A freshly funded American startup just deployed a Chinese open-source model into its production pipeline. The model handles code generation, routes customer queries, and summarizes contracts. The company pays nothing for the weights. Not a cent. No license fee, no token paywall, no revenue share. The model's developers — a Chinese AI lab with hundreds of engineers — receive zero direct compensation.
This is not a hypothetical. It is the current state of cross-border AI adoption. And it exposes a value capture failure that the market refuses to price. Logic doesn't lie: if a model is doing real work, it has economic value. But if no one pays for it, that value flows to the user, not the creator. The gap between usage and monetization is now the industry's most underreported vulnerability.
Read the code, ignore the roadmap. The code is Apache-2.0. The roadmap promises future enterprise services. The code wins. And the developers lose.
I spent 200 hours auditing yield farming contracts during DeFi Summer. The pattern is identical: open code, no economic capture. Smart contract vulnerabilities had clear exploit paths. This is the same — only the exploit is legal.
Context: The Weight of Free Weights
The story begins with a simple fact: Chinese AI labs — DeepSeek, Alibaba's Qwen, Zhipu's GLM — publish state-of-the-art models under permissive licenses. DeepSeek-R1 matches OpenAI's reasoning benchmarks on math and coding. Qwen-72B holds its own against Llama-3. GLM-4 handles tool calling reliably. All are available as open weights.
American companies, from stealth startups to Fortune 500 divisions, download these weights and integrate them into products. No API call. No usage metering. No payment. The license permits it. The economics incentivize it. Why pay $20 per million tokens to an American vendor when a Chinese model delivers 90% of the performance at zero marginal cost?
The result is a silent dependency. Crunchbase data shows over 4,000 US-based AI startups launched since 2023. A significant fraction rely on open-source models. Among those, Chinese models hold a measurable share — HuggingFace download statistics place Qwen and DeepSeek in the top ten globally. This is not a fringe phenomenon. It is mainstream infrastructure.
But the funding flows tell a different story. Chinese AI labs raised over $10 billion in 2024. Their revenue from Western users? Negligible. The market values them as research institutions, not as software companies with recurring revenue. This disconnect is ripe for forensic analysis.
Core: The Mechanics of Unpaid Labor
Let me reverse-engineer the value chain. A model is a static artifact — a set of weights, a tokenizer, a config file. Once released under an open license, the creator loses control over its distribution. The user downloads, deploys, and serves the model on their own hardware. No third party observes the inference. No billing system triggers. The usage is invisible to the developer.
This differs fundamentally from API-based models. When you call OpenAI's API, they log every token, meter usage, and invoice monthly. The model remains a service, not a product. Open weights transform the model into a commodity. The creator gets a one-time reputational boost, then zero ongoing revenue.
The perverse incentive is obvious: American companies offload R&D costs to Chinese labs while pocketing the margin. Consider a typical SaaS product. It uses Qwen for internal document parsing. The startup charges customers $50 per user per month. Its infrastructure costs include servers, but no model licensing fee. The margin is fat. The Chinese lab that trained the model? It paid for GPU clusters, data curation, and human feedback. That investment is amortized across zero paying users.
In my due diligence work, I call this a "silent extraction" — a mechanism where one party captures value without triggering any payment obligation. I saw it in crypto with MEV bots. They extract value from LPs without alerting anyone. The Ethereum network processes the transaction, but the value flows to the bot operator. Here, the American corporation is the MEV bot. The Chinese lab is the LP.
Volatility is just unpriced risk. The risk here is not price volatility. It is supply-chain seizure risk. If the US government bans Chinese weights, every company that built its product on Qwen faces a migration cost of millions. The open-source ecosystem has created a single point of failure — not in code, but in geopolitics.
The Structural Flaw in Open Source AI
The open-source AI movement inherited ideology from open-source software. But the economics are different. Open-source software like Linux was developed by companies who could charge for support. Red Hat proved that path. AI models demand far higher training costs. A leading model requires hundreds of millions of dollars in compute. There is no enterprise support contract large enough to recover that investment if the weights are free.
Chinese labs are attempting a hybrid: release a free base model, charge for API access. But the free weight version cannibalizes the paid API. Why buy API calls when you can run the model yourself? The cost of running a 7B parameter model on a T4 GPU is pennies per hour. For many tasks, that is cheaper than any API — including the Chinese labs' own API.
This is the fundamental mispricing. The marginal cost of inference on open weights is borne by the user. The fixed cost of training is borne by the lab. In a rational market, the user would pay a licensing fee. But there is no enforcement mechanism. The license is permissive. The code is immutable. The payment is voluntary — hence, it never happens.
Some argue that Chinese labs benefit indirectly. The usage data, the community feedback, the downstream applications — these create ecosystem lock-in. When enterprises eventually need custom support, they will contract with the lab. That is the Red Hat thesis. But Red Hat's support contracts were mandatory for mission-critical systems. With AI, a competent internal team can fine-tune LLaMA or Qwen without external help. The optionality reduces willingness to pay.
Contrarian: The Bulls Might Be Right
Let me steelman the optimistic view. The Chinese labs are not naive. They understand the value of penetration. By giving away billions of dollars of model value, they have achieved something the American incumbents cannot buy: global trust. Every American company that deploys Qwen in production is a proof point. That trust can be converted into cloud contracts. Alibaba Cloud and Huawei Cloud already bundle Qwen into their offerings. When an American company uses Qwen via a Chinese cloud provider, the revenue flows to the cloud vendor — which may redistribute part of it to the model team.
The indirect path is real. In 2025, Alibaba Cloud reported a 77% growth in international revenue, partly driven by AI services built on Qwen. The model is the loss leader; the cloud is the profit center. This is Amazon's strategy with AWS. Give away the product, sell the infrastructure.
Moreover, the "not getting paid" narrative overlooks the geopolitical value. Chinese AI models now influence how the American workforce writes code, summarizes documents, and interacts with AI. That soft power has strategic weight. It shapes the next generation of AI standards. The OECD and the United Nations are drafting AI governance frameworks. Chinese models are being tested and evaluated in those forums. The presence they have earned through free deployment gives them a seat at the table. That seat is not invoiced, but it is priceless.
There is also the data feedback loop. Every deployment generates real-world usage data. Even if the lab does not see the data, the open-source community does. The community identifies bugs, reports failure modes, and fine-tunes the model. This is unpaid contribution to the Chinese lab's engineering effort. In the crypto world, we call this a free audit. The community acts as an unpaid QA department.
Finally, the capital markets are shifting. Chinese AI labs are backed by sovereign funds and tech giants. They do not need to monetize directly. The parent company (Alibaba, Baidu, Tencent) absorbs the cost because the model strengthens their broader ecosystem. A free Qwen that destroys ChatGPT's download share is worth more than a paid Qwen that sells 10,000 API subscriptions. This is a strategic play, not a business failure.
The bulls have a point. The model developers are not starving. They are playing a longer game.
Takeaway: The Exploit Has No Patch
Here is the uncomfortable truth: the current arrangement has no sustainable endpoint. If American companies continue to use Chinese models without payment, the Chinese labs' incentive to release frontier models will diminish. Eventually, they will either close the weights or restrict commercial use. That will happen exactly when the American dependency is deepest. Volatility is just unpriced risk. The risk is now accrued, and the bill will come due.
My advice to investors: do not value Chinese AI labs as infrastructure providers. Value them as geopolitical arbitrage plays. The free lunch cannot last forever. The next release cycle will test whether the labs accept zero direct revenue again. If they do, the migration cost for American companies is not technical — it is strategic. Code is law, but only until someone forks the law.
Read the code, ignore the roadmap. The code says the weights are free. The roadmap says the enterprise version is coming. Logic doesn't lie. The market will eventually price this asymmetry.
I have seen this pattern before. In 2022, Terra's algorithmic stablecoin promised yield without risk. The code showed that yield was impossible under stress. I published a 40-page teardown. A year later, the system collapsed. The same logic applies here. The value capture mechanism is broken. The system will have to be rebuilt — or the free rides will stop, abruptly.
Those who understand the mechanism will position themselves accordingly. The rest will wait for the invoice.
