The pricing gap between Chinese and American AI models isn’t a rounding error. It’s a structural shift. And it’s happening whether we’re ready or not.
Yesterday I wrote about being entrenched in Claude Co-Work and whether Kimi K3 might make me rethink my setup. Today I want to talk about the money. Because the numbers are genuinely shocking.
Token economics in plain English
AI models charge by tokens. A token is roughly three-quarters of a word. When you ask an AI to write something, it generates output tokens. When you feed it information, it processes input tokens. Models charge per million tokens.
This is where things stand right now.
Anthropic’s Fable 5 costs $50 per million output tokens. That’s the model I use daily in Claude Co-Work. It’s powerful. It’s also expensive.
OpenAI’s GPT 5.6 Sol sits in a similar price range. Premium American models hover between $30 and $60 per million tokens.
Then there’s Kimi K3 from Moonshot AI in Beijing. $15 per million output tokens. Same ballpark of performance as Fable 5. A third of the price.
And if that’s not cheap enough for you, there’s DeepSeek. Their V4 model costs $0.87 per million tokens. Not $87. Eighty-seven cents. They’re practically giving it away.
Why Chinese models are so cheap
Several reasons. And none of them are “because the models are worse.”
First, power. China has invested massively in energy infrastructure. New data centres get built fast. In the US, new data centres face political resistance over grid strain and water use. Cheaper power means cheaper AI.
Second, the chip squeeze. US export controls were designed to slow China’s AI progress by cutting off access to advanced Nvidia chips. Instead, Chinese labs got efficient. They learned to squeeze more performance out of less capable hardware. That efficiency translated directly to lower running costs.
Third, Chinese companies can run on locally-made chips. For the price of one Nvidia chip, a Chinese AI company can buy ten Huawei chips. Different economics entirely.
Fourth, they’re willing to lose money. Chinese AI companies are sacrificing profit margins to grab market share. They want their models to become the default. Cheap now, dominant later.
And fifth, open-source. Almost every Chinese AI model is released as open-weight. That means anyone can download the model and run it on their own hardware. No licensing fees. No per-query charges. Just the cost of the hardware and electricity to run it.
Think of it like a recipe.
A normal AI model is like going to a restaurant. You order, they cook, you pay per meal. You never see the recipe. You can’t cook it at home. You’re dependent on the restaurant being open and the chef being available.
An open-weight model is like the chef publishing the exact recipe online for free. You can download it, buy your own ingredients, and cook it in your own kitchen. No one can cut you off. No per-meal charge. You just pay for your own groceries and electricity.
That’s why it matters for the trust question. If you run a model locally on your own hardware, your data literally never touches a server in Beijing. The model is just math running on your machine. The catch is you need serious hardware to run it — we’re talking expensive GPUs, not a laptop. So in practice most small businesses would still use the hosted API version, which is where the data flow concern comes back.
So why is something so valuable and nutso powerful FREE?
It’s a land grab. Same play as Android. Google gives the operating system away because if every phone manufacturer uses Android, Google owns the ecosystem. The money comes later. App store fees. Search ads. Data.
Kimi K3 is doing the same thing.
Most people can’t actually run it locally. You need expensive GPUs. So even though the model is free to download, most businesses end up paying for the hosted API anyway. The free version is a gateway drug.
Developers build on it. Learn it. Trust it. Recommend it. Once your whole workflow is built around a model, switching costs are real. I wrote about this in Part 1 — that’s exactly how I got stuck in Claude Co-Work.
And it undercuts US competitors. If Kimi K3 is free and nearly as good as Claude at $20 a month, every American AI company has to drop prices or lose users. It’s a competitive weapon disguised as generosity.
Then there’s data. The people who use the free hosted version are feeding their prompts and data back to Moonshot AI. That data improves the next model. Free users are basically unpaid beta testers.
And geopolitically, China wants global AI dominance. If developers worldwide are building on Chinese models, China gains influence over the global AI stack. It’s strategic, not just commercial.
So the free part is really a customer acquisition cost. They’re buying market share with it.
The GPU crunch
Kimi K3 was so popular that Moonshot AI had to pause new subscriptions after 48 hours. Demand overwhelmed their computing capacity. They literally couldn’t keep up.
That tells you two things. One, the demand for cheaper AI is enormous. People are hungry for alternatives to expensive American models. Two, Chinese labs still have hardware constraints. The chip export controls aren’t completely toothless.
What this means for small business
Right now, all top five most-used models on OpenRouter (a popular AI model marketplace) are Chinese. Tencent, Xiaomi, DeepSeek, MiniMax, and z.ai. Not OpenAI. Not Anthropic. Not Google.
Microsoft is already testing Kimi K3 for Copilot on Azure. GitHub Copilot added Kimi K2.7 as its first open-weight model at $0.95 per million tokens. The Chinese models are quietly embedding themselves into US tech infrastructure.
DoorDash is pushing lower-level work to Moonshot’s Kimi model. Their CTO said it delivers better quality at cheaper cost. That’s a US tech company publicly saying a Chinese AI model beats their American option on value.
If you’re a small business spending hundreds a month on AI subscriptions, this matters. The same capability is available for a fraction of the cost. The question isn’t whether the pricing gap is real. It clearly is. The question is whether you’re comfortable with where that capability comes from.
That’s tomorrow’s post.
The China Question series
Part 1 — I’m So Deep in Claude Co-Work I Can’t See the Surface
Part 3 — Should You Trust a Chinese AI With Your Business? I’m Genuinely Asking.
Frequently Asked Questions
Why are Chinese AI models so much cheaper than American ones?
Lower power costs, chip export controls forcing efficiency, willingness to sacrifice profit for market share, and open-source licensing all contribute to dramatically lower prices.
Is DeepSeek really only $0.87 per million tokens?
Yes. DeepSeek V4 is priced at $0.87 per million output tokens. It’s the cheapest frontier-level model currently available, though it runs on Chinese-made hardware and has its own performance trade-offs.
Are Chinese AI models actually being used by US companies?
Yes. DoorDash uses Moonshot’s Kimi model, Microsoft is testing Kimi K3 for Copilot, and all top five models on OpenRouter are currently Chinese.
Written by a human. Affiliate links may apply. Nuffin’s for free.
