# Economic Warfare & Chinese AI > Source: > Published: 2026-07-17 10:51:36+00:00 This summer, Chinese AI labs released two open-weight large language models comparable to the best closed American ones. The Chinese models released by [Z.ai, named GLM-5.2](https://z.ai/?ref=irregularwarfare.org), and by [Moonshot AI, named Kimi K3,](https://www.constellationr.com/insights/news/moonshot-ai-launches-kimi-k3?ref=irregularwarfare.org) represent the same central asset behind American AI model companies like Anthropic, xAI (a component of SpaceX), and OpenAI. Together, these American firms hold [multi-trillion dollar valuations](https://finance.yahoo.com/markets/stocks/articles/openai-now-considering-2027-ipo-013300058.html?ref=irregularwarfare.org), based on the premise that their AI models command a price for their capability. Yet, Chinese open-weight models of the same capability are now being given away, at no cost, for anyone to download and run on their own hardware. In the AI race, leading in model performance and the resources required to train and operate them is imperative. The fact that [China dominates in energy capacity](https://www.bloomberg.com/news/articles/2026-01-28/china-s-four-year-energy-spree-has-eclipsed-entire-us-power-grid?ref=irregularwarfare.org) but lacks the most sophisticated hardware and model architectures gives the United States an early, but waxing, lead. To counter that advantage, China has chosen a different strategy for competition. China is cultivating an AI ecosystem that builds cutting-edge open-weight models, replicated from pricey American ones, then distributes them for free. **Open-Weight v. Closed** To understand the Chinese strategy, let’s first review the three model distribution types. First, there is the closed model, which is the predominance of American AI. In these models, the underlying math (referred to as “weights") are a trade secret (think Opus 4.8, GPT 5.5, and Gemini). Second, there is open-weight. These trained model weights are released for anyone to download, run, and fine-tune on their own hardware. Lastly, there is the open-source model. The weights plus the training data and code are released as open source, so the model could be rebuilt from scratch. The important bit here is that China's GLM-5.2 ships as an open-weight model rivaling the capability of a closed model like Anthropic’s Opus 4.8, but it’s [licensed under MIT](https://en.wikipedia.org/wiki/MIT_License?ref=irregularwarfare.org). Written at the Massachusetts Institute of Technology in the 1980s but owned and controlled by no one since, it is a permissive software license that puts few restrictions on reuse. Choosing MIT is a unique and strategic choice for [Z.ai](https://z.ai/?ref=irregularwarfare.org). The terms let anyone use, modify, redistribute, or sell the model, with a single obligation to keep the original copyright notice attached. A more guarded license, like the one governing [Meta's open-weight model Llama](https://developer.meta.com/ai/?ref=irregularwarfare.org), keeps a leash by throttling large commercial deployments or fencing off certain uses. The MIT license keeps no leash at all, meaning [Z.ai](https://z.ai/?ref=irregularwarfare.org) published a frontier-class model royalty-free with a right to run it forever. **Economics of Frontier AI** A [frontier model](https://www.nvidia.com/en-us/glossary/frontier-models/?ref=irregularwarfare.org) is the most capable general-purpose AI system and is the leading edge of what's technically possible at a given time. These models require enormous compute and capital to train, and their resultant capabilities set the performance ceiling that other models are measured against. This includes capital costs of the hardware, but also large amounts of energy required to run model training. It’s held that China [lags behind American frontier models by approximately six months](https://www.economist.com/china/2026/06/21/china-is-having-another-ai-moment?ref=irregularwarfare.org). There are two phases to every AI model. First, there is the training. This is the massive one-time process of building and fine-tuning the model by feeding it data until it learns. Then, there is inference. This is the everyday act of running that finished model to answer a query. Both cost money and both use compute, but they sit on opposite sides of the economics: training is the sunk capital that creates the capability, and inference is the recurring cost of using it. In America, the training costs are underwritten by future token revenue. A [token is the unit of value](https://blogs.nvidia.com/blog/ai-tokens-explained/?ref=irregularwarfare.org) by which a closed lab converts a model’s capability into cash. A token itself is a chunk of text, usually a word or piece of a word, that a model reads and generates one at a time. AI companies use tokens to price usage. The business model stands when each query is metered and priced above its cost. A healthy spread will pay back the training investment and the inference cost, with profits left over. The average blended token cost is ever rising, and can be tracked [here](https://tokenpriceindex.com/?ref=irregularwarfare.org), currently [$2.47 per million tokens as of this writing](https://tokenpriceindex.com/?ref=irregularwarfare.org). The latest, most capable models, like Fable 5, a recent subject of US export restrictions, cost up to [$50 per million output tokens.](https://tokenpriceindex.com/compare?ref=irregularwarfare.org) However, this token model crumbles when the capability does not stay scarce enough to charge for. An open-weight model at the same level of capability as a costly, closed one removes the scarcity and sets the price spread to zero. China can do this because the government subsidizes model costs and unburdens it’s AI labs from requiring the return on invested capital that necessitates the American closed model AI ecosystem. China [cuts the energy costs of AI firms by up to half](https://www.ft.com/content/cad2cdd6-7cce-4de3-8710-977de667378c?syn-25a6b1a6=1&ref=irregularwarfare.org). These interventions drastically change the cost basis of Chinese AI. And now with the release of frontier-class open-weight MIT licensed Chinese models, the effect will be a bit like product dumping, except with information goods, the effect is permanent. By competing economically, not technologically, China hopes to constrain American AI labs to slow their pace of development so they may eventually leap-frog them. This is the Chinese strategy, and it directly undermines American AI labs and the investments that their projected revenues underwrite. **AI and Economic Warfare** Despite [export controls on the latest AI hardware](https://www.irregularwarfare.org/chokepoints-american-power-in-the-age-of-economic-warfare/), how has China caught up to American AI? To start, they used a technique called [knowledge distillation](https://en.wikipedia.org/wiki/Knowledge_distillation?ref=irregularwarfare.org). Instead of starting from scratch with the massive training, computation, and expertise required to build a frontier model, China first distilled its models from existing, closed American ones. In distillation, a smaller "student" model learns from a larger "teacher" model by studying its outputs: you query the teacher at scale, collect its answers, and train the student to reproduce them. The student inherits much of the teacher's capability at a fraction of the cost, because someone else already paid to discover it. The technique is a legitimate machine learning tool used to train lesser capable, smaller models. But in January 2025, OpenAI and Microsoft said they had evidence that [DeepSeek queried ChatGPT at scale ](https://www.reuters.com/technology/microsoft-probing-if-deepseek-linked-group-improperly-obtained-openai-data-2025-01-29/?ref=irregularwarfare.org)and trained on its outputs, a direct violation of terms of use that forbid using the models to build competitors. In February 2026, OpenAI escalated the claim to Congress, telling the House Select Committee on the CCP that [DeepSeek was routing around access restrictions to keep harvesting American model outputs](https://www.reuters.com/world/china/openai-accuses-deepseek-distilling-us-models-gain-advantage-bloomberg-news-2026-02-12/?ref=irregularwarfare.org), and that the practice was growing more covert. American labs spend billions developing the capabilities of its frontier models. A Chinese lab reproduces a usable fraction of it, then releases the output under a license that lets the world run it for free. The expensive half of the work, the discovery, is effectively expropriated. On the surface, America and China are fierce competitors in AI. However, beneath the surface, each is playing an entirely different game. Rather than racing the American labs to the frontier, China removes the profit that makes reaching it worthwhile. In doing so, China removes the profit motive for the American AI industry, which threatens to impede innovation – and thus U.S. leadership in the space. By shifting AI profit margins lower, China moves competition to an income statement. This is not traditional capitalism, where the competitor is trying to out-earn the incumbents. China is following Sun Tzu. The oldest principle in Chinese strategy is to [win without fighting](https://note.com/human_macaw1926/n/nc579142449ff?hl=en&ref=irregularwarfare.org): defeat the opponent before any battle by making the battle pointless. How do you compete with free? An open-weight frontier model forces a closed one to answer this question, and the implications for the American economy are far reaching – starting with the potential destruction of billions of dollars in wealth. The economic leverage from undermining the American AI ecosystem is immense. Closed-lab spending by the likes of Anthropic and OpenAI is the demand thesis underneath a hardware supercycle and a [$7 trillion-dollar investment trend](https://www.weforum.org/stories/2026/04/ai-investments-7-trillion-buildout-right/?ref=irregularwarfare.org). American AI firms justify the capital expenditure by the future profits from their closed models. These returns justify them to [pay 75% gross margin on](https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-first-quarter-fiscal-2027?ref=irregularwarfare.org) chips, with these boosted earnings accruing massive gains across the American stock market. Without a viable profit motive, the justification for investment dries up, and with it, U.S. leadership in the space. **Open-Weight American AI** China’s subtle and potentially powerful approach to combatting U.S. leadership in the AI signals an important development in economic warfare. Rather than try to catch up – or leapfrog – the advantage held by its adversary, China’s decision to freely distribute it’s models changes the rules of the game. China uses American models to train its own and then makes them freely available. More important than the erosion of U.S. market share, the Chinese strategy denies American companies the rewards they require to justify trillions of dollars in risk. Massive amounts of dollars in American investment towards AI development is distilled from its models at a drastically lower cost by Chinese ones, and then sold for free to the world. This setup will portend a revaluation which will reverberate throughout the capital markets. The obvious rebuttal is that none of this is warfare, it’s just the innovation cycle doing what it always does. Early on, capital burns while the market hunts for the right path and eventually, the innovation becomes commoditized. That happens to every technology, and it was likely coming for AI regardless. In this case, the “right path” trends toward low margins and open, own-your-own hardware deployment no matter who forced the issue. But in economic warfare, timing is the weapon. A margin collapse that arrives in 2032, after the American labs have earned back their capex, is a maturing industry. But facing one today, Chinese models now threaten a margin collapse in the middle of the AI build out after trillions have been financed to back supposedly scare models. China drove the world to converge towards open-weights. However, Chinese open-weight models do not have to be the global standard. [America has open models](https://en.wikipedia.org/wiki/Nemotron?ref=irregularwarfare.org), it just does not yet have a frontier one. The American AI strategy needs to shift from competing on model capability to competing on the open-layer with a frontier-class open-weight model of it’s own. Without one, the world risks fracturing into Western models and Chinese models. America still trains the best models on the planet. It should build the open model for the free world, and let this American model become the global standard.