China's AI Policies: Overview and US Comparisons

Introduction

As of 2026, the United States and China are the two most influential centers of global AI development, with comparable ambitions for scale and international reach. However, while their industrial strategies to that end share common foundations, they also rely on different drivers and measures of success that reflect the specificities of their political and industrial ecosystems.

In the United States, the development and deployment of AI over the last 5 years have followed a path primarily set by a few dominant commercial actors. From a regulatory perspective, prevailing narratives treat the technology as unprecedented and best understood by its largest developers. The federal government facilitates development through commercial agreements, grants, and cross-border data diplomacy, while using mechanisms such as regulatory exemptions to streamline providers’ operating conditions. Governance of these technologies emphasize voluntary commitments and results over process and relies on ex post litigation to address harms. Yet the government also retains a more direct lever of influence as both provider and arbiter of commercial contracts, particularly those with perceived implications for national security and international relations. Still, industrial strategy reflects an interdependence between the state and private actors rather than a governmanet-directed approach.

In China, conversely, AI extends a decades-old project of government-led digital industrial policy: the state directs infrastructure and research investment, encourages cooperation among private firms, adapts existing legal frameworks ex ante, and channels investment through state-owned enterprises and other public instruments. Documents such as the latest Five-Year Plan (15th, 2026–2030) have operationalized this approach, organizing strategy around increasing domestic capacity for the infrastructure and resources underlying AI development and fostering its integration across economic domains within existing regulatory frameworks; expanding data access domestically and across borders stand out as central levers.

This report surveys Chinese regulatory and organizational developments through May 2026, with comparisons to the US federal government following in each section; states are considered primarily through the lens of their interactions with federal authority. Part I introduces the industrial context, main policy documents, and legal concepts. Part II examines how the state and markets build the supply side of generative AI: compute and energy infrastructure, access to training data, and open-source development on domestic hardware. Part III turns to deployment governance, reviewing updates to digital competition laws for AI and requirements on model development and disclosure.

I. Background

I.A. Market Structures as AI Policy Context

China and the US have been pursuing distinct industrial policies over the last decade, particularly in the ICT and digital sphere. [1,2] Consequently, their industries have built momentum in different directions, and have distinct tools available to them with which to build the next stage of their development. In practice, this means that a comparative analysis of their industrial policies needs to account for these specificities: ignoring them and judging one country’s success by the other’s metrics leads to incongruous conclusions, such as US companies primarily attributing the performance of Chinese labs’ models to significant “distillation” even as their own accounting reveals very modest access to the supposedly copied models (see DeepSeek numbers in [3]).

One of the most conspicuous differences between the two ecosystems is the market structure underlying AI development and deployment, which stands out as a major determinant of industrial policy. Allan & Nahm [4] formalize one version of this relationship for green industry where they consistently observe that high market concentration and technological uncertainty correlate with more industry-led strategy and open-ended support for commercial actors rather than targeted, state-directed intervention. These factors map directly onto generative AI: its initial scaling in the United States was driven by dominant commercial actors in the digital economy, [5] and uncertainty around the technology was heightened both by the speed of that scaling and by strategic opacity that sustains concentration and limits accountability. [6]

Market concentration shapes the impact of policy tools differently across the financial, narrative, and scientific dimensions of technology development. For interventions that are structured primarily as fiscal support, Klinge et al. [7] show dominant actors may leverage increased cash reserves toward technological trajectories that favor rent extraction. Grill [8] shows how a limited set of incumbents with aligned interests have reinforced their central position by shaping “capabilities” discourse to steer commercial and regulatory attention. Mayer [9] further argues that under conditions of concentrated computational capacity and platform dominance, open resources flow disproportionately to actors best placed to exploit them at scale. Market structure suffuses each of these channels. Where participation is incumbent-dominated, fiscal, discursive, and public innovation tools more easily yield short-term returns that reinforce leaders. Where participation is more distributed similar tools more often support longer-horizon benefits and create greater scope for directing development.

I.A.1 US Ecosystem: Networked Concentrated Actors pushing AI

This economic concentration among a few private actors is readily apparent in the United States along several dimensions. The first such dimension is their weight on equity markets, which we review in more detail in Appendix A. In the US, the ten largest publicly traded companies by market capitalization as of September 10th, 2026, all had a major stake in large-scale AI, and accounted for over 35 percent of total listed equity across all economic sectors. A handful more AI-bound firms in the extended cohort of Appendix A, together with reported secondary marks for OpenAI, Anthropic, and Databricks, would put this share near 41%. A second dimension is cloud infrastructure, reviewed in Appendix B. In Q2 2025, Amazon, Microsoft, and Google together accounted for nearly 82% of cloud infrastructure revenue collected by US providers; a concentration that reflects the broader industrialisation of AI through cloud infrastructure dependence [10]. Similar dynamics involving the very same actors can be witnessed at the level of search engines, operating systems, and phones, extending vertical integration to these actors’ control of the data flows that power large-scale AI systems and the marketplaces for AI-powered applications [11].

These dominant positions are held together by cross-investment, networks of influence, and convergent strategies. Even as they compete for cloud market share, these firms align their financial interests through cross-investment in large model developers (OpenAI, Anthropic) and downstream start-ups, documented in detail in recent FTC review of cloud–AI partnerships [12]. Those ties often bind developers to hyperscaler platforms through cloud credits and compute-for-equity arrangements rather than cash investment, reinforcing a market structure that cements incumbents’ positions [13,14]. Beyond investment structure, these ties translate into shared influence sustained by networks across industry, academia, and government [15,16], as exemplified by their reciprocal defense of shared interests in regulatory settings [17]. Strategies for demand creation carry those alignments further: individual partnerships with management consultants end up normalizing a common commercial model of AI adoption that benefits the group [18,19], and joint funding by supposedly rival model providers of profession-scale training primes incoming workforce to depend on their commercial tools [20]. Together, these patterns of convergence weigh on industrial policy more heavily than market concentration alone, because the leading firms hold mutually beneficial priorities.

Concentration at the commercial layer does mask a research and open-source development environment that remains considerably more distributed. It includes thousands of projects across open-source model tooling, datasets, and deployment software maintained by hundreds of university labs, nonprofits, start-ups, and developers outside the largest commercial actors [21]. Large platform firms draw on this layer in their own research pipelines, leveraging and contributing to widely used stacks such as PyTorch and Transformers [22], and integrate externally originated methods into commercial development [23]. Open weights and transparent releases sustain the research paths, independent evaluation, and third-party scrutiny of social impacts that closed APIs only partially support [24–26], but industrial policy does not resource that function. It reaches that layer only through increasingly sparse research funding and the same firms’ releases [27], with no public provision of shared compute or post-training capacity on the scale of private actors [28]. Commercial dynamics reinforce the imbalance: production revenue and developer attention concentrate in closed proprietary APIs rather than in the open research economy [29].

The strongest federal instruments therefore attach most readily to those concentrated commercial actors, giving the government distinctive levers to exert influence over the technology: either to accelerate development along its current lines or to limit its deployment. In FY 2022 alone, major federal agencies obligated roughly $7 billion on cloud computing contracts, with additional spending in the billions going to just evaluating new AI models recently, channeling recurring spend to the same few hyperscalers that dominate the stack [30,31]. Grants and public-private partnerships likewise land most easily on companies already controlling access to the platforms large models depend on [12]. Because dominant firms already accumulate user data at scale, the government can readily give them access to massive troves of AI training data by simply clearing regulatory barriers [32]. The same concentration creates equally sharp chokepoints for restriction: because access to leading models already routes through a few providers [33], federal procurement rules, contract disputes, or national-security designations affecting leading large model developers Anthropic and OpenAI in 2026 have shown how pressure on one or two providers can cascade across the entire ecosystem [34–36]. AI industrial strategy nonetheless remains interdependent with those same concentrated actors [37], with federal influence operating largely as a choice between amplifying or interrupting their reach.

I.A.2 Chinese Ecosystem: Full-Ecosystem Managed Interventions

China shares some of the distinguishing features of this AI development ecosystem: large commercial actors such as Alibaba play a major role in AI cloud and platform services [38]; private capital is buying into large valuations for model-focused start-ups [39]; and vertical integration dynamics appear in both pre-generative AI digital platform structures and recent acquisitions [40]. However, both the intensity of market concentration in the country’s technology ecosystem and the primacy of the technology sector in its overall industry remain significantly lower than in the United States.

We review the equity and cloud shares in Appendices A and B. On the equity side, the main similarity between the US and China lies in their respective top spots, with two AI infrastructure companies, NVIDIA and CXMT, leading each country’s ranking. Technology companies with major AI stakes account for just three of the top 10 Chinese listed companies however (40% of the top-10 total weight, compared to all 10 on the US side), with the rest split between finance, consumer goods, and energy. The larger cohort of companies with a major AI stake adds up to 20.7% of total listed equity in mainland China (compared to 37.8% in the US), and accounting for the largest private actors would bring the proportion up to 25.8% (in large part due to ByteDance, compared to 40.8% in the US). Another notable difference is the ownership stake of the Chinese government in CXMT, with the Hefei city alone holding 36.8% and “roughly half” state ownership overall. On the cloud market side, a report on the Chinese domestic cloud market for Q2 2025 shows that the top three providers account for 52% of revenue shares (compared to 82% in the US), with State Owned Enterprises (SOEs) China Telecom and China Mobile making up 21.7% together. Alibaba’s role as the top provider of cloud services bears the strongest similarities with US-style stack integrators, but even within the specific market of AI, ByteDance, Huawei, Tencent, and Baidu also each hold meaningful cloud positions.

That gap can be partly read through two related features of the Chinese economy. First, a form of “state capitalism” in which the central government holds a minority but influential share of financial institutions and industry, as evidenced by its role in companies like CXMT, China Telecom, and China Mobile. Second, an approach to digital competition that has combined a “tech crackdown” to limit the power of the largest tech companies in the early 2020s with a subsequent ordoliberal-like philosophy leveraging ex ante regulation to preserve a contestable market structure [41], including revisions of the Anti-Monopoly Law and Anti-Unfair Competition Law [42–44]. Even as platforms such as WeChat are virtually ubiquitous in China [45] and deeply integrated with state and private actors [2], no single actor has to date managed to translate this digital hegemony into the same level of financial concentration achieved in the US.

The number of Chinese organizations that develop and release large AI models bears witness to that wider distribution of the Chinese AI ecosystem. We review major model developers in the US, in China, and internationally in Appendix C, accounting for both closed and open developers. We start with models of at least 500 billion parameters, where Chinese and US models alike top common benchmarks. As of September 2026, the US counts six developers with models of that size, four closed and proprietary models, and two open-weight. In contrast, twelve Chinese organizations have trained a model of that size, all of whom have released at least one as an open-weight model. We see broadly similar ratios for other size categories, with 8 US to 15 Chinese passing the 250 billion parameter mark, and 9 US to 16 Chinese at a 100 billion parameters or more: one of which is State-Owned Entreprise China Telecom. In addition to the commercial and regulatory context outlined above, this diversity is bolstered in part by more consistent sharing of technical innovation [46,47], and overall cheaper model training for similar benchmark performances [48].

Overall, these mechanisms paint a different picture of government influence on the trajectory of AI than in the US. A more distributed ecosystem, less vertically integrated actors, and less available capital for the majority of these actors require ecosystem-wid interventions, whose effectiveness depends on having a broader reach and on efficiently filling infrastructure gaps that are outside of the reach of most developers. The latter role is filled in part by SOEs and public investment in energy, data and hardware production to bolster AI development and deployment. Government control over the direction of AI development and the behavior of AI models is similarly diffuse. Conditional SOE financing and conditional permitting reinforce endeavors that follow state priorities, and filing requirements and digital content operations such as Qinglang shape training data and what deployed systems may produce, without requiring interventions that target specific provider [49,50]. The durability of these levers depends on maintaining a balanced market, which is a focus of recent updates to China’s digital competition laws. We review the governance instruments this approach depends on in §I.B, the investment in energy and compute in §II.A, strategies to provide data resources in §II.B, and reliance on open innovation §II.C. The competition updates are addressed in §III.A, and the filing and content rules in §III.B.

I.B. Governance Frameworks for China’s AI Strategy

China’s AI strategy framework includes aspects of industrial policy, data compliance, and competition law, as well as processes for leveraging fiscal funds and state-owned enterprises, which depend on . We briefly introduce several of the relevant governance concepts below to support discussion throughout the report.

II. Building up AI: Energy, Compute, and Data

II.A. Developing Energy and Compute Infrastructure

The Chinese Action Plan for the High-Quality Development of Computing Infrastructure (Gong Xin Bu Lian Tong Xin (2023) No.180) [62] emphasizes the need to develop “computing power”, defined as a “a new form of productivity that integrates information processing capabilities, network transmission capacity, and data storage capabilities which primarily delivers services to society through computing infrastructure.” The AI+ Strategy Guidance and 15th Five-Year Plan translate this priority into infrastructure targets: ultra-large “intelligent computing” clusters, coordinated green-power and compute deployment, and nationwide scheduling of compute resources. [51] To meet the energy needs of this new physical infrastructure, China has significantly ramped up its power capacity production across all energy technologies since 2021; its additions in the last five years are estimated to be equivalent to the total US capacity and the government plans to add six times as much over the next five years, [63] with the greatest increase coming in the form of wind and solar energy. [64]

To accelerate this development despite the low attractiveness of infrastructure projects to private developers, China has injected fiscal funds and developed specific financial instruments to promote AI infrastructure, following recommendations outlined in the AI+ Strategy Guidance to encourage “long-term, patient, and strategic capital”. [59] In November 2025, China allocated a new ¥500 billion ($70.3 billion) policy-based financial instrument to propel mainly tech-driven projects and urban renewal programs. [65] In January 2026, four central government ministries issued a Notice on Optimizing the Implementation of the Fiscal Interest Subsidy Policy for Equipment Renewal Loans (Cai Jin [[2026]] No. 2) [66]. AI-sector borrowers renewing equipment through bank loans receive a 1.5% interest subsidy on the associated fixed-asset loan principal. The central government additionally directs where data centers should be built by lowering administrative review times in locations chosen to maximize utilization and localized service. [67] Central SOEs are also building operator-run compute and data hubs aligned with the “East Data, West Computing” layout. For example, SOE China Mobile’s “Chengdu trusted data space” for other SOEs provides general and AI computing capacity for data storage, analysis, and model training. [68]

State-owned enterprises (SOEs) also play an important role in driving investment to AI infrastructure. During a meeting held by the State-owned Assets Supervision and Administration Commission (SASAC) of the central government in February 2026, central SOEs were called to actively expand effective investment in computing power and advance the coordinated development of “computing power + electric power”. [68,69] Subsequent reporting describes SASAC’s wider push to implement the “AI+” special action through digital and intelligent transformation, with sectoral deployments such as CHN Energy’s partnerships on smart power generation. [68] In January 2025, the Ministry of Industry and Information Technology (MIIT) and the Ministry of Finance jointly established the National AI Industry Investment Fund (the “AI Fund”). [70] With a total capital of ¥60 billion ($8.2 billion) raised from SOEs, the AI Fund mainly invested in AI chips, computing infrastructure, and core integrated compute components, exemplified by its January 2026 investment on Xheart, a company specializing in integrated circuit chip design and the R&D of autonomous driving technologies. [71,72]

US industrial policy shows related developments: in 2025, the White House put out an Executive Order aiming to facilitate permitting for data centers, primarily overriding previous environmental protection regulations; it differs from similar Chinese efforts however in that it does not require developers to follow any specific strategic directives. [73] In terms of direct investment, the 2022 CHIPS Act included awards in the form of grants and subsidies to support US semiconductor manufacturing, including $8 billion in direct funding to Intel. [74] The US government subsequently required taking a 10% ownership stake of Intel as a condition for having the remaining two thirds of the grant paid out, [75] following a model closer to Chinese SOEs, although this development was recently contested in court by other shareholders. [76] Funding from the US government is more commonly transferred to the technology companies building data centers through contracts for services, with around $7 billion in yearly cloud contracts estimated for FY 2022. [30] Together, these arrangements represent a strategic model characterized by interdependence and alignment between the state and major commercial actors, rather than centrally defined objectives enforced through binding conditional support encoded in specific instruments. [37]

II.B. Making Data Available for AI

II.B.1. Domestic Data Elicitation and Flows

II.B.1.a. Making Public Data Available for AI

Public data has been treated as a strategic input to AI and the Chinese digital economy at least since the State Council named data the “fifth factor of production” in 2020. [77] However, translating that designation into usable training data has required successive reforms to how public institutions hold, share, and commercialize datasets.

The regulatory framework for public-institution data has evolved in phases. In 2022, the “Data Twenty Measures” [78] introduced a “Three-Rights Separation” framework distinguishing resource ownership, processing and usage rights, and product disposal rights, moving from applying an indivisible bundle model of data property toward a modular regime in which processors can generate commercial value without privatizing public assets. The September 2024 Opinion on Accelerating the Development and Utilization of Public Data Resources [79] solidified this Authorized Operation model. The policy also started a shift from mandating institutions to enact “open data” policies for the sake of transparency toward a “data development” requirement directed at making the data economically valuable, authorizing specialized entities to manage and monetize public datasets under state supervision. In early 2025, the central government further scoped the commercial mechanisms for using data in such a way, directing institutions to assess a fair fee to dispose of the data for commercial use (public benefit use remains free) as well as invest in making procedures easier to follow. [80]

The responsibilities for implementing these procedures are shared between the central and local governments. The National Public Data Resource Registration Platform was launched for trial operations on 1 March 2025 to establish a unified national public data circulation ledger and formulate an evaluation logic for data transactions, with the goal of enhancing public data resource management and sharing across organizations. [81] This ledger connects data sources curated and managed by local governments and standardizes technical access requirements under a single user account. The local governments in turn have autonomy to decide what data to prioritize sharing, determine fee structures under centrally determined maxima, and make other context-dependent technical decisions. [82] Cities like Wuhan [83] and Guangzhou [84] for example have published their detailed guidance on public data resource transactions.

Data access for AI in the US follows a different model — combining federally supported research infrastructure, collaboration agreements with leading technology firms, and private licensing markets — rather than the nationally administered commercialization of public institutional data described above. For the largest commercial developers, training data is often already plentiful: vertically integrated cloud and platform operators that build models also control large proprietary collections of user-generated, search, and commerce data, and partnership arrangements among those incumbents can exchange additional inputs unavailable to smaller actors. [12] Universities, independents, and smaller firms also publish models, tools, and techniques, but at commercial scale development tends to concentrate among those same integrated operators rather than spreading across a broad base of API-scale commercial developers. The National AI Research Resource (NAIRR), established in January 2024 under Executive Order 14110 and transitioning toward a sustained national infrastructure, addresses some of these barriers in the context of scientific research by connecting US researchers and educators to shared compute, models, and datasets through competitive allocations and industry-contributed resources. [85] Additionally, the NSF’s Public Access Plan 2.0 and required Data Management and Sharing Plans extend open-science obligations to scientific data produced under federal research awards — mobilizing federally funded research outputs rather than the institutionally held public datasets China is authorizing for commercial use. [86] The US also aims to direct data to model developers through public-private partnerships. The Genesis Mission reflects a comparable effort to mobilize federal scientific data for AI, but channels access through agreements with a small set of established partners rather than through open registration; [87] whether that structure will extend beyond those incumbents is not yet specified in the mission’s founding instruments. [88] Supplemental data otherwise comes through private licensing deals, often at multi-million-dollars to nine-figure annual cost, [89,90] or through copyright litigation favoring well-capitalized developers. [91]

II.B.1.b. Easing Compliance with Data Security Laws

Section I outlines the Three Paths framework for data security between Chinese actors and international partners. The 2024 Provisions on Promoting and Standardizing Cross-Border Data Flow (the “New Regulations”) [92] are the instrument that has most directly recalibrated its application for commercial and AI-related cross-border flows, moving from a rigid “security-first” posture to a stated goal of balancing development and security requirements, through mechanisms such as raising thresholds for mandatory security assessments and introducing broad exemptions for commercial activities deemed routine. For most AI companies that are non-Critical Information Infrastructure Operators (non-CIIOs), the New Regulations increase the exemption threshold for non-sensitive PI tenfold, from 10,000 to 100,000 individuals [93] processed annually. This relaxation applies strictly to non-Sensitive Personal Information (non-SPI), whereas Sensitive Personal Information (SPI) remains under strict scrutiny.

The exemption regime builds on China’s Data Classification and Grading System, introduced in the Data Security Law. [94] Under this system, the New Regulations establish that data containing neither Personal Information (PI) nor “Important Data” is generally free to flow. While the definition of important data remains a source of ambiguity, the centralized government has produced further overall guidance [95] and tasked local governments and industry regulators [96] with maintaining specific “Important Data Catalogues.” Free Trade Zones (FTZs) are also experimenting with even more limited application of the three paths requirements, by restricting them to “Negative Lists” of even more specific data types.

A key additional provision in the New Regulations is the Data Transit exemption, [97] whereby data collected outside of China, that does not include Personal Information of Chinese citizens or “Important Data” under the grading system, and that is intended for processing and re-export does not trigger “Three Paths” requirements. This allows multinational corporations (MNCs) to move their data processing operations to China’s compute infrastructure with greatly facilitated compliance with China’s data security laws. Furthermore, the regulations include exemptions for transfers of PI classified as “truly necessary,” including contractual necessity, cross-border HR management, and emergencies.

II.B.2. Importing Data, Exporting Compute and Infrastructure

China’s global strategy treats international data and use cases as critical inputs to AI development, as reflected in the Global AI Governance Action Plan [98] and the Global Cross-Border Data Flow Cooperation Initiative [99] — which together promote cross-border data flows, shared digital platforms, international digital infrastructure buildout including data centers and computing systems, and standardization of use formats and deployment practices; extending the reach of China’s AI systems and data pipelines.

Investments in global digital infrastructure constitute the base layer of this approach. A central example is the Digital Silk Road (DSR), a component of the Belt and Road Initiative. This state-led framework focuses on digital and technological infrastructure including telecommunications networks, fiber-optic cables, data centers, cloud services, and smart city systems. It is financed by state-backed institutions such as the China Development Bank and the Export-Import Bank of China and large Chinese technology companies such as Huawei and Alibaba. [100] The Digital Silk Road improves digital connectivity across Asia and Africa, thereby expanding the global presence of Chinese technology companies and increasing China’s role in global data storage, transmission, and processing networks; as well as acting as a major lever of political influence. [101] Work with state organizations to contract Chinese firms for their infrastructure projects creates long-term technological relationships. [102] For example, Huawei has built telecommunications networks and data infrastructure across multiple African countries, where Western firms have been less active in large-scale infrastructure deployment. In 2024, Huawei partnered with MTN and China Telecom to expand 5G, cloud, and AI capabilities across the continent. [103]

The international deployment of AI technologies specifically is a key component of this strategy, as reflected in the Global AI Governance Action Plan. In July 2025, China proposed the establishment of a new global AI cooperation organization, along with expressing offers to share its development experience and products with other countries, particularly the “Global South”. [104,105] These efforts are aligned with the document’s framing of AI as a “global public good” and emphasis on sharing technological achievements internationally. [98] Chinese firms are acting on that mandate abroad. Alibaba Cloud opened a second Dubai data center in 2025 as part of a broader international infrastructure push. [106] Zhipu AI, one of China’s leading large language model developers, has expanded in Southeast Asia and the Middle East with localized government and enterprise offerings, often through Alibaba Cloud partnerships. [107] Autonomous-driving firms such as WeRide are running robotaxi pilots with Gulf authorities in Dubai and Abu Dhabi. [108] These deployments of AI systems abroad generate new data and operational experience which serve as inputs to domestic model development, in addition to bolstering China’s diplomatic reach.

Bipartite diplomatic relationships beyond direct exchange of commercial services also play a major role in the overall strategy. The China–EU High-Level Digital Dialogue [109] serves as a regular government-to-government platform for coordination on digital policy — including AI regulation, platform governance, and cross-border data flows — and has recently facilitated practical cooperation on industrial data transfer, including a mechanism to streamline the transfer of non-personal industrial data in sectors such as finance, automotive, and information technology. [110]

The United States has pursued related international approaches, but primarily as facilitation for private-sector export and market access rather than as state-defined deployment programs. General digital connectivity has relied on private-sector models such as SpaceX’s Starlink [111] or discontinued experimental programs such as Meta’s Aquila. [112] For AI, the Department of Commerce’s American AI Exports Program invites industry-led consortia to submit integrated technology packages — spanning hardware, data infrastructure, models, security measures, and applications — for partner markets; designated consortia may receive prioritized export-licensing review, financing referrals, and government-to-government facilitation, without the program setting sectoral adoption targets or requiring participation in a centrally coordinated overseas rollout. [113,114] Cross-border data governance abroad follows a related but distinct logic. Where China has promoted cooperative flow frameworks through standing policy instruments and bilateral forums such as those described above, US diplomatic guidance has directed posts to engage partner governments on data-localization rules that could restrict cross-border use of US cloud and AI services — a market-access orientation that supports private providers’ ability to operate across borders, rather than a cooperative flow regime articulated as part of a broader state integration strategy. [37,115]

II.C. Self-Reliance through Open-Source AI and Domestic Chips

Sustainable success for China’s AI industrial strategy also requires tailoring development and deployment to the country’s full technical stack and ecosystem. That push for self-reliance has been reinforced by US restrictions on advanced semiconductor exports, [116] pressing developers toward domestic chips and greater computational efficiency in pursuit of more performant models. An open-source and open-weight ecosystem has proven particularly effective to that end, allowing model developers to build on one another’s technical contributions rather than each starting from closed baselines. [46,117]

Policy operationalizes that ecosystem through initiatives such as the International Open-Source AI Cooperation Initiative, [60] which emphasizes lowering barriers to open AI development at home and encouraging global participation. National programs embed AI in education, industrial planning, and public governance; local governments add compute vouchers, tax concessions, and discounted data-center access so smaller actors can train or adapt open-weight models without large capital reserves. [118] Those instruments favor compute-efficient training and deployment under tighter resource constraints — DeepSeek’s R1 model is a notable example [46,119] — atop a compute stack in which private AI cloud remains gradated among platform operators [38] while state-directed hubs and scheduled infrastructure supply a distinct public stratum [120]. They also support a distinct commercial logic in which firms release open weights while monetizing deployment, integration, and industry-specific applications rather than closed API access alone. [121] Domestically, that shift is reflected in growing demand for application-layer AI in industrial and enterprise settings. [122] Abroad, open-weight release supports adoption of Chinese-trained models across a greater variety of computational infrastructure setups, spreading Chinese model and data formats and allowing downstream innovations to flow back to upstream developers.

On the hardware front, Chinese firms including Huawei have advanced domestic AI chips such as the Ascend series, now widely used in state-supported and enterprise applications. [123] Meanwhile, model developers have worked on ensuring that their AI systems run as efficiently as possible on domestically developed chips. Zhipu AI highlighted efforts to make the GLM 5 model inference run efficiently on a wide range of domestic chips at the time of release. [124] More recently, Meituan also announced it had trained a trillion-parameter model entirely on domestic chips, [125] and OpenBMB, a partnership with Tsinghua University, released an ultra-efficient open language model for edge devices trained on Huawei Ascend chips. [126]

The integrated approach of providing mandates and incentives for open-source development, domestic chip manufacturing, and mutual integration of the two has allowed Chinese developers to play to the strengths of their comparatively more distributed technological ecosystem, rather than simply attempting to catch up in the footsteps of their US counterparts with more limited resources. Outside rankings of open model releases count dozens of organizations, from the better-known names such as DeepSeek and Qwen through a long tail of startups, platform R&D arms, and university-linked communities, releasing notable and commercially viable open-weight models. [46,47] This approach is starting to bear fruit in the development of “token factories” serving these models with certified compliance and fully domestic infrastructure; such as the Shantou Free Trade Zone, [127] with more modularity across stack layers and more points of intervention for compliance than typical US deployment at comparable scale.

The United States follows a different industrial logic at commercial scale. Proprietary inference APIs dominate token usage among developers who switch models [33], and commercial AI infrastructure is concentrated among a small set of hyperscale cloud providers linked to leading model developers through partnership and investment patterns documented in the data-access section above [12]. An active open-weight community on platforms such as Hugging Face remains important for startups and smaller organizations [47]. Among its outputs are field-defining alignment methods such as Direct Preference Optimization [128], and fully open model families such as Olmo 3 that represent the main basis for fully reproducible and contextualized research [129]. The platform firms that dominate API access also release open-weight families into that commons [130,131]. Those outputs keep a research and startup layer alive. At commercial scale, however, the industrial logic remains the proprietary API, and the firms’ own open-weight releases extend it rather than replace it.

III. Addressing New Market Power Factors and Data Risks of AI

III.A. Adapting Platform and Competition Law to AI

III.A.1. Updates to the Anti-Monopoly and the Anti-Unfair Competition Laws

The development of AI is tied to the growing importance of a platform-based economy. Platforms’ broad user bases and digitally legible interactions make them a key driver of generative AI deployment; in addition to their existing uses of AI for content moderation, recommendation, and income optimization, particularly so in China where platforms mediate an even greater share of online activity. [132] China’s AI ambitions are therefore linked to the growth of its platform economy, as reflected in the 2025 Outline of the Government Work Report target that “core digital economy” industries reach 12.5% of GDP [133] and in local government support such as Sichuan’s commitment of ¥200 million in annual platform-economy fiscal support over three years. [134] At the same time, competition has moved beyond price as data, algorithms, and capital become increasingly major factors of market power under the new technical paradigm. This shift has been showcased for example by Alibaba’s ability to leverage its capital advantages over competitor Meituan, and the rapid expansion of platforms like TikTok (a short video platform owned by ByteDance) and Rednotes (an online community to share life experiences) supported by new modes of AI-driven interaction between e-commerce and personal digital lives. This evolution, which threatens the current balance of economic actors in China’s digital platform ecosystem, has necessitated amendments to the Anti-Monopoly Law and the Anti-Unfair Competition Law.

The 2022 revision of the Anti-Monopoly Law (AML) [42] adapted the existing framework to the platform economy. The legislature incorporated concepts such as data, algorithms, technology, and capital advantages, signaling that digital market competition requires tools beyond traditional industrial-era metrics. The AML updated provisions — including on hub-and-spoke agreements, abuse of dominance through data and algorithms, and below-threshold merger review — to recognize that control over data, user lock-in, and platform-level aggregation determines competitive significance more directly than traditional price factors. Furthermore, the inclusion of civil public interest litigation provides a collective mechanism for addressing the dispersed, low-value harms that individual users often face when their rights are infringed by platforms. [135] These efforts continue a trend started with the 2021 Anti-Monopoly Guidelines for the Platform Economy Sector [44] addressing the role of technical advantages and setting foundations for addressing specific concerns about how recent generative AI systems raise the risk profiles of new kinds of personal data. [136]

The 2025 revision of the Anti-Unfair Competition Law (AUCL) [43] complements the AML by focusing on unfair competitive conduct within digital platforms. While the AML addresses market structure and dominant power, the AUCL targets specific on-platform dynamics including digital fraud and misrepresentation of goods, imposition of unfair payment prices and conditions, and lock-in through technical means. This focus reflects a recognition that the same AI-enabled tools that improve efficiency can also facilitate exclusion, manipulation, and unfair advantage. It places greater compliance responsibilities on platform operators and strengthens penalties, addressing how large digital intermediaries may distort fair competition without necessarily fitting older monopolization models. Regarding AI specifically, the regulation explicitly prohibits training products on data scraped from the platforms, as well as using AI to distort markets by fabricating products or traffic. The 2025 AUCL also includes an extraterritorial jurisdiction clause (Article 40), [137] as a critical tool to enforce the use of approved channel for training models, both domestically and abroad.

Among other goals, these reforms can be read as stabilizing market structure: they encourage AI and data use in the digital economy while guarding against a single private actor acquiring complete ownership of any given sector, or using the technical characteristics of AI to bypass existing competition protections designed for a previous dominant paradigm.

III.A.2. Comparison of US and Chinese Cases

Both China and the United States have begun to recognize that the use of data, algorithms, and platform technologies may create new competition law problems in the digital economy. China’s approach is primarily rule-based, responding through statutory amendments and judicial interpretations (e.g., the 2022 AML revision and 2024 Judicial Interpretation of the Supreme People’s Court on monopoly civil disputes [138]). In contrast, the United States relies more heavily on case-driven enforcement actions and litigation under existing antitrust statutes (e.g., the Sherman Act and the FTC Act).

The issue of algorithmic collusion showcases these dynamics. Under the Supreme People’s Court 2024 interpretation, undertakings that use data, algorithms, technology, or platform rules to exchange information or coordinate conduct may be reviewed as monopoly agreements. This shows a growing recognition of the specific risks of algorithmic collusion, although there are still relatively few mature public cases in which AI algorithms themselves are the core issue. In the United States, this issue was addressed in the US v. RealPage case, [139] in which the DOJ argued that algorithmic pricing tools may facilitate coordination among market participants. The case ended in a settlement that included prohibition on different forms of data use for training and deploying AI models, particularly non-public data and data from competitor platforms — similar to provisions in the revised AUCL. The settlement also supported private class actions [140] by affected individuals seeking financial remedy.

Issues related to self-preferencing and access restrictions show similar dynamics. According to Articles 9 and 22 of China’s 2022 Anti-Monopoly Law, undertakings may not use data, algorithms, technology, capital advantages, or platform rules to engage in monopolistic conduct or abuse market dominance. These provisions provide a legal basis for addressing practices such as self-preferencing, ranking manipulation, or access restrictions; but in practice they have been operationalized through administrative decisions more than judicial judgments. In Qihoo 360 v. Tencent, Tencent blocked competing applications and restricted interoperability; Qihoo sued for abuse of dominance. [141] The court held that multi-homing costs are low and alternatives exist, and blocking does not automatically foreclose competition; and in general tends to reject a dominance finding based only on a lack of technical interoperability. Administrative decisions, on the other hand, have taken a stricter stance, such as by mandating Alibaba to open its data, payment system, and application access. [142] Conversely, the same question in the US has been addressed chiefly through litigation. In FTC v. Amazon [143], Yelp v. Google [144], and the Department of Justice’s litigation against Google [145], US authorities have been more willing to frame platform self-preferencing and distribution restrictions as concrete antitrust issues, although the remedies have been behavioral rather than structural and limited in scope. [146]

Efforts to address data-driven price discrimination show similar limits in both jurisdictions. The Supreme People’s Court’s 2025 platform-monopoly discussion recognizes “big data discrimination” as a competition concern, [147] but mature antitrust judgments remain rare. [148] Courts have rejected claims concerning “big data price discrimination against loyal customers” on grounds such as the user’s prior consent to data use, the view that price fluctuations were caused by factors other than data, the user’s ability to switch platforms, or insufficient technical evidence. [149] In the United States, according to the FTC’s 2024 surveillance pricing inquiry, regulators have also begun to examine the use of consumer data, behavioral information, and algorithmic tools in personalized pricing. However, in the absence of the direct use of competitors’ data (as in the RealPage case outlined above), the courts have tended to favor deployers of algorithmic pricing software. [150]

III.B. Disclosure Requirements and Content Rules for AI Technology

AI models developed in China are subject to substantial transparency requirements, implemented through the Cyberspace Administration’s (CAC) Internet Information Service Algorithm Filing System. [151] The system rests on three regulations — the Algorithmic Recommendation Provisions (2022), [152] Deep Synthesis Provisions (2023), [153] and Generative AI Interim Measures (2023), [154] and applies to services whose algorithms have “public opinion attributes” or “social mobilization capabilities” as per the Provisions on the Security Assessment of Internet Information Services with Public Opinion Attributes or Social Mobilization Capabilities, [155] including major content platforms, e-commerce and service apps, AI chatbots, and search engines. Filings center on providers’ behavioral logic, giving regulators specific information to support rule-making about technical systems.

Submissions are meant to help regulate algorithmic conduct, protect user rights, and surface risks such as algorithmic discrimination, information cocoons (often called echo chambers or algorithmic bubbles in US discourse), and the generation of false or non-compliant content. [152,153] Providers of algorithmic-recommendation services, deep-synthesis services (including technical supporters), and generative AI services that meet the criteria above (e.g., forums, blogs, or chat services that channel “public expression” or have “social mobilization capabilities”) must complete an Internet Information Service Algorithm Filing. According to CAC system guidelines, filing categories include generation and synthesis, personalized recommendation, ranking and selection, retrieval and filtering, and scheduling and decision-making; providers must display the filing number in a prominent location on their service platforms, link to the public filing record, and disclose the basic principles of their algorithmic mechanisms. [152] Generative AI services listed on the CAC filing registry include DeepSeek and Baidu’s Ernie Bot. [156]

Large language models deployed as generative AI services meeting the criteria above fall under the same filing system but face additional requirements under the 2023 Interim Measures, enforced by the CAC with MIIT and MPS. [154] LLM filings combine technical registration and security assessment. First, developers submit detailed disclosures to the CAC portal — including parameter scale, architecture, and the “core logic” governing retrieval and generation. Second, they must document training-data sources, demonstrate dataset legality, describe filtering of illegal or harmful content, and report domestic-to-international data ratios. Third, a mandatory Security Self-Assessment tests the model against a standardized corpus of sensitive prompts to ensure outputs align with “Socialist Core Values” and do not threaten national security or raise concerns related to “social stability”.

The US regime for development disclosure and content rules is more distributed. The nearest analogues are state laws such as California’s Training Data Transparency Act (AB 2013) [157] or the Colorado AI Act, [158] although developers are challenging both state regimes in court — xAI v. Bonta over California’s AB 2013 [159] and xAI over Colorado’s AI Act, where DOJ has intervened [160] — and compliance to date has often been formal rather than substantive (e.g., OpenAI’s AB 2013 training-data summary) [161]. On these topics, the US has recently directed at state-level requirements an approach parallel to its diplomatic engagement abroad on cross-border data rules (subsection II.B.2) [115,162] — executive preemption efforts that extend federal involvement in non-federal regulatory conditions at home, including direction to challenge state AI laws in court. Federally, policy has also relied on procurement standards requiring “truth-seeking and ideologically neutral” LLMs [36] and voluntary pre-release testing through early API access to NIST’s CAISI [163]; both focused on model behavior rather than development process or data disclosure.

Conclusion

China’s AI policy operates as a coordinated industrial strategy in which the state directs investment, sets adoption targets, and adapts pre-existing digital-governance tools ex ante — in contrast to a US approach that more often shapes AI-related constraints through litigation under statutes not originally designed for the technology and through the priorities of a concentrated set of private developers, while also steering scale through grants, commercial agreements, and recent executive action on providers’ operating conditions. Binding instruments and SOE-backed capital give the central government sustained leverage over infrastructure and sectoral uptake, filling gaps where private investment is slow or returns are uncertain.

Supply-side policy concentrates on computing power, energy, domestically viable chips, and data. China has scaled physical infrastructure through fiscal capital and SOE investment at a pace that reflects strategic priority rather than market timing alone. Recent data reforms work within — and extend — the PIPL/Data Security Law and classification-and-grading framework, raising exemption thresholds and introducing carve-outs for routine commercial flows and data transit while keeping sensitive personal information and “important data” under stricter scrutiny. Training-data access follows a distinct logic in which the state has authorized commercial use of public institutional datasets and pursued international data and deployment channels through initiatives such as the Global Cross-Border Data Flow Cooperation Initiative and Digital Silk Road investments. Open-source and open-weight development has been central to that strategy — enabling developers to share technical contributions, pursue compute-efficient training paths, and co-optimize models with domestic hardware across a more distributed ecosystem; US export controls on advanced semiconductors have reinforced that push toward domestic chips and efficiency-focused model work. The United States has pursued a parallel but distinct supply-side path — relying on platform-proprietary collections and private licensing for training data, steering capacity through grants, service contracts, infrastructure facilitation, and industry-led export packages rather than binding overseas industrial targets, and concentrating routed inference among a few closed commercial API providers.

As generative AI scales, the strategy also puts deliberate weight on maintaining the stability of the digital ecosystem as a condition for its effectiveness: market-structure stability through competition law, and information-order stability through upstream model and algorithm filing. Competition-law reforms address the market-power risks of scaled platform AI: AML and AUCL revisions explicitly treat data, algorithms, and capital as competition factors and seek a digital sector that can grow without any single private actor exclusively controlling a sector or the channels through which models are trained; taking a more proactive approach than the US focus on litigation under existing statutes. Separately, algorithm and generative-AI filing extends privacy, data-security, and content rules upstream — requiring disclosure, training-data documentation, and security self-assessment before launch, including testing against sensitive prompts to ensure model outputs follow China’s online content control regime. In the US regime, this is most directly compared with various state disclosure and anti-discrimination requirements and voluntary federal pre-release testing, rather than a unified upstream filing gate.

Together, these deployment rules show comparable ambition for AI expansion between the US and China, but a different architecture of control: process requirements and directed investment under state oversight versus reliance on concentrated private developers and ex post litigation once market structures have formed.

Acknowledgement

Research work on the Chinese regulatory ecosystem supporting this report was conducted during the LLM Pro Bono Intensive organized by the New York University School of Law’ Public Interest Law Center and hosted by Data & Society. In particular, we thank Brian J. Chen for helping organize and supervise this effort.

Appendix

A. Methodological Notes on Equity Market Share of AI

This appendix documents the market-structure metrics used in §I.A.1 and §I.A.2: equity-market concentration (listed cohorts, private secondary marks, sector composition).

A.1. Purpose and scope

We provide an evaluation of how much of a country’es economy as measured by total company value is tied to firms with major stakes in AI-relevant layers: including cloud, platforms, large models, AI silicon, or datacenter supply chains; starting with publicly listed companies. Additionally, we find that privately owned (non-listed) actors also weigh significantly on this balance both in the US and in China, with secondary markets putting them among the most valuable companies overall in their respective ecosystems; and while other valuable privately owned companies do exist beyond AI technology companies, the latter do tend to have uniquely high valuation-to-revenue ratios. As a result, we provide two values for the AI shares of each economy: the proportion of currently listed value of AI companies, and the proportion if AI companies if currently private high-value AI-relevant companies with a recent valuation based on secondary market transactions were to enter the market at this valuation. For the latter, we chose this counterfactual scenario rather than trying to estimate the total weight of privately owned companies across the economy for several reasons: 1. AI-relevant companies are much more likely than other compaies to have a recent valuation in secondaty markets, 2. such a valuation typically denotes a process to prepare for an IPO in the foreseeable future, and 3. we found the secondary markets to be a more defensible and comparable valuation across the board than alternatives that would apply to a broader share of the economy: in particular, recent public valuations of AI companies have showned that heuristics like revenue multiples fail to apply to the current paradigm. We provide both proportions for the US and China below, as well as sensitivity analyses across various other factors. In general, we find that our findings on the differences in market concentration hold across choices.

Inclusion rule for listed companies. A listed firm enters the equity cohort if it has material exposure to at least one of the following layers: hyperscale or AI cloud; large-model or AI-platform operations; AI silicon, memory, or accelerators; foundry or semiconductor equipment; or the datacenter physical stack (optics and interconnect) where AI is a primary growth driver. In the US, all ten top-valued companies fit squarely into that category, as do AMD, Intel, and Oracle among the next 10-15. Companies like Palantir, Lam Research, and Applied Materials meet a slightly broader reading of the same rule and are reported as sensitivity in §A.5 rather than in the canonical total. On the China ranking, the same rule selects CXMT, Tencent, Alibaba, Zhongji Innolight, Cambricon, Xiaomi, Eoptolink, SMIC, NAURA, Meituan, China Telecom, AMEC, and Baidu. China Telecom is included because it publishes large open-weight models (Appendix B); China Mobile is held for sensitivity. Hong Kong–only listings, including Z.AI, are not on the China ranking page and are treated in §A.5.

A.2. Sources for Company Valuations and Market Shares

For publicly listed equity, we compare their values as of mid-September 2026: September 10 for the United States and September 9 for China, based on available Web Archive snapshots. Private valuations use the latest information available at that date. The specific sources are described in the table below.

For privately held AI companies that meet the criteria outlined above of having a recent secondary market valuation, we use the valuations provided int he sources below. All private marks are point-in-time secondary prices, sensitive to deal structure, buyer pool, and illiquidity; they are not fully equivalent to daily listed caps, but reflect investors’ beliefs about the value of a company.

The oldest secondary valuation included is that of the Ant Group, which has had no further secondary stock sale events since 2023.

A.3. United States Equity Analysis

CompaniesMarketCap reports total US listed equity of $79.537 trillion on the September 10, 2026 ranking page.

Top 10 listed companies. All ten largest US-listed firms by market cap on the snapshot date are technology firms with major AI stakes (hardware, cloud, platforms, or models). Together, they account for over 35% of US listed equity.

Listed AI cohort. Several companies between positions 10 and 25 also meet the criteria for the AI cohort: in particular, Intel, AMD, and Oracle are a clear fit for the cohort given their focus on AI infrastructure. Other companies, such as Palantir, Lam Research, or Applied Materials also have business cases relevant to AI, although less directly related. We include them here for comparison and in the sensitivity analysis in §A.5. After these, several additional companies could arguably be included in an AI cohort, but their smaller valuation and AI-dependent share of revenue make them less obvious fits (e.g. Salesforce, ServiceNow, Dell, IBM, KLA, etc.) Including them would have further bolstered the AI share of the US economy by company value.

Privately held AI companies. OpenAI, Anthropic, and Databricks all define themselves primarily as AI companies, and while they do not have an official listed valuation, their post-money valuations at their latest funding rounds (or in the case of OpenAI a secondary tender) put them among the most valuable companies in the US. To account for their weight, we run a counterfactual analysis, where we update the denominator to reflect the total listed equity value if these companies entered the stock market at their secondary price. We make this adjustment in part because while other private companies in the US do have comparable or higher revenues, they overall are not raising venture capital and are not on a track to enter the stock exchange; whereas Anthropic and OpenAI have both filed S-1s to prepare for an IPO.

Summary: US AI value share

A.4. China Equity Analysis

CompaniesMarketCap reports total Chinese listed equity of $11.486 trillion on the September 9, 2026 China ranking page [165]. This aggregate follows CompaniesMarketCap’s “China” classification for mainland China, excluding Hong Kong listed company (we discuss this choice and its consequence further in §A.5).

Top 10 listed companies. Similarly to the US, the top ranked company is a semiconductor company, but the rest of the distribution diverges: where all 10 stops in the US are taken up by companies with a strong AI activity, the head of the Chinese distribution has a more even split between financial institutions and technology/industrial firms, with additional listings covering consumer good and energy: for an aggregated weight of ≈40% financial, ≈40% technology/platforms, ≈7% consumer, ≈13% energy/industrial.

Listed AI cohort. Several additional companies among the highest-value AI firms also meet the criteria outlined above for the AI cohort. One conspicuously missing company is Huawei: while it plays a major role in Chinese industrial policy, it is fully privately owned and has not raised private capital in a way that would provide a defensible valuation. We discuss the possible impact of adding Huawei to our analysis further in §A.5.

Privately held AI companies (and Z.AI). Companies ByteDance, DeepSeek, Moonshot, and the Ant Group have all released large AI models with performance comparable to those of their US counterparts in the last year, so we included them based on their secondary valuation using the same method as for OpenAI, Anthropic, and Databricks on the US side. Additionally, Zhipu AI/Z.AI is a major model developer listed in Hong Kong. We choose not to add the Hong Kong market directly as explained in §A.5, so we instead provide an analysis treating Z.AI similarly to private companies at its Hong Kong listed value.

Summary: China AI value share

A.5. Sensitivity Analysis and Limitations

CompaniesMarketCap also publishes a Hong Kong ranking (152 companies, total market cap $1.219 trillion on the August 30, 2026 snapshot). We exclude Hong Kong from the canonical denominators and cohort for two reasons:

1. Non-China issuers, before the shares are computed. The ranking is not a China market. Its largest names include AIA, Hong Kong Exchanges and Clearing, Swire Pacific, Jardine Matheson, and the Hong Kong property groups. That mix is enough to withhold the total from the China denominator.
2. Directional bias, after the shares are computed. Adding the exchange lowers both listed shares: the top ten from 30.7% to 27.8%, and the listed AI cohort from 20.7% to 19.4%. Leaving Hong Kong out does not create the paper’s contrast. The China side would look slightly less concentrated if it were included.

Hong Kong’s listed AI cohort. Four names meet §A.1: Lenovo ($47 billion, PCs, AI servers), China Unicom ($22 billion, public cloud), ASMPT ($9 billion, semiconductor assembly and packaging equipment), and Silicon Motion ($8 billion, SSD controllers). Those caps, $86 billion together, are the numerator add in the Hong Kong rows below.

Huawei. Huawei is central to China’s AI cloud and domestic silicon stack (see the Omdia shares in §B.2) but remains unlisted, with no consistently reported secondary mark across 2025–2026. A private estimate or a revenue multiple would not be reproducible, so the canonical totals exclude it. The sensitivity row below adds Huawei only to the counterfactual line, at $150–300 billion, between major listed equipment and cloud peers and below historic private peaks. Those bounds are for intuition only and are not used in the main text.

Sensitivity. The following table shows the impact of different choices on the AI value shares of the US and Chinese economies under the assumptions outlined above. We see that even comparing the most conservative US accounting (strictly defined cohort of publicly listed companies only) to the most generous Chinese version (counterfactual analysis including privately owned companies, Z.AI listed in Hong Kong, and the higher possible end of Huawei valuation), the US still shows a significantly higher AI concentration than China does at 37.8 vs. 27.5 points.

Additional limitations. In addition to aspects addressed in the sensitivity analysis above, we acknowledge two main limitations of the methodology. First, valuations of AI-relevant companies, whether on public or private markets, can be highly volatile: for example, between its peak in July and its September valuation, Z.AI lost over half of its listed value, and SpaceX also saw substantial variation after its IPO; so this analysis should be seen as a point in time rather than a stable accounting of concentration over a long period. Second, no adjustments are made for overlapping economic interest. Alibaba holds roughly one-third of Ant Group; hyperscalers hold stakes in OpenAI and Anthropic; Tencent and ByteDance invest across model labs. These ties are noted because they affect interpretation of “independent” concentration, not because double-counting can be resolved without arbitrary carve-outs.

B. Cloud market shares

The overall weight of AI companies in a country’s equity market provides one view of concentration: primarily focused on which entities are estimated by financial markets to have the most value. We can also look to a more specific measure of market power according to market share by focusing on a particular layer of the AI stack. Since cloud infrastructure-as-a-service ties in most of the actors involved in AI development and deployment, both in the US where three of the largest AI companies have significant cloud commercial activity and in China where AI companies are either owners or direct customers, we focus on this layer.

Even more so than for equity shares, canonical market shares for different cloud providers in the US, China, and internationally are particularly difficult to obtain. Where such comprehensive numbers are compiled, they are often only accessible through expensive purchases: for example, companies like the International Data Corporation make detailed reports on market shares of cloud infrastructure available to customers for upwards of $7,500. While this makes a detailed apples-to-apples comparison between US and Chinese markets, we can still rely on partial reporting from more public-facing organizations and reporting on the content of the commercially produced reports to paint a higher-level picture of both markets; enough to compare general concentration dynamics that underline each country’s industrial policy.

B.1. US cloud providers in the worldwide infrastructure market

For the US market, we rely on cloud market shares for Q2 2025 from a report published by outlet Computer Reseller News citing numbers from Synergy Research Group’s quarterly release.

The reported numbers cover world-wide market, not specifically US consumers. Therefore the concentration question we ask here is: of the cloud revenue collected by US-based companies, how much flows toward the main actors in the field? We start by collating market shares and actor locations mentioned in the report in the following table:

In order to get a sense of US-side concentration, we then re-normalize the share of each US company by the total share of US companies. While we do not have access to the ~11% in unnamed remainder, we feel reasonably confident that US alternatives together account for less than 2 percent of the total: Cloudflare total revenue that quarter was about 0.5% of overall Q2 cloud revenue ($512M of the $99B reported by Synergy Research Group), Rackspace about 0.7%, HPE about 0.55% (dividing ARR by four), and Digital Ocean at 0.2%. These represent the companies’ total revenue, not just through public cloud market. Using a 2% upper bound on US companies in the unnamed category above, we get the following normalized market shares for US companies:

Overall, this distribution shows a dominance of three main actors, with Amazon, Microsoft, and Google together accounting for nearly 82% of the cloud revenue of all US companies, followed quite far behind by Oracle, another member of the AI cohort we identified in §A.1.

B.2. China domestic revenue for public and AI cloud

On the Chinese side, we use two different sources. First, a third-party report by the Chinese Securities Times cites figures from the IDC report “China Public Cloud Services Market (H1 2025)” showing the general cloud market shares of selected companies, including major relevant State-Owned Enterprises (SOEs). Second, reporting by the South China Morning Post using figures from a report by Omdia gives us insights into the AI-specific market shares of the top 5 major Chinese providers for the same period. We collate these numbers below:

The public cloud market column above is the more directly comparable to the US numbers provided in §B.1, although they compere market share on domestic and international markets. Still, for overall cloud markets, the top three actors in China account for 52% of the market, compared to nearly 82% in the US; additionally, SOEs make up a notable 21.7% of total cloud revenue. The picture changes somewhat when focusing specifically on compute-heavy AI cloud, where different actors hold more significant market shares, but even then the top 5 together hold about 77% of the total market. The main point of convergence between both markets, however, remains the existence of a clearly leading actor: Amazon cloud with its nearly 39% share of the revenue captured by US companies, and Alibaba with its 26.8-35.8% of Chinese general and AI cloud markets.

C. Organizations Publishing Large Open Models

For a third view of concentration in AI, we look to the number of organizations in each country developing and publishing large AI models. Models, particularly Large Language Models (LLMs), are a common focus of policy discussions.

We review open-weight models of three size categories:

• 100B+ parameter models: these represent a significant effort to train and typically start running on cloud compute resources. While models in that category do not typically rank with closed APIs across all benchmark tasks, they constitute the backbone for several companies’ offering - as showcased by e.g. Cohere’s Command A Plus or Poolside’s Laguna-S-2.1
• 250B+ parameter models: as of 9/28/2026, the smallest model featured on the Artificial Analysis Intelligence Index, a widely used benchmark for a model’s “general-purpose” performance, is Z.AI’s GLM-5.3-Flash: a 321B-parameter model ranking 15th, ahead of Google’s best models and several other closed API models. We report models with at least 250B parameters
• 500B+ parameter models: most models that are typically described as “frontier” have at least 500 billion parameters. These models are the most resource-intensive to train and deploy, and the most likel to have state-of-the-art performance on benchmarks across the board.

In addition to open models, we also count organizations who develop and commercialize models that are likely 500B parameters or more even when they do not release them. At the time of writing, there are four US-based developers following this approach: Anthropic, Google, OpenAI, and xAI.

The following table provides a summary of the number of organizations per country developing models of the corresponding size, with the N (+M) format for the US lines denoting N organizations releasing the models and M organizations having closed models of that size. Even when counting closed models, China has about twice as many organizations as the US for each size category (a little less below 500B).

C.1. Methodology

The counts above start from the Hugging Face Hub. On 23 September 2026, the index of base models with at least 100 billion parameters returned 3,722 repositories (models index). The same request on the public models API uses num_parameters=min:100B and base_model_relation=base (Search the Hub). Of those repositories, 3,415 were uploaded on or after 1 January 2025. We use the upload date because the API does not record when a model was trained.

That filter is the starting set. Marking a repository as a base model drops those that declare a parent, but many re-uploads never do, and the 3,415 repositories belonged to 1,692 accounts. We drop a handful of repositories whose headers claim more than 10 trillion parameters, then treat repositories that share an exact safetensors parameter total as one set of weights. Within a set we drop third-party quantizations and conversions, and repositories whose names embed another laboratory’s model family unless the account is that laboratory. We keep the publishing organization rather than a personal account or a community mirror. A few repositories still marked as base models are post-trains of another laboratory’s weights; we remove those after reading the card, including Rakuten AI 3.0, Nex-N2.5-Max, Fx-Bio, and FengHe, along with dummy weights and randomly initialized shells. Copies of the same release in FP8, BF16, base, or instruct form then count as one row. The size we use is the main weight count, rounded to the nearest billion, rather than the sum of every format stored in the repository.

The cell in each column is that organization’s latest release at or above the threshold, uploaded on or after 1 January 2025. A newer and smaller model can therefore stand in for an older and larger one. Six organizations whose only model at this scale was uploaded in 2024 are left out: BAAI, Skywork, and XVERSE in China, Snowflake in the United States, and SB Intuitions and Preferred Networks in Japan.

One release never entered the 3,722 because the repository reports no parameter total. Mistral Large 3 is described in its card as trained from scratch, with 675 billion total parameters. We include the base checkpoint, uploaded 30 November 2025, on that basis. Mistral’s later Medium 3.5 release, at about 128 billion parameters, remains the entry in the 100B column.

The Hub does not record a country. We assign it from the model card, the organization’s own description, or a primary announcement of the release. The Institute of Foundation Models is counted in the United Arab Emirates, as part of MBZUAI in Abu Dhabi, even though it also has offices in California and Paris. Hong Kong is not split out from mainland China, and no organization in this set is a Hong Kong-only issuer of the kind discussed in §A.5.

C.2. List of organizations

Each row is one organization. A link is the latest public release at that size, and a dash means there is none. The same name can appear in more than one column when that release is still the latest above each cutoff. Italic cells are the closed developers named above, not repositories on the Hub. OpenAI’s 100B cell is its public gpt-oss-120b release; above that size it is counted with the other closed developers.

C.3. Limitations

1. Published models, and an assumption for closed ones. The table counts models that have been published. The italic cells for Anthropic, Google, OpenAI, and xAI assume that each has a model above 500 billion parameters. No closed developer outside the United States is added on the same basis.
2. The parameter count can be missing or inflated. Repositories with no safetensors total never enter the query, which is why Mistral Large 3 had to be added from its card. Another release with the same omission could still be missing. Packed formats, and headers that do not match the weights, can also make a raw total larger than the model. The closest calls are Upstage Solar-Open2 at 250 billion, just above the middle cutoff, and Alibaba’s Qwen3-Coder at 478 billion, just under the top one.
3. The date is the upload date. An organization qualifies on a repository it uploaded on or after 1 January 2025, including a later copy of a 2024 inventory. The cell itself may be a post-training release from the same organization, as with MiMo-V2.6-Flash and Inkling-Small, rather than the original pretraining run.
4. Country is assigned. The Hub does not return one. The assignments here are the organizations’ public headquarters. A lab whose only presence was a shell affiliate would be harder to place; that case did not arise in this set.