the 2 write ups are about different layers. the bull case is about capital spend/data center build. the bear case is about model revenue. model revenue at the enterprise level might not be a race to the bottom- kyc, aml, audit trails, hipaa compliance, regulatory approval will be gating functions.
OpenRouter token counts are not cleanly separated for US/EU hosted Chinese models and direct calls to Chinese APIs. This limits the exposure of data to the Chinese government. And OpenRouter doesn’t include locally hosted models.
Most of the OpenAI and anthropic traffic from us companies is through direct contracts and subscriptions, not aggregators like openrouter. Companies and developers go there for low level tasks. Could be an innovator’s dilemma but the data as of today is very flawed. It’s like reading only fans data to try to understand marriage.
Another thing that hurts the signal from OpenRouter is that many Chinese models, while cheaper per token, use significantly more tokens for the same output versus Open AI/Anthropic. Heavy token usage and thinking time is one of the chief complaints of GLM5.2, see: https://news.ycombinator.com/item?id=48668297. Openrouter also breaks some of the caching too (forcing higher token usage than going via native APIs). So, Openrouter token usage does provide trend/signal but the scale of the difference is probably off quite a bit if looking at the raw numbers in isolation.
The political hedge you end with is the sharpest part of this piece. Reading across 72 asset manager reports, Information Technology is the single most crowded overweight in the sector allocation, 71.4% of managers long, with only one dissenter quietly stepping aside in Q1.
The GS/Chris Woods split you describe maps almost exactly onto that positioning: the bull case is already in the price. What forces the unwind is not the argument. It is the leverage.
I asked Claude Opus an there are 6 Chinese LLM's competing with the hot new GLM 5.2 Zhipu from Knowledge Atlas. Short borrow was 43% on IBKR the other day and there were no shares available to borrow
One of your leitmotivs has been that capital has becoming scarce. Can the most indebted nation of the world (on net basis) follow the most expensive path to AI?
Hey Russell - possibly like u i am trying ro narrow down my intake of research to cut down the noise - any thoughts on Michael Chembalist at JPM (eye on the market). Pieces seem very well researched but i am not smart enough to understand if he / they are trying to cleverly back fill the data to justify a preferred jpm narrative / view. Intuitively it seems difficult to me that JPMs most read CIO / research guy takes a bearish view on any aspect of US markets… Although in his latest eye on the market is noting froth in memory and semis (possibly as close as he can sail to a sell suggestion).
the 2 write ups are about different layers. the bull case is about capital spend/data center build. the bear case is about model revenue. model revenue at the enterprise level might not be a race to the bottom- kyc, aml, audit trails, hipaa compliance, regulatory approval will be gating functions.
OpenRouter token counts are not cleanly separated for US/EU hosted Chinese models and direct calls to Chinese APIs. This limits the exposure of data to the Chinese government. And OpenRouter doesn’t include locally hosted models.
Thats fair enough... but even with that the trend is striking unless there was a methodology change
Most of the OpenAI and anthropic traffic from us companies is through direct contracts and subscriptions, not aggregators like openrouter. Companies and developers go there for low level tasks. Could be an innovator’s dilemma but the data as of today is very flawed. It’s like reading only fans data to try to understand marriage.
Another thing that hurts the signal from OpenRouter is that many Chinese models, while cheaper per token, use significantly more tokens for the same output versus Open AI/Anthropic. Heavy token usage and thinking time is one of the chief complaints of GLM5.2, see: https://news.ycombinator.com/item?id=48668297. Openrouter also breaks some of the caching too (forcing higher token usage than going via native APIs). So, Openrouter token usage does provide trend/signal but the scale of the difference is probably off quite a bit if looking at the raw numbers in isolation.
Interesting. That would match up better with markets and orice action...
The political hedge you end with is the sharpest part of this piece. Reading across 72 asset manager reports, Information Technology is the single most crowded overweight in the sector allocation, 71.4% of managers long, with only one dissenter quietly stepping aside in Q1.
The GS/Chris Woods split you describe maps almost exactly onto that positioning: the bull case is already in the price. What forces the unwind is not the argument. It is the leverage.
I asked Claude Opus an there are 6 Chinese LLM's competing with the hot new GLM 5.2 Zhipu from Knowledge Atlas. Short borrow was 43% on IBKR the other day and there were no shares available to borrow
Interesting - and it is a large market cap too!
One of your leitmotivs has been that capital has becoming scarce. Can the most indebted nation of the world (on net basis) follow the most expensive path to AI?
Keeping debt loads at bay requires ever rising prices....
Hey Russell - possibly like u i am trying ro narrow down my intake of research to cut down the noise - any thoughts on Michael Chembalist at JPM (eye on the market). Pieces seem very well researched but i am not smart enough to understand if he / they are trying to cleverly back fill the data to justify a preferred jpm narrative / view. Intuitively it seems difficult to me that JPMs most read CIO / research guy takes a bearish view on any aspect of US markets… Although in his latest eye on the market is noting froth in memory and semis (possibly as close as he can sail to a sell suggestion).
Will take a look