Why "Is China Winning?" Is the Wrong Question
Watch on YouTube · subscribe for the daily brief
Everyone asks if China is winning the AI race. Ben Lorica says that's the wrong question. China is stuck in 'involution', hyper-competition where 140 humanoid-robot makers undercut each other into the ground. Terrible for any single company. And it's building a technology ecosystem the West can't match, one bankruptcy at a time.
Your feed has been fed a steady diet of "China is catching up in AI" all year, and it is time for a smarter lens, because the headline framing gets the dynamic almost exactly backwards. The best read on your list today is Ben Lorica's "China's AI Involution," and it reframes the entire question in a way that actually helps you understand why your model prices keep falling and why Chinese open models keep showing up cheaper and better. This is a strategy story, not a threat story, and it is the most useful thing published today.
Start with the word, because it unlocks everything. Involution, from the Chinese term 内卷 (nei juan), describes hyper-competition where everyone has to work harder and improve constantly just to stay in place, while the rewards keep shrinking. Companies cut prices, add features, improve service, adopt every new technique, and then discover that all their competitors did the exact same thing. What looked like an advantage last quarter becomes the new baseline requirement this quarter. Nobody pulls ahead. Everybody runs faster. Margins bleed. Lorica's sharp observation is that AI, robotics, EVs, and e-commerce in China are all deep in this pattern, and that it complicates the comfortable Western story that better technology leads to higher productivity leads to better economics. In China right now, productivity is improving while margins fall. Those two things are supposed to move together. They are not.
The clearest picture is humanoid robots. At China's 2026 World AI Conference, more than 200 robotics companies showed up, and separate counts found more than 140 humanoid-robot manufacturers shipping more than 330 products. Let that sink in. One country, 140-plus companies, all building humanoid robots, all racing each other on price and features, in a market where the durable commercial uses are still being figured out. That is textbook involution, and Chinese industry participants are already using the word to describe it, especially in the lower-end categories where prices are collapsing. From the outside it looks like dominance: look how many robot companies, look how fast, look how cheap. From the inside it is a bloodbath where most of those 140 companies will not survive.
Here is the part that makes it genuinely important rather than just interesting, and it is the insight to actually carry with you. Lorica draws a crucial distinction: involution can produce terrible economics for individual companies while still strengthening the underlying technology ecosystem. When most of those 140 robot startups die, the founders and investors lose. But the component suppliers they created do not vanish. The manufacturing expertise does not vanish. The prototyping capacity, the deployment experience, the trained engineers, the driven-down component costs, all of that stays and compounds. So the West looks at China and sees either "they are winning" or "it is a bubble that will pop," and both are wrong. The truth is stranger and more durable: China is running a brutal, wasteful, company-killing process that is simultaneously building one of the strongest technology ecosystems on earth. The individual companies are the fuel. The ecosystem is what gets built by burning them.
For anyone building on AI, this reframes the whole summer's price story in one move. We have tracked DeepSeek, Kimi, GLM, MiniMax, and Qwen driving open-model prices toward the floor, the "cheaper became better" thread, the Tokenpocalypse, the routing arbitrage. Involution is the engine underneath all of it. Those models are cheap and relentless not because any one Chinese lab found a magic efficiency trick, but because they are all trapped in nei juan, forced to undercut each other quarter after quarter. That is fantastic news if you are a buyer of AI, because it means the downward pressure on model prices is structural and is not stopping, so plan your cost curve around models getting cheaper, not more expensive. It is a warning if you are a seller of anything AI-adjacent, because the same dynamic that crushed Chinese robot-maker margins is the dynamic your product may be walking into. When your feature becomes everyone's baseline in a quarter, you are in involution too, whether or not you use the word.
The rest of the day was builder-and-values texture. On the hardware-independence front, which matters more every week as the memory and GPU squeeze tightens, vLLM shipped a plugin for serving models on Tenstorrent hardware, one more crack in the assumption that AI only runs on Nvidia. It is niche today, but the direction, more silicon options, less lock-in, is the escape hatch this run keeps flagging. OpenAI announced it is supporting independent journalism in Ukraine, which is worth noting plainly as a values-and-PR move rather than a technical one. Damien Benveniste's "AI Digest: Model Validation, Agents, and Generative AI" is a solid practitioner roundup if you want a second synthesis of the week. And Simon Willison had a characteristically busy bench: llm 0.35, a video compressor, a Mercator to Equal Earth map projection tool, a "Creepy crawlies" post, and a quote from Jakub Pachocki, presumably pulling from yesterday's "An Alien Mind" essay, which tells you that piece is still rippling through the practitioner conversation a day later.
The through-line, one sentence: on September 7 Ben Lorica reframed the tired "is China winning AI" question by naming the real dynamic, involution, a hyper-competition where 140-plus humanoid-robot makers and a dozen AI labs undercut each other into margin oblivion, which is terrible economics for any single company and, at the same time, the engine building an ecosystem the West cannot easily match, and the direct reason the model prices you pay keep falling.
One thing to do today: use involution as a lens on two things this week. First, your cost planning: recognize that the relentless drop in open-model and API prices is structural, driven by Chinese hyper-competition that is not going to ease, so build your unit economics assuming intelligence keeps getting cheaper and design your architecture to swap in the next cheaper-and-good-enough model without a rewrite. Second, and more uncomfortable, turn the lens on your own product or team: are you in involution without realizing it, where every feature you ship becomes a competitor's baseline within a quarter and nobody actually pulls ahead? If so, the escape is not working harder inside the race, it is the thing this whole run keeps pointing at, building something a competitor cannot simply copy next quarter, the proprietary data, the accumulated judgment, the compounding loop. Involution punishes everyone who competes on the copyable. Find the part of your work that is not copyable, and move there.
Sources covered
1 story · 9 sources (0 news, 2 company, 7 voices)