Has China Found AI’s Cheat Code?

Earlier this month, one of America’s top AI companies made an unusual request.

It asked Congress for help.

Not because the company was running out of computer chips or because it needed more electricity. But because it believes America’s lead in artificial intelligence is becoming increasingly difficult to defend.

This is a different kind of AI bottleneck, and it might be a LOT harder to solve.

Until now, the biggest challenge facing AI companies has been building smarter AI.

But the next challenge may be stopping Chinese copycats from selling similar intelligence for a fraction of the price.

AI Copycats

Almost every major breakthrough during the AI boom has come with an enormous price tag.

OpenAI, Anthropic, Google (GOOGL), Meta (META) and xAI have collectively spent tens of billions of dollars training increasingly powerful AI models. That’s because building one of these frontier systems requires vast amounts of computing power, specialized chips and electricity.

And for a long time, that level of spending seemed like an unbeatable advantage.

After all, if it costs tens of billions of dollars to build the smartest AI in the world, how many companies can realistically compete?

That’s why DeepSeek shocked Silicon Valley and Wall Street last year.

The little-known Chinese startup released an AI model that rivaled many of America’s best systems. Even more surprising, it claimed to have built it using a fraction of the computing power experts thought was necessary.

The news erased hundreds of billions of dollars from AI stocks and raised an uncomfortable question:

What if building the smartest AI doesn’t always require starting from scratch?

Last week, that question came roaring back when Moonshot AI released its new open-source Kimi K3 model.

Many developers said it performed surprisingly close to America’s newest frontier models while costing a fraction as much to use. And the news sparked another selloff in AI infrastructure stocks.

Ironically, Moonshot soon had to stop accepting new subscribers because it ran out of computing power. Even the company disrupting the AI race couldn’t escape the infrastructure bottleneck.

But DeepSeek and Kimi raise an even bigger question.

How have Chinese companies closed the gap so quickly?

OpenAI accused DeepSeek of using its AI models to accelerate the company’s own development.

And Anthropic is now making similar claims against several Chinese companies, including Alibaba (BABA).

This concern led to Anthropic’s Head of Policy reaching out to Congress this month.

According to Anthropic, Chinese companies created roughly 24,000 accounts to access Anthropic’s Claude chatbot. Those accounts generated more than 16 million conversations with the system.

But Anthropic believes they weren’t ordinary customer interactions.

Instead, the company says the conversations were carefully designed to train competing AI models by learning from Claude’s responses.

In the AI industry, that process is known as distillation.

Rather than training a model entirely from scratch, researchers ask a more advanced AI millions of questions. Those answers become training data for another model.

So the original AI becomes the teacher, and the new AI becomes the student.

There’s nothing new about AI distillation. Google researchers first developed it more than a decade ago to build smaller, more efficient AI models.

Geoffrey Hinton, often called the “Godfather of AI,” helped pioneer the approach.

Image: Vaughn Ridley/Collision via Sportsfile

Back then, distillation was simply a way to build smaller, more efficient AI models. But today it’s becoming one of the industry’s biggest security concerns.

And it doesn’t just affect the AI race between the United States and China.

Earlier this year, Elon Musk was asked under oath whether his AI company, xAI, had ever distilled technology from OpenAI.

His answer was surprisingly candid: “Generally AI companies distill other AI companies.”

So the real issue today isn’t that distillation exists. It’s that nobody has figured out how to stop it.

But why does any of this matter?

Because it seems the AI industry could be running into a bottleneck that no amount of computing power can solve.

Over the past two years, we’ve watched the biggest technology companies in the world pour staggering sums into artificial intelligence.

And they did it on a simple assumption. The more you spend, the further ahead in the AI race you’ll stay.

But what if that spending advantage doesn’t last as long as everyone expects?

If competitors can dramatically shorten the learning curve by studying the behavior of frontier AI models, then building the smartest system may only be half the battle.

Protecting it is the other half. And that’s a much harder thing to do.

Here’s My Take

The AI battle is quickly becoming something much bigger than America versus China.

It’s becoming open source versus closed source.

Chinese companies are betting that powerful AI will become cheap and widely available. But OpenAI, Anthropic and the other big AI companies are betting that businesses will keep paying for models they trust.

Right now, that bet is still in its early stages. Deloitte recently found that only 11% of U.S. companies have AI agents in production.

That means most businesses haven’t made their choice yet.

At the end of the day, I think one of two things happens. Either the Trump administration bans American companies from using advanced Chinese AI models… or businesses willingly pay more because they trust OpenAI and Anthropic more than China.

That’s why Anthropic went to Congress.

Because protecting America’s AI lead could prove even harder than building it.