There was Xi Jinping, leader of one of the world’s most tightly controlled societies, standing before the World Artificial Intelligence Conference in Shanghai and presenting China as the new champion of openness.

Not political openness, certainly. Not an unrestricted internet, a free press or an uncensored public square. Xi was talking about artificial intelligence and China’s willingness to share it with the world through open models, international cooperation, technical assistance and access for developing countries.

The temptation is to laugh at the contradiction. How can an authoritarian country claim to lead an open movement?

Quite easily, as it turns out.

Political openness and technical openness are not the same thing. A government can control speech, information and political activity while encouraging developers to download, modify and deploy software. Open source describes how technology is distributed. It does not certify the values of the government or company distributing it.

China has not suddenly embraced openness as a moral philosophy. It has recognized openness as an instrument of economic and geopolitical power.

That makes Xi’s argument more important, not less.

Two Very Different Visions of Openness

The United States remains a more open society. Yet the most powerful American AI models are largely closed. Customers can access them through controlled applications and APIs, but they generally cannot possess the weights, inspect the training data or move the underlying intelligence somewhere else.

The model provider sets the price, determines the rules and retains the ability to modify or withdraw the product. The customer rents access to intelligence without owning it.

China can now point to that arrangement and offer another definition of freedom.

China is not offering the world political freedom. It is offering freedom from dependence on American AI companies—and betting that much of the world will consider that open enough.

At the Shanghai conference, Xi described AI as something that should be broadly shared rather than monopolized. He promoted open models, international governance and expanded access for developing nations. China also backed the creation of the World Artificial Intelligence Cooperation Organization, an alliance reportedly involving 29 countries and positioned as an alternative to the U.S.-led Pax Silica initiative and its emphasis on secure, trusted AI supply chains.

The message is carefully constructed. The United States restricts. China shares. America controls access to chips and models. China helps countries build their own AI capabilities. Washington talks about security. Beijing talks about inclusion.

There is plenty of hypocrisy embedded in that presentation. China maintains extensive information controls, and Chinese models can reflect state censorship and political restrictions. Beijing’s definition of open AI does not imply open training data, open governance or politically neutral outputs.

But geopolitical narratives do not have to be pure to be effective. They need enough truth to be credible.

American frontier models are closed. Many Chinese models are available as open weights. A company or government may be able to download a Chinese model, customize it, deploy it privately and avoid permanent dependence on an American API provider. Whatever one thinks of the government behind that model, the operational freedom can be real.

Kimi K3 Puts a Product Behind the Policy

Xi’s speech supplied the doctrine. Moonshot AI supplied the demonstration.

Moonshot’s new Kimi K3 is a 2.8-trillion-parameter, native multimodal model with a one-million-token context window. It was designed for long-horizon coding, knowledge work, reasoning and agentic tasks. The company says it performs competitively with the strongest American systems in several categories, while independent evaluations place it among the leaders in front-end coding and other technical benchmarks.

Moonshot says it will release the model’s complete weights on July 27, 2026. Assuming that happens, K3 will become the largest open-weight model yet released.

That does not mean K3 is indisputably the best model in the world. Benchmarks remain imperfect, and Moonshot’s claims will receive greater scrutiny once outside researchers can inspect and test the weights. K3 appears to trail the strongest proprietary models in some evaluations while outperforming them in others.

China does not need Kimi to win every benchmark.

It needs Kimi to be close enough.

A model that provides 90% or 95% of the capability of a leading proprietary system at a much lower price—and can be possessed, customized and deployed independently—will be more attractive to many organizations than a marginally superior model available only through someone else’s platform.

That is particularly true for countries concerned about sovereignty, companies handling sensitive information and developers who do not want their products tied permanently to the commercial decisions of a single American model provider.

The United States is competing to build the world’s best AI. China is competing to become the AI the rest of the world builds upon.

Those are not the same contest.

Open Weights Do Not Mean Hardware Independence

There is another contradiction inside China’s open AI strategy. Kimi’s weights may be open, but running them efficiently still requires extraordinary computing infrastructure.

A 2.8-trillion-parameter model is not something most developers will download onto a workstation. Depending on precision and quantization, the weights alone can require several terabytes of memory. Production deployment requires clusters of accelerators, fast interconnects, optimized kernels, serving software, power and cooling.

Today, Nvidia remains the strongest provider of that complete system.

Moonshot has demonstrated K3 performing optimization work across Nvidia processors and hardware from an unnamed alternative vendor. But that is not proof that the model already runs on Huawei Ascend or other Chinese accelerators with equivalent production performance. Nvidia’s CUDA ecosystem, software libraries and server architecture continue to provide enormous advantages.

That creates an almost perfect picture of China’s current position. Beijing can give away the intelligence layer, but Nvidia may still collect the toll when organizations try to run it at scale.

China’s longer-term goal is not difficult to discern. Chinese models must become increasingly portable to Huawei and other domestic infrastructure. Open models can help create the developer adoption, workloads and economic incentives needed to improve that alternative ecosystem. The model strategy and the semiconductor strategy are ultimately part of the same campaign.

China is giving away the intelligence layer before it has completely secured the infrastructure beneath it.

Then Comes the Distillation Charge

American AI companies are unlikely to concede that China has independently closed the frontier-model gap.

The distillation argument is already waiting.

In February 2026, Anthropic accused Moonshot, DeepSeek and MiniMax of conducting large-scale campaigns to extract capabilities from Claude. According to Anthropic, the three companies collectively generated more than 16 million exchanges through approximately 24,000 fraudulent accounts. Moonshot was allegedly responsible for more than 3.4 million of those exchanges, targeting coding, agentic reasoning, tool use, computer vision and other capabilities prominently associated with Kimi.

Those are serious allegations. If Anthropic’s account is accurate, this was not ordinary customer use. It was an organized effort to acquire the outputs of a competitor’s model, evade regional restrictions and violate the provider’s terms.

Yet distillation should not become a magic word that makes the rest of China’s engineering disappear.

Learning from the outputs of another model can transfer behaviors and accelerate development. It does not by itself explain K3’s scale, Kimi Delta Attention, Attention Residuals, million-token context window or the systems engineering required to train and serve a 2.8-trillion-parameter model.

Distillation may be part of Kimi’s story without being the entire story.

There is also an uncomfortable counterargument that American model providers will have to confront. They trained their systems using vast quantities of human-created material gathered from the internet, much of it without individual permission and some of it without a license. The companies will correctly argue that collecting public material for training is technically and legally different from creating fraudulent accounts to extract a competitor’s capabilities.

The distinction is real. It may nevertheless be lost in the geopolitical telling.

Much of the world may hear the American position this way: Our use of everyone else’s work was training. China’s use of our work is theft.

Before Washington acknowledges that China has turned open AI into a strategic advantage, it will argue that Beijing built that advantage by distilling closed American models. There may be truth in the accusation. There will also be considerable convenience in it.

Even if Kimi benefited from Claude, the strategic result remains. An American company kept its model closed. A Chinese company learned from it, legitimately or otherwise, and is preparing to release the resulting weights to the world.

China’s campaign for open AI may have been accelerated by intelligence extracted from America’s closed AI.

The Indispensability Trap Arrives at the Model Layer

All of this leads to a larger economic truth.

AI models are becoming indispensable and commoditized at the same time.

Businesses will not operate without intelligence embedded in their applications, workflows and decisions. Governments will treat access to models as a matter of economic capacity and national sovereignty. Developers will assume intelligence is available just as previous generations assumed the availability of electricity, networking and cloud computing.

Yet the more indispensable models become, the less willing customers and governments will be to let a few companies control them.

That demand produces alternatives. Alternatives create competition. Competition creates standards, substitutability and falling prices. Open weights accelerate the process by allowing customers to possess and modify the technology rather than merely renting access to it.

This is the Indispensability Trap at work. A technology can remain absolutely essential while its scarcity value, pricing power and differentiation steadily erode. The infrastructure continues to make the new economy possible, but lasting value migrates higher, lower or sideways in the stack—to applications, workflows, proprietary data, distribution, governance and the customer relationship.

American model companies are spending hundreds of billions of dollars attempting to make their intelligence indispensable. China appears willing to help them succeed—and then make that intelligence a commodity.

China’s Two Loops

In my forthcoming book, The Indispensability Trap, I describe China’s strategy through what I call China’s Two Loops.

The first is the digital loop. Models from DeepSeek, Alibaba’s Qwen, Moonshot and Zhipu spread globally through downloads, deployment, adaptation and developer adoption. Every organization that builds on one of these models extends the ecosystem and makes Chinese intelligence a more credible default substrate.

China does not have to collect the highest possible price for every token. Its model companies can compete for ubiquity rather than scarcity.

The second is the physical loop. China’s enormous manufacturing base gives it factories, industrial machinery, robots, supply chains and physical systems through which artificial intelligence can be deployed. Those systems generate operational data and experience that feed back into the development of embodied AI, industrial automation and intelligent machines.

The digital loop distributes Chinese intelligence. The physical loop gives that intelligence experience in the world. Each reinforces the other.

Kimi K3 strengthens the digital loop. Xi’s speech reveals that Beijing now understands that loop as an instrument of foreign policy.

China is not playing exactly the same game as the American model companies. OpenAI and Anthropic want to make their models indispensable while retaining control over access. China wants its models to become indispensable by surrendering some of that control and encouraging the world to build upon them.

Beijing is not trying to avoid commoditization. It may believe commoditization favors China.

Springing the Trap

China has not become an open country. It has discovered that openness can be a distribution strategy, an economic weapon and a source of geopolitical influence.

The United States remains more open politically while its most important AI companies remain closed technically. China remains closed politically while offering the world increasingly capable open-weight intelligence.

That is not a clean moral reversal. It is a strategic inversion—and one the United States should not dismiss merely by pointing out China’s hypocrisy.

China does not have to own the world’s best proprietary model to win influence over the next stage of AI. It may succeed by making models abundant, substitutable and increasingly difficult for any American company to control.

China is not escaping the Indispensability Trap. It is springing the trap deliberately, commoditizing the intelligence layer before American companies can convert their temporary lead into permanent control.

The danger is not that China’s claim to openness is perfectly sincere.

The danger is that America has made it plausible.