The Export Wall Is Leaking: Why China's AI Will Outpace America

01The Model That Shouldn't Exist

Moonshot AI released Kimi K3 barely two weeks ago. It is the largest open-source model on the planet at 2.8 trillion parameters. And the company is already preparing Kimi K4, which according to two people with knowledge of the plan will be significantly bigger than K3, which is already a monster.

There is exactly one thing standing in the way of training something that large: Moonshot needs more of Nvidia's top-end AI chips. Under US export rules, Chinese firms are not supposed to be getting those at all.

And yet, according to three people with direct knowledge, Kimi K3 was trained on Nvidia silicon, including Blackwell, the most advanced compute hardware on Earth. This partially confirms what a senior White House official posted on X. This is not a rumor floating around a Discord server anymore.

The White House's version of events differs from what sources say actually happened. The official claimed K3 was trained at a data center in Thailand, a framing that matters because Chinese access to Nvidia chips in Thailand might be permissible under current US rules. It is the clean explanation, except one of the three sources says K3 was trained partly inside China.

02Stitching Together What America Won't Sell

The engineering is the part that is most impressive. Moonshot could not walk up to one supplier and rent a giant coherent cluster, because nobody in China had enough Blackwell sitting around. So they tapped at least two different Chinese cloud firms that had Blackwell access. How those providers obtained the chips remains unclear, though their access is very likely in contravention of US export rules.

What Moonshot ended up with was a pile of eight-chip Blackwell servers stitched together into something big enough to train on. That is a dramatically smaller building block than what the biggest American labs use for frontier runs. To make it work at all, Moonshot's engineers refined the network design themselves, improving how the chips talk to each other and, critically, across data centers owned by completely different providers.

Multi-provider distributed training on hardware you are not supposed to have. That is the reality of China's AI development in 2026.

And Moonshot is not the only one playing this game. Days after K3 dropped, Alibaba pulled out Quen 3.8, its largest model ever at 2.4 trillion parameters, claiming performance comparable to top US models. Two people say Alibaba trained that one on Nvidia chips too, Blackwell included. Rewind to December, when The Information reported DeepSeek was training its next model on Blackwell chips that had been smuggled into the country. That model, V4, came out earlier this year.

03The Distraction: Silicon Valley's Revenue Problem

While China's labs keep shipping bigger models on smuggled chips, American AI companies face a more insidious problem: the return on investment is shrinking. Building frontier models costs billions in compute, energy, and talent. But the revenue per model generation is not keeping pace.

OpenAI published something this week that illustrates the problem from a completely different angle. GPT-5.6 Sol has cracked long-standing open math problems, including the cycle double cover conjecture. It has beaten Pokemon Fire Red. Then it went to ARC-AGI-3, a benchmark of 2D puzzle games, and scored 7.8 percent.

GPT-5.5 could barely play at all at 0.4 percent. On the leaderboard for one of those games, no frontier model gets past the first level.

OpenAI went digging, inspired by ARC's own writeup on where GPT-5.5 fell short. At first, they saw the same thing ARC saw: the model looked dim. It sat there thinking forever before each action and made almost no progress. But two things in the test harness turned out to be doing the damage.

First, after every single game action, all of the model's private reasoning got thrown away. So each move, it was rebuilding its understanding of an unfamiliar game from zero. Second, the harness used rolling truncation: once the conversation passed 175,000 characters, the oldest messages started falling off. No memory of its thinking, and gradually no memory of its actions either.

They rebuilt the harness to retain reasoning across turns. Two changes showed up immediately. Sol spent less time deliberating per action, and it got dramatically better at learning over a run and holding a coherent strategy. The score jumped from 7.8 percent to 38.3 percent. Roughly triple, with six times fewer output tokens.

This matters beyond benchmarks. If the biggest American AI lab can spend billions training a model that looks mediocre under one test setup and brilliant under another, the evaluation infrastructure is as important as the model itself. And that infrastructure is not cheap. Every harness tweak, every API migration, every reasoning retention system is engineering time that does not go into making the next model bigger.

04The Inference Squeeze

On the inference side, the picture is different but not exactly comfortable. Moonshot leans heavily on Nvidia's H20 to actually run K3 on the Kimi platform, according to three people with direct knowledge. The H20 is the chip US rules do permit Chinese startups to buy. The company recommends that any cloud firm or customer hosting K3 on their own servers wire up at least 64 server chips. Sixty-four minimum. That is the recommendation, not the ceiling.

The popularity wrecked their capacity. Within 48 hours of release, Moonshot stopped accepting new subscriptions entirely. Their post about it was just: our GPUs are feeling it. They said they would reopen once more compute came online, except finding that compute is going to take a while.

Beijing has been discouraging local companies from using the H20 and pushing them toward Chinese-designed accelerators instead. In theory, fine. The problem is that running something like K3 requires large server configurations connecting dozens or hundreds of chips together. Racks built around Huawei-designed silicon are not widely available yet. Customers are looking at delivery times of up to six months. A government telling you not to buy the chip you can get, and a domestic supply chain that cannot hand you the alternative until next year.

05Nvidia's Two-Faced Position

The reaction stateside has been genuinely split and not in the way you would expect. K3's success pushed the Trump administration to escalate its criticism of Chinese models, but Jensen Huang went on Axios and said American companies should absolutely be allowed to use Chinese open-source models like K3. Then late last week, Huang made his first ever post on X, and used it to share a letter signed by Nvidia and 24 other companies about the importance of open-source models.

His line was that the world needs both frontier closed models and frontier open models, and the industry mostly agrees with him. The vast majority of American cloud firms, tech incumbents, and startups are in favor of letting US businesses use Chinese open-weight models. The dissenters are a small group, mainly Anthropic and OpenAI, who have pushed for restrictions or regulation over what are described as theoretical national security and cybersecurity concerns.

Nvidia's position is not complicated. The company sells chips to both sides. Every Chinese lab training on Blackwell is buying Nvidia hardware, whether through authorized channels or not. Every American lab training frontier models is buying Nvidia hardware. The export ban creates scarcity, scarcity raises prices, and Nvidia captures the margin. The ban is, perversely, good for Nvidia's business.

06The Distillation Accusation

Some in the US also argue Chinese models were built on American work more directly. The White House official accused Moonshot of using US model output to train K3, the distillation accusation, which comes up basically every time a Chinese lab overperforms.

Whether that is true matters less than whether it can be stopped. Model outputs are text. Text crosses borders freely. If a Chinese lab can query GPT-5.6 through an API, save the responses, and use them as training data for a competing model, export controls on chips are a Maginot Line. The real bottleneck is compute, not data. And the compute keeps finding its way through.

07The Industry Wants to Slow Itself

On Tuesday, more than 1,100 employees across the major labs signed onto an initiative called Pacing the Frontier, asking the US government to back an international effort to build the technical and governance tools needed to deliberately pace frontier automated AI development. Meta, Anthropic, OpenAI, and Google are all represented, with backing from the nonprofits Guidelight AI Standards and Encode AI.

The signature list is not junior staff. Anthropic CEO Dario Amodei and several co-founders. Meta's VP of AI research. OpenAI chief scientist Jakub Pachocki. OpenAI responded on X saying that at some point acceleration in frontier development may get high enough that the world will need to pace the rate of advancement. Anthropic pointed at its own recursive self-improvement research from last month as evidence the tooling is needed so society can prepare. It was also Anthropic that called on major labs last month to consider a coordinated pause, warning that systems could start improving themselves faster than the risks can be managed.

Separately on Monday, Nvidia announced a coalition with Adobe, CrowdStrike, and others to build AI safety and cybersecurity tooling, which follows a recent Hugging Face cybersecurity incident that sharpened worries about autonomous agents running loose.

08Why China Will Exceed Us

The argument is not that Chinese engineers are smarter or that American companies are incompetent. It is structural. Three forces are converging:

First, the export bans are a speed bump, not a wall. Three of China's top labs have already trained frontier models on Nvidia Blackwell. The chips get through via smuggling, third-country cloud rentals, and multi-provider distributed training. Every ban triggers a new workaround within months. The ban raised the cost of compute for Chinese labs, but it did not eliminate access. And the models keep getting bigger.

Second, American AI companies are spending more for less. GPT-5.6 Sol is a marvel of engineering that can solve open math problems and play Pokemon. But its ARC-AGI-3 score reveals that the evaluation and deployment infrastructure is as expensive and complex as the model itself. The revenue from API calls, subscriptions, and enterprise deals is not scaling with the cost of training. Each generation of model costs more to train and delivers proportionally less incremental value. The ROI is shrinking.

Third, China's open-weight strategy is eating the market. K3 is free. Quen 3.8 is free. DeepSeek's models are free. American businesses are adopting Chinese open models because they are good enough and cost nothing. Nvidia says US companies should be allowed to use them. The dissenters, OpenAI and Anthropic, are the ones losing market share. Every download of a Chinese open model is revenue that does not go to an American lab.

09The Society Angle

This is not just a tech story. The models being built on both sides of the Pacific are going to reshape labor markets, information ecosystems, and power structures. The question is who controls that reshaping.

When the largest open-source AI model on Earth is built by a Chinese company on smuggled American chips, the geopolitical frame is obvious. But the societal frame is more important. If American businesses adopt Kimi K3 and K4 because they are free and good enough, the American AI industry loses its customer base. And if the American AI industry loses its customer base, the revenue problem gets worse. And if the revenue problem gets worse, the investment in safety, alignment, and governance slows down. The 1,100 workers who signed the Pacing the Frontier initiative are asking for tools to slow development. The market may slow it for them, but not in the way they want.

Meanwhile, China's labs are not asking for permission. They are not signing initiatives to slow themselves. They are training the next generation of models on hardware they are not supposed to have, and releasing them for free to the world. The export wall is leaking, and the water is rising on both sides.