Why China Gives Its Best AI Away: The Return Almost Everyone Misses

While you are still paying for the most expensive American model, your competitor may have moved 95 percent of its requests to a Chinese one. They will post the falling bill in a group chat. Everyone else will start testing the same model the next morning.

Most people notice the price. Far fewer notice the political return: before the model answers a single political question, Chinese technology has already saved you money and helped you ship. Your feeling about China has shifted, if only slightly.

That may be the most underrated meaning of open weights. Every model speaks with the accent of the country and institutions that made it. China is distributing more than intelligence. It is distributing its technology, standards, and view of the world.

The first return is the one almost nobody discusses

DeepSeek V4 Flash and GLM 5.2 use MIT licenses. Kimi K3 also permits broad use, modification, and deployment. Open weights are not the same as fully open source, and the chips and electricity required to run them are not free. But developers can put near-frontier capability into their own products without sending every request back to its creator. [3][4][5]

Some observers treat this as proof that the Chinese government has not understood the danger. I do not find that credible. A government that has materially increased China's power over several decades can make mistakes, but it is unlikely to miss the same obvious risk while DeepSeek, Z.ai, Moonshot, and Alibaba all move in the same direction. The policy is explicit. China's 2025 “AI Plus” plan called for a globally oriented open-source ecosystem, and its 2026 international cooperation plan called for sharing open AI and expanding access to compute. Openness is not a door regulators forgot to close. It is a door the government placed in the blueprint. [11][12]

The first return is political influence.

A model is never a neutral calculator. Its corpus, annotation, safety rules, and regulatory environment shape how it describes Taiwan, compares capitalism with socialism, and explains international conflict. Chinese rules require public generative-AI services to follow socialist core values. American models are just as inevitably shaped by their creators' political instincts and risk preferences. [14]

The accent can be measured. Researchers have found more Western-centric framing in GPT-4o and more alignment with Chinese state narratives in DeepSeek. Changing the proportion of political material in training data can directly change how a model evaluates Chinese institutions and leaders. A model does not need to praise a country in every answer to carry influence. It only needs to make certain facts feel central, certain words feel natural, and certain assumptions disappear from view. [15][16][17]

Years ago, I wrote about the political tendencies I observed in GPT-4. I also find Anthropic's framing of China unusually hostile. “Democratic model” and “anti-China model” are crude labels. The durable point is simpler: no frontier model comes without a worldview.

The most effective influence arrives before politics does. An Indian programmer finishes a project with a Chinese model. A European founder cuts an AI bill. A Japanese team builds a stable workflow. None of them received a political lecture. They simply experienced Chinese technology as useful.

Soft power is strongest when it does not look like soft power. It just helps you get something done.

That experience will not erase concerns about censorship, security, or dependency. It does create familiarity and goodwill that diplomatic slogans struggle to buy. It also gives India, Japan, Europe, and developing countries an alternative to placing every future AI product behind the APIs of a few American firms. Some of those countries are China's competitors. China is still willing to strengthen them because a market that American companies cannot monopolize is better for China.

The U.S. government has not banned open weights, and Meta has released important open models. The real contrast is that many leading American labs remain closed while a growing number of Chinese labs use openness as global distribution. [13]

The risks are real. Released weights cannot be recalled, safeguards can be removed, and models can be adapted for cyberattacks or disinformation. China is accepting those risks in return for adoption, ecosystem power, goodwill, and a larger voice in how the world understands technology. This is not charity or ignorance. It is a calculated political trade.

China would rather compete on power and infrastructure

Intelligence is becoming one of the foundational tools of the economy. If it remains locked inside the servers of a few American companies, a handful of private firms will decide its price, rules, and boundaries. The danger is not merely that “America controls AI.” It is that something this basic could be controlled by a small number of profit-seeking corporations.

Open weights tear down that tollbooth. Once models become standard components rather than scarce treasures, competition moves from owning the smartest weights to producing reliable tokens at the lowest cost. The battlefield shifts to chips, inference engines, data centers, cooling, grids, and power plants.

That is a contest China would rather enter.

China remains disadvantaged in advanced chips, and its electricity is not always the world's cheapest. In many industrial settings, however, it benefits from lower energy and construction costs. Its deeper advantage is scale and speed. By the end of 2025, China had 3.89 terawatts of installed power capacity. The IEA estimates that it invested more than $625 billion in clean energy in 2024 and already accounted for roughly a quarter of global data-center electricity consumption. Power plants, transmission lines, industrial supply chains, and data centers cannot be copied overnight. [7][8][9]

Compute constraints have also forced Chinese labs to care intensely about efficiency. DeepSeek V4 Flash has 284 billion parameters but activates only 13 billion for each token. Mixture-of-experts systems, quantization, support for domestic chips, and production inference have become priorities across the open ecosystem. [3][10]

Open weights do not guarantee that a token will be generated in China. A European company can run a Chinese model in an American cloud with Nvidia chips and European electricity. But openness compresses profit in the model layer and moves advantage toward the efficiency of the full stack. China is betting that once the model stops being the crown jewel and becomes a component, its power system, engineering speed, inference work, and supply chains will matter more than the walls around a closed lab.

Make the rival's scarce product abundant, then compete where you have scale. That is the second calculation.

Finally, pull the foundation from under hundredfold prices

The model market is brutal to second place.

Suppose the best model scores 10 and yours scores 9.5. Switching between them takes little more than changing an API endpoint, so users flow toward the 10. The leader gets the revenue, usage feedback, developers, and reputation, then uses those advantages to stay ahead. If the 9.5 remains closed, it can lose both the market and the information it needs to catch up.

Open weights are the fastest way to break that loop. Private deployments do not magically return user data to the model company. Public evaluations, bug reports, derivatives, hardware integrations, and developer discussions still create valuable feedback. The community also trains engineers and attracts researchers, turning one company's model into an ecosystem that thousands of people help expand.

Individuals may still choose the most intelligent and expensive model for the hardest problem. Production is different. Support classification, document extraction, code review, and stable workflows need to clear a quality threshold. Beyond that threshold, cost, speed, and reliability matter more than half a point on a leaderboard. Send the hardest five percent to the 10 and the other 95 percent to the cheap 9.5. Users may not notice. The bill will.

As of August 2026, DeepSeek V4 Flash costs $0.28 per million output tokens. Anthropic's Claude Fable 5 costs $50. The models are not equivalent, but a price gap of more than one hundred times reveals the enormous space between adequate intelligence and frontier intelligence. [1][2]

The commercial logic is sharp: even if users will not pay you yet, you can stop them from paying your rival so much. Once the leader can no longer charge the highest price for every task, the market opens again.

This also clarifies the spectacular revenues of leading American AI labs. Calling them fake is inaccurate because customers paid real bills. A hundredfold price difference is not a hundredfold profit difference either; training and inference burn money. OpenAI reportedly generated $5.7 billion in revenue while consuming $3.7 billion in cash in the first quarter of 2026. Anthropic was only approaching its first quarter of operating profit. [18][19]

What is “fake” in the looser sense is the part of the premium built on temporary scarcity. Open models give routine work a credible substitute. Leading labs can still charge high prices for scientific research, difficult reasoning, and critical agents. They simply lose the power to charge the same toll on every ordinary token. Profit does not disappear. Monopoly profit returns toward normal profit.

Openness is itself a business model. Hugging Face reported that Chinese models reached 41 percent of downloads over the preceding year, while the Qwen family produced more than 113,000 derivatives. Every fine-tune, deployment guide, and hardware optimization expands the original model's reach. Every engineer who learns it can become a future employee, customer, or partner. [6]

The three layers now connect. Open models carry Chinese technology and narratives around the world. Cheap models move the contest toward electricity, data centers, and inference efficiency. Broad adoption then weakens the power of a few American companies to set the price of access to intelligence.

China is pursuing its own advantage, not performing charity. But self-interest can still create public value. A better future is not an American monopoly replaced by a Chinese one. It is a world in which no company and no country owns the only road to intelligence.

If this competition makes tokens cheaper and gives small companies, universities, and poorer countries access to powerful intelligence, the largest winner will not be China alone.

It will be everyone who was never going to own a frontier lab.

References

1. DeepSeek API Docs, Models & Pricing, accessed August 2, 2026.
2. Anthropic, Claude Fable 5, including API pricing.
3. DeepSeek, DeepSeek V4 Flash model card, including architecture and MIT license.
4. Z.ai, GLM 5.2 model card and release post.
5. Moonshot AI, Kimi K3 repository and Kimi K3 License.
6. Hugging Face, State of Open Source on Hugging Face: Spring 2026.
7. National Energy Administration of China, 2025 national electricity statistics, January 29, 2026.
8. International Energy Agency, World Energy Investment 2025: China.
9. International Energy Agency, Energy and AI: Executive Summary.
10. Hugging Face, Architectural Choices in China's Open-Source AI Ecosystem, January 27, 2026.
11. State Council of China, Opinion on Deepening the “AI Plus” Initiative, August 2025.
12. State Council of China, China issues action plan on AI cooperation and development, July 17, 2026.
13. U.S. National Telecommunications and Information Administration, Dual-Use Foundation Models with Widely Available Model Weights Report, July 30, 2024.
14. Cyberspace Administration of China, Interim Measures for the Management of Generative Artificial Intelligence Services, July 13, 2023.
15. Pacheco, Cavalini, and Comarela, “Echoes of power: investigating geopolitical bias in US and China large language models”, Humanities and Social Sciences Communications, March 25, 2026.
16. Waight et al., “State media control influences large language models”, Nature, May 13, 2026.
17. Haslett et al., “Made-in China, Thinking in America: U.S. Values Persist in Chinese LLMs”, December 2025.
18. Reuters, “OpenAI burned $3.7 billion in first quarter of 2026, The Information reports”, June 16, 2026.
19. Reuters, “Anthropic closing in on first quarterly operating profit”, May 20, 2026.

The arguments and interpretations in this essay are the author's. Model availability, licenses, and API prices were checked on August 2, 2026 and may change.