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

A developer rarely tries a Chinese open model to make a political statement. They have a bill due at the end of the month, a launch deadline looming, and a long list of concrete problems to solve. If the model works, they recommend it to a colleague; if that colleague ships something with it, the model finds its way into another company.

Few people think of this as political influence. The conversation is about price and performance, yet what lingers is a plain, personal judgment: Chinese technology is genuinely useful. Before the model has answered a single political question, China's image has already shifted, however slightly, through ordinary use.

The deeper point: models carry a country's view of the world

DeepSeek, GLM, and Kimi have placed powerful model weights in the hands of developers around the world. Those developers can download, modify, and deploy them inside products of their own. [3][4][5]

Some people argue that the Chinese government may simply have failed to see the danger. I do not buy it. China has spent decades increasing its national power; it is hard to imagine several leading companies moving in the same direction while the government remains oblivious. Recent Chinese AI policy explicitly calls for a globally oriented open-source ecosystem. This looks deliberate. [11][12]

The political value of that choice goes well beyond putting Chinese model names inside more products. As more people around the world use Chinese models, they are not merely borrowing those models' capabilities. They are also borrowing their judgments about the world.

An LLM does not draw a sentence at random from a neutral pile of facts. Its training data, annotation, safety rules, and regulatory environment all find their way into the answer. On Taiwan, capitalism and socialism, or an international conflict, different models select different facts, words, and frames. China's rules require public generative-AI services to uphold socialist core values; American models likewise reflect the politics and risk preferences of their makers. [14]

The tendency is easiest to see in direct political questions. On Taiwan, one model may begin with national sovereignty while another begins with democracy and self-determination. On U.S.-China competition, one may describe “defending the international order” while another describes “containing China's rise.” The underlying facts may barely change, yet the reader has been placed at a different moral starting point. A few words of framing can, over time, lead people toward very different conclusions.

Researchers have measured the difference. GPT-4o more often frames issues through a Western lens, while DeepSeek sits closer to the Chinese government's position. Change the balance of political material in the training data, and the model's assessment of Chinese institutions and leaders changes with it. A model does not need to praise a country openly. Influence accumulates whenever some facts remain in the foreground and others are repeatedly left in the background. [15][16][17]

Years ago, I wrote about the political tendencies I observed in GPT-4. I also find Anthropic's treatment of China consistently more negative. Calling them “Democratic models” or “anti-China models” is too crude, but frontier models do carry traces of their development teams, training data, and social environments. There is no need to decide first which worldview is superior. The important point is that if the world relies on only a handful of American models, American frames can quietly harden into the default answer.

And the influence of an LLM reaches far beyond a few sensitive questions. More and more people ask models to summarize the news, explain history, compare political systems, recommend books, and draft reports. The model's answer may not be the final word, but it is often the first word a user sees. It decides where the story begins, which context deserves inclusion, which views count as “mainstream,” and which positions sound “extreme.” Whoever shapes that opening frame holds a kind of power once reserved for the media, schools, and cultural institutions.

This power also differs from traditional media in one crucial respect. A model does not broadcast from a distance to a crowd; it enters each person's work and private conversations, answering one individual at a time. Users voluntarily hand it their questions. They ask it to organize information, rewrite arguments, and help form judgments. The influence is therefore more intimate, and often harder to see.

More importantly, political trust is not created from nothing at the moment someone asks a political question. A model first writes a hundred correct pieces of code, catches a flaw in a contract, and saves the user real money. By the time that user asks how to understand China, America, or a war, trusting the model has become a habit. Trust earned through competence spills quietly into trust in judgment. That may be the deepest political return from opening Chinese models to the world.

An Indian programmer finishes a project with a Chinese model; a European founder cuts an AI bill; a Japanese team gets a workflow running. What they remember is that a Chinese model helped them. Real soft power is not a country shouting its virtues from the rooftops. It is people discovering, almost without noticing, that the country made something genuinely useful.

Goodwill built through daily use is difficult to manufacture with diplomatic slogans. India, Japan, Europe, and developing countries also gain another choice: they no longer have to place every AI product behind the APIs of a few American firms. Many of the strongest American labs remain closed, while Chinese labs are using open models to break that grip. Even if some of those countries become China's competitors, a market with more choices is still better for China. [13]

Openness has a cost. Once weights are released, they are difficult to recall, and safeguards can be removed. Those risks are in plain sight, yet China is moving ahead because beyond them lies a larger market, a wider developer base, and a degree of international influence that was once much harder to acquire.

Next comes the hard arithmetic of power and infrastructure

Intelligence is becoming a basic economic capability. If it remains locked inside the servers of a few American companies, a handful of firms will be able to set the price, rules, and boundaries for everyone else. Open weights break that scarcity first. As the gap between models narrows, competition moves down the stack toward chips, inference engines, data centers, cooling, grids, and generating capacity.

China is closing the chip gap quickly. In 2026, a scientific computing cluster built from 60,000 domestically produced AI accelerators entered service. At that scale, the specifications of a single chip no longer tell the whole story; cluster design, software, scheduling, and system-wide efficiency matter just as much. [20]

China's advantage in electricity is more direct. The International Energy Agency calls China the world's largest power system. In 2025, the country generated 10.58 trillion kilowatt-hours of electricity and reached 3.89 terawatts of installed capacity. Behind those numbers stand power plants already built, a transmission network still expanding, and industrial capabilities accumulated over decades. This heavy physical infrastructure is the foundation large-scale AI inference requires. [7][21][22]

Chinese labs also care intensely about how much work they can squeeze from each chip. DeepSeek V4 Flash has 284 billion parameters but activates only 13 billion for each token. The emphasis on mixture-of-experts architectures, quantization, and inference optimization begins with a simple cost equation. [3][10]

As models become cheaper, the price of a token depends less on any single chip than on the system around it. Model weights can be copied quickly; power plants, transmission lines, and industrial systems take years to build. China is steering the competition toward this deeper, heavier terrain, where its advantages have been patiently built over time.

Finally, the 9.5 model fights back

Suppose the best model scores 10 and yours scores 9.5. Switching takes little more than changing an API endpoint, so users vote with their feet and move to the 10. The leader takes the revenue, usage feedback, developers, and reputation, then turns those advantages into an even stronger lead. If the runner-up stays closed, it may not even get the feedback it needs to catch up.

Open weights give the second-place model another route. Private deployments do not automatically send user data back to the model company, but public evaluations, bug reports, derivative models, hardware integrations, and developer discussions still help improve the model. The open community trains engineers and attracts researchers, allowing an outmatched challenger to build an ecosystem that grows alongside it.

For personal use, people may still choose the smartest and most expensive model for the hardest problems. Production follows a different logic. Once a model is good enough for customer-support classification, document extraction, code review, or a stable workflow, insisting on the very best model for every request is simply overkill. Give the hardest work to the 10 and everything else to the cheaper 9.5. Users may not notice the difference; the bill will tell the truth.

As of August 2026, DeepSeek V4 Flash costs $0.28 per million output tokens, while Anthropic's Claude Fable 5 costs $50. The two models are not directly comparable, but a price gap of more than one hundred times is enough to make any serious company reconsider how much intelligence each task actually needs. [1][2]

This is the opening Chinese companies can exploit. They do not need to challenge the strongest model head-on in every category. Price and openness let them sidestep that contest and enter the vast territory where “good enough” is enough. Even if they cannot persuade users to pay them yet, they can stop those users from paying a rival quite so much. This is not a frontal assault; it attacks the economics underneath.

The revenues of leading American AI companies are real because customers paid real bills. But a hundredfold price difference is not a hundredfold profit difference; training and inference consume enormous amounts of cash. OpenAI generated $5.7 billion in revenue while burning $3.7 billion in cash in the first quarter of 2026, and Anthropic was only approaching its first quarter of operating profit. [18][19]

The issue is that some of that revenue depends on intelligence remaining scarce. Once open models offer a cheap, credible substitute for ordinary work, leading labs can still charge a premium for scientific research, difficult reasoning, and critical agents. What becomes harder is charging the same toll on every routine token. Open models are not trying to eliminate legitimate profit; they are attacking the extraordinary rents that a few companies collect from their position at the gate.

Nor does openness mean abandoning commercial interest. Hugging Face reported that Chinese models accounted for 41 percent of downloads over the preceding year, while the Qwen family produced more than 113,000 derivative models. A fine-tune here, a deployment guide there, one more hardware optimization: each contribution looks small on its own, but together they become an ecosystem that is extraordinarily difficult to reproduce. Engineers who learn those models may later become employees, customers, or partners. [6]

China is giving its weights to the world with a clear calculation of its own. It wants more people using Chinese models, more talent, more demand for infrastructure, and more goodwill and influence. Yet self-interest and public value can coexist. What the world receives from this competition is something it has badly needed: a wider range of choices.

The most hopeful future is not an American monopoly replaced by a Chinese one. It is intelligence flourishing in many hands. In that world, small companies, universities, and countries with fewer resources can afford to use powerful models, and ordinary people do not need permission from a single lab to use one of the defining tools of their age.

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.
20. Cyberspace Administration of China, China's largest scientific AI computing cluster enters service, April 14, 2026.
21. National Energy Administration of China, China's power industry advances in scale and sustainability, June 15, 2026.
22. International Energy Agency, Global Energy Review 2025: Electricity.

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.