The ledger does not care about national borders. It only records what moved, what was used, and what was never compensated. This week, Dimension Capital published a thesis that reads like a forensic finding: Chinese AI models are doing the work inside American production systems, and the model developers are not getting paid. Hype is a mask; the ledger is the face beneath it. The numbers tell a story that no press release can obscure.
Let me be precise about what we know. The original analysis, which I have dissected line by line, contains a single core assertion: American companies depend on Chinese AI models, and that dependency does not translate into revenue for the model creators. The report itself is thin on specifics. No model names. No benchmark scores. No licensing details. No payment figures. What it lacks in data, it makes up for in implication. The phrase "doing the work" suggests production-grade deployment, not experimental tinkering. These are not research demos. These are systems in the critical path.
I have spent twenty years watching this industry confuse narrative with reality. The current situation is a textbook case of value capture failure. Every transaction leaves a scar on the chain. The scar here is not on a blockchain ledger, but on the balance sheets of Chinese AI companies who have built the world's most widely adopted open-source models and monetized almost none of it.
The Context: Open Weights, Closed Wallets
To understand the paradox, you need the full picture. Chinese open-source models โ DeepSeek, Qwen, GLM, and others โ have achieved something remarkable. They have become default infrastructure for a significant portion of the global developer community. HuggingFace download statistics have shown Chinese models at or near the top of the charts for years. The Apache-2.0 and MIT licenses under which these weights are released remove every legal barrier to adoption. No fees. No restrictions. No friction.
This is not an accident. It is a strategy. Chinese AI companies chose the open-source path deliberately, understanding that the fastest route to global influence is not through a paywall but through ubiquity. The result is a strange inversion of the traditional software business model. In the old world, you sold the software and gave away the services. In the current AI landscape, Chinese companies give away the software and hope to sell something else later. The "something else" has not materialized at scale.
American companies, meanwhile, face a rational economic calculation. Why pay OpenAI or Anthropic for API access when a comparable open-weight model can be self-hosted for a fraction of the cost? Why accept vendor lock-in when you can control your own deployment? The answer, for many engineering teams, is that you do not. You download the weights. You fine-tune them on your own data. You integrate them into your product. And the Chinese developer who spent millions training that model receives nothing for the transaction.
Numbers have no emotions, only consequences. The consequence here is a structural misalignment between global technical influence and direct commercial return. The model developers have won the adoption war and are losing the revenue war simultaneously.
The Core: A Systematic Teardown of the Value Gap
Let me break this down the way I would break down a smart contract audit. The system has multiple layers, and each layer has its own failure mode.
Layer One: The License Trap. The open-source licenses that enabled global adoption also eliminated the obligation to pay. Apache-2.0 permits commercial use, modification, and redistribution without royalty. This is not a loophole. It is the explicit design. The model developers chose this license knowing full well that it would make direct monetization nearly impossible. The question is whether they understood the magnitude of the trade-off. Based on my audit experience, I can tell you that most teams underestimate the difficulty of converting open-source adoption into sustainable revenue. The conversion rate from user to paying customer in open-source software is typically in the low single digits. For AI models, where the marginal cost of serving an additional user is near zero, the incentive to pay is even lower.
Layer Two: The Deployment Bypass. The most common deployment pattern for these models is self-hosting. American companies pull the weights, run them on their own GPU clusters, and integrate them into their own pipelines. This bypasses the API revenue stream entirely. The model developer sees zero token-based income. The cloud provider โ AWS, Azure, Google Cloud โ captures the infrastructure spend. The company captures the value. The model developer captures nothing. This is the digital equivalent of a manufacturer giving away the factory blueprint and watching competitors build the same factory for free.
Layer Three: The Indirect Revenue Mirage. There is an argument that Chinese companies benefit indirectly. The argument goes like this: open-source adoption builds brand recognition, creates a developer ecosystem, generates feedback data, and establishes a foundation for future enterprise sales. This is true in theory. In practice, the indirect benefits have not translated into meaningful revenue. The enterprise sales cycles for Chinese AI companies serving Western customers are long, complicated by geopolitical friction, and frequently blocked by procurement teams who view Chinese technology as a compliance risk. The feedback data is real, but it is not monetized. The brand recognition is real, but it does not pay for compute.
Layer Four: The Cloud Intermediary Problem. Some revenue does flow back to China, but it flows through a different channel. Alibaba Cloud, Huawei Cloud, and Tencent Cloud all offer Chinese models to international customers. The revenue from these deployments is booked by the cloud providers, not the model developers. The model developers may receive licensing fees or internal transfer pricing, but the economics are opaque and the margins are thin. The cloud provider captures the customer relationship, the infrastructure margin, and the strategic position. The model developer is reduced to a component supplier in someone else's value chain.
Layer Five: The Valuation Disconnect. This is where the investment thesis becomes sharp. Chinese AI companies carry significant valuations based on their technical capabilities and market position. But those valuations are not supported by revenue. The market is pricing in a future monetization that has not yet arrived. If that future never arrives, the valuations are fiction. If it does arrive, the current prices are a bargain. The uncertainty is not about capability. It is about capture. Can these companies convert their technical dominance into financial returns before the capital markets lose patience?
I have seen this pattern before. In the early days of open-source software, companies like Red Hat built sustainable businesses by selling support, certification, and enterprise features on top of free software. The Red Hat model worked because the software was complex enough that enterprises needed help deploying and maintaining it. AI models are different. They are easier to deploy. The infrastructure is commoditized. The support requirements are lower. The Red Hat analogy breaks down precisely where it matters most.
The Contrarian Angle: What the Bulls Get Right
I am not in the business of one-sided analysis. The bulls have a case, and it deserves a fair hearing. The dependency that Dimension Capital highlights is not a weakness. It is a moat. American companies that have integrated Chinese models into their production systems have made a significant switching-cost investment. They have fine-tuned the models on proprietary data. They have built evaluation pipelines around them. They have trained their engineering teams on their quirks. Replacing these models with alternatives would require substantial time, money, and operational risk. That is the definition of lock-in.
The lock-in cuts both ways. If the American companies are locked into the Chinese models, the Chinese model developers are locked into the American market. But the asymmetry favors the model developers. They can serve the entire world. The American companies can only serve their own customers. The model developers have diversified their user base across thousands of companies. Each individual American company has concentrated its dependency on a single model family. In a disruption scenario, the model developer loses one customer. The American company loses its entire AI infrastructure.
There is also the data feedback loop. Every deployment of a Chinese model in an American production system generates real-world usage data. This data โ even if anonymized and indirect โ helps the model developers understand failure modes, edge cases, and performance characteristics that they would not otherwise see. This is a form of reverse R&D funding. The American companies are effectively paying for the privilege of improving the Chinese models through their own usage. The value transfer is not zero. It is just not monetary.
The geopolitical dimension adds another layer. The dependency that Dimension Capital describes is a form of soft power. Chinese models have penetrated Western infrastructure in a way that no Chinese software product has achieved before. This is not a trivial achievement. It changes the terms of the technology competition. The United States cannot simply ban Chinese AI models without disrupting its own companies' operations. The interdependency is a shield. It protects Chinese AI companies from the most aggressive forms of decoupling.
The Security Question: The Elephant in the Room
No analysis of this situation is complete without addressing the security dimension. American companies running Chinese models in their production environments are making a risk calculation. The risks are real: potential data exfiltration, supply chain vulnerabilities, compliance violations, and the possibility of model weights being updated with malicious code. The fact that these risks have not prevented adoption tells you something important. The economic incentives are overwhelming the security concerns.
This is not a sustainable equilibrium. At some point, a high-profile incident will occur. A Chinese model will be found to have a backdoor, or a data breach will be traced to a Chinese model deployment, or a regulatory body will force a separation. When that happens, the entire ecosystem will face a reckoning. The companies that built their infrastructure on Chinese models will face a painful migration. The model developers will lose their most valuable market. The trust that underpins the current adoption will evaporate.
I have seen this movie before. In the blockchain world, we call it a smart contract exploit. The code looks fine. The tests pass. The deployment is smooth. Then someone finds the edge case. The funds are gone. The reputation is destroyed. The same pattern applies here. The current adoption of Chinese models is a bet that no catastrophic failure will occur. It is a bet that the geopolitical environment will remain stable. It is a bet that the models will continue to improve without introducing new vulnerabilities. These are not safe bets.
The Investment Implications: Reading the Tea Leaves
Dimension Capital is not publishing this analysis for altruistic reasons. They are positioning their investors for a specific outcome. The thesis is clear: Chinese AI models are undervalued because the market is not pricing in their actual global influence. If the monetization problem is solved, the upside is enormous. If it is not solved, the downside is equally enormous. The asymmetry is the investment opportunity.
The path to monetization is not obvious. The most likely route is a hybrid model: open-source weights for adoption, paid enterprise features for revenue. This could include managed deployment, security auditing, compliance certification, custom fine-tuning, and priority support. The challenge is that American companies are unlikely to pay Chinese companies for these services given the geopolitical climate. The revenue would have to come from non-American markets. This limits the addressable market and extends the timeline.
There is another possibility. The Chinese government could mandate monetization. If Beijing decides that its AI champions need to be financially sustainable, it could impose licensing requirements or restrict the free distribution of weights. This would be a dramatic shift, but it is not impossible. The government has shown a willingness to intervene in strategic industries. AI is now a strategic industry. The question is whether the intervention would help or hurt the global adoption that has been so carefully cultivated.
The most likely outcome is a prolonged period of ambiguity. The models will continue to be used. The revenue will continue to be elusive. The valuations will continue to be debated. The market will continue to price in hope rather than evidence. This is the nature of emerging technology markets. The winners are not determined by who has the best technology. They are determined by who figures out how to capture value from that technology. The Chinese AI companies have won the technology race. They are losing the value capture race. The outcome of that second race will determine everything.
The Takeaway: A Reckoning Is Coming
The blockchain is never silent. Neither is the global AI supply chain. The current situation is a ledger that has not yet been balanced. Chinese AI models are doing the work. American companies are capturing the value. The model developers are waiting for a payment that may never arrive. This is not a sustainable equilibrium. Something will break.
The question is not whether the reckoning will come. It is what form it will take. Will it be a regulatory crackdown that forces separation? Will it be a monetization breakthrough that finally aligns value and revenue? Will it be a security incident that destroys trust? Will it be a quiet erosion of Chinese AI companies as their capital runs dry? Each path leads to a different outcome. Each outcome has different winners and losers.
I do not make predictions. I make observations. The observation here is that the current state of affairs is unstable. The dependency is real. The value transfer is one-directional. The incentives are misaligned. The market is pricing in a resolution that has not yet occurred. When the resolution comes, it will be sudden. It will be disruptive. And it will leave scars on the chain that no amount of narrative spin can erase.
Chaos is just unanalyzed data. The data here is clear. Chinese AI models are the most widely adopted technology that no one is paying for. That sentence will not remain true forever. The only question is how it changes.