Frontier AI’s Dinosaur Problem

Chinese AI companies cannot die because their government protects them, allowing them to survive a price war by winning the API battle while frontier AI firms like OpenAI and Anthropic face diminishing returns from running out of training data and burning money in a Cold War-like competition, according to an analysis by an unnamed commentator.

Unfortunately, I think the Chinese AI companies cannot die. Because their government protects them. Claude and ChatGPT never made it into the Chinese market. And the Chinese market is big enough to feed them well. Look at how well WeChat/Alipay/Pinduoduo are doing, despite having approximately 0% global market share. In a price war, they cannot win the subscription battle, but they can win the API battle. Humans subscribe to ChatGPT, but humans get tired. They need to sleep. They forget to open the app. These people subsidize the heavy users. So the economics work out for OpenAI. But APIs are used by apps. Apps don’t get tired. Apps never forget to make the HTTP request. If you sell the API at a loss, the more your customers use it, the more money you lose. This is why API pricing tends to be expensive https://platform.claude.com/docs/en/about-claude/pricing . And why everyone uses DeepSeek for their SaaS products. The frontier AI companies also have a problem. They have run out of text on the internet for training. So now they have to do a lot of weird things. Like download and transcribe YouTube videos https://www.theverge.com/2024/4/6/24122915/openai-youtube-transcripts-gpt-4-training-data-google and risk getting banned by YouTube , pirate ebooks to scan them https://www.theguardian.com/technology/2025/jan/10/mark-zuckerberg-meta-books-ai-models-sarah-silverman , destroy physical books to scan them https://arstechnica.com/ai/2025/06/anthropic-destroyed-millions-of-print-books-to-build-its-ai-models/ , generate synthetic training data. And they are hitting diminishing return. But the Chinese companies can just keep copying the homework for cheap. Their models were always about 6 months behind the frontier. Then Kimi K3 https://www.kimi.com/blog/kimi-k3 came out and people said it’s just as good as Claude Fable 5 https://twitter.com/arena/status/2077824029126504525 . Yeowch. What can Tom do in a cat-and-mouse game when Jerry is protected by Spike the bulldog? Then they open-source their models https://www.scmp.com/tech/tech-war/article/3361142/why-chinas-open-weight-ai-model-kimi-k3-sparking-anxiety-silicon-valley . Then smaller American companies distill from those https://stratechery.com/2026/whos-afraid-of-chinese-models/ . Distillation is the elephant in the room. Every major AI company’s terms of service prohibit it. But terms of service are not laws. If you violate them, the company can… ban your account. The Chinese AI labs have thousands of accounts. Terms of service are like DRM for LLMs. And we all know what happened with iTunes and DRM https://opensource.com/life/11/11/drm-graveyard-brief-history-digital-rights-management-music . Clayton Christensen said this many years ago. Low-end disruption https://en.wikipedia.org/wiki/Disruptive innovation Low-end disruption happens when the performance of the product overshoots the needs of the mainstream market 1. Time is on the side of the Chinese companies. Every year, the frontier models get more capable and The Chinese models don’t need to win. They just need to not lose. They keep copying the homework. They keep staying in the Chinese market. They keep open-sourcing. And they wait. Right now, Anthropic and OpenAI are in a Cold War. Look at how Anthropic keeps Fable 5 accessible on subscription plans https://simonwillison.net/2026/Jul/18/claude-make-fable-5-permanent/ . Look at how frequently OpenAI resets its usage limits https://codex-resets.com/ . They cannot stop burning money, because if you stop and Soviet Union doesn’t and gets results, you die. So they keep building bigger nuclear warheads. “We need to improve our security guardrails.” “Our competitor released a security model https://www.anthropic.com/glasswing .” “Damn, now we need one https://openai.com/daybreak/ too. Go build it.” Dinosaurs ruled the earth for 100 million years. They kept getting bigger. And their necks kept getting longer. Then a meteor hit the earth. The climate changed. The food chain broke. The dinosaurs who needed the most food died first. Then the tiny mammals inherited the earth. Christensen also said low-end disruption happens when the cost structure of the incumbent prevents them from truly moving downmarket. Anthropic and OpenAI have smaller models too. But their cost structures are still expensive frontier R&D and high-end brand positioning. ↩ https://zhenyi.gibber.blog/feed fnref1