Chapter 1: The Empirical Void in US-China AI Discourse

The Benchmark Illusion and National Capacity

For more than three years, national evaluations of artificial intelligence leadership have relied almost exclusively on public benchmark leaderboards. Metrics derived from evaluations such as LMArena, MMLU, GSM8K, and HumanEval are routinely cited by policymakers, intelligence analysts, and technology journalists as definitive scorecards of geopolitical standing. When an American laboratory releases a model scoring two points higher on graduate-level reasoning, commentators declare the United States insurmountable; when a Chinese laboratory closes the gap weeks later, headlines proclaim immediate parity.

This fixation on benchmark rankings produces a dangerous analytical distortion. Benchmark scores evaluate isolated algorithmic reasoning in static synthetic environments. They offer virtually no visibility into how models are deployed across commercial value chains, how effectively enterprises integrate automated systems into daily operations, or whether national industries are building enduring data feedback loops. An economy that excels purely at benchmark mathematics while failing to integrate AI into manufacturing, energy, and logistics risks misjudging its true industrial strength.

To move beyond this empirical void, strategic analysis requires a shift from macro benchmark posturing to micro firm-level observation. By cataloging what AI developer firms are actually building, what modalities they prioritize, and which economic sectors they serve, analysts can construct an accurate assessment of national technological competitiveness.

From Model Parameters to Economic Micro-Foundations

Artificial intelligence does not transform national economies through theoretical laboratory publications. Realized economic transformation requires capital expenditure, enterprise software customization, physical tooling reconfiguration, and organizational workforce adaptation. A nation possessing the most capable conversational chatbot gains little lasting geopolitical advantage if its domestic manufacturing firms lack the technical capability to deploy edge-based vision models on factory floors.

Furthermore, evaluating national AI capacity solely through frontier labs obscures the vast commercial ecosystem operating immediately below the frontier. While frontier model developers capture public attention, downstream developers translate foundational discoveries into tangible software applications, industrial robotics controls, and autonomous navigation stacks. These downstream commercial developers determine the true velocity of economic absorption.

Recognizing this empirical deficit, the RAND Corporation undertook a systematic empirical mapping of the commercial AI developer landscape in the United States and China. Rather than tracking model weights in isolation, this study measures the concrete corporate entities translating artificial intelligence into commercial products across both economies.

Chapter 2: Dataset Construction and Methodology

A Comprehensive Cohort of 1,181 AI Developer Firms

To conduct a rigorous comparative investigation, the authors constructed a bespoke microeconomic dataset of commercial artificial intelligence developer firms in the United States and China. A firm was defined as an active commercial entity that designs, trains, or substantially adapts artificial intelligence models for commercial release, internal enterprise deployment, or industrial automation. Purely non-technical service agencies and hardware component resellers without internal AI modeling capabilities were systematically excluded.

The resulting dataset comprises 1,181 verified commercial AI developer firms, consisting of 743 firms headquartered in the United States and 438 firms headquartered in China. Data collection was finalized in early 2026, capturing corporate registrations, patent filings, active open-source software repositories, venture financing rounds, commercial customer deployments, and regulatory filings from both jurisdictions.

This cohort represents the most extensive audited catalog of active AI developer firms compiled to date. By establishing consistent classification taxonomies across both geographic contexts, the dataset enables direct empirical comparison of technical capabilities, model architectures, commercial market targets, and ecosystem dependencies.

Multi-Agent Extraction and Expert Validation Pipeline

Assembling a microeconomic dataset across two distinct corporate reporting environments required an innovative data collection pipeline. Traditional registry scraping often misses fast-moving startup entities or introduces significant naming ambiguities between US parent holdings and domestic Chinese operating subsidiaries. To address these challenges, the research team deployed a multi-agent information retrieval framework powered by specialized LLM agents tasked with scanning technical documentation, code registries, patent databases, and local business registries.

The automated multi-agent extraction pipeline operated in structured stages: candidate entity identification, commercial product verification, model architecture classification, and deployment sector tagging. Each automated extraction was subsequently subjected to dual-analyst manual validation by bilingual research specialists with technical training in machine learning and knowledge of Chinese corporate filings.

The classification taxonomy categorized each firm along multiple dimensions: core model development (foundation model pre-training vs task-specific fine-tuning), supported modalities (text, vision, audio, multimodal, and embodied control), target market (digital enterprise services, industrial manufacturing, consumer applications, public sector), and deployment paradigm (cloud API, edge embedded, on-premises sovereign cluster). This multi-tier taxonomy ensures that nuanced architectural differences are accurately reflected.

Chapter 3: The Critical Divergence: Pure Software vs. Physical Embodiment

The United States: Dominance in Enterprise Software and Digital Work

While the architectural foundations of AI development in the United States and China exhibit remarkable convergence, the commercial destinations of those models diverge dramatically. In the United States, commercial AI development is overwhelmingly channeled toward digital knowledge work and enterprise software. 61.2 percent of US AI developer firms build pure-play digital software products, operating entirely within cloud and desktop environments.

The concentration of US commercial activity in enterprise SaaS, developer tooling, healthcare informatics, and defensive cybersecurity reflects the structural economics of the American tech sector. Decades of venture capital investment have established software-as-a-service as the benchmark for financial returns, characterized by gross margins exceeding 70 percent, rapid software deployment cycles, and minimal hardware depreciation risk. American enterprise buyers possess both the discretionary IT budgets and the digital infrastructure required to absorb high-margin cognitive software tools.

Consequently, American foundation models and vertical adaptations are predominantly optimized for text comprehension, coding syntax, analytical reasoning, and knowledge retrieval. While American laboratories have demonstrated breakthrough physical robotics systems in research settings, the vast majority of commercial capital flows into digital applications that substitute or augment office-bound knowledge workers.

China: Rapid Expansion in Industrial Edge and Physical Systems

In contrast to the software-centric focus of the American market, Chinese AI developer firms are heavily focused on physical embodiment and industrial infrastructure. Approximately 73.8 percent of Chinese AI developer firms in the dataset build models explicitly designed to interface with physical hardware, industrial manufacturing equipment, robotic actuators, autonomous transportation fleets, or smart electrical grids.

This structural orientation is anchored in China's position as the world's largest manufacturing powerhouse. Industrial enterprises face rising domestic labor costs, demographic transitions, and stringent quality control mandates across high-volume assembly lines. AI developer firms in China have found their most lucrative and predictable enterprise revenue not in generic productivity software, but in factory edge vision for defect detection, predictive maintenance algorithms for heavy machinery, automated warehouse logistics, and autonomous industrial haulage.

Furthermore, national industrial policy incentives have deliberately steered computational talent and capital toward physical manufacturing upgrades. Chinese government procurement programs, national manufacturing transformation funds, and specialized industrial testbeds prioritize AI deployments that enhance industrial output and technological self-reliance over consumer chatbots or digital marketing automation.

Chapter 4: The Strategic Moat of Embodied AI

The Physical Data Flywheel and Synthetic Data Limits

The divergence between purely digital software and physical embodied systems carries profound long-term strategic ramifications. In the digital domain, foundational models are rapidly exhausting high-quality human-generated text and code on the public internet. While synthetic data generation provides meaningful performance extensions, frontier labs acknowledge that synthetic self-training exhibits diminishing returns and error compounding over successive generations.

In contrast, embodied artificial intelligence operates in the infinite, non-replicable complexity of physical reality. Robots, autonomous vehicles, and industrial machinery interact directly with dynamic real-world environments, continuously capturing multimodal sensorimotor streams: high-resolution spatial video, LIDAR point clouds, tactile force feedback, torque telemetry, and thermal variations. This physical interaction data cannot be scraped from the open web or synthesized through digital simulation alone.

Firms operating physical systems establish proprietary data flywheels that strengthen with every operational hour. An automated assembly line operating 24 hours a day generates rare corner-case failure data that cannot be simulated artificially. Over time, this empirical interaction moat becomes virtually insurmountable for competitors lacking physical hardware deployed in the field.

Hardware Supply Chain Proximity and Iteration Velocity

Developing embodied AI systems requires tight co-design between algorithmic models and mechanical hardware. Model architectures must be tailored to specific actuator response rates, sensor latency profiles, thermal envelopes, and edge computing chipsets. When algorithmic updates occur, physical mechanical components often require rapid prototyping, machining, and field recalibration.

Chinese embodied AI firms benefit from geographic co-location within the world's densest advanced hardware supply chains. In regional manufacturing hubs such as the Greater Bay Area and Yangtze River Delta, a robotics team can design a custom actuator or sensor bracket in the morning, receive precision-machined prototypes by afternoon, and test real-time inference on the production line by evening. This rapid mechanical feedback cycle accelerates empirical progress.

In the United States, hardware prototyping cycles frequently encounter protracted supply chain lead times, with domestic manufacturing capacity hollowed out and overseas procurement introducing weeks of shipping latency. As a result, American researchers often spend significant effort optimizing simulation environments to bypass physical testing, while their Chinese counterparts iterate directly against real physical prototypes on active factory floors.

Operational Resilience Against Remote Disruption

The operational resilience of physical embodied systems represents an underappreciated geopolitical dimension. Purely digital software services rely heavily on global cloud infrastructure, continuous public internet connectivity, and foreign API service agreements. A domestic enterprise relying on a foreign SaaS platform for mission-critical workflows is inherently vulnerable to remote account termination, cross-border data transfer embargoes, or submarine cable interdictions.

Conversely, industrial embodied AI systems are engineered to operate autonomously on local physical infrastructure. Factory quality control models, robotics navigation kernels, and smart electrical grid dispatchers are deployed directly onto on-premises edge appliances and isolated industrial control networks. These systems function uninterrupted even in the event of complete severance from external cloud networks.

By embedding artificial intelligence directly into the physical machinery of core industries, China is constructing an automated industrial base with high operational durability. Even under severe geopolitical friction or export isolation, these embodied systems continue driving factory production, logistics throughput, and energy distribution without relying on foreign cloud APIs.

Chapter 5: Open-Weight Diffusion and Global Token Parity

OpenRouter Telemetry: The Global Rise of Open Weights

The international diffusion of artificial intelligence capabilities cannot be evaluated solely through domestic commercial markets. Globally, developers choose models based on three decisive factors: performance, control, and inference cost. Over the past two years, open-weight models developed by Chinese institutions (notably DeepSeek, Alibaba Qwen, 01.AI, and Zhipu AI) have experienced a dramatic surge in international adoption.

To quantify this phenomenon, the research team analyzed longitudinal routing telemetry from OpenRouter, one of the world's primary vendor-agnostic LLM routing platforms serving millions of global software engineers. In early 2024, proprietary US closed models (led by OpenAI, Anthropic, and Google) accounted for over 80 percent of daily routed token volume. By the first quarter of 2026, empirical telemetry reveals that Chinese open-weight models reached token volume parity with leading US closed models, each commanding approximately 50 percent of global routing traffic.

This shift represents a fundamental transformation in global developer mindshare. Software teams in Europe, Latin America, Southeast Asia, and Africa are increasingly adopting open-weight models for core production infrastructure. The ability to inspect model weights, fine-tune locally without vendor lock-in, and deploy on sovereign hardware clusters has made open-weight architectures the preferred substrate for international enterprise development.

Inference Economics: The 1/4 to 1/5 Cost Disruption

The primary catalyst behind this global migration is inference economics. As frontier artificial intelligence transitions from experimental prototypes to high-volume commercial automation, the operational expenditure of running billions of inference tokens daily has become the dominant cost center for modern digital enterprises. Software companies cannot build scalable businesses if every customer query consumes tens of cents in proprietary API fees.

Telemetry data demonstrates that Chinese open-weight models deliver reasoning capabilities comparable to tier-one US proprietary models at roughly 1/4 to 1/5 the inference cost. Hosted commercial API endpoints for frontier open-weight models routinely charge between $0.14 and $0.60 per million output tokens, compared to $3.00 to $15.00 per million tokens for premier American closed alternatives. When enterprises self-host open weights on commodity compute clusters, marginal inference costs fall even further.

This severe price disparity has initiated a commoditization wave across the global software industry. High-margin US software businesses that previously built simple wrappers around proprietary frontier APIs are experiencing margin collapse, while open-weight adopters build cost-resilient enterprise applications that can scale to billions of monthly active users without catastrophic cloud compute bills.

Chapter 6: Strategic and Policy Implications

Policy Recommendations for the United States

The findings of this empirical study indicate that US artificial intelligence strategy requires a fundamental recalibration. American policymakers must recognize that algorithmic leadership on digital benchmark leaderboards does not translate automatically into national economic resilience or industrial superiority. Relying exclusively on high-margin enterprise SaaS leaves critical industrial sectors exposed to international technological leapfrogging.

To preserve long-term technological competitiveness, the United States should actively incentivize the commercialization of embodied artificial intelligence and robotics. Policy mechanisms should include targeted federal investment in physical testing testbeds, tax credits for domestic manufacturers deploying edge-based automation systems, and streamlined regulatory pathways for autonomous industrial operations. Rebuilding domestic advanced manufacturing ecosystems is not merely an industrial policy imperative, but a foundational requirement for sustaining national AI leadership.

Furthermore, American defense and civilian planners must diversify their technical evaluations beyond digital language models. National AI readiness scorecards must incorporate microeconomic metrics tracking the installed base of robotic systems, physical sensorimotor data capture volumes, and industrial automation density across critical supply chains.

Strategic Opportunities and Bottlenecks for China

For China, the empirical data validates the strategic efficacy of national policies aimed at industrial integration and open-source diffusion. By focusing development efforts on smart manufacturing, robotics, and physical infrastructure, Chinese firms have carved out defensible technological positions anchored in the world's most versatile production supply chain. Furthermore, the global success of open-weight models has established significant international goodwill and developer adoption.

However, significant structural vulnerabilities persist within the Chinese AI ecosystem. Chief among these is continued exposure to advanced semiconductor export restrictions. While architectural optimizations and efficient fine-tuning techniques have allowed Chinese developers to deliver competitive frontier models with fewer resources, sustaining foundation model pre-training across increasingly large multimodal datasets will inevitably encounter hardware bottlenecks if domestic chip alternatives fail to match global leading-edge yields.

Additionally, Chinese developers must cultivate greater domestic enterprise willingness to pay for software innovation. While industrial enterprises readily invest in physical robotic hardware and machinery, willingness to fund standalone software licenses and high-tier cloud computing architectures remains constrained compared to Western markets. Fostering a healthy, self-sustaining domestic software market will be essential for funding next-generation basic research.

Rethinking Export Controls and International AI Governance

The empirical convergence in core algorithmic architectures and the proliferation of open-weight models reveal significant limitations in existing export control paradigms. Multilateral export restrictions designed around rigid computational FLOP thresholds struggle to restrict technological diffusion in an era where algorithmic efficiency gains, model distillation, and open-source parameter sharing allow developers to achieve frontier performance on modest hardware footprints.

Policymakers must transition from blunt hardware interdictions toward sophisticated whole-of-ecosystem monitoring. Restricting raw silicon shipments alone cannot halt the rapid iteration of embodied systems when physical training data and manufacturing feedback loops reside within domestic borders. Future non-proliferation and economic security frameworks must incorporate software optimization trajectories and industrial edge deployment realities.

Finally, the rapid militarization and industrial integration of physical artificial intelligence underscore the urgent necessity for bilateral US-China technical engagement on safety and catastrophic risk mitigation. As autonomous drones, industrial machinery, and critical infrastructure become increasingly reliant on machine learning kernels, technical protocols governing fail-safe actuation, unintended physical escalation, and sensor spoofing require clear channels of scientific communication between the two leading AI superpowers.

Conclusion: Moving Beyond the Frontier Benchmark Illusion

The empirical evidence gathered across 1,181 AI developer firms provides a necessary corrective to contemporary geopolitical discourse. The competition in artificial intelligence between the United States and China cannot be reduced to a binary sprint toward an elusive benchmark score. Instead, the global landscape is witnessing the emergence of two structurally differentiated ecosystems: an American paradigm optimizing cognitive software services for digital enterprises, and a Chinese paradigm embedding autonomous perception and control into the physical machinery of global manufacturing.

Neither model possesses an absolute, uncontested monopoly on future technological progress. The nation that successfully bridges its structural deficiency: whether by connecting digital cognitive software to physical manufacturing, or by scaling domestic high-performance silicon to sustain open-weight foundational research: will establish the standard for 21st-century technological leadership. For strategic analysts and corporate executives alike, navigating this landscape requires looking past the benchmark illusion to measure real-world commercial and industrial deployment.