B (China) has already accepted defeat at the "Lithography" layer. They are moving the war to the "Cluster" layer.
Here is the Alpha your dataset implies but doesn't explicitly state:
1. The "Supernode" is the New Unit of Compute System A optimizes for the Chip (Nvidia's margin). System B optimizes for the Rack (Huawei's CloudMatrix). When you can't make the transistor smaller (SMIC limit), you make the interconnect faster (optical backplanes) and the cooling cheaper (liquid cooling at scale). China isn't trying to build a Ferrari; they are building a High-Speed Rail network. Individual trains (chips) can be slower, but the system throughput is higher.
2. DeepSeek is a "Cost Discipline" Mechanism You noted the DeepSeek standard. The strategic intent here isn't just "good enough" performance. DeepSeek R1 is designed to commoditize Intelligence down to the price of electricity. It forces every domestic chipmaker to optimize for Inference Cost (Tokens/$), not Training Speed. Why? Because System B plans to embed this AI into everything from BYD cars to $10 toasters.
3. The Missing Variable: "Stranded Energy" Your "Tokens/Joule" metric is brilliant. But the killer equation is: (Tokens/Joule) × (Cost/Joule). System B has massive amounts of "stranded renewable power" in the West (East Data West Compute). They are routing these "good enough" chips to where electricity is effectively free. System A has better chips, but runs on an inflationary grid.
Don't watch the Fab. Watch the PUE (Power Usage Effectiveness) of the Guizhou clusters. That’s where the real "Sovereign AI" is being forged.
Hi Ryan, I've been very interested in reading your essays! I'm a Tech Reporter and former Asia Correspondent for the Swedish newspaper Svenska Dagbladet, spending a fair bit of time trying to cover the geopolitics in tech development. I'd be very interested in chatting to you about the diverging AI paths of China and the US – would you have time soon-ish? Please write me at per.liljas@svd.se Much appreciated!