A recent report explains a shortcut called distillation, and why it matters as the country argues over where to build the data centers that run AI.

Demand for artificial intelligence is growing fast. By one industry estimate, the amount of work Google’s AI systems handle has grown about 330 times since May 2024.

In September, Anthropic, the company that makes Claude, released a report describing how some Chinese companies are trying to catch up to the leading American labs without doing all the work themselves. The method is called distillation. Here is how it works: A company sends a leading American AI model millions of questions, saves the answers, and uses them to train its own model to answer the same way. Companies often use distillation on their own models. It becomes a problem when someone does it to another company’s model in secret, using fake accounts.

According to Anthropic, operators tied to Alibaba, one of China’s largest tech companies, sent Claude more than 151 million requests between May and July of this year. At the busiest point, that was close to 3 million requests a day from more than 3,500 fake accounts. When Anthropic shut down one group of accounts, the requests moved to another group. The report says DeepSeek, Moonshot, Zhipu and Xiaomi ran similar operations.

The copying causes problems beyond lost business. Anthropic says a copied model does not keep the safety protections built into the original, yet it can still help someone develop dangerous skills, including skills useful for cyberattacks. The report also found that some of these companies secretly passed their own customers’ requests to Claude without telling them. Those requests included private company documents and passwords, and some came through services that are popular in the United States and Europe.

At the same time, Americans are debating where data centers should go. In July, New York Governor Kathy Hochul ordered a pause of up to one year on new large data centers while the state writes new rules. Neighbors have fair questions about construction and about who pays for new power lines. Yet places that have welcomed data centers with proper safeguards have seen real benefits. Loudoun County, Virginia, is the largest data center market in the world. Data centers there paid about $894.5 million in property taxes in 2025, and the county says they pay about $26 in local taxes for every $1 of public services they use. That money helps keep homeowners’ tax bills down.

Staying ahead in AI depends on both of these issues: building enough capacity to run the most advanced models, and stopping others from copying those models through fake accounts. Neither issue belongs to one party. Congress can crack down on the services that sell access to American models under false identities. States can set clear rules for where data centers go and make sure host communities share in the benefits.