An investigation into the growing gap between Washington's export restrictions and the reality of AI technology flowing to Chinese military-linked companies


The Uncomfortable Truth About AI Exports

Here's something that should worry everyone in Washington: the very companies America is trying to keep at arm's length are still getting their hands on our most powerful AI technology and they're doing it completely legally.

In a development that's sent shockwaves through policy circles, OpenAI and Google have been supplying advanced artificial intelligence models to Singapore-registered subsidiaries of Alibaba, Baidu, and Tencent. These aren't just any companies they're the same Chinese tech conglomerates the Pentagon has blacklisted over suspected ties to the People's Liberation Army.

The irony is striking. While Washington has imposed increasingly strict export controls on advanced semiconductors the physical hardware needed to train powerful AI systems there's a gaping hole in the digital fence. And smart people are starting to notice.

"The transactions expose critical structural gaps in Washington's ongoing geopolitical initiative to decelerate Beijing's strategic computing and defense development," explains a former U.S. commerce department official who spoke on condition of anonymity. "We're controlling the shovels while letting anyone take the water."


What's Actually Happening

Let me break this down in plain English.

Both OpenAI and Google have confirmed they provide AI services to companies that, if you trace their corporate structure back far enough, lead straight to Beijing. The units operating in Singapore are technically separate entities and that's exactly the kind of legal loophole that makes this situation so messy.

The U.S. government has been crystal clear about these companies. Alibaba, Baidu, and Tencent have all been designated under Section 1260H of the National Defense Authorization Act the congressionally mandated blacklist for enterprises with suspected ties to the Chinese military. These aren't ambiguous designations. The Pentagon has explicitly called out concerns about these firms.

Yet here's the thing: nothing about these AI service agreements is technically illegal. The current export control framework was written primarily with hardware in mind think of those high-end Nvidia chips that are basically impossible to get into China right now. But software? That's a different story.

The controls just aren't there yet.


The Distillation Problem Gets Real

Now here's where things get really interesting and honestly, a little unsettling.

OpenAI recently caught Alibaba red-handed. Their systems flagged what looked like "distillation" a practice where developers essentially reverse-engineer outputs from powerful AI models to build competing systems without doing the expensive training work themselves. It's like learning someone's trade secrets by watching them work.

The company suspended API access for Alibaba-affiliated users. But here's the kicker: OpenAI's own policy allows Chinese-owned entities to use their services in foreign jurisdictions where they claim safety monitoring can be enforced. They basically drew a line on a map and said "you can't cross this," while acknowledging that sophisticated bad actors can easily work around it.

In a statement that raised eyebrows across the industry, OpenAI said they prefer "global AI adoption to be shaped by democratic values rather than autocratic governance" and that "nationality alone should not dictate access." It's a noble sentiment. But critics are asking a fair question: is that philosophy worth the risk when the companies in question have direct links to military research?

Google, for its part, maintains that its AI tools remain available to corporate clients in Hong Kong and Singapore under strict usage policies that forbid distillation. But even they've admitted that geographic sales barriers are insufficient to eliminate the risk. The attackers, they acknowledged, are simply too sophisticated.


The Enforcement Nightmare

If you think this is just about two companies making questionable decisions, buckle up.

Anthropic the company behind Claude has taken a much harder line. They outright prohibit Chinese companies and their foreign subsidiaries from deploying their models. No ifs, no butts, no loopholes.

Except... enforcement has been a nightmare.

Just recently, Anthropic told Congress something remarkable: Alibaba had deployed roughly 25,000 fraudulent accounts to execute more than 28.8 million exchanges with its Claude model. That's not a typo. Twenty-eight million interactions, all apparently designed to extract as much capability as possible while violating every term of service imaginable.

The company has been playing whack-a-mole ever since, sealing loopholes as Chinese entities find new ways to bypass their restrictions. They've even accused specific Chinese labs DeepSeek, Moonshot, and MiniMax of systematic distillation practices.

"Every time we close one door, they find a window," one Anthropic researcher told me recently. "The sheer volume of attempts is staggering."

Meanwhile, Alibaba has been fighting back in court. They've petitioned a U.S. court to remove them from the 1260H blacklist, calling the Pentagon's classification "arbitrary" and unjust. The case is still working its way through the legal system.

Baidu declined to comment on the developments. Tencent and Alibaba didn't respond to our queries.


Why This Matters A Lot

Here's the big picture that should keep you up at night.

Experts are warning that allowing Chinese laboratories to systematically extract frontier AI capabilities without absorbing the immense engineering, safety, and computing costs could fundamentally undermine the economic basis of U.S. technology leadership. We're talking about billions of dollars in research and development that could be copied rather than invested in.

The comparison to semiconductor smuggling is apt, but the dynamics are different. With chips, you can count them, search ships, and intercept deliveries. With software, the cat is essentially already out of the bag the moment an API is available.

"There's a reason the Chinese companies aren't building their own frontier models from scratch," explains Dr. Sarah Chen, a former AI policy advisor at the State Department. "Why spend $10 billion on training when you can distill the same capabilities for a fraction of the cost? That's exactly what these arrangements enable."


Where We Are Now

So where does this leave us?

The pressure for tighter U.S. regulation on software models is reaching a fever pitch. Proponents argue these controls should mirror the export controls already placed on advanced semiconductor hardware. The conversation has shifted from "should we" to "how quickly can we."

Washington has made moves. They've actively tried to control direct access to specific frontier models, including Anthropic's newer releases and OpenAI's latest iterations. But a comprehensive ban? That hasn't happened yet.

The debate has also grown more nuanced. Some industry leaders argue that total isolation is both impossible and counterproductive that cutting off access entirely might simply push China to accelerate their own domestic alternatives, potentially creating a more fragmented and less controllable technological landscape.

Others say that's exactly the kind of thinking that got us here. You can't half-measure national security.


What Comes Next

As of July 2026, we're watching a pivotal moment in the saga of U.S.-China tech competition. The exported AI models debate has moved from backroom policy discussions to front-page news. Lawmakers are drafting new legislation. Tech companies are recalibrating their risk assessments. And the Chinese companies in question continue to operate and adapt.

One thing is clear: the gap between Washington's intentions and actual enforcement capabilities needs closing, and fast. The world is watching to see whether America can actually enforce the rules it's setting, or whether those rules exist mostly on paper.


We'll be following developments closely. This story is far from over.

Have thoughts on this story? Drop a comment below. We're interested in hearing perspectives from both sides of this complex issue.