China's AI Models Are 95% Cheaper, and the Sanctions Backfired

It's Wednesday, September second. A few months ago I was looking at my AI bills and thinking, this is stupid. Why am I paying this much to ask a chatbot questions? So I tried one of the Chinese alternatives. DeepSeek. The quality was the same. It wrote code just as well. It answered questions just as clearly. And it cost ninety-five percent less.

I had the same reaction you probably just had. Chinese. Cheap. We are trained to assume that means worse. But Chinese AI models now account for sixty-one percent of all AI usage worldwide. They match or beat American models on coding, math, and reasoning. They cost roughly a twentieth of what OpenAI and Anthropic charge. And today, DeepSeek is reportedly closing a funding round that values the company at seventy-four billion dollars.

This is not a sale. This is a structural difference that the US government accidentally created. And it raises a question that most coverage skips: when you use these models, where does your data actually go? A token is just a chunk of text, roughly three quarters of a word. So a million tokens is about seven hundred fifty thousand words, or a couple thick novels. When you ask ChatGPT a question, OpenAI charges you based on how many tokens go in and how many come out.

OpenAI charges roughly two dollars for every million words you send it and fourteen dollars for every million words it sends back. Anthropic is three dollars in and fifteen out.

DeepSeek charges twenty-eight cents for input and forty-two cents for output.

That is not a typo. It is roughly one twentieth the price. For the same quality. Independent benchmarks show DeepSeek and Qwen, which is Alibaba's AI, matching or beating the American models on coding, math, and reasoning. DeepSeek's model scored ninety-seven percent on the standard math reasoning test. These are not budget alternatives that get you eighty percent of the quality. They are competing at the top.

So how did a company from Hangzhou that spent six million dollars training its model match competitors that spent hundreds of millions? In October 2022, the US banned exports of the most advanced AI chips to China. NVIDIA could not sell its best GPUs to Chinese companies. The logic was simple: no hardware, no AI.

Chinese labs responded by getting more efficient. They had to.

DeepSeek trained its model on a few thousand slightly limited NVIDIA chips that were still allowed. Training cost: five point six million dollars. American labs spend two hundred to five hundred million dollars to train a single model. Same result. One fiftieth the cost.

How? An architecture called Mixture of Experts. Instead of firing every part of the model for every question the way GPT-5 does, DeepSeek only activates a small fraction, the parts that are actually relevant to what you asked. Imagine a hospital where every doctor in the building has to weigh in on every patient. DeepSeek has the reception desk route you to the right specialist. Same answer. Eighty percent less work.

This is not a trick they can match with a better cloud contract. It is baked into the design.

The other half is that DeepSeek and Qwen release their models as open weights. Anyone can download and run them. A compressed version fits on a single high-end graphics card. The electricity is less than a penny per million words. No subscription. No data going anywhere.

The sanctions meant to choke Chinese AI produced the most efficient AI designs in the world. Here is the part most people miss. The privacy setups of these companies are not even close to equivalent.

DeepSeek stores everything on servers in China. Every message you type, every file you upload, your email, your phone number, your IP address. The policy even covers keystroke patterns, a behavioral signal that a VPN cannot hide. Chinese law governs that data, which means Chinese law enforcement can compel access to it.

When the app went viral in January 2025, the response was immediate. Italy banned it within three days. More than a dozen European countries opened investigations. Governments from the US to Australia banned it from official devices. No consumer AI product has been banned by this many governments this fast.

But here is the thing. The American companies are not saints. OpenAI trains its models on your ChatGPT conversations by default, unless you know to dig into settings and turn it off. Most people never do. Your chats are training data. OpenAI also keeps them for up to three years.

Anthropic's Claude is the outlier in the consumer's favor. It does not train on paid subscriber data. It deletes conversations after thirty days. At twenty dollars a month, it is the strongest consumer privacy deal in AI right now.

The most important distinction, though, is between the app and the model. The app, any app, sends your data to some company's servers. The model, if you download the open-weight version and run it yourself, sends your data nowhere. DeepSeek the app is a privacy risk. DeepSeek the model, running on your own machine, is as private as a Word document. The data never leaves your computer. A policy designed to contain Chinese AI backfired so completely that the company it was supposed to stop is now worth seventy-four billion dollars and its models power the majority of AI usage on the planet.

The privacy tradeoff is real, but it is also avoidable. If you trust yourself more than you trust any of these companies, you can run the open-weight model locally and keep your data off every server. That was not possible five years ago. It is trivially possible now.

This is the pattern every time. Linux beat proprietary operating systems. PostgreSQL ate Oracle's lunch. In every infrastructure wave so far, open and cheap has won. AI looks like it is following the same script, just faster. Stay informed, stay curious, and we'll see you tomorrow.