A week ago I posted one line that got more argument than anything else I have written this year. The AI bubble can burst while AI keeps improving. The post did 23,096 impressions, 465 reactions and 150 comments, and most of the replies wanted to fight the premise. People kept reading it as a claim that AI is about to get worse. That is not what I said. The models get better every month. The thing that can break is the money underneath them, and those are two different stories.
I run Growth Cab, a go to market advisory. We put AI inside outbound, research and reporting for B2B teams selling contracts above fifty thousand dollars a year. So this is not a stock call and I am not a macro investor. I care about the question in the title for a boring operational reason. Most of my clients now run part of their revenue engine on top of a frontier model, and the price of that model is set by a market that has started to behave strangely. When people ask me whether the AI bubble can survive a price war with China, they usually mean it as doom. I read it as a planning problem.
in
“The AI bubble can burst while AI keeps improving. The warning sign isn't weaker AI. It's slower revenue growth.”
Can the AI Bubble Survive a Price War With China?
Start with how a bubble actually breaks, because the headline version hides it. A bubble does not need the product to get worse. It needs the growth rate to slow. Look at housing before 2008. FHFA data show US home prices were still at record highs in 2006. Nothing had crashed. What had changed was the annual growth rate, which had fallen from around fifteen percent to around eight percent. Federal Reserve research later described how, once that growth slowed, borrowers who were counting on refinancing could no longer do it. Loans began failing before house prices ever fell. The crack showed up in the growth rate first, quietly, while the headline number still looked fine.
AI is a different asset, and the wiring still rhymes. The enormous spending on data centers only pays back if AI revenue keeps growing fast enough to justify it. Flat is not enough. It has to stay fast. That is the number that has to hold. And the fastest way to slow revenue growth has nothing to do with a weaker model. It is a cheaper one that does the same job well enough.
This is where China enters. In July, Moonshot AI released Kimi K3, a 2.8 trillion parameter open model that ranked at or near the top US systems on several coding and agent benchmarks while running at roughly a third of the cost of the western frontier. Kimi is not alone. Reporting through August described a full price war inside China, with API tokens trading at something like an eighty percent discount to US rates and close to a thousand models fighting over the same customers. When a capable model gets that cheap, it does not have to win on quality. It only has to be good enough that your customer asks why they are paying four times more for a few extra points on a benchmark.
I watch this happen inside client accounts already. A team runs its enrichment and its first draft outreach on a top tier model, because that is what it started on. Then someone tests a cheaper open model on the same task, gets a result that is ninety percent as good at a quarter of the price, and the swap takes an afternoon. Multiply that one decision across every company that builds on AI and you get the thing that actually threatens the bubble. Sales grow more slowly because prices fall. Investors expect less. The next funding round gets harder. Better AI rescues none of that once the revenue line bends.
The Warning Sign Is the Growth Rate
This is why I keep telling people the warning sign to watch is not model quality. Model quality is going straight up and it is the most visible number in the whole industry, which is exactly why it distracts everyone. The number that matters is the revenue growth rate at the companies selling AI. It is less exciting, it moves slowly, and it is the one that went quiet in 2006 while everyone was still admiring record prices. If that rate keeps falling while spending keeps climbing, you have the same divergence that broke housing, dressed in newer clothes.
What This Changes for the Rest of Us
If you are an operator rather than an investor, the useful move is to stop treating your model choice as permanent. Three things I now do at Growth Cab because of all this. I keep the model layer replaceable, so the prompts, the data and the workflows live outside any single provider and I can move to a cheaper model in a day if the gap opens. I price my own service on the outcome I deliver rather than on the tool I use, because the tool is going to keep getting cheaper and my client will find that out on their own. And I keep asking one question. If the cost of intelligence falls to almost nothing, what is still scarce? The answer never changes. The data you own, the distribution you built, and the trust you have with a buyer. None of those get cheaper when a Chinese lab ships a good open model.
Where This Argument Is Weak
I owe you the holes, because I got sharp pushback on the post and some of it landed. The housing analogy is not clean. A house is a leveraged asset that millions of people borrowed against, and a failed refinance forces a sale into a falling market. A model API is nothing like that. If Kimi gets cheap, buyers simply save money, and cheaper inputs can grow a market instead of shrinking it. That is the honest counter and it might be right. Lower prices could pull in so much new usage that total revenue keeps climbing even as the price per token drops. Nobody knows which force wins yet.
The second weak spot is that cheaper is not free to adopt. Kimi K3 is a 2.8 trillion parameter model. Running it yourself takes serious hardware, and most companies will never do that. They will keep paying a US provider for the convenience, which protects those prices more than a raw benchmark suggests. So the price war is real, and the distance between a cheaper number on a chart and a cheaper bill inside a real company is wider than any doom post admits.
So my honest answer to the title is not a clean yes or no. The bubble survives if AI revenue keeps growing fast enough to outrun the falling price. It cracks if the price falls faster than the volume grows. What I am sure of is where to look. The signal is not the model, which keeps getting better in public. It is the revenue growth rate at the companies selling AI, the quiet number that bent before 2008, and it is worth checking every quarter from here.
Every Thursday, AI Frontier gives you one signal, my read on it, and one practical play from the AI and go to market systems we run inside Growth Cab, all in under five minutes. Drop your email below and confirm your subscription to get the next edition. And if you think I am wrong about where this argument breaks, tell me on LinkedIn. I answer everything.
