Beyond the Thucydides Trap: WAIC 2026, China’s AI Civilization Bet, and the American Order It Still Needs

July 22, 2026

Overview

At the 2026 World Artificial Intelligence Conference in Shanghai, Mr. Xi did more than endorse artificial intelligence as another strategic industry. He presented a political and economic architecture for an alternative AI order: open-source models rather than exclusive access, shared capacity rather than a small club of compute-rich states, cultural plurality rather than a single language system, and multilateral governance rather than rules written only by the countries that reached the frontier first.

That vision has real industrial foundations. China reported that its core AI industry exceeded RMB1.2 trillion in 2025 and included more than 6,200 companies. WAIC itself gathered more than 1,100 companies, over 3,000 exhibits, and more than 300 global product debuts across an exhibition area exceeding 100,000 square meters. Huawei displayed an increasingly integrated domestic compute architecture. Alibaba, Baidu, and Tencent reported measurable AI-related revenue or commercial usage. Chinese robotics companies shipped thousands of humanoid units while building supply chains that can lower hardware costs faster than most Western competitors.

Yet industrial scale is not the same as a self-sustaining AI civilization. An industry can build models, data centers, robots, and international institutions while still failing to answer four harder questions: Who pays for the intelligence after the subsidies end? Can enterprises earn enough to replace the capital they consume? Will foreign users trust the technical and political system behind the product? And can China expand AI supply without recreating the excess-capacity dynamics that have already generated trade conflict in electric vehicles, batteries, solar equipment, steel, and other strategic industries?

The analysis proceeds along two connected axes. The first is China's AI civilization declaration: the system Mr. Xi described at WAIC, built around open models, industrial deployment, multilingual knowledge, sovereign access, and a Shanghai-centered governance network. The second is the emerging U.S.-China AI civilization divide: a Chinese system organized around state coordination, open-weight diffusion, manufacturing, and Global South capacity building versus an American system organized around private frontier laboratories, advanced accelerators, hyperscale clouds, controlled model access, and trusted-partner exports.

Those two axes are then tested against corporate reality. Company revenue, capital expenditure, model distribution, compute architecture, robot shipments, customer structure, and startup economics reveal whether the declaration can become an operating system rather than remain a geopolitical narrative. The analysis also applies the warning made by Enrico Fardella and Sergey Radchenko in Foreign Affairs: China rose through an international system it increasingly seeks to revise, and it may weaken the markets and institutional openness on which its own expansion still depends.

The resulting picture is neither propaganda nor collapse. China already possesses much of the industrial machinery required to build a distinct AI civilization. Its strongest capabilities are production, integration, cost reduction, physical deployment, and policy coordination. Its least proven capabilities are durable application revenue, reciprocal market openness, international trust, and the conversion of state-backed capacity into demand that can survive without permanent subsidy.

Structural Judgment

China's WAIC declaration becomes a durable alternative AI system only to the extent that industrial capacity can be converted into economically competitive compute, recurring demand, and trusted international deployment; without that conversion, scale remains capacity rather than a self-sustaining order.

Anchor Set

The Western institutional anchor is the U.S.-centered control of frontier accelerators, hyperscale cloud infrastructure, export licensing, and trusted-partner technology networks. The non-Western anchor is China's already deployed combination of open-weight models, domestic cloud platforms, industrial procurement, Global South training programs, and the 29-country World Artificial Intelligence Cooperation Organization (WAICO) framework reported by Reuters. Neither anchor is merely rhetorical: both are supported by companies, capital expenditure, procurement systems, and formal institutions.

The physical and industrial anchor is harder to change. Both systems remain constrained by advanced fabrication, high-bandwidth memory, packaging, electricity, cooling, network reliability, field maintenance, customer budgets, and productive utilization. Political narratives can alter access and incentives, but they cannot remove those material requirements.

Axis One: China's AI Civilization Declaration

WAIC is often described as China's answer to the major technology conferences of the United States. That description is now too narrow. The 2026 conference combined a product exhibition, a national industrial showcase, a summit on global governance, and the launch of international institutional machinery. Its theme, "AI Partnership for a Brighter Future," was not only promotional language. It defined how Beijing wants other countries to interpret the next phase of technological competition.

Mr. Xi organized the challenge around access, control, culture, and governance. His framework rejected permanent AI concentration in a small group of countries, favored wider participation by latecomers, emphasized human control of autonomous systems, criticized the broad use of national security to restrict technology, and called for greater representation of underrepresented languages and cultures. Governance, in this framework, would be coordinated through institutions that include developing countries rather than written only by the states and companies controlling frontier chips and models.

These ideas are connected. Open-source models reduce the cost of entry. Lower-cost access increases the number of states capable of deploying AI. Broader deployment creates demand for local data, languages, standards, and governance. That in turn gives China a reason to build training programs, application centers, model ecosystems, and international organizations. At WAIC, China pledged 5,000 AI training opportunities for developing countries over five years and proposed cooperation centers linked to ASEAN, the African Union, the Arab League, CELAC, the Shanghai Cooperation Organization, and BRICS. A founding agreement for a World Artificial Intelligence Cooperation Organization was signed by 29 countries.

The strategic logic is clear. The United States currently holds the strongest position in advanced accelerators, frontier laboratories, hyperscale cloud infrastructure, and the software layers that surround them. China cannot easily displace that structure by reproducing it at the same price and under the same rules. It therefore has an incentive to redefine the contest around different advantages: open weights, lower inference costs, sovereign deployment, industrial integration, hardware manufacturing, and state-to-state capacity building.

In this model, AI becomes more than software. It becomes an exportable development package. A country might receive a local data center, a Chinese model adapted to its language, training for engineers and officials, an industrial application for mining or agriculture, and a governance framework that emphasizes sovereignty and non-interference. The package competes not only with an American model API but with the entire American proposition for how intelligence is supplied.

The Four Operating Layers of the Declaration

An alternative route to compute access

The first promise is access without a privileged political relationship with Washington or a permanent contract with a small number of American cloud companies. China presents open models and domestic compute systems as a path around that concentration.

This is not purely altruistic. Every model family that becomes widely deployed creates dependencies around toolchains, fine-tuning methods, developer communities, security procedures, and compatible hardware. Open-source distribution can therefore serve the same strategic function that proprietary cloud platforms serve for American companies: it creates an installed base. The difference is that China lowers the entry price and gives governments the option to host the model inside their own jurisdiction.

A production-centered theory of AI

The second promise is the entry of AI into the physical economy. Chinese policy places unusual weight on factories, vehicles, electric grids, logistics systems, scientific laboratories, medical equipment, and robots. This gives China a potential advantage that cannot be measured by chatbot rankings alone.

American AI developed first through advertising platforms, cloud software, consumer applications, and knowledge work. China's industrial system offers a different route. A model can be trained on factory inspection, warehouse handling, autonomous driving, equipment maintenance, or energy dispatch. It can then be combined with domestic sensors, motors, batteries, power electronics, vehicles, and production lines. The value is created by integrating intelligence into manufactured systems rather than charging only for software seats.

A claim to represent the Global South

The third promise is political. China argues against permanent authority for the countries possessing the largest models to define safety, culture, and acceptable knowledge for everyone else. In this framing, multilingual corpora and locally controlled deployment become matters of civilizational representation.

There is a real problem underneath the rhetoric. English-language data, American corporate preferences, and the legal assumptions of advanced economies are overrepresented in global AI systems. Many languages lack high-quality digital text. Local knowledge may be filtered as noise because it is not represented in the formats used by frontier laboratories. A country that helps build local language models can gain both goodwill and structural influence.

A governance architecture headquartered outside the West

The fourth promise is institutional. Beijing is no longer satisfied with participating in institutions whose agendas are largely set elsewhere. The Shanghai-centered cooperation organization announced at WAIC is designed to make China a convening power for AI governance, especially among countries that believe Western export controls and safety frameworks preserve an existing hierarchy.

The significance is not that one new organization will immediately set global rules. The significance is that compute, models, training programs, standards, and diplomacy are being assembled into a coherent system. China is trying to become the place where developing countries can obtain both technology and a political explanation for why control of that technology need not remain concentrated in a small group of advanced economies.

How Chinese and Western Media Read the Declaration

The same conference produced four very different interpretations. Chinese policy media presented an emerging architecture of shared development. Chinese technology reporting exposed the scarcity, financing pressure, and platform dependence beneath that architecture. Business Insider emphasized the strategic split between Chinese open-weight models and the predominantly closed American frontier. The BBC link included in the research material adds a broader British media lens, placing the speech inside a contest over governance, security, and international credibility.

These perspectives are analytically distinct. Each sees a different operating layer of China's AI civilization. Taken together, they provide a better test than either official optimism or external skepticism alone.

The Chinese policy-media reading: openness as a four-layer operating system

A CCTV-affiliated commentary by Yuyuan Tantian translated Mr. Xi's speech into four linked domains: compute and models, applications and scenarios, data and corpora, and rules and governance. This interpretation is important because it shows how Chinese policy communicators understand "openness." It is not limited to publishing model weights. It describes an operating system for international development.

At the compute layer, the commentary described China's goal as "connection" rather than one-way technology export. At the application layer, it argued that WAIC had moved beyond a model-centered exhibition toward concrete industrial problems. At the data layer, it presented multilingual corpus construction as a way for underrepresented societies to participate in defining machine knowledge. At the governance layer, it argued for jointly built rules rather than rules inherited from the countries that reached the frontier first.

The commentary also supplied operating examples. GeoGPT was said to serve 55,000 researchers across 145 countries. Nearly 200 procurement groups reportedly attended WAIC, with approximately one-third coming from overseas. Malaysian rubber-processing workers were described as using Chinese small-model algorithms, while a Chinese "digital brain" was presented as supporting grid operations in Chile.

These examples strengthen the claim that China is building more than a domestic model industry. They also create a clear audit requirement. The number of countries reached does not reveal the value, duration, or financing of the contracts. Researcher registrations do not reveal sustained production usage. A foreign pilot does not show whether the customer can maintain the system without Chinese engineering support. "Connection" becomes a durable public good only when utilization, local capability, cost, reliability, and governance survive after the demonstration phase.

The 36Kr reading: WAIC's H4 startup area exposed the private cost of a national strategy

China's startup press saw a different WAIC. A 36Kr Intelligent Emergence field report described WAIC's H4 startup area as physically difficult to find, located below the main exhibition halls and separated from the largest corporate displays. The spatial arrangement became an economic metaphor. National champions occupied the visible surface. Small companies competed below it for a few square meters, a few minutes of attention, and the possibility of another funding conversation.

The startup area inside WAIC's H4 hall contained 175 projects, including only 22 positions for one-person companies selected from nearly 700 applicants. The implied acceptance rate was about 3 percent. That number measures more than conference selectivity. It reveals the size of the founder supply relative to the number of credible distribution channels available to them.

The operating details are more informative than the slogans. CookiePi charged RMB1,000 for a paid minimum-viable-product test. More than 80 percent of participants reportedly said they would be "very disappointed" if they could no longer use the product, and some sessions lasted roughly two hours. Yet the team still did not know how to price a hardware product whose ongoing token consumption would rise as capabilities expanded.

Mulan AI faced a different constraint. A previously approved financing round was withdrawn, trained employees were recruited by better-funded rivals, and investors asked how the company could survive if a larger video platform copied its workflow. SlashVibe planned an initial production run of about 5,000 devices, but the WAIC demonstration unit was still an unfinished engineering sample held together in part with temporary tape. The product may be early rather than weak, but production intent and validated demand remain different things.

The one-person-company examples were even more compressed. One founder built a WAIC event directory that attracted more than 2,000 registrations by promoting it through more than 50 meetup groups. Another acquired more than 20,000 users for a WeChat bookkeeping tool in two months, yet charged only RMB30 for lifetime access and earned little more than living expenses.

36Kr's separate review of capital flows behind WAIC exhibitors showed the other side of the market. Selected model, embodied-AI, and chip companies raised more than RMB100 billion over the preceding 18 months, but the money was highly concentrated among a small number of leaders. Strategic investors, industrial companies, and state-linked funds increasingly replaced or joined conventional venture capital.

This is not evidence that China's startup system is failing. It shows that the system is sorting rapidly. AI has reduced the cost of building a demonstration, but it has not reduced the scarcity of capital, distribution, customer trust, or defensible market access. The national strategy can produce many founders while still concentrating durable power in a few platforms and state-supported champions.

The Reuters reading: WAIC became a declaration of a rival global order

Reuters interpreted the event at a different level from the product and startup reporting. Its report described the speech as Mr. Xi's clearest articulation yet of China's ambition to shape global AI governance. In this reading, WAIC was no longer only a showcase of Chinese models, chips, robots, and industrial applications. It became a formal statement that Beijing intends to influence who receives AI capability, which institutions write the rules, and whether Washington remains the default center of technological authority.

Reuters identified open-source AI as the central diplomatic instrument. Mr. Xi presented unequal access to intelligence as a potential source of new historical injustice and offered training, cooperation centers, and technical support to developing countries. Chinese open-weight models were therefore framed not simply as cheaper products, but as global public goods that could support a political alternative to U.S.-controlled cloud and semiconductor access.

The institutional evidence made the argument more concrete. Reuters reported that the China-created World Artificial Intelligence Cooperation Organization had signed up 29 countries, while 35 countries supported Washington's AI Opportunity Statement. Kazakhstan was reportedly the only country listed in both initiatives. The numbers are not a geopolitical scoreboard, but the limited overlap suggests that two different governance and technology networks are beginning to form.

Institutional signal reported by Reuters	China-centered framework	U.S.-centered framework
International initiative	World Artificial Intelligence Cooperation Organization	AI Opportunity Statement and the broader Pax Silica strategy
Reported participation	29 countries	35 countries
Primary organizing language	Open source, capability building, development, and broader participation	Innovation, secure supply chains, controlled frontier access, and trusted partners
Reported overlap	Kazakhstan was the only country listed in both initiatives

Reuters also exposed a contradiction inside the Chinese declaration. It reported that Beijing was considering restrictions on overseas access to some leading Chinese models. That possibility does not invalidate China's open-source strategy, but it demonstrates that Beijing may use national-security controls once domestic models become sufficiently strategic. The difference between the two systems is therefore not that only the United States restricts technology. It is the layer at which each country chooses to impose control and the political language used to justify it.

Finally, Reuters connected AI access with AI safety. Mr. Xi called for early-warning systems, emergency-response mechanisms, and safeguards against autonomous systems escaping human oversight. This complicates the simple claim that China favors access while the United States favors safety. Both systems claim to pursue access and control; they differ over where control resides, how widely capability is distributed, and which institutions judge acceptable risk.

The Business Insider reading: openness is the clearest strategic divergence

The U.S. edition of Business Insider identified openness as the most visible strategic difference between the two AI powers. It pointed to DeepSeek, Moonshot AI's Kimi K3, and Z.ai's GLM 5.2 as prominent Chinese open models, while OpenAI and Anthropic continue to operate primarily through closed systems. Meta remains the major American exception through the Llama family, although Business Insider noted that not every Meta model is released openly.

The article also captured the security logic behind the American position. Anthropic has argued that unrestricted model release can create dangerous uses that cannot later be revoked or moderated. Mr. Xi's position at WAIC was the inverse: wider access, cooperation, and open source were presented as requirements for preventing intelligence from becoming the property of a small group of states and companies.

Business Insider also supplied an important counterexample to a simple China-open, America-closed narrative. Beijing ordered Meta to unwind its acquisition of Manus, the Chinese-founded AI agent company that had relocated to Singapore. The episode shows that China can promote open model distribution while restricting ownership, capital movement, data, and control over strategically important firms.

The relevant distinction is therefore not moral openness versus moral closure. It is where each system places control. The United States tends to preserve control at the model, cloud, chip, and export-license layers. China can open model weights while preserving control at the market-access, data, ownership, information, and state-governance layers.

The BBC reading: governance language under geopolitical pressure

The BBC Chinese report included in the research material is useful because a British media lens sits outside the direct American national narrative while remaining inside the Western security, regulatory, and alliance system. It encourages a different question: not simply whether China or the United States is winning, but whether China's language of shared governance can survive contact with strategic rivalry.

From that perspective, the tension is larger than open and closed models. China asks other countries not to generalize national security, while the United States and its allies view advanced compute, autonomous systems, infrastructure software, and model access as potential military and intelligence assets. Beijing emphasizes sovereignty and cultural plurality, while foreign users may ask how those principles coexist with domestic censorship, state access, cybersecurity reviews, and restrictions on cross-border information.

The British angle also prevents the analysis from reducing the world to two self-contained blocs. Europe, the United Kingdom, Southeast Asia, the Gulf states, and many developing economies may want Chinese model prices, American chips, local data control, and European-style regulation at the same time. Their choices will be transactional and layered rather than civilizationally pure.

The BBC lens therefore adds an international credibility test. China's governance proposal becomes influential not when foreign governments repeat its language, but when they are willing to place critical infrastructure, public data, and long-term technical dependence inside systems linked to Chinese suppliers.

The Foreign Affairs reading: China may weaken the system that enabled its rise

Foreign Affairs provides the broadest American strategic critique in this article. Enrico Fardella and Sergey Radchenko do not argue that China's industrial rise is fictitious. Their warning is that the rise occurred inside an international order China increasingly seeks to revise: Western consumer markets absorbed Chinese output, international firms transferred expertise, global capital financed expansion, and an open trading system allowed Chinese companies to scale.

In their account, Beijing has pursued a dual-track strategy. China uses the existing order when openness accelerates domestic development, while building parallel institutions, supply chains, standards, and political relationships that reduce Western leverage. The approach was highly effective during China's ascent. It becomes more unstable as China's scale begins to damage industries and political coalitions in the markets on which Chinese growth still depends.

This framework changes the reading of the WAIC declaration. Open-source models, lower-cost infrastructure, Global South training, and a Shanghai-centered governance organization can be genuine contributions to broader AI access. They can also become instruments for shifting the center of technical dependence away from U.S.-led institutions and toward Chinese platforms, standards, financing, and diplomatic networks.

Foreign Affairs therefore asks a different question from Reuters or Business Insider. Reuters identifies the construction of rival institutions. Business Insider identifies the open-versus-controlled model strategy. Foreign Affairs asks whether China's method of expansion will provoke the protectionism, security restrictions, and market fragmentation that eventually reduce the external demand required to sustain that expansion.

The article's most important economic conclusion is that China cannot resolve every domestic demand problem through greater production. If investment, industrial capacity, and savings grow faster than household consumption, the resulting surplus would require absorption abroad. A sustainable Chinese AI system therefore depends not only on the socialization of intelligent production, but also on a wider distribution of the income and productivity gains created by intelligence.

What the four perspectives reveal together

The Chinese policy narrative explains the intended architecture. The 36Kr reporting exposes the entrepreneurial cost of building it. Reuters shows the declaration becoming an international institutional contest. Business Insider identifies open source as the main competitive instrument while questioning where control actually resides. The BBC perspective relocates the debate from national ambition to international trust. Foreign Affairs adds the macroeconomic constraint: an AI system built through expanding supply still depends on domestic demand and continued access to external markets.

Together, they provide five connected tests for China's AI vision: technical deployability, commercial survivability, institutional reciprocity, international trust, and macroeconomic absorption. The corporate data in the following sections measure how far each condition has advanced.

Axis Two: The U.S.-China AI Civilization Divergence

China's declaration matters because it is not simply another national AI plan. It proposes a different mechanism for organizing intelligence. Reuters described the WAIC address as Mr. Xi's clearest articulation yet of an ambition to shape global AI governance, while the creation of WAICO connected that ambition to a formal diplomatic platform. The American system and the Chinese system both seek global scale, but they distribute power, cost, and dependency through different layers.

This divergence follows the broader framework developed in Hi K Robot's analysis of a two-operating-system world. The central claim of that framework is that institutional boundaries are becoming hard constraints rather than temporary friction. AI, capital, supply chains, data, and standards may still cross borders, but they increasingly do so through routes defined by security alignment, regulatory compatibility, and sovereign control.

The institutional split is already visible. China's 29-country WAICO emphasizes development, open models, and broader participation by the Global South. Washington's 35-country AI Opportunity Statement and the wider Pax Silica strategy emphasize secure supply chains, frontier capability, innovation, and trusted-partner access. The small reported overlap between the two groups suggests that the model competition is becoming a competition between networks of states, standards, infrastructure, and suppliers.

Open-weight diffusion versus controlled service access

China's most visible competitive instrument is the open-weight model. DeepSeek, Alibaba's Qwen family, Moonshot AI's Kimi K3, and Z.ai's GLM series allow developers and governments to inspect, adapt, and deploy models without remaining permanently inside a single vendor API. Kimi K3, for example, is presented by Moonshot AI as a native multimodal model with 2.8 trillion total parameters and a one-million-token context window.

The leading American laboratories use a different method. OpenAI and Anthropic retain control of their most capable models and sell access through APIs, cloud partnerships, and managed products. That structure preserves intellectual property, safety intervention, usage monitoring, and recurring revenue. Meta's Llama family provides a partial American open-weight countermodel, but Business Insider noted that Meta itself is selective about which systems it releases.

The divergence is therefore economic as well as ideological. China's route lowers the price of adoption and helps a model spread into jurisdictions that want sovereign hosting. The American route concentrates control but creates clearer recurring revenue and tighter integration with enterprise clouds.

State-coordinated industrial integration versus private hyperscale concentration

China's system coordinates ministries, local governments, state enterprises, universities, cloud platforms, chip suppliers, and manufacturers. Huawei can connect Ascend processors, Atlas SuperPoDs, CANN software, networking, and government or enterprise customers. Alibaba can connect Qwen, Model Studio, Alibaba Cloud, T-Head chips, Taobao, logistics, and payment systems. Baidu can connect ERNIE, AI Cloud, search, advertising, and Apollo Go. Tencent can connect Hy3 with Yuanbao, CodeBuddy, WorkBuddy, ima, and Marvis, while financing the transition with cash generated by games, advertising, payments, and WeChat.

The United States concentrates comparable power in private hyperscale platforms. In the quarter ended March 31, 2026, Microsoft reported $82.9 billion in total revenue, $54.5 billion in Microsoft Cloud revenue, and 40 percent growth in Azure and other cloud services. Microsoft also said its AI business had surpassed a $37 billion annual revenue run rate, up 123 percent year over year. That single company's AI revenue base is larger than the reported AI revenue of most national ecosystems.

This scale changes the competitive meaning of private enterprise. OpenAI, Anthropic, Google DeepMind, Meta, Microsoft, Amazon, and Nvidia are not small market actors operating outside state power. Their capital expenditure, cloud distribution, data-center networks, semiconductor procurement, and developer ecosystems give them quasi-infrastructural influence. The American system is market-led, but it is not decentralized.

Physical deployment versus software and cloud monetization

China's comparative advantage is strongest where intelligence enters manufactured systems. Huawei Atlas, Baidu Apollo Go, Unitree H1 and G1, AGIBOT humanoids, industrial vision systems, electric vehicles, batteries, sensors, motors, and factory automation form a path from model to machine.

The United States remains stronger in frontier accelerators, cloud software, enterprise subscriptions, and model services. Nvidia's accelerator and software stack, Microsoft Azure, Amazon Web Services, Google Cloud, OpenAI, and Anthropic monetize intelligence through high-value digital infrastructure before it reaches a physical machine. American companies such as Tesla and Figure AI are also pursuing humanoids, but China's electronics and machinery supply chains provide a faster route to low-cost production.

The divergence is not simply software versus hardware. It is the order in which value is captured. The American stack attempts to monetize intelligence first through chips, cloud, APIs, and enterprise software. The Chinese stack is more willing to use low-cost models and subsidized infrastructure to accelerate adoption, then capture value through manufacturing, integration, services, and standards.

Universal-access diplomacy versus trusted-partner exports

Mr. Xi framed AI access as a development right and opposed the expansion of national-security restrictions. China's cooperation centers, training programs, weather systems, language projects, and open models are designed to make that argument operational.

The American approach is also an export strategy, but it is organized around controlled access. The White House's AI Action Plan calls for exporting complete American AI technology packages, including hardware, models, software, applications, and standards, while reducing reliance on systems associated with strategic competitors. The objective is diffusion inside a trusted network rather than universal technological access.

The two systems are diverging, but neither is internally pure

China's model companies compete fiercely, private capital still matters, and many Chinese developers depend on international research and tools. The United States uses export controls, subsidies, public procurement, defense contracts, and energy policy to shape its nominally private system. Meta releases open models; Chinese regulators restrict strategic transactions. American clouds support models from multiple countries; Chinese models can run on American hardware.

The emerging divide is best understood as a difference in default architecture rather than a complete separation. China defaults toward state-coordinated production, open-weight diffusion, sovereign deployment, and multilateral development language. The United States defaults toward privately controlled frontier capability, paid cloud access, advanced semiconductor control, and alliance-based security.

The distinction becomes harder once AI moves from copilots into command layers. As developed in Hi K Robot's analysis of the U.S.-China decision-infrastructure divide, systems that allocate resources, prioritize maintenance, control logistics, support intelligence operations, or trigger institutional workflows are no longer ordinary software products. They become sovereign decision stacks. At that layer, identity, permissions, auditability, procurement rules, cloud jurisdiction, and political legitimacy matter as much as model quality.

Structural layer	China-centered model	U.S.-centered model
Model access	DeepSeek, Qwen, Kimi, and GLM open-weight distribution	OpenAI and Anthropic closed services; Meta selectively open
Compute	Huawei Ascend and Atlas systems, domestic cloud and state procurement	Nvidia-centered accelerators integrated with Azure, AWS, and Google Cloud
Monetization	Cloud infrastructure, industrial integration, hardware, and lower-cost diffusion	High-margin chips, APIs, enterprise subscriptions, and hyperscale cloud
Physical AI	Apollo Go, Unitree, AGIBOT, manufacturing and urban deployment	Tesla, Figure AI, Nvidia Isaac, software and frontier-model integration
International expansion	Global South training, sovereign hosting, application centers, and standards	Trusted-partner full-stack exports, security controls, and alliance networks
Primary political claim	Broad AI access and joint governance	Secure, controllable, and commercially sustainable frontier capability

Corporate Reality Test: Can China Operationalize the Declaration?

The two-system divergence establishes the strategic direction. The following company evidence tests whether China possesses the operating machinery to carry its side of the divide. Durability depends on a sequence in which domestic compute supports models, models generate usage, usage creates revenue, revenue finances continued development, physical deployment creates productive demand, and overseas adoption survives political and commercial scrutiny.

Feasibility Test One: Can China Supply Enough Compute?

Every vision of open AI eventually reaches the same physical constraint: intelligence requires chips, memory, networking, electricity, cooling, and capital. Open weights cannot compensate for a compute system that is unreliable, too expensive, or too difficult to program. The first feasibility test is therefore not whether China can produce impressive models under laboratory conditions. It is whether Chinese firms can supply a usable compute stack at scale.

This is the infrastructure-dependency problem examined in Hi K Robot's NVIDIA Matrix analysis. NVIDIA's strategic position does not come from accelerators alone. It comes from the CUDA-centered coordination of hardware, memory, networking, compilers, libraries, orchestration, cloud deployment, developer habits, and enterprise trust. The comparison with COMAC is useful because manufacturing an aircraft did not automatically reproduce the certification, maintenance, financing, training, and logistics system surrounding Boeing and Airbus. In the same way, manufacturing an AI accelerator does not automatically reproduce the operational environment surrounding NVIDIA.

Huawei is turning a chip deficit into a systems-engineering problem

Huawei's Atlas 950 SuperPoD represents the most important Chinese response to restrictions on advanced accelerators. The 1,024-card system is specified at 1 EFLOPS of FP8 compute, 256 TB of globally addressed memory, and approximately 3 microseconds of round-trip interconnect latency. Huawei also says more than 750 Atlas 384 systems have been deployed, supported by more than 3,000 partners, over 7,000 industry solutions, and more than 2,000 core government and enterprise customers.

These are vendor-reported deployment figures rather than independently audited market-share data, but they demonstrate that the Ascend strategy is no longer confined to prototype systems. The architecture treats large numbers of domestic processors as one logical machine, using high-bandwidth interconnects and coordinated memory to reduce the penalty created by weaker individual chips.

This approach is rational. If China cannot reliably obtain the best foreign accelerator, it can compensate through scale, interconnect design, scheduling, lower-precision formats, and closer integration between hardware and software. The strategy resembles China's response in several other infrastructure industries: accept a component disadvantage, then use manufacturing depth and system-level engineering to narrow the usable-performance gap.

But a larger cluster is not a free substitute for a better chip. More processors can mean higher power consumption, more network components, more failure points, and more complicated software optimization. The most important commercial measurements are therefore not theoretical peak operations. They are useful tokens per kilowatt-hour, uptime under sustained workloads, memory efficiency, programmer productivity, migration cost from CUDA-based environments, and total cost per completed training or inference task.

Huawei has a strong position because it can integrate chips, servers, networking, storage, software, and customer relationships. Smaller domestic accelerator companies face a harder problem: persuading cloud operators and model developers to support additional software stacks while competing for scarce advanced packaging, memory, and fabrication capacity. A policy can create many chip vendors; it cannot guarantee that customers will tolerate fragmented toolchains.

The NVIDIA infrastructure-dependency framework therefore changes the substitution test. Progress cannot be measured only by peak FLOPS or the number of domestic chips shipped. The more relevant indicators are migration time from CUDA, framework compatibility, cluster uptime, inference economics, engineering productivity, developer adoption, and the willingness of enterprises to standardize long-term workflows around Ascend, CANN, MindSpore, or other Chinese environments.

China does not need complete semiconductor independence to build a large AI system

The argument is sometimes framed too absolutely. China does not need to reproduce every American or allied semiconductor capability before it can deploy AI at national scale. It needs enough secure supply for priority workloads, enough software compatibility to keep utilization high, and enough imported or domestically produced hardware to prevent severe bottlenecks.

That threshold is lower than full technological independence but higher than a successful product demonstration. The 2026 reality is mixed: Chinese firms can operate large clusters, train competitive models, and deploy inference widely, but they remain constrained by advanced fabrication equipment, high-bandwidth memory, electronic design automation, packaging, and the mature developer ecosystem surrounding American accelerators.

This is why the U.S.-China AI contest is inseparable from the broader network of strategic chokepoints discussed in Hi K Robot's analysis of fragmented globalization. A model may be national, but the production system behind it remains international.

Preliminary judgment

China has sufficient compute capability to sustain a large domestic AI economy and to offer meaningful capacity to selected overseas partners. It has not yet demonstrated that a fully domestic stack can match the efficiency, software maturity, and scalability of the leading U.S.-centered ecosystem across all workloads. The compute foundation is viable, but the cost of sovereignty remains high. This is the first material boundary between the declaration and the U.S.-China divergence: China can build an alternative stack, but it cannot yet assume that the alternative is equally efficient.

Feasibility Test Two: Is AI Producing Revenue or Only Capacity?

Official statistics place China's core AI industry above RMB1.2 trillion. That number establishes scale, but it does not answer how much cash customers are paying for model access, agents, cloud inference, or physical AI. "Industry scale" can include hardware, software, research activity, and adjacent businesses; it is not equivalent to audited AI revenue or profit.

The most useful evidence comes from companies that report financial results. Baidu, Alibaba, and Tencent reveal three different operating models: an AI company trying to replace a weakening legacy engine, a cloud and commerce platform building a full-stack commercialization loop, and a cash-rich super-app company using established profits to fund new AI products.

Baidu: infrastructure demand is growing faster than applications

Baidu reported RMB13.6 billion in revenue from its core AI-powered business in the first quarter of 2026, up 49 percent year over year. AI Cloud Infrastructure contributed RMB8.8 billion, up 79 percent, while GPU Cloud revenue increased 184 percent. By contrast, AI Applications produced RMB2.5 billion, approximately flat from a year earlier.

This is one of the clearest windows into China's AI economy. Enterprises are paying for compute, hosting, and infrastructure. They are not yet increasing their spending on standalone AI applications at the same rate. The market has proven demand for the means of producing intelligence more clearly than demand for many of the final products.

Baidu's numbers also show the cost side. The company said cost of revenue increased partly because of AI Cloud, while operating cash flow was RMB2.7 billion. The AI transition is real, but infrastructure growth consumes equipment, electricity, depreciation, and working capital. Revenue growth is not the same as a high-return business.

Apollo Go provides a second test. The service delivered 3.2 million fully driverless rides in the first quarter, up more than 120 percent year over year, and cumulative public rides exceeded 22 million by April. That is genuine deployment at a scale few autonomous-driving programs have reached. Yet ride counts alone do not prove unit profitability. The decisive figures would include fare revenue, vehicles per remote operator, utilization by city, insurance cost, maintenance, depreciation, local incentives, and the capital required to enter each overseas market.

Apollo Go therefore demonstrates that China can move AI into the physical economy. It does not yet demonstrate that global robotaxi deployment will generate returns comparable with the capital invested.

Alibaba: the strongest evidence of a commercial flywheel

Alibaba offers a more integrated commercial structure. In the final quarter of fiscal 2026, Cloud Intelligence revenue reached RMB41.6 billion, external revenue growth accelerated to 40 percent, and AI-related products accounted for 30 percent of external revenue. Alibaba said annualized AI-related product revenue exceeded RMB35.8 billion, while the customer base for Model Studio grew eightfold year over year.

The company also reported that more than 100,000 of its self-developed Zhenwu processors were deployed on Alibaba Cloud's public platform and that more than 60 percent of T-Head compute capacity served external customers. These are company-reported figures, but they show how Qwen, Model Studio, cloud infrastructure, and proprietary chips are being connected into a commercial stack rather than operated as isolated research projects.

This is important because Alibaba can connect every layer of the stack. Qwen models attract developers. Alibaba Cloud sells training and inference. Enterprise agents create recurring workloads. Taobao, Tmall, Amap, Fliggy, Alipay, and logistics services provide distribution and operating data. Proprietary T-Head chips can be consumed internally even if they do not become independent merchant products.

The structure is close to what a sustainable Chinese AI platform would need: open models for adoption, a commercial cloud for monetization, and a large domestic ecosystem for use cases. The unresolved question is margin. Rapid cloud growth can still destroy value if accelerator costs, price competition, and model development consume most of the incremental revenue.

Alibaba's model also illustrates why open source is not necessarily opposed to corporate power. Open weights can function as customer acquisition. The user receives freedom at the model layer, while the platform earns money from compute, storage, orchestration, security, databases, and enterprise integration.

Tencent: mature cash flow subsidizes the new frontier

Tencent reported first-quarter revenue of RMB196.5 billion, non-IFRS operating profit of RMB75.6 billion, capital expenditure of RMB31.9 billion, and free cash flow of RMB56.7 billion. It also disclosed that non-IFRS operating profit would have been RMB84.4 billion excluding the revenues, costs, and expenses associated with its new AI products.

The gap is a useful measure of transition cost. Tencent can fund models and agents because games, advertising, payments, social networks, and business services generate large amounts of cash. It can also deploy AI inside existing products with more than a billion users instead of paying to build distribution from zero.

That advantage is not available to most Chinese model startups. A standalone laboratory faces talent and compute costs while charging prices low enough to compete with subsidized offerings from the largest platforms. The likely outcome is consolidation. China may retain many application companies, but the capital-intensive foundation layer will tend to concentrate around firms with cloud infrastructure, state backing, strategic investors, or cash-generating consumer platforms.

Preliminary judgment

China has moved beyond an AI economy consisting only of research and government projects. Cloud infrastructure is generating substantial revenue, and physical deployments are measurable. However, the revenue is strongest where customers buy compute or where incumbent platforms use AI to protect existing businesses. Standalone application economics remain less proven. The American model currently converts frontier capability into cloud and software revenue more directly; the Chinese model is proving deployment scale faster than it is proving application margins.

Feasibility Test Three: Can Open Source Become a Global Export System?

Open source is the center of China's international proposition. DeepSeek, Qwen, Kimi, GLM, MiniMax, and other model families make Chinese capabilities easier to inspect, adapt, and deploy than a service that exists only behind a foreign API. At WAIC, Moonshot AI introduced Kimi K3, which the company describes as a native multimodal model with 2.8 trillion total parameters and a one-million-token context window, reinforcing the message that Chinese laboratories intend to compete on both capability and accessibility.

The strategic value of this approach is larger than direct model revenue. A model adopted by a telecommunications operator, university, ministry, bank, or manufacturer can create demand for Chinese engineering support, cloud services, accelerators, cybersecurity products, databases, and application software. It can also normalize Chinese technical standards.

For governments in the Global South, local deployment solves several problems at once. Sensitive data can remain inside the country. The model can be fine-tuned for local languages. Costs can be lower than repeated use of a frontier American API. Political leaders gain the appearance of digital sovereignty rather than dependence on a foreign cloud provider.

But open source also weakens the supplier's control. A developer can download a Chinese model and run it on American chips, a European cloud, or a domestic data center. The adoption may expand China's intellectual influence without producing Chinese corporate revenue. An open model becomes a durable export system only when the surrounding stack is competitive enough that users voluntarily buy the rest of it.

Four conversion points determine whether openness becomes power

The first conversion point is from downloads to sustained usage. Public attention can produce millions of downloads without producing production workloads.

The second is from usage to paid compute. If the model runs primarily on infrastructure supplied by other countries, the Chinese developer may receive prestige but limited revenue.

The third is from compute to standards. A model family becomes strategic when developers build tools, data formats, safety procedures, and enterprise systems around it.

The fourth is from standards to political trust. Durable adoption depends on governments believing that the supplier will maintain the model, disclose important vulnerabilities, respect local data rules, and continue supporting the deployment during a diplomatic dispute.

China is strongest at the first conversion point and increasingly competitive at the second. The third is under construction. The fourth remains uncertain because technical openness coexists with a tightly controlled domestic information system and with geopolitical concerns about data access, censorship, and state influence.

Open models do not automatically create an open market

Mr. Xi's critique of technological exclusion has force because American export policy explicitly treats compute as a strategic asset. Yet foreign governments and companies can reasonably ask whether Chinese openness is reciprocal. Can foreign foundation models enter the Chinese market with the same functional scope? Can foreign cloud providers access comparable customers and data? Are training sources, filtering rules, and government requirements transparent enough for critical infrastructure?

Model openness, market openness, and institutional openness are different properties. China may be more open than some U.S. laboratories at the level of downloadable weights while remaining more restrictive at the levels of information, market access, ownership, and political accountability. As noted in the Reuters section, the boundary between model diffusion and state control can shift once a capability becomes strategically valuable.

Preliminary judgment

Open source gives China a credible path to global relevance even when it does not control the most advanced chip. It is likely to win substantial adoption, especially where cost and sovereign hosting matter more than access to the absolute frontier. Whether that adoption becomes a durable Chinese-led ecosystem depends on the competitiveness of the surrounding cloud, hardware, support, and trust architecture. The central test is whether open weights become an operating ecosystem rather than remain a distribution statistic.

Feasibility Test Four: Can Physical AI Escape the Demand Gap?

Physical AI is where the Chinese vision appears strongest and where the Foreign Affairs warning becomes most relevant. China has the supply chains, engineering labor, factories, batteries, motors, sensors, and local government support required to produce robots at speed. It may therefore transform model intelligence into machines faster than a software-centered economy can build manufacturing depth.

The numbers are striking. More than 13,000 humanoid robots were shipped globally in 2025. Chinese companies accounted for roughly 85 percent. AGIBOT and Unitree each shipped more than 5,000 units. Unitree confirmed that its 2025 humanoid deliveries exceeded 5,500 units. Associated Press reported that Unitree generated about RMB1.7 billion in revenue and more than RMB278 million in profit.

These are not merely prototypes. They show that Chinese manufacturers can standardize components, manage suppliers, reduce assembly cost, and deliver physical products. Local parts also make Chinese humanoids at least 20 percent cheaper on average than foreign alternatives, according to estimates cited by Associated Press.

Yet supply has moved ahead of proven demand. Many current buyers are government entities, research institutes, power plants, data centers, schools, and exhibition operators. Robots that dance, box, greet visitors, or perform controlled demonstrations can generate orders without being economically superior to human labor in an unstructured workplace.

The decisive metric is not shipments but productive hours

Humanoid companies often report units sold, contracts signed, or production targets. For AI civilization, the more meaningful measurement is productive autonomous hours completed per machine without human intervention.

A robot can be inexpensive and still fail commercially if it requires frequent teleoperation, operates too slowly, has limited battery life, damages equipment, or cannot adapt to a changing environment. Factory buyers will compare the complete cost with a conventional industrial robot, a redesigned production line, or a human worker. Household buyers will compare it with much cheaper single-purpose devices.

The commercial threshold therefore includes reliability, safety certification, maintenance networks, task-learning speed, and measurable labor savings. China's manufacturing advantage lowers the hardware threshold. It does not eliminate the intelligence and service threshold.

This is the same structural divide examined in Hi K Robot's U.S.-China AI robotics analysis: China enters embodied AI with industrial density, supplier continuity, lower-cost hardware, and more deployment environments, while the United States enters with cognitive leverage, frontier models, simulation, software platforms, and capital willing to fund generalization. The durable advantage will belong to the system that closes the loop between deployment, operational data, model improvement, and redeployment faster.

China's advantage may be data generated through deployment

Even low-value early deployments can become strategically useful if they generate operational data. A robot placed in a warehouse, retail store, factory, or laboratory produces examples of failure and recovery that improve future models. State-backed procurement can therefore function as a data acquisition mechanism as well as an industrial subsidy.

This creates a powerful loop: policy creates early demand, manufacturing lowers cost, deployment creates data, data improves control models, and improved models expand the addressable market. The loop is plausible. The risk is that public procurement continues to validate production volume after private customers have concluded that the machines do not yet produce sufficient value.

Preliminary judgment

China is well positioned to lead the manufacturing phase of physical AI. It has not yet proven that current humanoid demand is broad enough to absorb the announced capacity. The sector is both a genuine industrial advantage and a candidate for excess supply.

Inside WAIC's H4 Startup Area: The Reality of Chinese Startups

The national narrative at WAIC was expansive, but 36Kr's reporting from WAIC's H4 startup area provided a more revealing view of the operating economy. The area was difficult to find beneath the main exhibition halls. Founders with unfinished prototypes competed for small exhibition spaces, investor attention, media coverage, and customer introductions. The physical hierarchy of the venue mirrored the capital hierarchy of the industry.

WAIC's H4 startup area contained 175 projects. Only 22 positions were reserved for one-person companies, selected from nearly 700 applicants across eight regional competitions. The approximate 3 percent acceptance rate shows how many founders are entering AI relative to the number of institutional gateways capable of giving them visibility.

Those who did enter still had to divide their time between operating a booth, meeting investors, finding customers, joining media interviews, and studying competitors. WAIC temporarily compressed financiers, industrial partners, developers, reporters, and potential customers into one space. For an early-stage company, the conference was less a celebration than a three-day acquisition channel.

CookiePi: strong user attachment, unresolved token economics

CookiePi, created by a team led by a former founding employee of SenseTime, placed interactive AI characters inside toys and household objects. Before WAIC, the company had raised tens of millions of yuan in seed financing. It also ran a paid RMB1,000 minimum-viable-product test rather than distributing the product entirely for free.

The early engagement signal was unusually strong. More than 80 percent of tested users reportedly said they would be very disappointed if they could no longer use the product. Some used it for roughly two hours at a time, limited more by battery life than by interest. The company even disabled use while charging as an anti-addiction measure.

Yet the economic question remained unresolved. The hardware could be sold once, but inference, personality memory, content generation, safety monitoring, and feature upgrades would create recurring token costs. High engagement could increase cost faster than revenue unless the company found a subscription, usage, content, or hybrid pricing model that users would accept.

Mulan AI: working technology without protected distribution

Mulan AI built a video-marketing agent that combined a canvas, reusable workflows, consistent digital assets, scene generation, and editable production steps. The product addressed a real workflow rather than offering only prompt-to-video generation.

Its difficulty was market structure. A financing round that had passed internal approval was later withdrawn. Employees trained by the company were recruited by better-funded competitors. Investors asked what would happen if a larger platform such as Liblib reproduced the feature set.

The question captured a central weakness of the application layer. A small company can innovate first and still lose because a platform owns traffic, cloud credits, creator relationships, and distribution. Mulan AI's response was to look for customers and overseas niches rather than compete directly for the same domestic mass market.

SlashVibe: the cost of proving a category before a platform validates it

SlashVibe was an external control device designed to help users invoke and organize computer agents. To prepare for WAIC, its team spent about half a month connecting the embedded hardware and software into a complete demonstration. The device was still an unfinished engineering sample, with temporary tape securing part of the connection, while the team planned an initial production run of approximately 5,000 units.

Investors had repeatedly questioned whether a dedicated AI interaction device was necessary. After a major American AI company introduced a specialized input product, the investor question changed from whether the category existed to how SlashVibe would compete.

The example reveals a recurring startup paradox. Before a large company enters, the category is considered unproven. After a large company enters, the category is accepted but the startup's scarcity is questioned. The startup pays the cost of discovering demand without necessarily receiving the value of validating it.

One-person companies: product creation is cheap, market access is not

One selected founder built a secure execution environment for agents and won enough orders to keep the project alive. He also created a WAIC event directory that attracted more than 2,000 registered users by promoting it through more than 50 conference meetup groups.

Another founder built a bookkeeping service inside WeChat. By repeatedly posting on Xiaohongshu and using the comment sections of popular posts, the product accumulated more than 20,000 users in two months without a conventional marketing budget. But lifetime membership cost only RMB30, and the resulting income was still little more than enough for rent and basic expenses.

These examples demonstrate that AI lowers the cost of software production and enables tiny teams to reach meaningful user counts. They also show why user count is not the same as a business. Low prices, lifetime plans, dependence on another platform, and founder-led promotion can produce adoption without durable cash flow.

What WAIC's H4 startup area adds to the feasibility test

The evidence from WAIC's H4 startup area complicates both the official and Western narratives. China is not only a top-down state project. It contains genuine entrepreneurial experimentation, paid testing, product iteration, and founders willing to accept extreme uncertainty. At the same time, the ecosystem is not broadly equal. Capital, traffic, cloud resources, and institutional sponsorship remain concentrated.

AI capability may be becoming more accessible, but business survival is not. The likely structure is a wide application layer built on top of a much narrower foundation layer controlled by cash-rich platforms, telecommunications groups, industrial champions, and state-backed infrastructure providers.

The Foreign Affairs Structural Test

The enterprise evidence shows that China's AI project is neither imaginary nor fully self-sustaining. Huawei, Alibaba, Baidu, Tencent, Moonshot AI, Unitree, AGIBOT, and hundreds of smaller companies have created real products, revenue, deployment, and industrial learning. The Foreign Affairs framework tests whether those achievements can scale without undermining the external economic and institutional environment on which they still depend.

China's AI stack remains embedded in the order it seeks to surpass

Fardella and Radchenko's first argument is historical. China's rise did not occur outside the Western-led international system. Reform and opening connected Chinese labor, factories, capital formation, and state capacity to foreign technology and consumer markets. China did not merely receive benefits passively; it used them strategically and built formidable domestic capabilities. But the external system was still part of the production function.

The same pattern appears in AI. Chinese laboratories build on globally published transformer research, open-source software, international developer tools, foreign scientific collaboration, and semiconductor supply chains that include non-Chinese equipment, memory, packaging, and design knowledge. Chinese models may run on domestic accelerators, but they also circulate through global repositories and are frequently deployed on Nvidia-based infrastructure outside China.

This does not mean China cannot develop autonomous capacity. Huawei's Ascend and Atlas architecture, Alibaba's Zhenwu processors, domestic model families, and local cloud platforms show substantial progress. It means that the transition from interdependence to a separate civilization stack is a costly continuum rather than a completed event.

The dual-track strategy is becoming an AI architecture

The second Foreign Affairs argument concerns institutional behavior. China has historically used existing international institutions while building parallel mechanisms that increase its strategic room. In AI, the same dual track is visible.

Chinese companies continue to participate in global research, international cloud markets, open-source communities, multinational supply chains, and standards discussions. At the same time, Beijing is supporting domestic accelerators, sovereign cloud infrastructure, Chinese model ecosystems, national data rules, WAICO, Global South training centers, and technical standards that can operate with less dependence on the United States.

Mr. Xi's WAIC declaration functions as the public doctrine of this dual-track system. China is not withdrawing from global AI. It is using global openness to expand Chinese capability while constructing an alternative center of gravity that could survive greater technological separation.

AI can reproduce excess capacity even when software is cheap to copy

The third argument is economic. China's industrial model frequently combines high investment, local government competition, subsidized finance, and ambitious production targets. When domestic demand is insufficient, companies seek overseas markets. Foreign governments then react when imports threaten local employment, industrial capability, or strategic autonomy.

Foundation-model software differs from steel, solar modules, or electric vehicles because the marginal cost of distributing another model copy is low. But an AI civilization is not made only of model files. Its physical layers are capital intensive:

• accelerators, servers, networking, memory, and advanced packaging;
• data centers, substations, generation, storage, and cooling systems;
• robotaxis, humanoids, industrial robots, sensors, and motors;
• cloud credits, model training, inference, security, and technical support;
• industrial parks, laboratories, procurement programs, and local subsidies.

Each layer can be overbuilt. Local governments can finance overlapping intelligent-compute centers without sufficient utilization. Cloud platforms can sell inference below full economic cost to gain developers. Humanoid companies can announce production far above verified private demand. State enterprises and research institutions can become early buyers whose orders prove manufacturing capacity but not long-term customer economics.

Efficiency and subsidy can coexist

The Western debate often treats Chinese price advantages as either genuine efficiency or unfair support. Corporate reality suggests that both can be true.

China's manufacturing clusters reduce component costs, shorten supplier feedback loops, and accelerate product iteration. Unitree and other robot makers benefit from domestic batteries, motors, controls, machining, electronics, and assembly. DeepSeek and other model developers have demonstrated meaningful engineering efficiency. These are real advantages.

At the same time, inexpensive land, policy-directed finance, public procurement, cloud credits, local subsidies, and profits transferred from mature platform businesses can reduce the price visible to the customer. Tencent can absorb AI investment through games, advertising, and payments. Alibaba can use commerce and cloud scale. State-linked capital can fund strategic firms through long development periods.

The correct test is not whether support exists. Every major AI power supports its strategic industries. The test is whether the resulting product eventually covers its full economic cost through voluntary demand.

The European and Global South markets cannot absorb unlimited AI supply

Foreign Affairs emphasizes that China's industrial strategy depends on other economies remaining willing and financially able to absorb surplus output. The United States has already restricted Chinese access to advanced chips and scrutinizes Chinese digital infrastructure. Europe is more divided but is expanding tools for economic security, data governance, procurement controls, and industrial protection.

The Global South is central to Mr. Xi's alternative because many countries need lower-cost technology and want greater sovereignty. Yet these markets have limits. Governments may welcome training, models, and pilot systems but lack the budgets to sustain large cloud bills, robot fleets, or imported infrastructure without concessional financing. A project financed by Chinese policy capital can create deployment without creating an independent local demand base.

This is why contract quality matters more than the number of countries reached. The relevant distinctions are:

• paid commercial contracts from grants or government assistance;
• recurring service revenue from one-time construction;
• locally maintained systems from projects dependent on continuing Chinese support;
• productive usage from conference demonstrations and pilot programs;
• voluntary standards adoption from standards bundled with financing or infrastructure.

Market openness becomes harder to demand when reciprocity is limited

China wants foreign countries to remain open to Chinese models, robots, cloud systems, vehicles, and technical standards. Yet foreign AI companies face extensive content controls, cybersecurity reviews, data restrictions, licensing requirements, and market-access barriers inside China.

This produces the paradox at the center of the Foreign Affairs critique. China asks the global economy to preserve the openness needed for Chinese expansion while reserving greater state discretion inside its own system. The earlier Reuters example illustrates that China's openness remains conditional when a technology is treated as a strategic asset.

The more important Chinese AI becomes to foreign infrastructure, the more likely host governments are to impose local data rules, ownership limits, security certification, source-code reviews, procurement restrictions, or domestic-content requirements. China can reduce resistance through price and financing, but it cannot assume that economic efficiency will override strategic anxiety indefinitely.

From socialized intelligent production to socialized intelligent prosperity

The deepest Foreign Affairs argument is domestic. China's economic imbalance is not simply that it produces too much. It is that household income and consumption do not absorb a sufficiently large share of what the production system can create.

AI can intensify that imbalance. Intelligent factories, autonomous vehicles, cloud agents, and robots can expand output while reducing labor demand in some occupations. If the gains accrue mainly to platforms, industrial champions, local governments, and capital owners, the economy may become more productive without creating enough customers to purchase the additional output.

A sustainable AI system therefore depends on a demand-side transmission mechanism. Its productivity gains would need to appear as some combination of higher household income, lower essential-service costs, broader ownership, improved public services, new employment, or shorter working time. Without that transmission, China can socialize intelligent production while leaving intelligent prosperity concentrated.

What the Foreign Affairs critique does and does not prove

The critique does not prove that China's AI strategy will fail. It does not prove that every low-cost Chinese product is subsidized, that open-source models are disguised geopolitical weapons, or that Western protectionism is automatically justified.

It establishes a structural boundary: China's AI civilization becomes less sustainable if it depends on continuously expanding supply while domestic demand remains weak and foreign markets become more defensive. The better China becomes at building AI capacity, the more important it becomes to create reciprocal markets, credible governance, and a broader internal distribution of the gains.

Foreign Affairs therefore adds the missing economic mechanism to the two main axes of this article. Mr. Xi's declaration explains what China wants to build. The U.S.-China divergence explains why the systems are separating. Corporate evidence shows what already works. The Foreign Affairs test explains what could cause the Chinese system to collide with its own success.

What This Means for the United States

China's path matters to the United States not because every Chinese model, robot, or governance initiative constitutes an immediate threat. It matters because Beijing is assembling an alternative method for distributing intelligence. The competition is therefore moving beyond who trains the most capable model. It is becoming a contest over which country controls the default infrastructure through which other societies obtain compute, models, industrial systems, standards, and technical dependence.

The United States could retain the frontier and lose the distribution layer

The American system remains stronger at the highest-value layers of the current stack. Nvidia accelerators and CUDA shape the compute environment. Microsoft Azure, Amazon Web Services, and Google Cloud operate global infrastructure. OpenAI and Anthropic monetize controlled model access. Microsoft reported $54.5 billion in quarterly cloud revenue, 40 percent growth in Azure and other cloud services, and an AI business above a $37 billion annual revenue run rate.

Those figures show that the United States has already built a far larger commercial AI engine than China. They do not guarantee that American systems will become the default everywhere. Chinese open-weight models can be downloaded, adapted, and hosted without permanent dependence on an American API. Huawei, Alibaba, and Chinese telecommunications or infrastructure partners can combine models with servers, networking, training, financing, and local deployment.

The strategic risk is therefore asymmetric. The United States can remain ahead in frontier capability while losing influence over the middle of the market: governments, universities, manufacturers, and regional cloud providers that value affordability, local control, and sovereign hosting more than the absolute best benchmark result.

Physical AI could move value toward China's strongest industrial layer

The American advantage is concentrated in chips, cloud, software, and frontier research. China's advantage is strongest where intelligence enters manufactured systems. Baidu's Apollo Go, Huawei's industrial AI stack, Unitree's H1 and G1, AGIBOT's humanoids, electric vehicles, sensors, motors, batteries, and factory equipment form a path from model capability to physical deployment.

If foundation models become more interchangeable and inference prices continue to fall, a larger share of value may migrate away from the model itself and toward integration, manufacturing, maintenance, data generated in the field, and control of the installed base. That transition would favor the parts of the Chinese economy that already know how to manufacture complex products at scale.

The relevant American comparison is not only OpenAI against DeepSeek or Anthropic against Kimi. It is Nvidia Isaac, Tesla, Figure AI, cloud robotics, industrial automation, and the broader ability of American firms to connect advanced models with a competitive manufacturing and service ecosystem. A country can lead in intelligence while allowing another country to define the machines through which that intelligence reaches the world.

Export controls create delay and substitution at the same time

American controls on advanced accelerators, semiconductor equipment, and selected technology transfers can slow China's access to frontier capability. The Huawei Atlas strategy also shows the second-order effect: restrictions create a protected domestic market for Chinese alternatives and give customers a reason to tolerate higher migration costs.

The relevant American measure is not whether a restriction produces pain, but whether the period of delay it creates is more valuable than the domestic substitution, political legitimacy, and ecosystem consolidation it creates for China. A control that preserves a durable technological gap has a different strategic value from one that merely transfers revenue from an American supplier to a Chinese replacement.

This does not mean that unrestricted exports would preserve American influence. Advanced compute can support military, intelligence, surveillance, cyber, and autonomous systems. The structural problem is that national-security controls and commercial diffusion are now inseparable. Every restriction changes both the security balance and the market architecture.

The Global South is becoming a test of system design, not diplomatic language

China's proposition to developing countries combines open models, training, sovereign deployment, infrastructure, and non-exclusion. The American proposition combines higher-performing technology, advanced clouds, security safeguards, commercial contracts, and access through trusted networks.

Many governments will not make a permanent ideological choice. They may use Nvidia hardware, a Chinese open model, a domestic data center, European privacy rules, and local applications at the same time. The system that becomes dominant will be the one that is easiest to finance, maintain, localize, and upgrade.

The United States is vulnerable where its offering is too expensive, too closed, or too dependent on continuous access to a small number of hyperscale providers. China is vulnerable where low prices conceal subsidies, local support is weak, data governance is opaque, or technical dependence becomes politically unacceptable. The Global South will test both systems against operational value rather than their preferred narratives.

American strength depends on the attractiveness of the full stack

The U.S. advantage is not any single company. It is the combination of Nvidia, Microsoft, Amazon, Google, OpenAI, Anthropic, Meta, semiconductor allies, research universities, venture capital, enterprise software, and global developer communities. China's strategy is designed to separate selected layers of that stack from one another.

That allied industrial base is especially important in Physical AI. Hi K Robot's analysis of the United States as the AI brain and Japan as the industrial body argues that Japan remains structurally important in semiconductor materials, precision production equipment, servo systems, machine vision, factory automation, robotics, power systems, and heavy industrial throughput. The American stack is therefore strongest when U.S. software, models, chips, and capital remain tightly connected to Japanese, Taiwanese, Korean, and European manufacturing capabilities.

Open models can reduce dependence on American laboratories. Domestic accelerators can reduce dependence on Nvidia. Sovereign clouds can reduce dependence on Azure, AWS, and Google Cloud. Chinese robotics can shift value toward manufacturing. WAICO and development programs can reduce dependence on Western governance forums.

The American system remains strongest when these layers reinforce one another and foreign users perceive the resulting package as capable, affordable, secure, and politically sustainable. It becomes weaker when access is restricted without a credible alternative for ordinary commercial users, or when private concentration makes the system appear less open than the Chinese alternative.

The central American question is not whether China can be stopped

China already possesses enough compute, models, capital, manufacturing, and state capacity to remain a major AI power. The more useful question is which parts of China's path can become globally durable and which remain constrained by cost, demand, reciprocity, and trust.

From an American perspective, the most consequential Chinese success would not be a temporary lead on a model benchmark. It would be the creation of a low-cost, open-weight, sovereignly deployable, physically integrated AI system that other countries can operate without depending on American clouds, licenses, or political approval.

The United States does not need every country to reject Chinese technology. Its structural position is stronger when foreign governments and companies continue to view the American stack as the more productive, trustworthy, and economically sustainable foundation. The competition is therefore over attraction and embedded dependence as much as denial.

Conditions Under Which the Chinese Vision Remains Viable

China's open AI system remains feasible only under several simultaneous operational conditions.

Economically Competitive Domestic Compute

Broad commercial adoption depends on Huawei and other suppliers reducing the total cost of useful training and inference, including power, networking, failures, software migration, and engineering labor. Sovereign hardware that remains permanently more expensive is likely to stay concentrated in strategic workloads.

Conversion of Open-Source Adoption Into Paid Services

Commercial durability depends on revenue from cloud compute, enterprise integration, support, security, fine-tuning, and applications. Download counts and benchmark rankings cannot finance continuous frontier development by themselves.

Demonstrable Labor Economics in Physical AI

Commercial validation depends on productive hours, error rates, maintenance cost, deployment time, and customer savings becoming observable. Government demonstrations can accelerate learning but cannot substitute indefinitely for private return on investment.

Local Capability in Overseas Deployment

Training programs and cooperation centers gain legitimacy when they produce local engineers, locally governed data, and applications that continue operating without constant external intervention. Legitimacy weakens when projects become tied packages that create opaque dependence on Chinese suppliers.

Greater Reciprocity in Market and Institutional Access

China's criticism of exclusion becomes more persuasive when foreign companies receive clearer access to Chinese customers, data rules become more predictable, and governance mechanisms reveal how models are filtered, monitored, and updated.

Transmission of AI Productivity Into Household Income or Lower Costs

An AI economy cannot be sustained only by industrial investment. Enterprises depend on customers, and customers depend on purchasing power. The long-run viability of China's model is linked to whether automation improves household welfare rather than only expanding production.

A Feasibility Scorecard

Structural capability	Current assessment	Operational evidence	Primary constraint
Manufacturing and supply-chain integration	High	Large robot output, electronics depth, vehicles, batteries, motors, sensors, and rapid cost reduction	Risk of capacity expanding before profitable demand
Domestic AI compute	Medium	Huawei SuperPoD architecture and large domestic cloud deployments	Advanced chips, memory, software maturity, power efficiency, and packaging
Cloud commercialization	Medium to high	Baidu AI Cloud growth, Alibaba Cloud acceleration, Tencent's AI-supported business services	Capital intensity, price competition, and uncertain incremental margins
Standalone AI applications	Medium	Rapid product formation and platform distribution	Flat or early revenue, weak moats, token costs, and platform copying
Open-source global reach	High	Multiple competitive model families and strong demand for sovereign deployment	Conversion from adoption to Chinese revenue and standards
Physical AI deployment	Medium to high	Robotaxi scale and leading humanoid shipment volume	Unit economics, reliability, maintenance, and private customer demand
Global governance legitimacy	Medium	WAIC institutions, developing-country training, and Global South partnerships	Reciprocity, censorship, data trust, and geopolitical resistance
Complete separation from the U.S.-centered technology system	Low to medium	Growing domestic alternatives across models, chips, and cloud	Semiconductor tools, global research, developer ecosystems, and foreign markets

Analytical Compression

Structural Anchor

The central structure is not a simple contest between two national technology sectors. It is a struggle over where control resides across the AI civilization stack. The United States concentrates power in frontier chips, hyperscale clouds, closed model services, enterprise software, and alliance-based supply chains. China is building leverage through open-weight diffusion, domestic compute, manufacturing integration, state-supported deployment, and Global South institutions.

Within Hi K Robot's wider map, this article connects five existing structural models: the two-operating-system world, the sovereign decision-infrastructure divide, the industrial-versus-cognitive robotics competition, the U.S.-Japan brain-and-industrial-body alliance, and the NVIDIA-centered AI infrastructure-dependency problem. WAIC gives those previously separate layers a single Chinese declaration: models, compute, industry, institutions, and global governance are being assembled into one civilization-scale operating system.

China does not need to surpass the United States at every layer to alter the global order. It can gain influence by making selected layers cheaper, more sovereign, and easier to deploy. The U.S. position is strongest when the American system remains the environment through which the most valuable and trusted forms of intelligence are created, financed, and scaled.

Counterfactual Compression

If not X: If China's WAIC declaration were only rhetoric rather than the early form of an operating system, its compute, cloud, open-model, robotics, and international-governance layers would remain isolated demonstrations.

Then Y would have to be simultaneously true: Huawei's reported compute deployments would not produce a usable domestic stack; Baidu, Alibaba, and Tencent would fail to convert AI usage into recurring commercial revenue; open Chinese models would not attract sovereign deployment; and Chinese robotics supply chains would not accumulate operational data through real deliveries.

But Y contradicts observable constraints: Company-reported revenues and deployments are already material, open-weight Chinese models are circulating internationally, WAICO has an identifiable institutional membership, and Chinese industrial supply chains are already lowering the cost of physical deployment. The remaining uncertainty is not whether the system exists, but whether demand, trust, reciprocity, and profitability are sufficient to sustain it over the next 5–15 years.

Epistemic Humility

Several forms of uncertainty limit any confident judgment. Company figures are reported under different definitions of AI revenue and are not always independently audited at the product level. Vendor claims about deployed clusters, partner counts, robot shipments, and model adoption may measure activity without measuring utilization or profitability. Public procurement does not prove private demand. Model downloads do not prove sustained usage. Conference participation and institutional membership do not establish treaty-level alignment.

The relationship between open weights, genuine openness, and geopolitical control is also unstable. Governments can support open models while restricting data, ownership, market access, or selected exports. The United States and China are still interdependent across research, software, hardware, capital, and customers, even as both construct more separable systems.

This analysis therefore evaluates structural direction rather than predicting a fixed outcome. China's AI system is sufficiently developed to affect American strategy and global market choices, but its final form remains contingent on costs, demand, trust, policy reactions, and enterprise execution.

Alternative outcomes remain possible if constraints shift. This reflects current observable trajectories, not inevitability. Structural balance may change under new technological or policy regimes.

Conclusion: Industrial Capacity Is Real; Market Conversion Remains the Constraint

The public evidence supports a bounded conclusion. Mr. Xi's WAIC declaration rests on real industrial and institutional capacity: domestic compute systems, cloud platforms, open-weight models, physical-AI supply chains, startup formation, training programs, and a new international organization. These elements are sufficient to create a durable Chinese-centered alternative in selected layers of the global AI stack.

The strongest evidence comes from operations rather than speeches. Baidu reports billions of yuan in AI infrastructure revenue and millions of driverless rides. Alibaba reports faster cloud growth, substantial AI-related revenue, and deployment of proprietary processors. Tencent has the cash flow and distribution to finance a prolonged transition. Huawei reports large-scale Atlas deployments. Unitree and AGIBOT have moved humanoid robots beyond isolated prototypes into thousands of delivered units.

The constraint is conversion. Infrastructure revenue is stronger than standalone application revenue. Platform profits can subsidize AI expansion while startups face weak bargaining power and high compute costs. Robot shipments are growing faster than independently verified private demand. International adoption does not automatically produce recurring Chinese revenue, and open weights do not remove concerns about data governance, security, ownership, or reciprocity.

Over the next 5–15 years, the structural question is whether China's production and deployment advantages generate enough recurring demand, household purchasing power, international trust, and commercially sustainable services to reproduce the capital consumed by the system. The Foreign Affairs critique remains relevant because external markets may become more defensive as Chinese capacity expands. Reuters adds the institutional layer: that capacity is increasingly connected to a rival governance network rather than operating only through commercial exports.

The United States faces the inverse constraint. Its system can remain dominant at the frontier while losing portions of the distribution layer if access remains expensive, concentrated, or politically conditional. Export controls can delay Chinese capability, but they can also strengthen demand for domestic substitution and alternative institutions.

The structural judgment is therefore narrower than a prediction of civilizational victory. China has demonstrated the capacity to build a distinct AI production and distribution system. Whether that system becomes self-sustaining depends on its conversion of scale into demand, profitability, reciprocity, and trust.

Scope, Method, and Limitations

This is an analytical, educational, and non-commercial assessment of industrial and institutional feasibility over a 5–15-year horizon. It is not investment advice, a prediction of political outcomes, an endorsement of either Chinese or American technology policy, or a policy recommendation.

Company figures are drawn from public reporting and are used to understand operational scale. Reporting definitions differ across companies, and the figures are neither perfectly comparable nor equivalent to product-level audited AI revenue. Deployment counts, model downloads, institutional memberships, and announced partnerships indicate activity, but do not independently establish utilization, profitability, or durable geopolitical alignment.

Sources

Core English-Language Sources

• Official WAIC materials: full keynote address and conference scale and exhibits
• Foreign Affairs — “China Is Sabotaging the World That Enables Its Rise”
• Reuters — China as the champion of a new global AI order, WAICO, and competing U.S.-China governance frameworks
• Business Insider — China's open-source AI strategy and the U.S. competition
• Baidu — First Quarter 2026 Results
• Alibaba Group — Fiscal 2026 cloud and AI commercialization results
• Tencent — First Quarter 2026 Results
• Huawei — Atlas 950 SuperPoD specifications and deployment data
• Associated Press — China's humanoid production capacity and demand gap
• Microsoft — Fiscal 2026 third-quarter cloud and AI results
• The White House — America's AI Action Plan

Chinese-Language and Additional Media Perspectives

• Yuyuan Tantian / CCTV — The official-policy interpretation of openness, applications, data, and governance
• 36Kr — CookiePi's product design, paid testing, financing, and commercialization
• 36Kr / IT Juzi — Capital concentration behind WAIC exhibitors
• BBC Chinese — WAIC and China's global AI governance message

Reproduction is permitted with attribution to Hi K Robot(https://www.hikrobot.com).