one way to look at it:
Toyota ™
- Market cap: ≈ $226.8 billion (USD)
- Revenue: ≈ $334.3 billion (USD)
- Net profit: ≈ ¥2.36 trillion → ≈ $13.0 billion
Ferrari (RACE)
- Market cap: ≈ $95 billion (USD)
- Revenue: ≈ $8.4 billion (USD)
- Net profit: > €1 billion → ≈ $1.15 billion
but more importantly: perhaps it doesnt have to be a competition. everyone wins either way.
Well these companies have a very specific business model which is to rent out LLMs as a service, and that business model is now going away.
in openAI’s case, 70% of their cost is related to researching, training, and aquiring hardware to train new SOTA models.
hypothetically, they could just put a pause on that anytime and their profit would skyrocket.
i know they wouldnt do that.
more likely just lower their prices a bit to remain competitive.The whole justification for charging their sky high prices is that their model is genuinely more capable than others. If they put a pause on R&D, then DeepSeek will outpace them in a few months charging a fraction of what they do.
The real problem with the American AI business model is that LLMs are a generic commodity. It’s very hard for any one vendor to differentiate themselves from others. And in such scenarios, the profit trends to zero. There’s a race to the bottom where people keep lowering their margins because everyone else is.
So, this is why American labs want to create a cartel and control prices. They can’t afford the margins to drop to the cost of inference because they’re on a hook for insane amounts of money. So, they need the profit rate to stay high to justify themselves as a business.
Meanwhile, Chinese companies don’t have the same problem because they didn’t spend like mad on this tech. And price of electricity in China is also around a third of what it is in the US.
These charts are very illustrative incidentally:
https://hai.stanford.edu/news/inside-the-ai-index-12-takeaways-from-the-2026-report
Even the low tier models have seen huge ongoing improvements, both much smarter and much faster, on the exact same consumer hardware. Great news for anyone who doesn’t want to see these big US corporations form an AI compute oligopoly.
Indeed, the low end is what I’m most excited about. Qwen 3.8 27b is already comparable to Opus 4.5, so if next version is as good as Fable, that’s good enough for vast majority of things people need to do. And while I’m really thankful that Chinese companies choose to release their models in the open, its far better to run your whole tool chain locally end to end.
Absolutely hilarious. Another race to the bottom. Enshittifying the shit. Oh did you think as a consumer you were even gonna get the primo slop? The choice slop? No, you get the affordable grade slop that will continue to be ratcheted down as the price squeeze towards profitability continues.
