153

Dong Jielin

In the script of history, technological victories are often folded into national glory and the market value of giants, while the cost of change quietly falls on the shoulders of every individual.

On July 24, 2026, an open letter titled "Open Weights and American AI Leadership" sparked a new round of debate in Silicon Valley. NVIDIA, Microsoft, Meta, IBM, Palantir, HuggingFace, and a group of cloud computing, cybersecurity, and venture capital firms have jointly called on Washington not to prematurely restrict open weight models that can be downloaded, modified, and deployed autonomously. The open letter describes the open model as the foundation of American innovation, competition, security, and technological sovereignty.

This letter does not directly name China, but it cannot be read in isolation from China. In the past year, China's ability to open up weighted models and global adoption rates have rapidly increased, and Washington is also discussing whether to restrict or even prohibit Chinese models from entering the US market. Anthropic, who has not signed an open letter, has become the most prominent cautious faction: it emphasizes the biological and cybersecurity risks that cutting-edge models may bring, advocates mandatory testing of high-performance models, and strengthens control over chip and model distillation.

On the surface, this is a governance debate about whether openness is safer, but in reality, it is also a debate about interests and industrial positioning. Nvidia and cloud platforms hope to widely spread their models to drive computing power and hosting demand; Meta hopes that the model layer can become a low-cost complementary product, consolidating the distribution, advertising, and application ecology; Harness and the application company hope to retain the ability to replace underlying models while mastering customer data, processes, and entry points; Companies that rely on closed source models for charging place greater emphasis on the risks of loss of control and irreversible diffusion of capabilities. Developers, workers, consumers, and capital owners are also maintaining their tools, positions, choices, or investment returns. This does not mean that security concerns are false. Once the open weights are released, they are difficult to retrieve, and closed source systems may also be attacked, abused, or fail in situations where they cannot be checked by the outside world.

The role of undertaking grand narratives and considering the overall interests of the United States belongs to Washington. On the one hand, the White House defines "winning the AI race" as an economic competitiveness and national security goal, and requires national security agencies to adopt both state-of-the-art closed source models and open technologies; On the other hand, a risk identification and pre release communication mechanism for cutting-edge models has been established, while leaving room for restrictions based on intellectual property, cybersecurity, and competition with China.

This debate reveals a often confused question: What scoreboard should we use to understand the outcome of the AI competition between China and the United States? Model ranking measures a category of technological products; The company cares about the moat and profit flow; The government cares about technological leadership, industrial control, national security, social regulation, and global standards; Individuals are concerned about whether AI will expand or compress their income, opportunities, choices, safety, and dignity. Therefore, when discussing the AI dispute between China and the United States, it is necessary to simultaneously check the three scoreboards of companies, countries, and individuals, whose results can be reversed from each other.

1、 Dismantling the six layers of ecology:

The great power game is far more than just a model competition

AI is not just a model, but a set of interdependent ecosystems. Huang Renxun summarized it into five layers: electricity, chips, infrastructure, large models, and applications; Chamath Palihapitiya further separates data and control (Harness) into a single column, forming a six layer structure, as shown in Figure 1 of the 22 layout. The unique Harness layer is responsible for coordinating the operation of different models and agents; If the model is a horse, the harness and reins determine how it works. The forefront of technology and value accumulation may not necessarily be on the same level: models determine the upper limit of intelligent supply, while durable customer lock-in may occur at the harness and application layers. The AI strength of a major country depends on whether the entire ecosystem is robust and forms a closed loop, rather than any single point ranking at any level. Different levels will provide different opportunities for capital, entrepreneurs, and individuals. At present, the most active areas of entrepreneurship in Silicon Valley are concentrated in the relatively light asset Harness layer and application layer, where entrepreneurs attempt to transform general intelligence into concrete business value; The advantages of large capital are more reflected in capital intensive fields such as infrastructure construction, continuous model training, and platform ecosystem expansion.

A small number of top talents can earn high premiums at the chip, model, and platform levels through their scarce abilities. At the same time, although users enjoy the dividends brought by the popularization of AI capabilities and price reductions, they may also come at the cost of personal data, attention, and platform dependence; As workers, they may also face risks of job substitution, income differentiation, and decreased bargaining power.

By comparison, China and the United States currently have their own strengths: the United States has the most advanced chips, the strongest cutting-edge closed source models, strong private capital, and a high-quality enterprise software market; China has more abundant electricity, faster infrastructure construction, stronger manufacturing organizational capabilities, and temporary advantages in the global diffusion of open weight models. American workers acquire the strongest tools earlier and bear the pressure of high wage job displacement earlier; The use of models by Chinese workers is cheaper, but whether it can be translated into salary, entrepreneurship, and career advancement still depends on the market and system.

Recently, model capabilities have been gradually converging, but the physical foundations supporting AI development in China and the United States have not yet approached each other synchronously. The core bottlenecks faced by the two countries are not on the same level: for China, the main constraint comes from advanced AI chips - both affected by US export controls and limited by the performance, production capacity, and supporting ecology of domestic chips, while China has stronger power generation and engineering construction capabilities. Although the United States has relatively abundant high-end chips, it faces infrastructure constraints such as power supply, grid approval, and data center construction.

These constraints at different levels will ultimately be reflected in the construction speed of AI data centers and the deployment scale of effective computing power. As shown in Figure 2, based on the unified caliber of peak FP16 computing power for scenario estimation, the AI computing power of the United States in 2025 will be about 16 times that of China; This gap may widen to about 25 times by 2027. There is significant uncertainty in the long-term forecast, but it is still a relatively stable judgment that the United States will maintain an order of magnitude lead in the next three to five years.

(Figure 2 data explanation: China's 725.3 EFLOPS and 74.1% year-on-year growth rate in 2024 come from IDC and Inspur Information, which can be used to infer approximately 417 EFLOPS in 2023.); The data for 2025 and the first half of 2026 comes from the Ministry of Industry and Information Technology and the State Council Information Office. The data for 2023-2025 in the United States is estimated based on EpochAI's chip and data center data, using H100 equivalent computing power and the global share of major cloud vendors. 2026-2027 is the scenario forecast. Although the data of the two countries are unified as peak FP16EFLOPS, the original caliber is different and only suitable for judging the order of magnitude. ）

Constraints not only determine 'how much can be done', but also change 'how to choose'. In an environment with relatively abundant computing power, American companies tend to expand their scale and build super large clusters, but electricity and grid connection will limit speed; Chinese companies are developing efficiency routes such as distillation, sparsification, quantification, and inference optimization under chip constraints. The constraints are thus written into the business model of the enterprise and the solution of the first generation of engineers, and even if they are removed in the future, path dependence may not be easily remedied.

The constraints between the two countries will ultimately translate into individual costs: the benefits for the United States may be concentrated among a few technology and capital owners, while the costs of China's redundant construction, adaptation, and catch-up may be transferred to individuals through public capital, profit pressure, and labor intensity.

2、 The Paradox of Open Weights:

Distribution advantage does not equal business victory

Chinese big model manufacturers such as Qianwen, Deep Search, Dark Side of the Moon, and Zhipu have quickly entered the global developer community through open weighting, low-cost calling, and toolchain construction. Against the backdrop of gradually approaching model capabilities and continuously decreasing call prices, open weighting has lowered the threshold for developers to try, modify, and deploy models independently, and has become the most internationally influential competitive approach for Chinese AI.

The HuggingFace report shows that in the past year, Chinese sourced models accounted for 41% of the platform's model downloads; As of April 2026, the Qianwen family has accumulated nearly one billion downloads. On the calling side, Deep Search has become the largest supplier of Open Router with approximately 16.3% token share. Different platform perspectives cannot be merged, but together they indicate that the Chinese model has gained significant global distribution advantages. For China, open weighting not only brings model adoption, but also expands technological influence and strategic space.

But there is a paradox here: the overseas distribution and use of China's open model still heavily relies on infrastructure led by the United States. The weight is propagated through HuggingFace, and the inference service is accessed through platforms such as Open Router and Vercel, with the underlying layer commonly running on the NVIDIA system. Chinese companies provide models, while American ecosystems provide developer access, computing power, cloud, routing, and payment channels; The larger the usage of Chinese models, the more likely it is to increase the revenue of American infrastructure companies.

Note that downloading does not equal deployment, deployment does not equal continuous use, and call volume does not equal revenue. According to Vercel data, China's open model accounts for about one-third of its gateway token usage, but only corresponds to less than 4% of user spending. Open weighting can bring developers, reputation, and ecological access, but it also weakens the ability to directly charge, making it easier for customers to copy, modify, or replace models.

Therefore, countries can gain influence from technology diffusion, while model companies have to bear the costs of training, toolchains, and low-priced APIs. A considerable amount of value may flow to chip, cloud, inference, routing, and harness and application companies. China is currently winning the advantage of model distribution, rather than the entire value chain; Whether it can be converted into company revenue and complete national technological capabilities depends on whether it can move from downloads to deployment, from usage to payment, from productivity improvement to revenue distribution.

For users, open weighting reduces the barriers to learning, experimentation, and entrepreneurship, as well as decreasing reliance on a single closed source platform; But acquiring ability does not necessarily mean an increase in income. Developers may still lack customers and distribution entry points, workers may increase output but may not necessarily share productivity dividends, and ordinary users may also bear the risks of data breaches, fraud, and model abuse.

3、 When the token price drops sharply,

The laws of competition will be rewritten

In the past few years, the most significant change in the AI industry has not only been the improvement of model capabilities, but also the rapid decline in the price of obtaining equivalent intelligence. As shown in Figure 3, in the past three years, the price of tokens required to achieve similar capabilities has decreased by more than two orders of magnitude. If this trend continues, months of technological leadership will become increasingly difficult to automatically translate into lasting business value.

(Figure 3 Data Explanation: From 2023 to 2026, data synthesis model vendors will publicly disclose API prices and perform price performance analysis for Artificial Analysis and EpochAI, representing representative token prices required to achieve approximately GPT-4 level capabilities.). Related research shows that the price of reasoning with the same ability decreases by about 5-10 times annually, but the task differences are significant. The stress tests for 2028 and 2030 are based on a decrease of one order of magnitude every two years, and are not predictions. ）

The decrease in token prices is not the result of the model layer alone, but a product of the collective progress of the entire ecosystem: chip performance, cluster interconnection, and utilization continue to improve, quantization, sparsity, inference decoding, caching, and routing reduce the computational requirements for inference, cloud platforms improve device utilization, and model architecture and training methods enable smaller models to achieve capabilities that were only possible with larger models in the past. In recent years, the rate of price decline for equivalent To ken capabilities has been much faster than the hardware performance improvement explained by Moore's Law.

Market competition has further accelerated price reductions. As more and more models reach similar capabilities, manufacturers find it difficult to maintain high prices solely based on "intelligence", and can only compete for users through price reductions, free quotas, and product bundling. Therefore, the price reduction of tokens is not only a result of technological progress and competition, but also indicates that some general capabilities are showing a trend towards commercialization; But if there is an AGI level breakthrough that is difficult to replicate, the scarcity and pricing power of the model layer may rise again.

At the end of July 2026, OpenAI made GPT-5.6Luna the default free model for ChatGPT Free and Go users, and lowered the API input and output prices to $0.2 and $1.2 per million Tokens, respectively, indicating that top US companies are starting to compete for mass users, developers, and global distribution entry points with free products and significant price reductions. At the same time, multiple Chinese model manufacturers have announced price increases, and the new price for Deep Search will take effect on August 17th.

Low price tokens do not equal low costs. It is necessary to distinguish between the token price paid by the customer, the cost of achieving a certain model capability, and the unit effective computing power cost of completing the same training or inference workload. The former will be affected by subsidies and competition, while algorithm optimization can reduce the second term, but it does not necessarily make the underlying computing power system cheaper. To measure the economy of underlying computing power, the total cost of ownership (TCO) should be compared, including chip and depreciation, power and cooling, interconnection and storage, data center facilities, as well as software adaptation and operation; Only by dividing by the actual effective workload completed by the system can comparable unit effective computing power costs be obtained.

Chinese AI companies have certain advantages in manual labor, engineering construction, and algorithm efficiency (such as distillation, quantization, sparsification, etc.), which can reduce adaptation and operation expenses. However, in the key link of effective computing power at the bottom level, domestic platforms still face shortcomings in single card performance, chip interconnection, cluster stability, and software maturity, which often require more hardware, power, and engineering resources to complete the same tasks. The lower utilization rate also makes it difficult to fully convert theoretical computing power into usable throughput; If you choose the NVIDIA platform directly, you will face reduced specifications, export licenses, and scarcity premiums. Due to the overall disadvantages of these software, hardware, and ecological aspects offsetting some efficiency advantages, some publicly available system comparisons show that the unit effective computing power cost of domestic AI stacks may still be significantly higher than that of Nvidia platforms when completing the same effective training or inference workload.

The recent price increase by Chinese model companies is more likely to be the result of multiple factors working together: subsidy reduction, rising training and service costs, limited capacity, shareholders demanding revenue and profits, and full stack TCO gradually reflected in product prices. The low price of the Chinese model is not entirely without an efficiency foundation, but the magnitude of the price advantage is greater than the underlying cost advantage, and the difference must be paid at the expense of profits and capital returns, so price increases are inevitable.

4、 Where does the value settle? The reshuffle of application ecology, market, and capital

Without AGI level breakthroughs to widen the model gap, the added value of AI may shift more towards mastering customer data, business processes, specific scenarios, and distribution entry points at the Harness and application layers. How does harness and application determine the ability to enter the business, reach users, and accumulate value; The market determines payment demand and revenue space; Capital determines the flow of resources and the form of enterprises.

Harness and applications control customer stickiness and entry points

Harness is responsible for integrating the model into enterprise data, permissions, software, and business processes, and transforming the output into executable results; Applications package capabilities into products that users can directly adopt, mastering specific scenarios, interactions, distribution, and customer relationships. Enterprise data and processes are difficult to migrate, and user habits, brands, and channels are also difficult to replicate. Therefore, even if the underlying model can be replaced, Harness and applications may still form stable stickiness.

From the perspective of national competition, Harness determines how intelligence enters the business processes of enterprises and public organizations, while applications determine who controls user entry, usage scenarios, and data feedback loops. Therefore, these two levels are the key to transforming model capabilities into real productivity, economic value, and organizational capabilities, and are also important landing points for data control and data sovereignty.

For individuals, Harness and applications together determine whether AI is an assistant, supervisor, or replacement. Data permissions, performance indicators, and automated boundaries will change employees' autonomy and bargaining power; Product design, recommendation logic, and business model affect what tools individuals can use, how they are reached, and at what cost they pay. Whether an individual can benefit depends not only on whether they can use the model, but also on whether they can retain professional judgment, customer relationships, choices, and income distribution rights.

Market determines commercial compound interest and monetization potential

The market determines how much revenue AI can generate after landing. American companies have high software budgets, mature subscription habits, and expensive white-collar labor. As long as AI can replace some software functions or reduce the working hours required by programmers, analysts, lawyers, or customer service, companies have a clear incentive to pay; The income can also support the next round of computing power, research and development, and service investment, forming commercial compound interest. This is the market foundation for coding agents to become AI "super applications" in the United States.

China has a large number of application scenarios and strong deployment capabilities, but enterprise software budgets are usually low, and the cost of mental labor is relatively low. Currently, there have been no super enterprise applications. If the cost of purchasing tokens, modifying software, and reshaping processes is higher than adding labor, customers will lack the motivation to pay. The government enterprise market can provide early customers, but customized projects are difficult to replicate at low cost like standard software. The poor quality of the Chinese and American markets is also an important reason for Chinese model companies to go global as soon as possible. In addition, the consumer end can also rely on advertising, e-commerce and flow cashing - China's super applications in the Internet era appear in the personal consumer end.

Corporate compound interest does not equal individual compound interest. The adoption of AI by enterprises may not only increase employees' output and income, but also reduce job positions, leaving productivity dividends mainly to shareholders. Whether an individual can benefit depends on whether their skills are scarce, whether they have customer relationships or capital, and how the additional income is distributed.

Capital determines the growth direction of enterprises

The funding for the AI ecosystem in the United States mainly comes from venture capital, public capital markets, and massive capital expenditures by super tech companies. Its resource allocation typically favors cutting-edge technology, high growth, and the ability to generate commercial revenue on a global scale. Chinese AI enterprises are funded by national industrial funds, Internet companies, local governments and venture capital institutions; In addition to commercial returns, some capital also undertakes goals such as technological autonomy, industrial localization, and national strategic capacity building. It is worth noting that the impact of government capital is often not limited to its direct shareholding ratio: subsidies, procurement, access, policy signals, and subsequent financing may all amplify its influence on corporate decision-making. Different capital structures shape different corporate goals and drive companies towards different customers, products, and development paths. When model capabilities converge and training and service costs remain high, capital is more likely to flow to enterprises that have control over cloud, traffic, customers, data, harness, and application entry points.

For individuals, the direction of capital allocation will ultimately translate into career opportunities, income returns, and risks. The US system may create a small number of highly profitable research, engineering, and entrepreneurial positions, while also rapidly laying off, acquiring, or integrating when growth falls short of expectations, resulting in highly concentrated and volatile individual opportunities. The employment opportunities in the Chinese system may be more concentrated in the landing of government and enterprise projects, industrial deployment, and localized implementation, but the space for individuals to obtain excess returns and cross market mobility is relatively limited, and career prospects are also more susceptible to the influence of policy direction, government procurement, and local investment cycles.

5、 Future Industry Simulation: Regarding Open Weights

Three Predictions on Restructuring and AI Territory

Prediction 1 | US Open Power Camp Expands, Global Token Share Sees Major Shuffle

On August 10, 2026, Meta announced the opening of MuseSpark 1.2 flagship model weights, becoming a clear signal for the United States to respond to the Chinese camp with high-quality open weight models. The next thing to observe is how many top-level models will follow, and who will receive more token distribution, developers, and ecological entrances after opening up.

This prediction includes three layers of judgment: firstly, the United States will have more top-level models with open weights, and promote a significant increase in the token share of American open models in major routers, cloud platforms, and AI gateways; Secondly, China's open models such as Qianwen, Deep Search, and Dark Side of the Moon will compete more fiercely with each other, while the direct commercial pressure on top American companies mainly comes from the low-priced competition model in the United States; Thirdly, the US model will still be selectively open rather than fully open, and cutting-edge capabilities with higher risks of biosecurity, cybersecurity, or autonomous agency may continue to remain behind closed source controlled APIs.

Falsification criterion: In the next 18 months, the token share of US sourced open models on major global platforms has not significantly increased, or Chinese models have significantly squeezed the revenue of US model companies.

Prediction 2 | The model layer will experience two years of intense competition before entering an integration period

In the next two years, Chinese and American large model companies will still be in a stage of fierce competition and sustained investment. By the end of 2028, if there is no AGI breakthrough that can reconstruct the competitive landscape, as the model's capabilities further converge and the differentiation space shrinks, and the costs of training, reasoning, and continuous iteration remain high, some companies lacking stable income, capital support, or unique capabilities may find it difficult to sustain. At that time, both the Chinese and American model layers may begin to experience bankruptcy, mergers and acquisitions, and integration.

If a company achieves a decisive AGI breakthrough first, the scarcity of the model layer will rise again. In this situation, integration will still occur, but it will mainly manifest as accelerated concentration of capital, talent, and customers towards technology leaders, while laggards will be quickly eliminated.

Falsification criterion: After the end of 2028, there have been no multiple verifiable model layer mergers, acquisitions, or independent business terminations in either China or the United States ecosystem, and independent model companies still generally rely on model revenue growth.

Prediction 3 | In the future, the Chu River Han border between China and the United States in the AI world will be close to the Internet

Today's Internet territory is roughly divided into the United States platform leading area, China's independent system, and a mixed zone that uses both Chinese and American technologies. The AI frontier is likely to replicate this sector map, and in addition to model technology, chips, cloud, payments, app stores, data rules, government procurement, and security standards will also jointly determine the technology ownership.

This boundary will directly affect which tools individuals can use, whether data can cross borders, where skills can be used in which markets, and where entrepreneurial products can be sold. In the short term, the United States may not necessarily ban Chinese models, as the value they create through the catfish effect, open weights, local deployment, and supplier selection may still outweigh market and security costs; But this tolerance is conditional. Once the Chinese model causes visible security, industrial, media, or political issues and geopolitical factions gain the upper hand, policies may escalate from restrictions to bans.

This boundary is not formed in perfect free competition, and the hand of the government is never absent: the United States focuses on national security and market access, China focuses on model access and exports, content and social control, and Europe emphasizes risk grading. Different regulatory directions have collectively increased the institutional costs of cross-border deployment.

Criterion of falsification: If Europe, India and other countries form a third pole that does not depend on the United States, or China's technology stack becomes dominant in many large economies that Chinese Internet companies have not occupied, this judgment needs to be revised.

6、 Final victory and questioning of three scoreboards

In the future, both China and the United States are likely to announce victory in AI competition, but the winning structures of the two countries are completely different. The United States is more likely to define victory based on capital returns, scientific discoveries, and individual capacity expansion; China is more likely to define victory based on its industrial scale, technological diffusion, national organizational capabilities, and ability to shape social order. The advantages of the United States are more likely to accumulate into measurable and compound commercial wealth, while China's advantages are more reflected in its national capabilities and strategic influence that are difficult to directly value. The same round of technological revolution may strengthen the original directions of two systems and thus shape two different future societies.

However, regardless of the type of victory, the "three scorecards" of companies, countries, and individuals influence each other but cannot replace each other. A US AI company can achieve astonishing profits and market value, but it may also replace a large number of white-collar jobs and concentrate productivity dividends; China can use AI to improve its manufacturing and governance capabilities, but the employment, income, occupational safety, and autonomy of many ordinary people may not necessarily improve synchronously, and may even be damaged as a result. If the company wins, the country may not necessarily win; If the country wins, ordinary people may not necessarily win either.

More importantly, what kind of future will each country use AI to build? Who receives the productivity dividend? Who bears the technical risks and consequences? Has the enhancement of national capacity simultaneously expanded opportunities, choices, and dignity for ordinary people?

In this unstoppable wave of AI, everyone should often ask themselves: Am I the winner?