Silicon Valley's Favorite Pricing Trick Doesn't Work on AI, and Cheap Chinese Models Are Why

I want to flag upfront that this is my read on where the AI market is heading, not a settled fact. But I keep running the numbers and landing in the same place, so here's the argument.

The playbook, as I understand it

Amazon sold books and e-books below cost for years to bury the competition, then leaned on the position it had bought once independent bookstores and smaller retailers were gone. Uber ran the same shape of play in ride-hailing: subsidize fares below what a ride actually costs, force Lyft and local taxi operators into a price war they couldn't win, then let fares drift back up once the alternatives had shrunk. Neither company invented this. What made it work in both cases is the same thing: once you've burned the competitors down, the customer's options are gone with them, and there's nowhere else for the money to go except back to you.

I think every major AI lab raising money in the last three years has been pitching some version of this same story to investors, whether they said it out loud or not. Burn billions on compute, subsidize inference well below what it actually costs to run, get developers and enterprises dependent on your API, and eventually raise prices once you're the default choice nobody wants to migrate away from.

Why I don't think that story survives contact with what's actually happening

The problem, as I see it, is that the thing standing between OpenAI and Anthropic and their competitors isn't a warehouse network or a fleet of drivers. It's a model, and a model can be copied, downloaded, and improved by someone else for free.

OpenAI lists GPT-5.5 at $5 per million input tokens and $30 per million output tokens. Anthropic's Claude Sonnet 4.6 runs $3 and $15. DeepSeek's V4 Flash lists at $0.14 per million input tokens and $0.28 output, and its V4 Pro model comes in at $0.435 and $0.87. That's not a modest discount, that's the same category of task at somewhere between a tenth and a fortieth of the price. One widely cited workload comparison put the same job at $4,811 through Claude, $3,357 through ChatGPT, $1,071 through DeepSeek, $948 through Moonshot's Kimi, and $544 through Zhipu's GLM. That's not a coupon. That's a different cost structure entirely.

And this isn't a one-time undercut that gets fixed once OpenAI matches the price. DeepSeek, Alibaba's Qwen team, Moonshot, and Zhipu keep shipping new open-weight versions that close the capability gap further with every release, and each one is downloadable and self-hostable the day it drops. You can't starve out a competitor by outspending them when the competitor's whole model is sitting on Hugging Face for anyone to pull down, run, and never pay a subscription for again.

The switching cost is the part I think everyone's underpricing

Amazon's predatory pricing worked because once a bookstore closes, it doesn't reopen. Uber's worked because building out a rival driver network from scratch takes years and real capital. The switching cost was the whole mechanism.

For a foundation model, the switching cost for a lot of real workloads is closer to changing an API key. Lindy, a San Francisco AI assistant startup, reportedly moved workloads from Anthropic to DeepSeek and said it saved millions of dollars, while still routing through a US-based provider to keep its data domestic. Reporting this year has pointed to a meaningful share of traffic on OpenRouter, the API marketplace that lets developers swap models with a config change, now going to Chinese open-weight models. If the thing you're switching away from is a single line in a config file, the lock-in that predatory pricing depends on never actually forms.

I think that's the whole story in one sentence: the moat that made the old playbook work was never really the low price, it was what the low price bought you, permanently gone competitors and customers with nowhere else to go. Neither of those conditions holds when the competing product is open-weight, improving every few months, and backed by companies that don't need this specific product line to ever turn a profit.

Why I don't think the labs can just wait this out

OpenAI is reportedly weighing steep token price cuts of its own, and Anthropic is expected to follow, both while quietly preparing IPO paperwork. I don't think that timing is a coincidence. It's hard to sell public investors on a story about durable pricing power in the same quarter you're cutting prices because a lab in Hangzhou or Beijing just made your flagship model look overpriced by a factor of ten.

The honest version of the pitch these companies are making to investors, as I read it, was always "trust us to be the eventual toll booth." I don't think you get to be a toll booth on a road that a competitor can pave for free right next to yours, and keep repaving, indefinitely. Companies like DeepSeek and Alibaba aren't racing to be the most profitable AI vendor in the world. That's arguably a national strategic project for them, not a startup trying to earn back its burn rate, which means they have no obvious reason to ever raise their prices the way the old playbook requires.

I'm not arguing OpenAI or Anthropic are going out of business. Their frontier models still lead on plenty of hard benchmarks, and enterprises with real compliance and data-governance requirements aren't rushing to self-host a model out of Shenzhen. But the specific bet that got them funded, that scale plus capital would eventually buy the kind of pricing power Amazon and Uber got, looks a lot shakier to me than it did two years ago. I think that's the quiet story sitting underneath every one of these price cuts, and it's worth watching what happens to both companies' IPO pitches once investors start asking the same question I keep asking myself.

Sources: pricing comparisons drawn from published rate cards at OpenAI, Anthropic, and DeepSeek; workload cost comparison and IPO context via Crypto Briefing; enterprise adoption reporting via TechRepublic, citing Axios reporting on Lindy's migration to DeepSeek and OpenRouter usage share; background on Amazon and Uber pricing strategy from public antitrust and business reporting.