Two charts showing extremely important, widely held AI assumptions are outdated:
Assumption 1: Chinese models are cheaper & more efficient. In fact, at almost every level of intelligence finds US models cheaper per task. Chinese models might be cheaper per token, but US models tend to be MORE efficient with tokens.
Everyone seizes on papers from Chinese labs with architectural improvements for efficiency and assumes US labs are undisciplined. But ask yourself what OpenAI/Anthropic/Google have behind the scenes that would cut down on their #1 cost. And then add that they have more efficient chips. Should not be surprising that US labs can undercut!
You can't undercut free, though, as anyone can download CN open weight models and use w/o paying the labs anything. BUT unless you have your own compute (few firms do), the days of firms running Chinese models for the cost of compute are numbered or over because Chinese AI labs are reportedly asking for 30% cut from cloud providers serving their models. This is likely main way these models are accessed, and the costs are going to be passed on.
Assumption 2) US company spending on AI is spiraling out of control.
finds that companies are getting better at picking the right cost/capabilities tradeoffs and getting better deals from competition between Anthropic and OpenAI. Use continues to rise, but at a cost that is leveling out. Controlling cost will put less pressure on firms to move to Chinese models, esp if they aren't cheaper anymore!
Conversation
Post your reply
Thank you Martin! Those are really good points. I have actually just published an article with Epoch where I touched on some of the points you made, e.g.:
Quote
Epoch AI
@EpochAIResearch
Replying to @EpochAIResearch
Case study: GLM 5.3 Flash
Aug 26: Zhipu launched the model with open weights
Aug 27: 12 other providers on OpenRouter
Sep 15: 26 other providers, 8 pricing below Zhipu
Over this period, Zhipu’s share of the model’s tokens fell from 88% to 22%, even as total volume grew 17%
Retries and tool calls belong in cost per task, too. A cheap first answer can be an expensive finished job.
Discover more
Sourced from across X
I haven't posted in a while so here are 20 takes about AI, in no particular order.
1. It's very telling that the default question in tech and journalism is "how worried are you about X" and fighting over what to worry about. There's a deeply anxious cultural backdrop in Western
Quote
Séb Krier
@sebkrier
Replying to @tyler_m_john and @itaisher
I agree that verbal probability terms are inherently too elastic to carry calibrated meaning on their own, but I think this isn't exactly my concern here. My issue isn't with "using numbers in general" but how this is ultimately used, where, and and what for.
In the case of
They arrested a man for doing his bloody job??????
Quote
Tom Nuttall
@tom_nuttall
Bloody hell. A former head of the BND, Germany's foreign intelligence service, has been arrested on suspicion of espionage. sueddeutsche.de/politik/hannin
Hot take: From about GPT-5 onward, if humans and LLMs disagree in a data labeling task, when you look at the actual disagreeing cases, it's almost always that the LLM is right and the human(s) are wrong OR the humans also substantively disagree with each other
"there has been basically no increase in car sales in China in over a decade" is one of those sentences that sounds like it can't possibly be true but turns out to in fact be true
Quote
Brad Setser
@Brad_Setser
China has the capacity to produce 50-55 million cars --
The domestic market has fallen from ~ 25m cars to ~20 m cars this year
China could easily export 2-3m cars to the US AND its 1.5m cars to Europe
Quote
Blume Industries CEO Balding 大老板
@BaldingsWorld
One of my favorite parts of this is most everyone missing the second order effect: Europe wouldn't be doing this if they weren't getting hammered by China; they wouldn't be getting hammered by Europe if US didn't raise tariffs even more.
In other words: a) increased US tariffs