How Should America Respond to China’s AI State Mobilization?

2.1 AI as a strategic asset

This article treats AI as a strategic asset. A technology becomes a strategic asset when it no longer remains a discrete industry but functions as an enabling layer connecting a country’s economic, scientific, security, human-capital, financial, and infrastructural capabilities.

AI competitiveness is not determined by model accuracy or benchmark scores alone. Developing and operating large models requires:

• semiconductors and large-scale computing resources;
• reliable and affordable electricity;
• high-quality data;
• researchers and engineers;
• cloud platforms and data centers;
• demand for deployment in business and the public sector;
• patient investment capital; and
• a commercial ecosystem that turns models into products and services.

Seen in this way, AI strategy is not simply the sum of technology policy and industrial subsidies. Even where the state supports R&D and infrastructure, sustained development becomes difficult if firms cannot create viable revenue models. Conversely, a company may continue investing despite short-term losses when it can draw on cash flow from established businesses, patient capital, and rapidly expanding markets.

2.2 Classification of evidence

The analysis separates its evidence into four categories.

First are official and authoritative documents issued by the Chinese Communist Party, China’s central government, the Ministry of National Defense, the National Development and Reform Commission, and related bodies. These sources are used to identify the Chinese government’s stated objectives and strategic perceptions.

Second are official earnings releases and investor materials from Chinese and global companies. These are used to examine revenue, cash flow, R&D spending, business structure, and management’s account of corporate strategy.

Third are policy and audit materials from the White House, the U.S. Departments of Commerce, Treasury, Defense, and Energy, NIST, and the Government Accountability Office. These sources show how the U.S. government interprets China’s AI and semiconductor development as an economic and security issue and how it is designing policies involving compute, capital, infrastructure, alliances, and cybersecurity. Executive-branch policy documents are treated as evidence of policy intent, not as independent proof of policy performance.

Fourth are reports from international news organizations such as the Financial Times and Reuters. Claims about local-government support that cannot be directly verified in official documents are used cautiously, with their source and verification status stated explicitly.

6.1 The integration of AI computing and energy

The fourth pillar is the integration of computing-power facilities with energy infrastructure. Training and serving large models requires not only semiconductors but also reliable electricity and cooling. As model use grows, recurring inference and data-center operating expenses may exert greater pressure on profitability than the one-time cost of a training run.

An official policy explanation associated with China’s National Development and Reform Commission identifies closer coordination between AI and energy, reliable power for computing facilities, consumption of green electricity, market-based power trading, multi-year green-power contracts, and integrated planning of computing and electricity as major priorities.^[4]

These policies seek four effects. First, they connect computing facilities to regions with large renewable-energy resources and improve supply reliability. Second, they encourage data centers to participate in green-power markets and long-term contracts, reducing exposure to electricity-price volatility. Third, they enlist computing facilities in demand response and ancillary services, potentially improving grid efficiency. Fourth, they coordinate the location of computing and energy facilities to reduce transmission and operating costs.

The official material does not establish a nationwide policy that cuts electricity prices for AI data centers by 50 percent or supplies power free of charge. What can be verified officially is a policy direction emphasizing compute-power coordination, green electricity, participation in power markets, and stable supply.

6.2 Reports of electricity-cost support by selected local governments

In November 2025, the Financial Times reported that local governments in regions including Gansu, Guizhou, and Inner Mongolia had offered support capable of reducing electricity costs by as much as 50 percent for certain large data centers using domestically produced AI chips.^[5]

The report suggests that local data-center attraction policies may be combined with efforts to cultivate domestic semiconductors. If the use of Chinese chips is a condition for cost support, electricity subsidies can operate not merely as regional investment incentives but also as a tool for expanding the domestic chip ecosystem.

The “up to 50 percent” figure, however, is not a nationwide standard announced by the central government. The Financial Times relied on anonymous sources and stated that the relevant local governments and the NDRC did not respond to requests for comment. Reuters repeated the report while explicitly noting that it could not independently verify it.^[6]

The claim must therefore be read with five qualifications:

• it concerns reports about selected local governments;
• it applies to particular large data centers;
• the use of domestic chips was reportedly a condition;
• it has not been verified as a nationwide central-government policy; and
• the underlying local implementation documents and disbursement records have not been directly examined.

Even with these limitations, the report demonstrates that electricity is becoming an important instrument of AI competition in China. Because power and cooling account for a substantial portion of data-center operating costs, access to affordable and dependable energy can strengthen the price competitiveness of model developers and cloud providers.

7.1 The cost structure of generative AI

The cost of a generative AI business does not end after a model has been trained once. Firms must repeatedly finance:

• pretraining on new models and data;
• post-training and alignment;
• inference services;
• cloud and data-center capacity;
• data acquisition, cleaning, and labeling;
• researchers and engineers;
• safety evaluation and regulatory compliance;
• marketing and consumer acquisition; and
• enterprise implementation and technical support.

More users can produce more revenue, but they also increase inference volume and expense. When companies use free access or very low API prices to attract users, growth itself can enlarge cash outflows.

Chinese model developers have relied on low API prices and free or inexpensive services to win developers and enterprise customers. DeepSeek’s official pricing, for example, lists very low per-token rates.^[7] Prices can change, but the policy illustrates the importance of cost competition and rapid diffusion in the Chinese market.

Low prices encourage adoption and enlarge the developer ecosystem. At the same time, they make it even harder to recover R&D spending through model-usage fees alone. An explanation of corporate survival must therefore examine sources of capital and revenue beyond direct API sales.

7.2 Cross-subsidization by large platform companies

Large platforms such as Alibaba, Tencent, and Baidu have a financial structure fundamentally different from that of a pure model start-up. They generate substantial cash flow from e-commerce, advertising, gaming, search, fintech, and enterprise services.

Even if their AI divisions do not become independently profitable in the short run, these companies can reinvest cash from established businesses in R&D, semiconductors, data centers, and cloud infrastructure. This is a form of cross-subsidization.

Alibaba’s fiscal-year 2025 results showed that AI demand helped raise quarterly cloud revenue growth to 18 percent and that revenue from AI-related products had recorded triple-digit growth for seven consecutive quarters. At the same time, rising spending on cloud infrastructure reduced quarterly free cash flow by 76 percent year on year and annual free cash flow by 53 percent.^[8]

These figures reveal both sides of the investment cycle. AI demand is producing real cloud revenue, but the infrastructure required to serve it is extraordinarily expensive. Alibaba can absorb this investment through cash generated by e-commerce and platform operations, changes in its asset structure, and improvements in operating efficiency.

Tencent stated directly in its 2025 earnings materials that cash-generative core businesses financed the recruitment of top AI talent and greater investment in AI infrastructure. It also used AI to improve advertising targeting, engagement in games, content creation, and cloud services.^[9]

The return on Tencent’s AI investment therefore does not appear only as API revenue. Better recommendation systems can increase advertising revenue and efficiency; AI can reduce game-development costs or deepen user engagement. If AI raises revenue and margins in established businesses, the group has a reason to invest even when the model business itself remains unprofitable.

Baidu is attempting to shift toward AI cloud and AI-based applications as its legacy search and advertising operations weaken. The company reported that 2025 revenue from AI cloud infrastructure reached approximately RMB 19.8 billion, an increase of 34 percent from the previous year. Subscription revenue from AI-accelerator infrastructure and AI-powered marketing services also grew.^[10]

For large platforms, AI is both a new line of business and an enabling technology for defending and reconstructing the old ones. Their investment decisions reflect expected effects across e-commerce, advertising, gaming, search, cloud computing, and enterprise services rather than the near-term income statement of a single model division.

7.3 Growth capital and tolerated losses at independent model companies

Independent model developers lack the cash flow of established platforms, making external capital, public markets, and rapid revenue growth more important.

MiniMax’s 2025 results illustrate the position of such companies. Revenue rose 158.9 percent to $79 million, more than 70 percent of it generated overseas. Gross profit reached $20.1 million, and gross margin improved to 25.4 percent. Yet adjusted net loss was approximately $250.9 million, while R&D spending increased to roughly $252.8 million.^[11]

The company thus achieved rapid sales growth and improving gross profitability, but not enough to fund R&D and market expansion. Firms in this position tolerate current losses through investors’ expectations about future market position and technological value, an eventual public offering, and subsequent fundraising.

They can attract capital because investors see foundation models not simply as chatbots but as prospective platforms for industrial software, content, robotics, cloud computing, and enterprise workflows. Technical capability, users, the developer ecosystem, enterprise customers, and access to overseas markets may matter more than current profit.

This structure is nevertheless fragile. If capital markets tighten or revenue growth slows, sustaining R&D becomes difficult. Long-term survival requires improvement not merely in the number of free users but in paid subscriptions, enterprise contracts, international revenue, recurring revenue, and unit economics.

7.4 Diversification of revenue models

Chinese AI firms pursue several monetization paths:

1. Usage-based APIs. Businesses and developers pay according to consumption. This supports market expansion, but intense price competition can compress margins.
2. Enterprise deployment. Models are installed in a company’s internal or dedicated cloud environment and adapted to industry data and workflows. Providers can earn implementation fees, usage charges, maintenance revenue, and fees for additional development.
3. Consumer subscriptions and services. Premium chatbots, video and image generation, productivity tools, educational applications, and content services produce subscription or usage revenue.
4. Cloud and infrastructure revenue. Rather than monetizing only the model, a provider earns revenue from GPUs, AI accelerators, storage, data processing, and cloud consumption. This path particularly benefits large platforms and cloud firms.
5. Indirect returns inside existing businesses. AI can improve advertising, e-commerce, search, gaming, finance, customer service, and logistics by raising sales, time spent, conversion rates, or operating efficiency.
6. Overseas markets. Firms may escape intense domestic price competition by reaching consumers and enterprises with greater willingness to pay. MiniMax’s high share of overseas sales demonstrates the importance of this route.

Large platforms enjoy advantages in cloud revenue and indirect returns from established businesses. Independent model developers depend more heavily on APIs, subscriptions, enterprise implementation, and international markets.

7.5 Lowering costs and creating early markets through local policy

Chinese local-government support is not limited to cash grants. Compute vouchers, model vouchers, access to data and corpora, R&D assistance, industrial parks, demonstration projects, and public demand can reduce early costs and market-entry risk.

China’s published compilation of 2025 regional measures reported that Shanghai offered compute, model, and corpus vouchers, while Shandong subsidized eligible projects through measures connected to computing power, models, data, and application scenarios.^[12]

Hangzhou’s 2025 AI policy established an annual pool of RMB 250 million in computing-power vouchers and provided eligible firms with support for portions of computing and model-service expenses. R&D, specialized models, open-source work, application demonstrations, corporate growth, listings, and industrial finance were also eligible.^[13]

Such policies can:

• lower model-training and inference costs;
• reduce the expense of data and model access;
• absorb part of the risk of R&D investment;
• provide opportunities for validation in industrial settings;
• create initial customers and reference revenue; and
• improve a firm’s credibility with lenders and investors.

Government support need not cover a company’s losses forever. Its economically defensible role is to reduce high fixed costs and weak demand during market formation. During the support period, firms must improve their technologies and products and expand private-sector revenue.

7.6 Open source and cost efficiency

The open-source ecosystem and engineering for efficiency are additional reasons Chinese firms can sustain development.

Alibaba has opened models in the Qwen family to encourage adoption by developers and businesses. Offering models free of charge or under open licenses can sacrifice some direct fees. But a larger developer ecosystem can increase demand for Alibaba Cloud, enterprise support, and dedicated services.

An open-source strategy can:

• attract developers and researchers;
• improve models and identify errors;
• diffuse a wide variety of applications;
• deepen enterprise customers’ technical dependence;
• generate cloud and enterprise-support revenue; and
• expand international technological influence.

Compute efficiency is equally important. Developers use mixture-of-experts architectures, quantization, caching, smaller models, and inference optimization to deliver the same or similar performance with fewer resources. This is not merely a research achievement; it is a business strategy for lowering the cost per service and improving long-term profitability.

7.7 An integrated view of corporate survival

The reasons Chinese AI companies continue developing technology despite weak profitability can be summarized as follows.

First, large platforms cross-subsidize AI with cash flow from established businesses. Second, AI raises revenue or efficiency in cloud computing, advertising, gaming, search, e-commerce, and enterprise services. Third, independent model firms draw on venture capital, strategic investors, and capital markets. Fourth, firms diversify revenue across APIs, enterprise implementation, consumer subscriptions, cloud services, and international markets. Fifth, local governments reduce computing, model, and data costs, absorb part of the R&D risk, and provide early demand. Sixth, open-source ecosystems and efficiency engineering pursue diffusion and better unit economics at the same time. Seventh, the strategic expectation that AI will shape future industries and national security encourages investors and governments to tolerate near-term losses.

It is therefore misleading to explain the survival of Chinese AI firms through subsidies alone. Patient investment is sustained by an ecosystem that combines national strategy, platform cash flow, private and public capital, local policy, cloud infrastructure, industrial demand, overseas markets, and technical efficiency.

8.1 Not an absence of markets, but a party-state that places strategic priorities above them

It is inaccurate to describe China simply as a country without a market economy. Private firms, price competition, venture investment, stock markets, consumer choice, and overseas revenue all operate extensively in its AI industry. Beijing calls the system a “socialist market economy” and stresses the combination of market mechanisms with the role of government.

In strategic industries, however, the state is far more direct than in conventional market economies. China’s leadership officially emphasizes a “new whole-of-nation system” for mobilizing nationwide resources to achieve breakthroughs in core science and technology. In April 2025, Xi Jinping called for using the advantages of this system in AI as well.^[14] The central government sets long-term goals and concentrates the resources of central and local agencies, national research institutes, universities, state-owned enterprises, and leading private firms on strategic tasks.

The persistence of China’s AI sector should therefore not be reduced to a binary choice between “the market” and “subsidies.” The system is better understood as a hybrid: market competition disciplines firms and technologies, while the party-state selectively organizes strategic direction, capital, land, electricity, data, research institutions, and early demand. Markets create rivalry, while the state socializes part of the risk of failure and lengthens the investment horizon in priority fields.

8.2 Five policy instruments enabled by the Chinese system

8.3 Difficult to compare precisely, but exceptionally broad in scope

Democracies can also use industrial policy and large subsidies; American support for semiconductor research and manufacturing is a leading example. The claim that only authoritarian governments can subsidize strategic industries is therefore false. The key difference is not whether subsidies exist, but whether budgets, beneficiaries, and performance standards are transparent; whether legislatures, auditors, courts, and competition authorities can review them; and whether failed programs can be terminated and funds recovered.

Differences remain in scale, breadth of instruments, and mechanisms of control. CSIS combined direct subsidies, tax incentives, below-market credit, and government investment funds to estimate China’s 2019 industrial-policy spending at approximately 1.73 percent of GDP—higher than in the other major economies it compared.^[16] Using listed-company financial data and land records, IMF researchers estimated a broad fiscal-equivalent cost of about 4 percent of GDP annually, including direct subsidies, tax preferences, preferential credit, and below-market land.^[17]

The two estimates cover different programs and use different methods. They should neither be directly compared nor treated as official statistics. They do, however, agree that Chinese support extends well beyond budgetary subsidies to credit, land, taxes, and state funds. During the WTO’s 2024 Trade Policy Review, members also questioned the transparency and completeness of China’s subsidy notifications; some argued that the programs distorted global markets and encouraged excess capacity.^[18]

8.4 Why democratic systems cannot—and should not—copy the model wholesale

Large-scale industrial support is not impossible in a democratic market economy, but it requires more justification and control. Legislatures must authorize budgets. Support for selected firms may face scrutiny under competition law, subsidy rules, and trade law. Land use and data-center siting are affected by environmental review, federalism, property rights, and community acceptance. Auditors, the press, courts, and elections impose accountability for policy failure.

These constraints are not merely inefficiencies. They protect public finances, competition, property, and civil rights. China may have an advantage in speed and scale because it has fewer veto points and can transmit central directives quickly through local governments and the state-owned sector. Yet weak disclosure and external oversight also allow mistaken investments to continue on a larger scale.

The United States should not respond by copying authoritarian control or promising unlimited subsidies. It can learn from the consistency of long-term strategy, the connection of computing, electricity, data, finance, and early demand, and the speed of policy experimentation. But budget transparency, competitive neutrality, judicial review, independent audit, community consent, and clear termination conditions are not defects in the American innovation system. They preserve long-term efficiency and political legitimacy.

8.5 The costs and sustainability of authoritarian mobilization

IMF researchers acknowledge that Chinese industrial policy may have supported development in selected industries and technologies, but estimate that subsidies encouraged overproduction and that policy-induced factor misallocation reduced total factor productivity by approximately 1.2 percent.^[17] The IMF’s 2025 Article IV consultation on China likewise concluded that state-led, debt-dependent investment and difficult-to-justify industrial support contributed to weaker productivity, financial vulnerabilities, and excess supply in some tradable sectors.^[19]

This does not mean that Chinese mobilization is always inefficient. In electric vehicles, batteries, solar equipment, and telecommunications—industries in which scale and learning effects matter—state support can accelerate technological accumulation and supply-chain formation. In AI, patient investment in computing, talent, and public-sector demonstrations can build foundations that markets may underprovide on their own.

Evaluating the system only through its successes, however, creates survivorship bias. The calculation must also include subsidized firms that failed, underused industrial parks, duplicated local projects, debt, and environmental costs. Authoritarian mobilization is not a formula for rapid growth without cost. It is an institutional arrangement that enables society to bear the risks and losses of industrial transformation more broadly while making long-term bets in strategic fields.

9.1 The relevant unit of analysis is an industrial system, not an individual model

The most important analytical error for the United States to avoid is equating a single benchmark result—or the profitability of one company—with China’s aggregate capability. Individual Chinese AI firms can fail, and particular products can lose money. Yet national strategy, platform cash flow, local cost support, state finance and infrastructure, a vast domestic market, and manufacturing demand distribute risk across a broader system.

America’s competitor is not a single Chinese chatbot. It is a system linking models, chips, cloud platforms, data centers, electricity, telecommunications equipment, manufacturing, public procurement, and military technology. An American response cannot therefore be completed through model regulation or semiconductor export controls alone.

In 2026, the GAO proposed assessing U.S. AI competitiveness across private and public investment, talent attraction, the regulatory environment, and computing infrastructure.^[30] The framework reflects a central reality: the relevant competitive capability is not an algorithm in isolation, but a country’s capacity to develop, deploy, and diffuse technology.

9.2 The economic challenge: low-cost diffusion and ecosystem capture

Price competition and open-source releases can weaken near-term profitability while rapidly attracting global developers and enterprise users. Releasing weights or code and charging very low prices may reduce direct model revenue, but it can increase cloud consumption, enterprise implementation, adoption of domestic chips, and international technical influence.

The threat to American interests does not arise only if a Chinese model leads every benchmark. A model that is “good enough” and substantially cheaper can spread through public services, telecommunications, manufacturing, education, and finance in emerging economies. When low prices and localization are combined with state-backed financing, superior American technology may not automatically translate into market power.

The United States must therefore complement frontier-model leadership with deployable packages of secure hardware, cloud services, models, cybersecurity, industrial applications, and financing. Executive Order 14320 of July 2025 directed the creation of a program to export a full U.S. AI technology stack—including hardware, models, software, applications, and standards.^[28] The initiative recognizes that competition concerns financing, deployment cost, maintenance, and standards as much as raw model performance.

9.3 The security challenge: conversion of civilian AI into military and surveillance power

AI is intrinsically dual use. Computer vision, language models, autonomous systems, cloud infrastructure, data analytics, and advanced chips can increase civilian productivity, but they can also support intelligence analysis, command and control, surveillance and reconnaissance, cyber operations, and unmanned systems.

The U.S. Department of Defense’s 2025 report on Chinese military developments assessed that China was accelerating work on advanced military technologies, including military AI.^[21] Because this is an official U.S. threat assessment, it cannot prove that the entire Chinese AI industry serves military purposes. It does, however, explain the American security rationale for refusing to treat transfers of frontier computing resources as ordinary commerce when the PLA regards AI as central to future capability.

Controls should be based on technical performance, end users, end uses, and concrete links to military, intelligence, or surveillance organs—not on broad suspicion of Chinese nationality or all Chinese firms. Indiscriminate exclusion can weaken the talent base of American universities and companies while producing discrimination and political backlash.

9.4 The normative challenge: diffusion of authoritarian AI standards

AI competition also concerns surveillance, privacy, freedom of expression, public procurement, and international standards. When Chinese digital infrastructure and AI systems are paired with demand from authoritarian governments for monitoring and censorship, technical standards can spread together with modes of political control.

America’s 2025 AI Action Plan called for governance in international standards bodies consistent with innovation and American values, and for countering authoritarian influence.^[20] But declarations about values will not generate trust by themselves. American AI products must provide concrete protections for privacy, cybersecurity, local data sovereignty, explainability, interoperability, and the ability to switch providers.

America’s advantage is not that its government can designate one national model or champion company. It lies in competing firms, leading universities, venture capital, immigrant talent, and an alliance network. If Washington damages this open structure in the name of competing with China, it will imitate China’s weaknesses while sacrificing its own strengths.

10.1 Strategic objective: limit military diversion and authoritarian dependence—not China’s development itself

The purpose of U.S. strategy should not be to stop all Chinese AI research or commercial services. Such a goal is neither feasible nor internationally legitimate. The objective should be to restrict the ability of China’s military, intelligence services, and high-risk surveillance systems to accelerate their capabilities with American and allied technology, while preventing world markets from becoming dependent on authoritarian AI infrastructure.

Controls should clearly define advanced-compute thresholds, semiconductor-manufacturing equipment and critical subsystems, military and intelligence end users, specified high-risk AI applications, and forms of capital or managerial support. The broader and more ambiguous the rules become, the more evasion, allied resistance, and dilution of enforcement resources they may produce.

10.2 Advanced-compute export controls: from performance thresholds to end-user and location verification

The United States has used the Export Administration Regulations and entity listings to limit China’s access to advanced semiconductors and supercomputing capacity. In 2025, the Department of Commerce moved to rescind and stop enforcing the worldwide tiered AI Diffusion Rule while issuing separate guidance concerning Chinese advanced chips, the use of U.S. AI chips for training or inference by Chinese models, and diversion risk.^[22]

In May 2026, BIS guidance reiterated that exports of specified advanced-computing items to entities headquartered in Country Group D:5 or Macau—or whose ultimate parent is headquartered there—may require a license even where the receiving entity is located in a third country.^[23] This development illustrates a shift from controls based only on shipping destination toward scrutiny of ownership and ultimate end users.

Future policy should strengthen three capabilities. First, it should use location verification and remote attestation for advanced chips and servers where technically and legally appropriate. Second, it should verify the final use and customers of data centers in high-risk transshipment jurisdictions. Third, it should align rules with allied controls on equipment and components so that American companies do not lose sales while the restricted supply is simply replaced. The AI Action Plan likewise called for location verification, combined intelligence and Commerce enforcement, and complementary allied measures.^[20]

10.3 Domestic semiconductor capacity: the productive base that makes controls credible

Export controls work only while the United States and its allies preserve technological leadership and productive capacity. If American R&D and manufacturing erode, the strategic advantage that controls are intended to protect disappears.

The CHIPS and Science Act allocated $50 billion to Commerce Department semiconductor programs. NIST describes $11 billion for the domestic semiconductor R&D ecosystem and $39 billion in incentives for facilities and equipment in the United States.^[25]

Policy cannot end with a factory count. It must encompass advanced packaging, materials and equipment, power semiconductors, electronic-design automation, manufacturing workers, and the path from research to high-volume production. Assistance should include milestones for output, R&D, and workforce development, while preventing the government from permanently absorbing corporate losses. Staged payments, private-capital matching, and clawback provisions can improve discipline.

10.4 Preventing the strategic transfer of American capital and managerial know-how

An overseas investment in an advanced-technology firm delivers more than money. Investors can provide managerial assistance, talent and customer networks, reputational validation, and access to subsequent financing. Treasury’s Outbound Investment Security Program designates China, Hong Kong, and Macau as countries of concern and prohibits or requires notification of selected transactions in semiconductors and microelectronics, quantum information technologies, and AI. The program took effect on January 2, 2025.^[24]

Its scope should not be expanded so broadly that ordinary commercial investment and legitimate research collaboration are obstructed. Rules should clearly identify technologies, firms, and transaction types directly linked to military, intelligence, or surveillance capabilities and apply transparent review standards and national-interest exceptions. Ambiguous rules can induce excessive risk aversion among U.S. investors while affected companies substitute capital from third countries.

10.5 Data centers, the power grid, and permitting: removing America’s physical bottlenecks

The competition does not end when chips are secured. A shortage of data-center sites, transmission capacity, generation, transformers, cooling, specialized labor, or community acceptance can delay American model development and deployment.

An LBNL analysis publicized by the Department of Energy estimated that U.S. data centers consumed approximately 4.4 percent of national electricity in 2023 and could consume about 6.7 to 12 percent by 2028.^[26] This is a scenario range based on equipment shipments, operations, and efficiency—not a single deterministic forecast. Failure to address power constraints can turn data-center investment into higher local rates, grid congestion, and community opposition.

Executive Order 14318 of July 2025 directed federal agencies to consider loans, guarantees, grants, tax incentives, and offtake agreements for qualifying data centers and related electricity, semiconductor, and network projects, while accelerating federal permitting.^[27] Speed is necessary, but environmental review and community rights should not be dismissed as obstacles. Large users should bear the generation and transmission costs attributable to their loads, with support tied to demand response, on-site generation, storage, efficiency standards, and disclosure of water use.

10.6 Allies and global markets: build both a control coalition and a supply coalition

The semiconductor supply chain cannot be completed by the United States alone. Design, manufacturing equipment, materials, fabrication, and packaging are distributed across allied economies. If Washington acts alone while allied suppliers fill the gap, controls become ineffective. Yet unilateral extraterritorial rules can also produce allied political resistance and accelerate efforts to develop alternatives to American technology.

An alliance strategy must supply credible alternatives as well as align restrictions. Executive Order 14320’s export initiative seeks to bundle hardware, cloud services, models, cybersecurity, applications, and standards into a full-stack offer for allies and partners.^[28] The United States should add finance, local data protection, workforce development, interoperability, and exit rights so that American systems become a more trustworthy and economical option than Chinese packages.

10.7 Cybersecurity and AI defense of critical infrastructure

The AI competition includes the protection of the power grid, telecommunications, finance, hospitals, government systems, and data centers. Adversaries can use AI for vulnerability discovery, phishing, malware, and influence operations; defenders can use it to detect anomalies and remediate software vulnerabilities.

A June 2026 executive action prioritized AI-enabled cyber defense for national-security and federal information systems and directed the creation of information-sharing arrangements with critical-infrastructure operators.^[31] The approach may give government access to frontier capabilities developed in the private sector, but it must also manage classified information, model supply chains, adversarial inputs, and dependence on individual vendors.

10.8 Talent and research openness: protect America’s largest asymmetric advantage

The United States has benefited from an open research ecosystem that attracts researchers, entrepreneurs, and students from around the world. If competition with China produces nationality-based suspicion and exclusion, American universities and firms will weaken their own human-capital base.

Research security should focus on undisclosed funding relationships, links to military or intelligence bodies, contractual confidentiality, export-controlled technology, and actual conduct—not ethnicity or national origin. Fundamental research and open science should remain as open as possible, while access controls and audit trails are strengthened in clearly sensitive areas such as advanced chip design, manufacturing processes, military model evaluation, and classified data.

10.9 Open-source strategy: turn openness and diffusion into American power

If the United States responds to the spread of open Chinese models by broadly restricting publication, it could weaken its own developer ecosystem and research transparency. Nor does every open model create the same security risk. Open small and medium-sized models, research tools, and safety-evaluation infrastructure can strengthen both innovation and security.

Policy should distinguish risk according to capability, dangerous functionality, training compute, and ease of modification. Frontier models with high-risk capabilities warrant more stringent cyber and biological misuse assessments and security measures, but ordinary research and commercial use should not be placed under a general preapproval regime. The June 2026 White House action similarly promoted voluntary collaboration on advanced AI cyber capabilities while explicitly declining to create mandatory federal preapproval for model development or release.^[31]

10.10 Democratic discipline and exit conditions for industrial policy

If the United States responds to Chinese subsidies by reimbursing every AI and semiconductor company’s costs, it may weaken competition and innovation. American industrial policy should focus on identifiable market failures and national-security externalities: enormous fixed costs, fragile supply chains, basic research, and infrastructure with broad spillovers.

Programs need measurable targets for capacity, technical performance, private investment, supply-chain diversification, workforce development, energy efficiency, and delivery schedules. Payments should stop or be clawed back when milestones are missed. Performance data should be disclosed while trade secrets and security information remain protected. Independent GAO audits, congressional oversight, and competition review should operate alongside financial support.

11.1 Controls may accelerate Chinese self-reliance and cost innovation

Restrictions on advanced semiconductors can raise China’s training and inference costs, but they also increase the strategic value of domestic chips, efficient model architectures, software optimization, and open-source development. Washington should not assume that controls can permanently halt Chinese progress. Their purpose is to buy time in which the United States and its allies expand technological and manufacturing advantages.

11.2 U.S. corporate revenue and the R&D base may weaken

Lower sales in China may reduce the research budgets of American semiconductor and equipment companies. Short-term revenue, however, cannot always take priority over military-diversion risk. Costs and benefits should be reassessed by technology generation so that controls do not expand unnecessarily to general-purpose legacy products.

11.3 Diversion and inadequate enforcement capacity can make rules ineffective

As controls grow more complex, transshipment, shell companies, third-country data centers, and rented cloud capacity may increase. In 2025, the GAO identified a need for better workforce planning and interagency information sharing in BIS licensing and enforcement.^[29] Investigators, data analysis, overseas enforcement, and allied customs and corporate-information cooperation matter as much as issuing new rules.

11.4 A coalition weakens if allies are treated only as targets of pressure

Cooperation will be difficult to sustain if U.S. rules impose costs on allied companies without supplying market access, finance, or technological alternatives. Allies should be treated as co-designers of rules and co-suppliers of trusted technology, not simply as jurisdictions expected to accept American restrictions.

11.5 America can damage its own openness while responding to authoritarianism

Political loyalty tests, nationality-based exclusion, excessive secrecy, and the designation of favored national champions would imitate features of the system America is trying to counter. The response must operate within constitutional rights, due process, academic freedom, competition, and federalism. Improving slow procedures is not the same as removing checks and balances.