# The AI Stack Tracker Assessing how the US stacks up against China across each layer of the modern AI value chain. ## Summary ### State of the Stack The United States and its allies maintain a commanding lead across most layers of the AI supply chain, particularly in semiconductor manufacturing equipment, fabrication, and AI accelerator production. China holds advantages in critical mineral refinement and power infrastructure, foundational inputs that could become increasingly important as AI scaling continues. Export controls have successfully limited China's access to cutting-edge chips, but the gap at the frontier model layer is narrowing. The race is not over. --- # Critical Minerals China dominates global critical mineral refinement, controlling approximately 70% of rare earth processing and significant shares of other semiconductor-essential materials. **Measure:** Net Refinement Capacity for Semiconductor-Critical Minerals **US & Allies Score:** 4.0/10 **China Score:** 8.0/10 **Leader:** China ## Key Metrics - **Mining share:** Mining Share (%) = (Country mine production / Global mine production) × 100 - **Refining share:** Refining Share (%) = (Country refined output / Global refined output) × 100 ## What matters in this layer AI hardware pulls demand from a narrow set of inputs: rare earth elements for magnets, gallium and germanium for compound semiconductors, cobalt and lithium for storage and power systems, and high-purity copper and aluminum for electrical and thermal management. ### Concentration risk Processing and separation steps concentrate in a small number of jurisdictions. This makes export controls and permitting delays systemically important. ### Spec and quality Semiconductor-grade materials require stringent purity and consistency. The constraint is often qualification and process integration, not only tonnage. ## Recent Developments ### China Dominates Rare Earth Processing *1 day ago | Supply Chain* China controls approximately 70% of global rare earth element processing capacity, creating a critical dependency for advanced semiconductor and AI hardware manufacturing. ### US Announces Lithium Processing Initiative *3 days ago | Policy* The Department of Energy has announced $2.8 billion in funding for domestic lithium processing facilities, aiming to reduce dependence on Chinese refinement operations. ### Australia-US Partnership on Critical Minerals *1 week ago | Partnerships* A new bilateral agreement between Australia and the United States aims to establish secure supply chains for critical minerals, including lithium, cobalt, and rare earth elements. ### China Restricts Gallium and Germanium Exports *2 weeks ago | Trade* Beijing has implemented export controls on gallium and germanium, two materials essential for advanced chip production. This move is seen as retaliation for US semiconductor restrictions. ### Strategic Stockpile Expansion Planned *1 month ago | Policy* Congress is considering legislation to expand the National Defense Stockpile of critical minerals, including gallium, germanium, and other elements essential for advanced chip production. --- # Power Infrastructure China has significantly more slack in its power grid, with massive overcapacity that can be directed toward AI data centers. The US faces grid constraints that limit rapid AI infrastructure scaling. **Measure:** Available Grid Capacity (Power Overhang) for AI Data Centers **US & Allies Score:** 3.0/10 **China Score:** 9.0/10 **Leader:** China ## Key Metrics - **Power overhang:** Available Capacity = Dispatchable surplus after reserves + operational constraints - **Net nuclear change:** Net Added GW = New nuclear capacity commissioned − retirements ## What matters in this layer Power availability sets the ceiling for AI compute scaling. Permitting, grid capacity, and generation mix determine what can actually be built and operated. ### Interconnection queues Delays for new large loads and generation create a practical ceiling even when projects are funded and equipment exists. ### Firming and reliability AI clusters need predictable, high-uptime power. Dispatchable capacity and grid stability matter more than annual averages. ### Time-to-build Permitting, transmission buildout, and supply chains for transformers and switchgear set the pace of scaling. ### Geography Compute wants cheap power and cooling, but constraints are local. Regional bottlenecks can dominate national totals. ## Recent Developments ### China's Grid Overcapacity Enables AI Expansion *2 days ago | Infrastructure* China's massive investment in power generation has created significant overcapacity, with an estimated 400 GW of slack that can be rapidly deployed for AI data centers without major grid upgrades. ### US Data Center Expansion Faces Grid Bottlenecks *1 week ago | Constraints* Major US hyperscalers report waiting times of 3-5 years for new power connections, with some projects delayed due to insufficient grid capacity in key regions. ### Nuclear Power Revival for AI *2 weeks ago | Energy* Several US tech companies are exploring partnerships with nuclear power providers to secure reliable, carbon-free electricity for AI data centers. Microsoft has signed a deal to restart Three Mile Island. ### China Adds 350 GW of New Capacity Annually *3 weeks ago | Analysis* China's annual power capacity additions dwarf all other nations combined, with a mix of coal, renewables, and nuclear providing the foundation for continued AI infrastructure growth. --- # Semiconductor Equipment Leading-edge chips rely on a dense ecosystem of lithography, etch, deposition, metrology, and inspection tools. US and allied firms dominate the most advanced toolchains. **Measure:** Advanced Lithography and Leading Process Tooling **US & Allies Score:** 10.0/10 **China Score:** 2.0/10 **Leader:** US & Allies ## Key Metrics - **Tooling dependence:** Dependence Index = (Critical tool steps supplied by constrained vendors / Total critical tool steps) × 100 - **Process capability:** Capability = (Achievable CD control + overlay + defectivity) at target throughput for leading nodes ## What matters in this layer Lithography is the headline, but sustained advantage comes from the full stack: deposition, etch, metrology, inspection, and the software and service ecosystem that keeps fabs running at high utilization. ### EUV as a hard gate EUV tools (and their optics, sources, and resists) set the feasible frontier for patterning. Export controls and spare‑parts access can translate directly into node ceilings. ### Service and yield learning Process windows are discovered in production. Rapid iteration depends on field service, metrology feedback loops, and tight vendor integration. > Next: we can add a “critical path” diagram of tool steps for a leading node once arrows/flows are introduced. ## Recent Developments ### EUV Lithography Remains Concentrated *Recent | Lithography* Extreme ultraviolet (EUV) scanners are essential for state-of-the-art nodes. Tight export controls and complex supply chains keep the frontier concentrated among a small set of allied suppliers. ### China Accelerates Domestic Tool Substitution *Recent | Industrial Policy* China continues investing in domestic etch, deposition, and metrology. Progress is fastest in mature-node tooling, with slower advancement at the leading edge. ### Metrology and Inspection as a Bottleneck *Recent | Yield* As nodes shrink, defect detection and overlay control become more challenging, increasing the strategic value of metrology and inspection systems. --- # Fabrication Leading-edge chipmaking remains concentrated in allied foundries, while China is constrained by access to cutting-edge tools and yields at the smallest nodes. **Measure:** Leading-Edge Node Capacity (<= 7nm equivalent) **US & Allies Score:** 9.0/10 **China Score:** 3.0/10 **Leader:** US & Allies ## Key Metrics - **Leading‑edge capacity:** Share (%) = (Leading‑edge wafer starts / Global leading‑edge wafer starts) × 100 - **Effective output:** Effective Output = Wafer starts × Good‑die yield × Die per wafer (at target performance bin) ## What matters in this layer Frontier accelerators and high‑end CPUs are constrained by access to the best nodes and stable high yield. Concentration creates systemic risk: a single geography can determine global compute growth. ### Yield learning and ramp speed The strategic advantage is how fast a fab can move from first silicon to high‑volume manufacturing. Tooling, metrology, and process discipline compound. ### Geopolitical concentration When leading capacity is concentrated, disruption risk becomes a supply‑chain constraint. Resilience requires both geographic diversification and compatible equipment ecosystems. > Next: we can attach the fabrication podium/wafer‑stack visualization as a dedicated page and link to it here. **Related:** [Related: Critical Minerals](/layer/critical-minerals) ## Recent Developments ### TSMC Continues to Lead at the Frontier *1 week ago | Foundry* TSMC remains the dominant producer of leading-edge logic chips, with strong demand from AI accelerator vendors and hyperscalers. ### Intel Foundry Expands Advanced Packaging and Logic Roadmap *2 weeks ago | Investment* Intel is investing heavily in its foundry roadmap and advanced nodes, with an emphasis on US-based capacity and secure supply chains. ### SMIC Advances on Mature Nodes, Faces EUV Gap *3 weeks ago | Constraints* China's leading foundry continues to improve production on mature nodes and uses multi-patterning for smaller geometries, but remains limited by the EUV tool gap. ### Geopolitics Keeps Capacity a Strategic Asset *1 month ago | Strategy* Regional incentives and security concerns continue to drive new fab announcements and reshoring efforts across the US, Japan, and Europe. --- # Advanced Packaging TSMC's CoWoS (Chip on Wafer on Substrate) technology is the critical bottleneck for AI chip production, enabling the integration of multiple chiplets and HBM memory. **Measure:** CoWoS & Advanced Chiplet Integration Capacity **US & Allies Score:** 8.0/10 **China Score:** 4.0/10 **Leader:** US & Allies ## Key Metrics - **Packaging throughput:** Throughput = (Packaged units/week) constrained by CoWoS lines, bump/TSV steps, and test capacity - **Substrate constraint:** Constraint = min(ABF substrate supply, interposer capacity, packaging tool availability) ## What matters in this layer Advanced chiplets trade monolithic scaling for integration complexity. Dominance comes from proven process recipes, low defectivity across large interposers, and a supply chain that can expand without breaking yields. ### CoWoS capacity as a gate Even with ample wafer starts, GPU shipments can bottleneck on CoWoS throughput. Expansions depend on equipment and qualified operators, not only capex. ### Yield compounding Multi‑die packages multiply yield loss modes. Small defect improvements compound into large effective output gains, especially for large HBM‑heavy packages. ## Recent Developments ### TSMC CoWoS Capacity Constrains AI Chip Supply *3 days ago | Manufacturing* TSMC is aggressively expanding CoWoS capacity, but demand continues to outstrip supply. The company expects to double capacity by 2025. ### Intel Advances Foveros 3D Packaging *1 week ago | Technology* Intel's Foveros technology enables face-to-face chip stacking, offering an alternative to TSMC's CoWoS for advanced AI chip integration. --- # Memory High Bandwidth Memory (HBM) is essential for AI training. SK Hynix and Samsung control over 95% of global HBM production, with chronic supply shortages. **Measure:** High Bandwidth Memory (HBM) Production Share **US & Allies Score:** 9.0/10 **China Score:** 3.0/10 **Leader:** US & Allies ## Key Metrics - **HBM share:** HBM Share (%) = (Supplier HBM output / Global HBM output) × 100 - **Stack yield:** Stack Yield ≈ Π(Die yield per layer) × TSV/bump yield × packaging yield ## What matters in this layer HBM capacity, binning, and reliability determine effective accelerator throughput. When HBM is short, compute is short. The constraint is as much packaging and qualification as it is DRAM wafer capacity. ### Supplier concentration A small set of firms controls most HBM output. This amplifies shocks, makes long‑term contracts valuable, and turns ramp decisions into strategic levers. ### Packaging coupling HBM availability depends on advanced packaging throughput and substrates. Memory cannot be analyzed in isolation from the packaging layer. ## Recent Developments ### SK Hynix Dominates HBM Market for AI *2 days ago | Memory* SK Hynix controls over 50% of the HBM market, with supply agreements locked in with NVIDIA and AMD through 2026. ### Samsung Ramps HBM3E Production *1 week ago | Manufacturing* Samsung has qualified its HBM3E memory with major GPU manufacturers and is rapidly scaling production. ### Micron Breaks Ground on New York Memory Fab *2 weeks ago | Investment* Micron has begun construction on a $100 billion memory fab complex in New York, strengthening US domestic production. --- # AI Accelerators The US dominates AI accelerator production through NVIDIA's near-monopoly on training GPUs. Export controls have severely limited China's access to cutting-edge chips, though domestic alternatives are emerging. **Measure:** Effective FLOPS Production Capacity (Peak x MFU) **US & Allies Score:** 9.0/10 **China Score:** 3.0/10 **Leader:** US & Allies ## Key Metrics - **Effective compute produced:** Effective FLOPs ≈ (Units shipped) × (Peak FLOPs) × (Realized MFU) - **System throughput:** Throughput is limited by min(Compute, Memory bandwidth, Interconnect, Power) ## What matters in this layer Dominance comes from the full stack: architectures, compilers, supply contracts, and the ability to scale systems. Export controls and allocation policies can redirect global compute flows. ### Supply chain coupling Accelerators are the intersection of leading fabs, advanced packaging, and HBM. A constraint in any upstream layer appears here as shipping delays. ### Software moat Compilers, kernels, and libraries determine real MFU. Software ecosystems are sticky and convert hardware advantage into durable platform power. > Below is a first‑principles embedded module that turns “Effective FLOPs produced” into a compact, scroll‑driven comparison. ## Recent Developments ### NVIDIA's H100 Dominates AI Training *2 days ago | Hardware* NVIDIA's H100 GPU continues to be the gold standard for large-scale AI model training, with US-based hyperscalers deploying hundreds of thousands of units in their data centers. The company controls over 80% of the AI training chip market. ### Export Controls Limit Chinese Access to Advanced GPUs *5 days ago | Policy* US export controls have effectively blocked Chinese entities from acquiring the most advanced AI accelerators, forcing domestic alternatives that lag 2-3 generations behind in performance and efficiency. ### Huawei's Ascend 910C Gains Traction *1 week ago | Competition* Despite sanctions, Huawei has developed the Ascend 910C AI accelerator using older process nodes. While performance remains below cutting-edge US chips, domestic adoption is growing among Chinese AI labs. ### NVIDIA Blackwell Architecture Ships *2 weeks ago | Product Launch* NVIDIA has begun shipping its next-generation Blackwell architecture GPUs, offering 4x the training performance of H100 for large language models. Demand far exceeds supply. ### Google TPU v5 Powers Gemini Models *3 weeks ago | Technology* Google has deployed its fifth-generation Tensor Processing Units (TPUs) across its data centers, optimized specifically for training and serving large language models like Gemini. --- # AI Factory AI factories are massive data centers purpose-built for training frontier AI models. US hyperscalers have deployed the largest clusters, though China is rapidly building capacity. **Measure:** Installed AI Training Compute Clusters **US & Allies Score:** 8.0/10 **China Score:** 5.0/10 **Leader:** US & Allies ## Key Metrics - **Installed training capacity:** Capacity = (Accelerators installed) × (Average utilization) × (Time online) - **Cluster power density:** Power Density (kW/rack) drives cooling architecture, floor design, and failure rates ## What matters in this layer The limiting factors are often mundane: permitting, transformers, chilled water, and network lead times. Operators who can compress these timelines turn capital into compute faster. ### Build velocity Time from site selection to first training run determines advantage. Standardized designs, supply contracts, and execution discipline compound over repeated builds. ### Operations and reliability Achieving high utilization requires strong SRE practices, failure recovery, and workload scheduling. Reliability is a strategic capability, not a back‑office concern. ## Recent Developments ### xAI's Colossus Cluster Goes Live *1 week ago | Infrastructure* Elon Musk's xAI has brought online the Colossus cluster with 100,000+ H100 GPUs, making it one of the largest AI training installations in the world. ### Meta Plans 2GW Data Center Campus *2 weeks ago | Expansion* Meta has announced plans for a 2 gigawatt AI data center campus, representing one of the largest single-site compute deployments ever planned. ### China Builds National AI Compute Network *3 weeks ago | Strategy* China is constructing a national network of AI compute centers, pooling resources across state-backed entities to maximize utilization of domestic chips. --- # Platform Cloud platforms are the interface through which most organizations access AI capabilities. US hyperscalers dominate globally, though Chinese platforms lead domestically. **Measure:** Cloud AI Infrastructure & Developer Platform Reach **US & Allies Score:** 7.0/10 **China Score:** 5.0/10 **Leader:** US & Allies ## Key Metrics - **Distribution reach:** Reach = (Active developers) × (Enterprise penetration) × (Regions served with compliant offerings) - **Compute allocation:** Allocation Share (%) = (Platform accelerator hours delivered / Total accelerator hours delivered) × 100 ## What matters in this layer The platform layer turns hardware into accessible capability: orchestration, networking, storage, monitoring, and security. Platforms also set the policy surface through export compliance and customer screening. ### Managed AI stack Turnkey training and inference services reduce friction and pull demand. Tooling quality and reliability directly affect adoption. ### Procurement advantage Platforms that can secure supply (accelerators, networking, power) can gate downstream innovation and attract the best workloads. ## Recent Developments ### AWS Expands AI Infrastructure *3 days ago | Cloud* Amazon Web Services continues to expand its AI infrastructure, offering access to NVIDIA GPUs, custom Trainium chips, and a growing suite of foundation models. ### Alibaba Cloud Deploys Domestic AI Chips *1 week ago | Strategy* Alibaba Cloud is increasingly deploying domestic AI accelerators across its infrastructure, reducing dependence on restricted US technology. ### Microsoft Azure AI Demand Surges *2 weeks ago | Business* Microsoft reports unprecedented demand for Azure AI services, with enterprise customers rapidly adopting Copilot and other AI-powered tools. --- # Frontier Model The gap between US and Chinese AI models is narrowing. DeepSeek's recent breakthroughs demonstrate that China can achieve competitive performance despite hardware constraints. **Measure:** Benchmark Performance, Capabilities & Adoption **US & Allies Score:** 7.0/10 **China Score:** 6.0/10 **Leader:** US & Allies ## Key Metrics - **Capability index:** Index = weighted benchmark performance + reasoning/agentic evaluations + reliability and safety metrics - **Adoption:** Adoption = (Active users + API volume + enterprise deployments) adjusted for retention and switching costs ## What matters in this layer Frontier performance is increasingly shaped by data quality, training infrastructure, and iteration speed. Safety work, evaluation, and deployment channels determine how capability translates into real influence. ### Iteration velocity Access to stable compute and strong engineering enables rapid experiments, faster post‑training, and better product integration. Speed compounds. ### Safety and policy surface Evaluation, alignment, and governance shape what can be deployed and where. Regulatory compliance becomes a feature of the product, not an afterthought. ## Recent Developments ### DeepSeek V3 Achieves Breakthrough Efficiency *1 week ago | Research* DeepSeek's V3 model demonstrates frontier-level performance trained with significantly less compute than Western equivalents, suggesting algorithmic innovations that may partially offset hardware disadvantages. ### OpenAI Releases GPT-4.5 *2 weeks ago | Product* OpenAI's latest model shows improved reasoning capabilities and multimodal understanding, maintaining its position at the frontier of AI capabilities. ### Anthropic Claude Sets New Benchmarks *3 weeks ago | Benchmarks* Claude's latest version demonstrates state-of-the-art performance on coding, analysis, and complex reasoning tasks. ### Chinese Open-Source Models Gain Traction *1 month ago | Adoption* Alibaba's Qwen and other Chinese open-source models are seeing increased global adoption, particularly in regions with data sovereignty concerns. --- # Testing and Safety Science The US leads in AI safety research and testing infrastructure, with dedicated organizations like Anthropic, OpenAI's safety teams, and government initiatives focused on responsible AI deployment. **Measure:** AI Safety Research Capacity & Testing Infrastructure **US & Allies Score:** 8.0/10 **China Score:** 5.0/10 **Leader:** US & Allies ## Key Metrics - **Internal safety spend ratio:** Safety Spend (%) = (Internal safety team budget / Total lab expenditure) x 100 - **Pre-release testing window:** Testing Window = Time from training completion to public release (US labs: ~2-6 weeks, Chinese labs: often <24 hours) - **Inference overhead for safety:** Safety Overhead (%) = (Compute for moderation & filtering / Total inference compute) x 100 — US major labs: 5-13%, Chinese labs: negligible - **External safety ecosystem scale:** Ecosystem = (Third-party red-teamers + Auditors + Safety data vendors + Independent labs) x (Revenue + Government funding) ## AI Testing and Safety Science Landscape AI Testing and Safety Science can largely be split into two dominant branches: internal safety departments within the major labs, and external safety teams that exist to serve frontier AI developers. The relationship between these branches—and the tension over talent, funding, and influence—shapes how safety outcomes evolve across both nations. ## Internal Safety Departments These exist within the major labs or other AI organizations releasing products and applications. At the frontier, these groups might be split into teams that focus on alignment, red-teaming, evals/preparedness, interpretability, and incident tracking. ### Internal safety spend as % of lab expenditures Rough ballparks of what fraction of total lab spend goes toward safety science and testing. The aim is to convey both total amount and relative amount spent—though adjacency between safety and capability work can cloud the picture. ### Safety-relevant research releases Number of safety-relevant research releases by the 5 leading labs per country, excluding system/model cards and focusing on other safety-adjacent findings. May also capture the magnitude of those findings via citation counts and awards. ### System cards and RSP habits Percentage of frontier releases accompanied by a safety card, extent of the card, existence of a Responsible Scaling Policy, and number of amendments. Chinese labs rarely publish cards and few maintain internal protocols. ### Inference overhead for moderation and filtering US major labs spend 5–13% of inference compute on additional safety layers (moderation, filtering, classifiers). Chinese labs spend negligible amounts on moderation. Many Chinese releases are open-weight and served elsewhere with varying levels of safety layers. ### Pre-release testing windows Chinese labs regularly ship open models within 24 hours after training completes. Some US labs perform suites of evals and external verification, sitting on models approximately 2–6 weeks before release. ### Safety benchmark leaderboard An amalgamation of safety benchmarks and jailbreaking scores looking at models from each country’s five leading labs, drawing from sources like the CAIS Safety Leaderboard and standardized red-teaming evaluations. ## External Safety Teams These exist to serve frontier AI developers, working to advance the safety outcomes of their products. Such groups take on many shapes and sizes, particularly as their relationships with AI labs continue to mature. ### Independent Research Labs Groups that spin up out of academia, government agencies, or other nonprofits. These organizations focus on the dominant frontier AI safety categories: alignment, red-teaming, evals/preparedness, interpretability, and incident tracking. ### Third-Party Red-Teaming Organizations that pursue contracts to uncover vulnerabilities at scale. Not all are for-profits—many come from nonprofit backgrounds (METR, Humane) or government (AISI). They focus on jailbreaking as a service, alignment evals, uplift studies, and model weight security. ### External Auditors A rapidly emerging category that performs process, managerial, and control checks. Whereas red-teamers focus on model behavior, auditors inspect the organizational mechanisms that shape those behaviors. ### Safety Data Vendors Another emerging category: teams that curate safety data for later training runs. These offerings are highly targeted, designed to enhance the security and reliability of specific AI releases. ## External Safety & Testing Metrics While the external wing complements the AI labs, there is tension in both nations. Talent and funding flow much more strongly to the labs and their internal safety teams. The disparity here is not common in any other industry. ### AI safety talent pool Size, quality, and flow of the AI safety talent pool across tiers: undergraduate, post-graduate, and professional. ### External safety partner revenue Total revenue of external safety partners across sub-branches: independent labs, red-teamers, auditors, and data vendors. ### Regulatory pressure for third-party oversight While movements at the state level have rebalanced this topic, there remains more mandatory oversight in China. Yet the focal point of that testing continues to revolve around content moderation. ### Government funding Direct funding to organizations like CAISI/CnAISDA or indirect grants to other safety organizations. ### Standards and certifications Count of published standards and number of certifications. While early attempts may border on safety theater, efforts to standardize safety protocols deserve credit. ### Incident tracking Using AIID as a source alongside the apparent rise of a Chinese equivalent. Key questions: has each nation built capacity to track incidents, and in reported harms, what was the AI weapon of choice? > This layer is designed for maximum interactivity and scrollability—quick-hitting metrics, accurate reporting, and memorable supporting graphics. Planned visualizations include: funding flow charts, heatmaps of system card releases, inference overhead diagrams, pre-release timeline comparisons, alignment score dashboards, talent pool tiers, and incident bulletin boards. **Related:** [Related: Frontier Models](/layer/frontier-model) ## Recent Developments ### NIST AI Safety Institute Expands *1 week ago | Policy* The National Institute of Standards and Technology has expanded its AI Safety Institute, developing comprehensive testing frameworks for frontier AI systems. ### China Announces AI Safety Regulations *2 weeks ago | Regulation* Beijing has introduced new regulations requiring safety testing for large language models before deployment, though enforcement mechanisms remain unclear. ### Academic Safety Research Grows *3 weeks ago | Research* US universities and research institutions continue to lead in AI safety research, with significant funding from both government and private sector sources. --- # Curated Data Training data is a critical input for frontier AI models. The competitive landscape depends on data quality, diversity, licensing, and regulatory frameworks governing data collection and use. **Measure:** High-Quality Training Data Availability & Access **US & Allies Score:** 6.0/10 **China Score:** 7.0/10 **Leader:** China ## Key Metrics - **Effective data advantage:** Advantage = (Unique high-quality tokens) × (Domain coverage) × (Licensing clarity) - **Synthetic data multiplier:** Effective Supply = (Real data) + (Synthetic data × Quality factor) ## What matters in this layer As model architectures converge, training data quality and curation are becoming primary differentiators. Access to proprietary, well-labeled, domain-specific data can determine which models achieve breakthrough performance in specialized areas. ### Data quality over quantity The shift from “more data is better” to “better data is better” is accelerating. Carefully curated, deduplicated, and high-quality datasets produce measurably stronger models at lower training cost. ### Regulatory landscape Differing privacy frameworks (GDPR, China’s PIPL, US state laws) shape what data is available for training. These regulatory asymmetries create distinct advantages and constraints for each ecosystem. ### Synthetic data AI-generated synthetic data is increasingly used to supplement real-world datasets, particularly for rare domains, code generation, and mathematical reasoning tasks. ### Licensing and provenance Legal challenges around training data usage are intensifying. Clear data provenance and licensing are becoming competitive advantages as litigation and regulation increase. ## Recent Developments ### Data Quality Becomes a Differentiator *1 week ago | Strategy* As model architectures converge, the quality and curation of training data is emerging as a key differentiator for frontier AI labs. Companies investing in proprietary, high-quality datasets are seeing outsized returns in model performance. ### China's Data Advantage in Specific Domains *2 weeks ago | Analysis* China's large internet population and different privacy frameworks provide access to vast datasets in areas like e-commerce, social media, and manufacturing, creating advantages for domain-specific AI applications. ### Synthetic Data Generation Gains Traction *3 weeks ago | Technology* Both US and Chinese AI labs are increasingly using synthetic data generation to supplement real-world training data, potentially reducing the importance of raw data access over time. --- # AI Applications & Products The application layer is where AI capabilities are translated into products and services. US companies lead in global enterprise AI deployment, while China dominates in domestic consumer AI applications. **Measure:** Commercial AI Product Deployment & Market Reach **US & Allies Score:** 8.0/10 **China Score:** 6.0/10 **Leader:** US & Allies ## Key Metrics - **Commercial reach:** Reach = (Active users) × (Revenue per user) × (Markets served) - **Product velocity:** Velocity = (AI features shipped per quarter) × (User adoption rate) × (Retention) ## What matters in this layer The application layer determines who captures the economic value of AI. Competitive advantage comes from distribution, user feedback loops, domain expertise, and the ability to ship AI-powered features rapidly while maintaining reliability and trust. ### Enterprise AI integration US tech giants are embedding AI deeply into enterprise workflows: coding assistants, document analysis, customer service automation, and business intelligence. This creates sticky adoption and recurring revenue. ### Consumer AI ecosystems China’s AI applications are deeply woven into daily life for over a billion users through super-apps, e-commerce, content platforms, and smart city infrastructure, generating massive data feedback loops. ### Autonomous systems Self-driving vehicles, robotics, and autonomous agents represent the next frontier of AI applications. Both the US and China are investing heavily, with different regulatory approaches shaping deployment speed. ### Global distribution US companies have significant advantages in global distribution through cloud platforms and existing enterprise relationships, while Chinese apps are gaining traction in Southeast Asia and other emerging markets. ## Recent Developments ### Enterprise AI Adoption Accelerates in the US *3 days ago | Enterprise* US enterprise AI adoption continues to accelerate, with companies like Microsoft, Google, and Salesforce embedding AI capabilities across their product suites. Enterprise spending on AI tools is projected to exceed $100 billion annually. ### Chinese AI Apps Dominate Domestic Market *1 week ago | Consumer* Chinese AI applications, particularly in e-commerce, content recommendation, and autonomous driving, are deeply integrated into daily life for over a billion users, creating massive feedback loops for model improvement. ### AI Coding Assistants Transform Software Development *2 weeks ago | Developer Tools* AI-powered coding assistants from US companies are fundamentally changing software development workflows, with adoption rates exceeding 70% among professional developers. --- # Talent The researchers, engineers, and operators who design, build, and run every layer of the AI stack. Talent is arguably the scarcest resource in the entire supply chain—the global pool of people capable of pushing the frontier numbers in the low thousands, concentrated at a handful of labs. **Measure:** AI Research and Engineering Talent Pool **US & Allies Score:** 7.0/10 **China Score:** 5.0/10 **Leader:** US & Allies ## Key Metrics - **Frontier researcher pool:** Pool = (PhD-level ML researchers) × (Publications at top venues) × (Lab affiliation concentration) - **Engineering capacity:** Capacity = (Infrastructure engineers + Data scientists + Safety researchers) × (Retention rate) ## What matters in this layer Talent is arguably the scarcest resource in the entire supply chain. The global pool of people capable of pushing the frontier numbers in the low thousands, concentrated at a handful of labs. Researchers, engineers, and operators design, build, and run every layer of the AI stack. ### Concentration at frontier labs A small pool of elite ML researchers is highly concentrated at Anthropic, OpenAI, Google DeepMind, Meta, xAI, and a few other labs. They design architectures, run training, and push the capability frontier. ### Infrastructure engineering Platform engineering requires specialized talent to build and maintain distributed training infrastructure, optimize GPU utilization, and develop ML toolchains. Operating AI-scale data centers requires expertise spanning electrical engineering, cooling, networking, and cluster management. ### Safety and alignment expertise Safety science requires specialized expertise in alignment research, interpretability, red-teaming, and evaluation methodology—a nascent discipline with very few experienced practitioners worldwide. ### Global competition for talent Both the US and China are actively competing to attract and retain top AI talent. Immigration policy, research funding, compensation, and lab culture all factor into where the best researchers choose to work. ## Recent Developments ### Frontier Lab Researcher Concentration *1 week ago | Research* A small pool of elite ML researchers is highly concentrated at Anthropic, OpenAI, Google DeepMind, Meta, and xAI. These individuals design architectures, run training, and push the capability frontier. ### China Expands AI Talent Pipeline *2 weeks ago | Education* China is producing more STEM PhDs than any other country and has made significant investments in AI education and research. However, many top Chinese researchers continue to work at US labs. ### Immigration Policy Shapes AI Talent Flows *3 weeks ago | Policy* US immigration policy remains a critical variable in AI talent competition. Visa backlogs and policy uncertainty have pushed some researchers to consider labs in the UK, Canada, and other countries. ---