Chinese companies are using distillation technology to transfer knowledge from American models and narrow the gap to six months, reducing costs by up to 90%.

An AI company may spend billions of dollars on computing, data, and researchers to build an advanced model. But what if a competitor could use that model’s answers to train its own model and acquire some of its capabilities at a lower cost?

This is where distillation becomes important. It is a well-known AI technique that, in simple terms, uses the outputs of a more capable model to help train another model. The technique itself is not necessarily unlawful, but the dispute begins when the outputs of a commercial model are used extensively and without authorization to develop a competing model.

That is precisely what has placed “distillation” at the heart of the escalating competition between American and Chinese AI companies.

Millions of Conversations Turn into Training Data

In June 2026, Anthropic accused operators linked to Alibaba and its Qwen lab of using approximately 25,000 fraudulent accounts to conduct more than 28.8 million exchanges with the Claude model between April 22 and June 5.

Anthropic said the aim was to extract capabilities from Claude that could be used to develop competing models. Alibaba did not comment on the allegations in the report published by Reuters at the time.

But the issue later expanded. In a report published by Anthropic in September, the company said it had detected a larger campaign between May and July that it attributed to Alibaba. The campaign involved more than 151 million exchanges, with activity peaking at nearly 3 million exchanges per day. It said that, according to its investigation, some of the data was used to help develop reasoning capabilities in Qwen models. These figures and findings remain based on an investigation conducted by Anthropic itself.

Why Is This an Economic Issue?

The economic value of an advanced model does not lie in the software alone. Behind it are massive investments in data centers, chips, energy, researchers, and training and testing processes.

Therefore, when another model can learn from millions of answers produced by an advanced model, it may benefit from some of the knowledge that required substantial investment to build, without repeating every step of the development process from scratch.

Put more simply:

The first company pays the cost of discovering the knowledge, while the competitor tries to reduce the cost of learning it.

Here, distillation shifts from a technical matter to an economic question concerning who bears the cost of innovation and who can turn existing innovation into a competing model at the lowest cost and in the shortest time.

But Distillation Does Not Create an Advanced Model from Scratch

It is important not to overstate what distillation can do.

It does not eliminate the need for data, computing, researchers, and infrastructure, nor does it mean that the receiving model will become an exact copy of the original. But it can accelerate the acquisition of specific capabilities, particularly when carried out at scale and using high-quality outputs.

That is why model outputs themselves have become an economic asset that should be protected, just as companies protect software, patents, and trade secrets.

From Selling Intelligence to Protecting Intelligence

This paradox creates a new challenge for AI companies.

They want to make their models available to millions of customers and developers because usage generates revenue. At the same time, however, they need to monitor usage patterns that may be intended to collect massive quantities of outputs for training a competing model.

Anthropic says it uses systems to detect campaigns relying on fraudulent accounts and block them when discovered. This shows that competition in AI is no longer about who builds only the most powerful model, but also about who can protect the knowledge produced by its model.

The Takeaway

In the traditional economy, a company might protect its factory, patent, or manufacturing secrets.

In the AI economy, the model’s answers themselves may become part of a company’s intellectual capital.

That is why the “distillation war” reveals a deeper shift in technological competition: the race is not only about who has the best model, but also about the cost of accessing capabilities, the speed of developing them, and the ability to protect them from becoming training material for other competitors.

This may be the most important economic equation:

Lower learning costs + faster development = greater pressure on advanced-model owners.

This, more than the idea of “copying a model,” is what makes distillation an economic issue worth watching.