China is winning the AI deployment race while Europe debates the rules

The most useful way to understand artificial intelligence in China in 2026 is to stop counting frontier models and start counting installations. The country that two years ago was widely described as a fast follower now runs the largest, densest deployment of AI systems on earth, across hospitals, factory floors, ride-hailing fleets, school classrooms, payment apps, and street-level surveillance. The interesting question is no longer whether China can build a competitive large language model. It can, and it has built several. The question is what happens when a state and a private sector both treat AI as infrastructure to be pushed into every sector of the economy at once.

China’s AI story is now about deployment, not just models

That distinction matters because the Western conversation about AI has fixated on the frontier: which lab has the smartest model, who reaches artificial general intelligence first, how large the next training run will be. China is running a different race. Its State Council “AI Plus” initiative, approved in August 2025, sets explicit penetration targets rather than capability milestones: AI adoption across key sectors is meant to reach 70 percent by 2027 and 90 percent by 2030, with the stated goal of a “fully intelligent economy and society” by 2035. These are diffusion targets, and they reveal where Beijing thinks the advantage lies.

The numbers behind that bet are already substantial. China’s domestic AI market is valued at close to 600 billion yuan, supported by an ecosystem of more than 4,500 enterprises spanning chips, algorithms, data, platforms, and applications. One industry survey found that about 34 percent of job functions at Chinese companies are already fully integrated with AI tools, compared with 30 percent globally. The Brookings Institution has described China as running “multiple AI races” at once, with the deployment race further advanced than in any other country.

Europe, by contrast, has spent the same period building a legal architecture for AI before building much of the AI itself. The EU AI Act, the world’s most comprehensive horizontal AI law, entered its phased enforcement in 2024 and 2025. The continent has world-class research and a handful of genuinely strong companies, but its share of global AI investment and its deployment depth both lag far behind. This article looks at where AI is actually used across China, how deep that use runs sector by sector, and what an honest comparison with Europe shows. The answer is not that China is uniformly ahead. It is ahead in the dimension it has chosen to compete on, and that choice has consequences Europe is only starting to confront.

The national plan driving the build-out

China did not arrive at this position by accident. The current build-out is the product of a planning sequence that goes back almost a decade. In 2017, the State Council issued the New Generation AI Development Plan, which named the goal that still anchors policy today: becoming the world’s primary AI innovation center by 2030. That document also proposed, unusually early, a system for AI security supervision and assessment, which is why China’s regulatory apparatus matured alongside its commercial sector rather than after it.

The 2025 “AI Plus” plan, formally the Opinions on Deeply Implementing the Artificial Intelligence Plus Action, deliberately echoes the “Internet Plus” strategy of 2015. The logic is the same: take an enabling technology and push it into every offline activity until the activity itself is restructured around it. Internet Plus produced Meituan, Didi, and livestream commerce. AI Plus is intended to do the same thing one layer up, embedding model-driven automation into manufacturing, public services, governance, science, and consumer life. The plan prioritizes six domains: science and technology, industrial development, consumer services, public welfare, governance, and international cooperation.

Money follows the plan. In January 2025, China launched an $8.2 billion National AI Industry Investment Fund. The broader $138 billion National Venture Capital Guidance Fund targets AI-adjacent fields including robotics and what Chinese policy documents call “embodied intelligence.” Local governments in Hangzhou, Beijing, Shanghai, and Shenzhen have layered their own state funds and procurement programs on top. The RAND Corporation, assessing this approach in 2025, concluded that while there will be significant waste, the funding will support a growing start-up ecosystem, particularly at the application layer where China’s strengths are clearest.

The reach of state coordination is the structural feature Europe cannot replicate. Beijing can direct public procurement toward domestic AI hardware, create captive demand for chips, and order entire categories of institutions to adopt AI on a fixed timeline. When the National Health Commission mandated that all tertiary hospitals integrate AI-assisted diagnosis by 2025, that was not a suggestion. When the Ministry of Education declared 2025 the first year of “smart education,” provinces moved. This top-down capacity to manufacture demand is the engine behind the deployment numbers, and it is the single largest difference between how AI spreads in China and how it spreads in market democracies.

The targets attached to AI Plus give the strategy a clarity that slogans usually lack. The State Council, approving the initiative in August 2025, set out to reach roughly 70 percent AI adoption across major sectors by 2027, 90 percent by 2030, and what it called a fully intelligent economy and society by 2035. By the time the plan was formalized, Chinese assessments suggested the country had already covered a large share of the distance toward the 2027 figure, which is why the document reads less like a wish list and more like a codification of a build-out already under way. Setting an explicit adoption percentage as a national objective is itself revealing, because it frames AI not as a frontier to be discovered but as an input to be distributed, the same way an earlier generation of planners treated electrification or rural broadband.

The scale of the ecosystem the policy is feeding is now substantial. China is home to something on the order of 4,500 AI enterprises, and the domestic core AI industry has been valued at around 600 billion yuan, with forecasts that put the broader market on a path from roughly 21 billion dollars in 2024 toward 200 billion dollars by the early 2030s. Those figures sit below American private investment in absolute terms, but they understate the effect, because Chinese state spending, procurement, and local-government funds operate outside the venture-capital channel that dominates Western accounting. The money that matters most in China is not always the money that shows up in a startup-funding league table, and that mismatch is one reason outside observers repeatedly underestimated how quickly deployment would scale.

The model layer powering Chinese AI

Underneath every deployment sits a layer of foundation models, and this is where China’s progress has been most visible to the outside world. The turning point was January 2025, when Hangzhou-based DeepSeek released its R1 reasoning model and its chatbot briefly overtook ChatGPT as the most downloaded free app in the United States. What made R1 a shock was not only its quality but its reported cost. The model was trained for roughly six million dollars using a few thousand export-compliant chips, a fraction of the hundred-million-dollar-plus budgets behind comparable Western systems. DeepSeek demonstrated that frontier-adjacent capability could be reached through architectural efficiency rather than brute-force compute, and it did so with open weights that anyone could download and run.

DeepSeek is not alone, and that breadth is the point. Alibaba’s Qwen family, integrated with Alibaba Cloud and the company’s commerce and productivity products, has become one of the most widely used open-weight model families in the world, supporting more than a hundred languages and tool-calling for autonomous agents. Zhipu AI’s GLM line, Moonshot’s Kimi, ByteDance’s Doubao, Tencent’s Hunyuan, and Baidu’s Ernie round out an unusually crowded field. By early 2026, independent benchmark trackers placed several Chinese models among the strongest open-weight systems available, with GLM-5 and Qwen variants leading the open category and DeepSeek’s V4 release matching the best open-source options in agentic coding while trailing only top closed models in general knowledge.

The strategic choice running through all of this is openness. Where the leading American labs largely keep their best models closed, Chinese labs have leaned into open weights. The effect has been to spread Chinese models through the global developer base. Chinese open-source large language model downloads on Hugging Face grew roughly 340 percent year-over-year, and by mid-2025 Chinese models held six of the top ten trending spots on the platform. One Andreessen Horowitz partner estimated that around 80 percent of US startups building derivative AI products were using Chinese base models. Premier Li Qiang framed this at the 2025 World Economic Forum as China being “willing to share indigenous technologies with the world.” The framing is generous, but the leverage is real: a model that becomes the default substrate for thousands of downstream products embeds influence in a way no closed system can.

There are limits worth stating plainly. The absolute frontier, by most measures, still belongs to closed models from OpenAI, Anthropic, and Google, and the best Chinese open-weight row typically trails the top proprietary systems by several points on aggregate benchmarks. Chinese models also carry hard-coded restrictions on politically sensitive subjects, which constrains them for some uses and reflects the content-governance rules under which they are released. But the gap has closed faster than most 2024 forecasts predicted, and for the overwhelming majority of commercial tasks, the difference between the frontier and a strong open Chinese model is no longer decision-relevant. That is precisely the condition under which mass deployment becomes economically rational, and it is why the model layer and the deployment layer reinforce each other.

Factory floors became the first proving ground

Manufacturing is where China’s AI ambitions are least abstract. The country installs the majority of the world’s industrial robots each year, and the figure for 2025 was around 276,000 units, well over half of the global total. On top of that base, AI has been layered through computer vision for quality inspection, predictive maintenance that flags failing equipment before it stops a line, digital twins that simulate production before it happens, and generative systems that help with design and scheduling. One widely cited industry estimate put manufacturing AI adoption in China at roughly 67 percent, the highest sector penetration of any major economy.

The policy scaffolding is specific rather than aspirational. Shanghai’s “AI plus manufacturing” roadmap, covering 2025 to 2027, sets out to integrate AI solutions across 3,000 manufacturing enterprises, develop ten industry benchmark models, create a hundred benchmark smart products, and establish around ten model “AI plus manufacturing” factories. That kind of numbered target, attached to a city government with the means to fund and enforce it, is what turns a national slogan into installed capacity. The World Economic Forum, surveying China’s industrial AI in early 2025, pointed to digital twins, predictive maintenance, and generative AI as the technologies reshaping manufacturing, transport, retail, energy, and healthcare in parallel.

What makes the manufacturing story coherent is the supply chain underneath it. The Yangtze River Delta and the Shenzhen hardware cluster give Chinese AI deployment something European and even American projects often lack: the ability to prototype, fabricate, and iterate physical systems within a two-hour logistics radius. A robotic arm design that takes months to source and build elsewhere can be turned around in days in Shenzhen. This vertical integration is the unglamorous foundation of China’s edge in applied, embodied AI, and it explains why the next frontier, humanoid robots, emerged from the same industrial base rather than from a software lab.

The density numbers underline how far ahead the base already is. China crossed the threshold of more than 470 industrial robots per 10,000 manufacturing workers, overtaking Germany and Japan and trailing only South Korea, and it did so while still installing more new units each year than the rest of the world combined. The robots are no longer confined to automotive lines. Electronics, metals, plastics, food processing, and increasingly textiles and pharmaceuticals have absorbed them, which spreads the AI layer that sits on top of the hardware across the whole industrial economy rather than a few flagship plants. Penetration that broad is what turns automation from a productivity story into a structural cost advantage, because the savings compound across every link of a supply chain rather than at a single stage.

The most striking expression of this is the spread of so-called dark factories, plants that run with the lights off because almost no humans are on the floor. Xiaomi’s smartphone facility and several appliance and battery plants have moved toward near-total automation, with AI scheduling production, vision systems catching defects, and autonomous vehicles moving parts between stations. These are still a minority of Chinese factories, and the marketing around them outruns the reality in places, but the direction is unambiguous. China is using AI not to assist factory workers at the margin but to redesign the factory around the assumption that intelligence, not labor, is the scarce input worth optimizing for.

The humanoid robot push from demo to production line

For most of 2024, humanoid robots were a stage act, useful for demonstrations and viral clips but absent from real work. In 2025 that changed, and it changed mostly in China. According to Counterpoint Research, around 16,000 humanoid robots were installed worldwide in 2025, and Chinese firms accounted for more than 80 percent of them. Shanghai-based AgiBot led the global market with roughly 30 percent of installations, followed by Hangzhou’s Unitree at around 27 percent. UBTech and Shenzhen’s Leju each held about 5 percent. The top five suppliers, most of them Chinese, took roughly 73 percent of the global total.

The deployments are early and concentrated in controlled settings, mostly data collection, research, logistics, manufacturing, and automotive work, but they are real production rather than prototypes. AgiBot reported shipping more than 5,000 units of its X2 and G2 robots from a Shanghai factory across hospitality, entertainment, manufacturing, and logistics. UBTech’s Walker series is already working on automotive assembly lines. The demand side reads like a roster of China’s industrial champions: BYD, the world’s largest electric-vehicle maker, battery giant CATL, and contract manufacturer Foxconn together represent appetite for tens of thousands of systems.

Policy and supply chain are again doing the heavy lifting. The 2025 Humanoid Robot Action Plan, issued by the Ministry of Industry and Information Technology with five other ministries, targets 100,000 deployed humanoids by 2027. By late 2025, the National Development and Reform Commission counted more than 150 humanoid robot companies in China, with the sector growing at over 50 percent a year and a projected market scale near 100 billion yuan. Unitree manufactures its own motors, reducers, and sensors in-house, which is how it has driven unit costs down far enough to make commercialization plausible. Industry capacity targets for Unitree and AgiBot alone in 2026 exceed 75,000 units a year, more than the entire Western supply combined.

The leaders themselves are candid about the gap that remains. Unitree founder Wang Xingxing compared the current state of the field to the period one to three years before ChatGPT: the direction is clear, but the breakthrough has not arrived. He defined that breakthrough precisely, saying a true “ChatGPT moment” will come when a robot dropped into an unfamiliar real-world setting can complete about 80 percent of tasks just by following voice or text instructions. The hardware is racing ahead of the general-purpose intelligence needed to make it broadly useful. China’s advantage here is manufacturing scale and cost, not yet cognitive capability, and whether that scale converts into a durable lead depends on solving the software problem that no one, in any country, has solved yet.

The structural bet behind the humanoid push is the one running through every Chinese AI deployment: build the physical capacity first and let the intelligence catch up, on the assumption that whoever already makes the bodies will be best placed to deploy the brains once they arrive. Europe has serious robotics expertise, in Germany’s industrial-automation firms above all, but no humanoid effort at anything close to this scale and no comparable supply chain feeding it, which means that if the software breakthrough Wang describes does arrive, the manufacturing base ready to absorb it will sit overwhelmingly in China. That is the logic that makes the demo-to-production-line shift matter beyond the novelty of walking machines, and it is why the humanoid race is being watched as a leading indicator of where embodied AI will be deployed first.

Robotaxis running at full city scale

Autonomous ride-hailing is the clearest case of a Chinese AI deployment operating at genuine commercial scale on public roads. Baidu’s Apollo Go is the central example. By late 2025 the service was completing more than 250,000 fully driverless rides a week, a figure on par with what Alphabet’s Waymo reported in the United States, and it had passed 17 million cumulative rides. Apollo Go operates across 22 cities, including Beijing, Shanghai, Wuhan, Shenzhen, and Hong Kong, with international testing in Dubai and Abu Dhabi and partnerships to expand into Europe and beyond.

The economics are what make the Chinese effort distinctive. Apollo Go’s sixth-generation vehicle, the Apollo RT6, reportedly costs under 30,000 dollars, with the seventh generation expected to come in under 20,000 dollars. That is a fraction of the cost of the sensor-laden vehicles many Western operators run, and it is the reason Baidu can talk credibly about profitability. The Wuhan operation is near break-even, and Baidu founder Robin Li has projected that by 2030 the operating cost of a robotaxi ride in the United States could fall to around 25 cents per mile, unlocking a five-to-sevenfold surge in ride-hailing demand.

Safety data, which is the gating factor for any expansion, has so far supported the case. Baidu reports that Apollo Go’s accident rate is roughly one-fourteenth that of human drivers, and that the fleet has logged one airbag-deployment incident for every 10.1 million kilometers driven, with no major accident involving human injury or death. Apollo Go has racked up well over 120 million driverless miles, which is a meaningful evidence base by the standards of the industry.

The expansion is not frictionless, and that matters for an honest account. In April 2026, after a widespread malfunction in Wuhan left passengers stranded and disrupted traffic, Chinese authorities reportedly suspended the issuance of new autonomous driving permits, at least the second such pause. The episode is a useful corrective to the narrative of unstoppable rollout. China deploys faster than anyone, but it also runs into the same reliability and trust problems everyone else does, and its regulators have shown they will hit the brakes. The difference is that the baseline scale is already enormous, so even a slowdown leaves Chinese robotaxis operating at a volume European cities are nowhere near approaching. The contrast with Europe, where driverless commercial services barely exist, is stark, and Baidu’s Lyft partnership to launch in Germany and the United Kingdom in 2026 means Chinese autonomy may reach European streets before any homegrown European service does.

Surveillance remains the most mature application

If deployment scale is the measure, the most fully realized AI application in China is surveillance. Cameras with facial recognition and object detection are routine on urban streets, and the systems behind them have moved well beyond passive recording. They feed into integrated “city brain” platforms that fuse video, sensor data, and records to track events in real time. A December 2025 report covered by CNN described AI being used to predict public demonstrations, monitor the moods of prison inmates, and recommend, in at least one Shanghai system, whether prosecutors should arrest or seek suspended sentences for suspects. China’s Supreme Court has urged all courts to develop competent AI systems for use in trials and administrative work.

The commercial ecosystem behind this is dense and specialized. Companies such as SenseTime, Megvii (Face++), Yitu, and CloudWalk build the software and hardware that government agencies, banks, and transport operators rely on. The state is the dominant buyer, and that procurement both funds the firms and gives Beijing leverage over the sector’s direction. China’s surveillance AI is the rare domain where the country is not catching up to anyone; it set the global standard for scale and integration, and it exports the underlying technology to other governments, with documented adoption in countries including Iran and Saudi Arabia.

The regulatory picture is more nuanced than the surveillance narrative usually allows. On June 1, 2025, China’s first dedicated rules on facial recognition took effect, jointly issued by the Cyberspace Administration and the Ministry of Public Security. The Measures require meaningful consent, restrict mandatory facial scans for routine services, and demand encryption, access controls, and auditing. They were prompted in part by public backlash, including an incident in Shanghai where hotels mandated facial scans. These rules constrain commercial and private use far more than they constrain the state, which retains broad latitude for public-security applications, and the technology remains linked to social-management systems. Still, the existence of a privacy framework complicates the simple image of an unregulated surveillance state and signals that even Beijing sees reputational and social costs in unchecked biometric collection.

For a European reader, the surveillance case is the sharpest illustration of why “ahead” is the wrong frame without qualification. China is technologically ahead in deploying population-scale biometric AI, but that lead is partly a function of a political system that tolerates uses Europe has chosen to prohibit. The EU AI Act classifies most real-time remote biometric identification in public spaces as prohibited or high-risk precisely to prevent this category of deployment. The gap here is not a measure of European technical weakness. It is a measure of different choices about the relationship between the state and the individual, and reading it purely as a technology race misses the point.

Hospitals adopted AI faster than regulators could frame it

Healthcare shows how quickly a deployment can scale in China once a capable open model meets a national mandate. After DeepSeek released V3 in December 2024 and R1 in January 2025, hospitals could run a high-performing model locally, behind their own firewalls, without sending patient data to an external provider. The response was immediate. Within roughly three months, DeepSeek had been deployed across dozens of tertiary hospitals, beginning in Shanghai and spreading nationwide. By early 2025, reporting documented deployment across nearly 90 tertiary hospitals, with the National Health Commission having mandated that all tertiary hospitals integrate AI-assisted diagnosis by 2025.

The applications are concrete. Shanghai’s Ruijin Hospital, working with Huawei, launched a pathology AI model called Ruizhi that automates slide analysis with a daily capacity of around 3,000 slides. Huashan Hospital tested DeepSeek’s full model on an internal network to keep data secure. Shanghai Fourth People’s Hospital built a localized deployment integrating a knowledge base of more than 30,000 typical cases and regional treatment guidelines. Across these sites, AI is being used for imaging analysis, pathology, clinical decision support, and workflow automation that reduces the documentation burden on clinicians. A Stanford evaluation of clinical AI models ranked DeepSeek R1 first among nine leading systems, with a win rate of 66 percent.

The risks are being run in real time, and Chinese researchers are unusually frank about them. A 2025 study in Frontiers in Public Health flagged three categories of medical liability risk: product liability, diagnosis-and-treatment damage liability, and medical ethics liability. The central concern is automation bias, the tendency of overworked clinicians to defer to an AI recommendation, combined with the absence of a clear liability framework for when an AI-assisted diagnosis goes wrong. China’s smart-healthcare market has been growing at a compound annual rate above 50 percent, but the legal scaffolding has not kept pace, and academics have proposed solutions such as liability insurance pools jointly funded by AI vendors and hospitals.

The healthcare case captures the broader pattern in miniature. A capable, cheap, locally deployable model appeared. A national mandate created instant demand. Hospitals adopted at a speed that would be unthinkable in a European system bound by medical-device regulation, GDPR, and cautious procurement. The benefits, in diagnostic support and efficiency, are real and arriving now. The risks, in accountability and over-reliance, are also real and being deferred. Whether that trade is wise is a separate question from whether it is fast, and on speed there is no contest.

Payments and commerce folded AI into the super-app

China’s consumer internet was already the most concentrated and mobile-first in the world before generative AI arrived, and that gave its platforms a distribution advantage no Western company can match. WeChat, with over a billion users, already combined messaging, payments, commerce, and services. Alipay sat at the center of daily transactions. When large language models matured, the platforms did not need to acquire users or teach new habits. They needed only to fold AI into surfaces hundreds of millions of people already touched every day.

The result is what Chinese firms now call agentic commerce. In January 2026, Alibaba upgraded its Qwen chatbot so users could complete transactions inside the interface, comparing tailored recommendations from Taobao and the travel platform Fliggy and paying through Alipay without leaving the chat. ByteDance pushed its Doubao assistant to handle ticket bookings autonomously through Douyin’s commerce features. Tencent embedded its Hunyuan models directly into WeChat and Tencent Cloud rather than launching standalone products, turning model advances into assistants inside chats, meetings, and documents. Ant Group has rolled out AI-powered payment systems across its network.

The financial-services sector, broadly defined, holds the largest share of China’s AI market, driven by early adoption for fraud detection, algorithmic trading, customer-service automation, and risk management. E-commerce and payments have reportedly reached effectively full AI deployment, with adoption widespread since early 2025. Mobile-first commerce gives China an implementation rate, around 84 percent, that leads the world. The capital behind this is significant: Alibaba alone pledged over 50 billion dollars for cloud and AI development, and spent 123 billion yuan, roughly 17 billion dollars, on capital expenditure in 2025.

AI deployment across major Chinese sectors in 2025 and early 2026

The table shows breadth rather than depth at the frontier, and that breadth is the story. China is not winning every sector by capability, but it is deploying across all of them at once, which is a different kind of advantage.

These platforms also expose a real constraint. The same data scale that powers Chinese commerce AI comes with content-governance rules and political sensitivities that shape what the assistants can say and do, and monetization pressure has pushed some firms, including Alibaba and Zhipu, to release their newest models in closed form and to raise prices. The era of uniformly free, fully open Chinese models is already giving way to something more commercially conventional. Even so, the integration of AI into the everyday transaction layer of an entire economy is further along in China than anywhere else, and it is happening at a velocity that Western platform regulation, by design, slows down.

The hardware and energy base under the surface

Every claim about China’s AI lead has to be tested against one stubborn constraint: compute. AI runs on advanced semiconductors, and this is the dimension where China is genuinely behind, not ahead. Stanford’s 2025 AI Index estimated that the United States holds roughly nine times China’s AI compute capacity, and US export controls have been tightening that gap rather than closing it. Since late 2022, Washington has restricted the sale of Nvidia’s most capable chips to China, repeatedly revising the rules as Nvidia designed compliant variants like the H20, which itself was later subjected to licensing requirements.

China’s answer is Huawei’s Ascend line. The 910C is the current workhorse, used for inference by DeepSeek and other labs, and Huawei has plans to roughly double output toward 600,000 units, with the broader Ascend line targeting as many as 1.6 million dies in 2026. A more powerful 910D, aimed at matching Nvidia’s H100, has been in testing. On paper this looks like emerging self-sufficiency. The reality is more constrained. Huawei’s process technology remains stuck at around 7 nanometers because its foundry partner SMIC cannot access the most advanced lithography equipment, and China still depends on stockpiles of foreign high-bandwidth memory it cannot yet produce domestically at scale.

Revealingly, Chinese AI developers themselves have preferred Nvidia chips even in performance-degraded form. In 2024, Chinese companies bought around a million Nvidia H20 chips against an estimated 450,000 Huawei Ascend 910B units, and only a handful of state-backed firms such as iFlytek, SenseTime, and China Mobile used Huawei silicon for training. Developers resist switching because Nvidia’s degraded chips still outperform Huawei’s in important respects, and because Nvidia’s CUDA software ecosystem is hard to leave. Huawei is also a competitor to many of its potential customers, which adds commercial friction.

This is where DeepSeek’s significance extends beyond its models. Its real achievement was making architectural innovations below the level of CUDA, squeezing more capability out of constrained hardware. If that efficiency work were turned toward strengthening Huawei’s Ascend chips and its CANN software stack, it would pose a far more serious challenge to Nvidia’s dominance. For now, the picture is a paradox: China leads the world in deploying AI applications while remaining dependent on a hardware base it does not fully control. Export controls have not stopped Chinese AI, but they have shaped it, forcing an emphasis on efficiency and inference over the raw-scale training runs that define the American frontier. The long-term question is whether constraint breeds durable advantage or simply a ceiling, and that question is genuinely unresolved.

That hardware constraint rests on top of an energy story, and here the balance tilts back toward China. AI infrastructure is, at bottom, an energy story, and China has approached it with the same build-first instinct it brings to everything else. The country has poured resources into data centers, often powered by renewable energy and connected through its ultra-high-voltage transmission network, which moves electricity efficiently across long distances from generation sites to demand centers. This grid capacity is itself increasingly managed by AI. In March 2025, Hangzhou launched City Brain 3.0, built on the DeepSeek-R1 model, integrating traffic management, city patrolling, and smart-grid coordination for ultra-high-voltage transmission into a single platform.

The energy advantage cuts in China’s favor in a way that is easy to overlook. Training and running large models consumes enormous power, and the cost and availability of that power is becoming a binding constraint on AI everywhere. China’s combination of large-scale renewable build-out, cheap industrial electricity, and a state-directed grid gives it more room to expand compute capacity on the energy side than Europe has, where household and industrial electricity prices in the second half of 2025 sat near record highs and energy is the first and least avoidable limit on any AI infrastructure ambition.

The build-out has not been perfectly disciplined. China overbuilt data-center capacity during the speculative rush of 2024 and 2025, leaving a glut of underused facilities. By late 2025 the government was conducting an assessment to regulate and optimize existing resources and sell off unused compute, and it remained unclear whether those facilities could be repurposed productively. The data-center glut is a reminder that state-directed investment produces waste as well as scale, and that not every gigawatt of planned capacity translates into useful AI work.

For the broader comparison with Europe, energy is one of the clearest structural asymmetries. Europe’s AI ambitions, including the gigafactory plans discussed later, run into an energy-cost wall that China simply does not face to the same degree. A continent that has chosen higher energy prices, for environmental and geopolitical reasons that are defensible on their own terms, has made AI infrastructure more expensive to operate. That is not a technology gap. It is a structural cost difference, and it compounds every other disadvantage on the deployment side.

Schools turned into a national AI delivery system

China has decided that the way to secure an AI future is to build the workforce for it from primary school upward, and it is doing so with the same scale and central coordination it applies to industry. Beijing became the first provincial-level region to make comprehensive AI education compulsory, and in the 2025 to 2026 school year more than 1,400 primary and secondary schools in the capital began providing at least eight class hours of AI instruction per year across all grades, starting with first-graders. The Ministry of Education declared 2025 the inaugural year of China’s smart-education initiative, and an April 2025 State Council and MOE guideline ordered every province to weave AI into teaching methods, textbooks, teacher training, and campus infrastructure by 2030, with interim milestones in 2027.

The infrastructure behind this is already the largest of its kind. China’s national Smart Education portal hosts the world’s largest digital education platform for basic education, with more than 110,000 high-quality resources by one count and around 170,000 courses and AI-curated micro-lessons by another, recently upgraded with an AI search engine, video summarizer, and adaptive features. Curriculum is tiered: foundational concepts and basic coding in elementary grades, data analysis and problem-solving in middle school, and advanced applications, innovation projects, and ethics in high school. Elite schools push further, with classes on embodied intelligence, brain-computer interfaces, and autonomous driving developed in partnership with the Chinese Academy of Sciences and firms like Huawei.

The goals are partly economic and partly social. Beijing is betting that AI tools can both raise the productivity of the next generation of workers and help close China’s substantial educational inequities by distributing high-quality instruction to under-resourced regions. The state sees AI education as a way to manufacture a national capability advantage over time, and the timeline, with universal basic AI access targeted for 2030, is characteristically aggressive.

There is a control dimension that a complete account cannot omit. A November 2025 official teachers’ guide included guidance on using AI to instill “correct morals,” such as building moral-education case libraries and generating ethical situations to shape students’ value analysis and behavior. The same platform that democratizes access also centralizes influence over what is taught and how. For European observers, the contrast is instructive. The EU has its own AI Continent Action Plan emphasizing worker skills, and several member states are integrating AI into classrooms, but nothing approaches the scale, speed, or central direction of the Chinese rollout, and nothing carries the same explicit ideological framing. Whether China’s approach produces a more capable workforce or simply a more uniformly trained one will take a decade to judge, but the investment in human capital is real and it is enormous.

Agriculture and logistics carry the quieter deployments

The headline AI applications attract attention, but some of the most economically consequential deployments are the least visible. Logistics is a prime example. China’s e-commerce volume created the conditions for the world’s most automated warehousing and delivery systems, and AI now optimizes routing, demand forecasting, sorting, and last-mile delivery across networks that move billions of parcels. The same companies building agentic commerce, Alibaba and JD.com among them, run logistics arms where AI-driven optimization is mature rather than experimental. China’s strategic framing, repeated across policy documents, is that AI should be embedded in practical, sector-specific applications rather than concentrated in a single general-purpose chatbot, and logistics is where that philosophy pays off most quietly.

Agriculture is following a similar path. AI-assisted crop monitoring using satellite and drone imagery, predictive systems for pest and disease management, precision irrigation, and yield forecasting are spreading across a sector under pressure from an aging rural workforce and food-security concerns. The low-altitude economy, a category Chinese policy has explicitly promoted, ties agricultural drones to broader ambitions in aerial logistics and monitoring. These deployments rarely make international news, but they are where AI touches the parts of the economy that feed and supply 1.4 billion people, and they reflect the same pattern: scale, state encouragement, and a supply chain capable of producing the necessary hardware cheaply.

Energy management, building automation, and resource planning round out the quieter layer. The thread connecting all of them is that China is not waiting for AI to be perfect before putting it to work on unglamorous problems. A demand-forecasting model that cuts warehouse waste by a few percentage points, deployed across thousands of facilities, produces more aggregate economic value than a marginally smarter chatbot. The diffusion of good-enough AI into ordinary economic activity is precisely the kind of advantage that compounds, and it is the dimension where the gap with a more cautious Europe is widest and least discussed.

The low-altitude economy deserves a closer look, because it shows how China bundles AI deployment with industrial policy in a way Europe has no equivalent for. By designating drones, electric vertical-takeoff aircraft, and aerial logistics as a strategic sector, Beijing created a single policy umbrella that pulls together agricultural spraying, parcel delivery, infrastructure inspection, and emergency response, all of them increasingly coordinated by AI flight-planning and computer-vision systems. Agricultural drones already treat a large share of China’s farmland, and the same airframes and autonomy stacks feed straight into the delivery and inspection markets, so a single manufacturing base serves several applications at once. That bundling is the mechanism that lets a deployment scale faster than any individual use case would justify on its own.

Logistics shows the same compounding from a different angle. The warehouses behind the country’s e-commerce giants run on AI that forecasts demand down to the district level, pre-positions inventory before orders arrive, and routes autonomous vehicles and sorting robots through facilities that handle volumes no human-coordinated system could match during peak shopping festivals. What looks like a retail convenience story is really an industrial-AI story, because the optimization that shaves a percentage point off fulfillment cost across billions of parcels a year is worth more than almost any consumer-facing model, and it is invisible precisely because it works. Europe has capable logistics operators, but none sit inside an e-commerce ecosystem of comparable density, so the data flywheel that improves these systems spins slower.

Civil-military fusion runs through the whole stack

AI in China cannot be cleanly separated into civilian and military categories, because policy explicitly fuses them. The civil-military fusion strategy channels university research and private-sector innovation into defense applications, and AI is central to it. The People’s Liberation Army describes its goal as “intelligentized warfare,” using AI to improve decision-making, battlefield awareness, logistics, and cyber operations. The same models, chips, and talent that power commercial deployment feed into military programs, and the same firms often serve both.

This dual-use character has two consequences worth drawing out. The first is that it complicates any external assessment of China’s AI capability, because the most advanced military applications are not publicly documented and the commercial figures understate the full scope of state investment. The second is that it shapes how other countries respond. US export controls are justified in large part by the difficulty of preventing commercially acquired compute from being repurposed for military ends, and the civil-military fusion doctrine is the reason that argument carries weight.

Europe is not absent from military AI, and the comparison is more even here than elsewhere. Germany’s Helsing has become a serious developer of defense AI, with its Centaur agent autonomously piloting a Saab Gripen fighter on test flights over the Baltic and the company expanding into AI-driven drones and submarines. Its partnership with Mistral, in which Mistral models help Helsing-powered systems make targeting decisions, shows European AI firms moving into exactly the dual-use territory China has institutionalized. The difference is structural rather than technical: China fuses civil and military development by design and at scale, while Europe does so through individual companies and partnerships, constrained by national defense budgets and the absence of a unified European defense procurement system.

The strategic point for the broader argument is that AI capability is now a component of national power in a way that makes the China-Europe comparison about more than economics. A country that can field intelligentized military systems, embed AI across its industrial base, and export surveillance technology to aligned states is accumulating a form of leverage that a continent organized around regulation and individual-rights protection is not built to match. That is not a verdict on which approach is wiser. It is an observation that the two systems are optimizing for different things, and that the deployment-first approach produces capabilities the rules-first approach is designed to constrain.

The mechanism that makes fusion effective is the same supply chain that powers commercial deployment, which is what gives the doctrine its reach. A drone autonomy stack refined for parcel delivery, a vision model trained for industrial inspection, or an efficiency technique developed to stretch constrained chips can each be redirected toward a defense application without starting from scratch, because the underlying components are shared. The civilian deployment boom is therefore not separable from the military build-up; it is the feedstock for it, and that is precisely why export-control regimes treat consumer-grade compute and open models as strategic rather than merely commercial. The same openness that lets Chinese models spread across the Global South also lets capability diffuse in directions the original developers neither intended nor controlled.

For the European reader, the uncomfortable implication is that the continent’s deployment caution carries a security cost as well as an economic one. The industrial base that would let Europe field autonomous defense systems at scale is the same base it has been slower to build on the commercial side, and the talent, compute, and manufacturing depth feed both. A Europe that lags in factory automation and embodied AI is, by the same token, a Europe with a thinner foundation under its defense-AI ambitions, however impressive individual firms like Helsing prove to be. That linkage between commercial deployment depth and hard-power potential is the part of the comparison that turns an economic story into a strategic one.

Beijing’s regulatory model shapes how AI spreads

The standard Western framing treats Chinese AI as essentially unregulated, racing ahead because nothing holds it back. That is wrong, and understanding why matters for the comparison with Europe. China regulates AI heavily, but it regulates differently, in a way calibrated to enable deployment while controlling content and political risk. The 2023 Interim Measures for Generative AI Services set obligations for AI providers, including service registration, model filing with the Cyberspace Administration of China, content governance, and safety checks. Services that influence public opinion must register, and AI tools must display model names and filing numbers.

The defining feature is that Chinese AI regulation prioritizes control over content and social stability rather than protection of individual rights against the state. Models carry hard-coded restrictions on politically sensitive topics, which is a regulatory requirement, not an accident. The June 2025 facial-recognition Measures constrain commercial and private use while leaving state security applications largely intact. This selective stringency is the regulatory expression of the same logic that runs through everything else: enable the uses the state wants, constrain the uses it fears, and move fast on both.

The contrast with the EU AI Act is fundamental. The European law is risk-based and rights-protective, prohibiting some uses outright, such as social scoring and most real-time public biometric identification, and imposing heavy obligations on high-risk systems in areas like law enforcement, critical infrastructure, and education. It is horizontal, applying across all sectors, and it is designed to be predictable. Supporters argue that this predictability accelerates compliant deployment by telling developers exactly what is permitted. Critics, including a June 2025 European Parliament committee motion, argue the opposite: that weak investment and too much regulation are causing the EU to fall further behind.

Both systems are coherent on their own terms. China’s regulation is built to enable a deployment-first economy under firm political control. Europe’s is built to protect citizens and embed values into the technology before it scales. The Chinese model produces faster diffusion and more state power. The European model produces stronger rights guarantees and slower adoption. Calling one “ahead” of the other requires first deciding what the technology is for, and that is a question of politics, not engineering.

Europe measured against the same yardstick

Holding Europe up against the same metrics produces an uncomfortable picture, and it is worth stating without softening. On investment, the gap is severe. In 2021 the European Union accounted for only about 7 percent of global AI investment, against 40 percent for the United States and 32 percent for China. The trend has not reversed. Europe attracted roughly 8 billion dollars in private AI investment in 2025, which placed it second globally and, by some measures, ahead of China’s 5 billion in private funding that year, but a distant second to the United States, which pulled in around 109 billion. The absolute scale of the American lead reframes the China-Europe contest as a competition for second place rather than first.

On models, Stanford’s 2025 AI Index counted 40 notable AI models from US-based institutions in 2024, against 15 from China and just three from Europe. On compute, the same index put the United States at nine times China’s capacity and 17 times Europe’s. These figures describe a self-reinforcing cycle: more compute produces better models, which attract more talent and investment, which buys more compute. Europe sits at the bottom of all three rankings, and the structural reasons, fragmented capital markets, limited access to equity, and high implementation costs for the small and medium enterprises that make up much of the economy, are deep rather than incidental.

There is a sharper way to put the comparison that matters for this article’s question. On the frontier-model axis, both China and Europe trail the United States, but China trails by less and is closing faster. On the deployment axis, China is comprehensively ahead of Europe, embedding AI across sectors at a depth Europe has not approached. So when the question is whether China is technologically ahead of Europe, the honest answer in 2026 is: yes, in most of the dimensions that can be measured, and decisively so in deployment, while the two are more comparable in pockets of frontier research and specific verticals.

The caveat that keeps this from being a simple verdict is that “ahead” depends on the yardstick. Europe is not trying to win a deployment race or a compute race. Its strategy, examined next, is built on the premise that trust, safety, and rights protection are themselves competitive assets, and that a continent does not need hyperscale parity to remain a relevant actor. That premise is contestable, and the investment figures suggest it is being tested under pressure, but it is a deliberate position rather than simple failure. The numbers show China ahead on the metrics China chose to compete on. They do not settle whether Europe’s different choices will look wise or naive in a decade.

Europe’s approach rests on a different foundation

Europe’s AI strategy is organized around a conviction that the others do not share: that the way to compete is to make AI trustworthy by law first and capable second. The EU AI Act is the clearest expression of this, the world’s most comprehensive horizontal AI framework, structured by risk tier. It requires transparency in chatbot interactions, imposes strict oversight on high-risk uses in law enforcement, critical infrastructure, and education, and prohibits the most dangerous applications outright. The phased enforcement that began in 2024 and 2025 is meant to give the market predictability about what is and is not allowed.

The argument for this approach is more than defensive. Trust is a genuine variable in adoption, and Europe has more of a trust deficit to manage than China does. Surveys show that only around 36 percent of people in the Netherlands see AI as beneficial, with similar skepticism elsewhere, rooted in data-privacy concerns and a historical wariness of unchecked technological power. A regulatory regime that guarantees rights and limits the worst abuses may unlock adoption among users and institutions that would otherwise resist, and it may produce AI systems that are more durable in democratic societies precisely because they were built within constraints. One framing calls this “regulation as architecture” rather than mere oversight.

Europe is also investing, not only regulating, though the scale is modest against the competition. The AI Continent Action Plan, the InvestAI initiative aiming to mobilize 200 billion euros including 20 billion for AI gigafactories, the planned Cloud and AI Development Act intended to at least triple EU data-center capacity within five to seven years, and AI-literacy requirements across institutions together amount to a serious if belated industrial program. The European Chips Act and the IPCEI framework, which assembled 21.8 billion euros across 14 member states for microelectronics, show that coordinated investment is legally and politically possible.

The weakness in the European model is not the philosophy but the execution gap between ambition and capacity. Mobilizing 200 billion euros is a real commitment, but it is being deployed into a race where competitors spend by an order of magnitude more, and it must overcome a fragmented internal market, 27 separate procurement systems, and an energy-cost disadvantage that makes every data center more expensive to run. The bet is that Europe can be a relevant AI actor without winning the compute race, by preserving sovereignty over sensitive workloads and key chokepoints rather than chasing parity. Whether that bet pays off is the central open question of European technology policy, and the deployment evidence from China is the strongest argument that it might not be enough.

The compute and energy gap Europe cannot wish away

The single hardest problem for European AI is the same one China faces, only worse: access to compute, and the energy to run it. The Bruegel think tank, analyzing Europe’s position in 2026, made the structural point bluntly. Unlike China, the EU has no coordinated mechanism to direct public procurement toward European AI hardware or to generate the revenues that fund domestic chip development. European demand is fragmented across 27 member states, each acting largely alone, and policymakers have limited leverage over the private-sector demand that powers a hardware-improvement cycle. There is no realistic prospect of Europe switching to its own advanced AI-chip production in the short to medium term.

China, for all its own chip constraints, has something Europe lacks entirely: the ability to manufacture demand. Beijing can order state-owned enterprises and public institutions to buy domestic chips, creating the captive market that funds the next generation. Europe’s most plausible path is not chip independence but a two-track strategy, continuing to buy from Nvidia and other US suppliers while building selective domestic capacity, the same approach Bruegel observed China using. The Important Projects of Common European Interest framework and the European Chips Act give Brussels tools that did not exist when Airbus was created, but assembling and directing that capacity across a fragmented union is slow.

Energy is the constraint underneath the constraint. Europe’s AI gigafactory plans envision facilities with more than 100,000 advanced chips each, drawing enormous amounts of power, in a market where second-half 2025 household electricity prices sat near record highs around 0.29 euros per kilowatt-hour. A continent that has chosen higher energy costs for sound environmental and security reasons has also chosen to make AI infrastructure more expensive to operate than it is in China or much of the United States. One critical assessment framed the gigafactory ambition as Europe deciding to socialize a hugely expensive arms race in a market where others already outspend it by an order of magnitude.

The realistic European position, argued by analysts who are sympathetic rather than dismissive, is that Europe does not need to win the compute race the way Washington or Beijing do. It needs to ensure compute can be acquired on acceptable terms, that sensitive workloads can run in trusted sovereign environments when necessary, and that upstream European chokepoints, the lithography expertise embodied in ASML above all, are preserved and upgraded rather than allowed to decay. That is a defensible strategy of strategic relevance rather than parity. But it is a strategy that concedes the deployment race to China at the outset, and the question is whether a continent can stay relevant in AI while ceding the ground where AI actually meets the economy.

European companies punching above the continent’s weight

The aggregate picture is bleak for Europe, but it would be a mistake to conclude that the continent has no AI capability. It has a cluster of genuinely strong companies, and they cluster, tellingly, in the niches that suit Europe’s constraints: efficiency, specialization, defense, and science rather than hyperscale general-purpose models.

Paris-based Mistral is the standout, having built competitive models with far fewer computing resources than its American rivals, which is precisely the kind of efficiency-led approach Europe’s compute disadvantage forces. Mistral has become Europe’s clearest answer to the question of whether the continent can produce a frontier-adjacent lab, and its survival and growth matter symbolically beyond its market share. Germany’s DeepL dominates a valuable translation niche. Black Forest Labs leads in image generation. The UK’s Wayve specializes in AI for autonomous vehicles. Germany’s Helsing, discussed earlier, has built a serious defense-AI business that now extends to autonomous drones and submarines.

Europe’s research strength is real and occasionally world-leading. Google DeepMind, headquartered in London, produced AlphaFold, whose protein-structure prediction earned Demis Hassabis and John Jumper a share of the 2024 Nobel Prize in Chemistry, a contribution to drug discovery and biology that no Chinese lab has matched. This points to where Europe’s comparative advantage genuinely lies: scientific AI, where deep domain expertise and academic excellence matter more than raw compute scale, and where the payoff is measured in breakthroughs rather than deployment volume.

The pattern these companies form is the realistic shape of European AI: not a broad deployment economy like China’s, but a set of sharp specialists occupying defensible niches. The strategic case for Europe is that a portfolio of leaders in efficiency, translation, defense, autonomous driving, and scientific AI can sustain relevance and economic value without competing across the board. The risk is that niches are vulnerable. A specialist in translation or image generation can be overtaken by a general-purpose model that absorbs the capability as a feature, and several of these firms depend on American compute and, increasingly, Chinese open-weight base models. Europe’s champions are impressive, but they are islands in an ecosystem that the aggregate numbers show is being outbuilt on both sides.

The deployment gap matters more than the model gap

The most important finding in any honest China-Europe comparison is that the gap that matters is not the one most commentary focuses on. The model gap, who has the smartest system, is real but narrowing and arguably less consequential than assumed, because good-enough models are now cheap and abundant. The deployment gap, how deeply AI is woven into the actual economy, is where the difference is decisive, and it is where China is furthest ahead of Europe.

The evidence is consistent across sources. Roughly 34 percent of job functions at Chinese companies are already fully integrated with AI tools, against about 30 percent globally, and Europe sits below the global average on most adoption measures. More than 90 percent of Chinese enterprises use open-source technologies, and the country reached an estimated 70 percent of its 2030 industry-deployment target by mid-2025. Europe, by contrast, struggles with adoption among the small and medium enterprises that dominate its economy, which cite high upfront costs and binding credit constraints as barriers, problems rooted in fragmented capital markets that no AI policy alone can fix.

China and Europe on the metrics that define the race

The table makes the asymmetry concrete: Europe is competitive with or ahead of China on private investment and arguably on research quality per model, but it trails badly on deployment depth, which is the dimension that converts AI into economic and strategic advantage.

The deeper reason deployment matters more is that it compounds. Each deployed system generates data, operational experience, and feedback that improve the next deployment. A country running 250,000 driverless rides a week learns things about autonomy that no amount of frontier research substitutes for. A health system with AI in 90 hospitals accumulates clinical experience a cautious system never gets. Deployment is not just the end use of AI; it is an input to the next round of capability, and China’s lead in deployment is therefore self-reinforcing in a way the model rankings do not capture. Europe’s slower diffusion does not just mean less AI in the economy today. It means less of the learning that drives the capability of tomorrow, which is the mechanism by which a deployment gap quietly becomes a capability gap.

The data and talent inputs behind the lead

A recurring explanation for China’s AI strength is data: a population of 1.4 billion, the most mobile-integrated consumer economy in the world, and fewer privacy constraints on collection than in Europe. There is truth in this, but it requires qualification to be useful. China does have one of the fastest-growing data ecosystems globally, and it has built institutions to exploit it, including a National Data Administration created to improve data interoperability and treat data as a strategic input to innovation. The sheer size of the domestic market gives Chinese firms an ideal environment for testing and deploying AI, because any application can reach scale quickly.

The advantage is most real where data volume directly improves the product. Surveillance systems get better with more faces and more scenarios. Commerce recommendation engines improve with more transactions. Autonomous-driving systems improve with more miles. In these domains, China’s combination of scale and permissive collection produces a genuine edge that Europe’s privacy regime, GDPR above all, deliberately limits. Europe’s data-protection framework is a constraint on AI development that is also a deliberate protection of citizens, and the trade-off is explicit rather than accidental.

But the data-advantage story is often overstated, and the limits matter. Raw data volume is not the same as useful training data. Frontier model performance has been driven more by data quality, architecture, and compute than by sheer quantity of consumer data, which is why American labs with less access to a captive billion-user market still lead at the frontier. Chinese models also operate under content rules that degrade their data in politically sensitive areas, since training and outputs must conform to governance requirements. A model that cannot engage freely with certain topics is working with a constrained information environment regardless of how much data underlies it.

The honest assessment is that China’s data advantage is real for applied, consumer-facing, and surveillance AI, and far less decisive for frontier reasoning capability. The countries leading the frontier are not the ones with the most consumer data; they are the ones with the most compute and the strongest research ecosystems, which is why the data narrative explains China’s deployment strength better than it explains model capability. For Europe, the lesson is that GDPR’s constraints hurt less at the frontier than the deployment numbers suggest, because the frontier was never primarily a data-volume game. Where Europe’s data constraints bite hardest is precisely in the applied deployment race it has already largely conceded.

Talent is the other input that data analysis usually gets paired with, and the picture there is more even than the deployment numbers suggest, with a twist that complicates the simple China-ahead story. By volume, China dominates the pipeline. The country produces around 41,200 AI research papers a year against roughly 28,400 from the United States, and by some counts nearly 36 percent of global AI publications and about 50 percent of the world’s AI researchers now originate from China. On raw numbers entering the field, no country comes close.

The twist is that a large share of that talent does not stay or does not work for China. An estimated 68 percent of Chinese AI PhDs relocate to the United States, drawn by a salary premium often cited at around 185,000 dollars versus 67,000, producing a net gain of top researchers for the American ecosystem. The United States also maintains a quality edge, with higher average citation impact per paper, reflecting the concentration of frontier work in a handful of American institutions and companies. China trains the most AI talent, but the United States captures a disproportionate share of the best of it, and that brain drain is one reason the American frontier lead persists despite China’s larger pipeline.

Europe sits in the most exposed position on talent. It has excellent universities and produces strong researchers, but it suffers its own brain drain, losing talent to both American compensation and the gravitational pull of where the compute and capital are. The frictionless academic-to-industry pipeline that feeds American labs, Stanford and MIT and Berkeley flowing directly into OpenAI, Anthropic, Google, and Microsoft, has no European equivalent at the same scale, and China’s state-directed system, for all its volume, also struggles to replicate it.

The strategic implication is that talent is not a fixed national asset but a flow, and the flows currently favor the United States above both China and Europe. For the China-Europe comparison specifically, China’s advantage is in quantity and Europe’s residual strength is in quality at the top of specific fields, particularly the scientific AI where DeepMind and European research institutions excel. But quantity, combined with a state willing to direct it toward deployment, is a powerful thing, and it is part of why China can staff a deployment effort across every sector simultaneously in a way a talent-constrained Europe cannot. The brains that stay in China are pointed at applied problems at scale, which is exactly the race China has chosen to win.

The DeepSeek moment and what it actually changed

It is worth slowing down on DeepSeek, because the episode in early 2025 changed the terms of the entire debate and is frequently misremembered. The shock was not that a Chinese lab had built a capable model. It was the combination of three things at once: the model was strong on reasoning benchmarks, it was released with open weights under a permissive license, and it had reportedly been trained for around six million dollars on export-constrained hardware. R1 scored 97.3 percent on the MATH-500 benchmark and ranked among the top systems for mathematical reasoning, and the chatbot built on it briefly became the most downloaded free app in the United States in January 2025.

The strategic lesson the world drew was that the relationship between spending and capability was looser than assumed. If a model trained for single-digit millions could compete with systems that cost over a hundred million, then the moats around the leading labs were shallower than their valuations implied. DeepSeek’s real contribution was engineering: architectural innovations, including an efficient mixture-of-experts design, that extracted more capability per unit of compute, made below the level of Nvidia’s CUDA software layer. That is the kind of efficiency work that matters most for a country operating under hardware constraints, and it is why analysts argued that if DeepSeek turned its efficiency expertise toward Huawei’s Ascend ecosystem, the competitive threat to Nvidia would intensify sharply.

The episode also reset expectations about openness. By releasing strong models openly, DeepSeek and the labs that followed, Alibaba, Zhipu, Moonshot, made Chinese systems the default substrate for a large share of global AI development. The downstream effect, that a meaningful fraction of Western startups now build on Chinese base models, is a direct consequence of the open-weight strategy DeepSeek helped popularize. It converted a model release into a form of soft infrastructure power.

The corrective, equally important, is that the DeepSeek moment did not mean China had caught the frontier. The best closed models from American labs remained ahead on aggregate capability, and DeepSeek’s subsequent releases, while strong, slotted into a competitive open-weight field rather than overturning it. What DeepSeek changed was the perception of the gap and the economics of competing, not the underlying ranking at the top. For Europe, the moment carried a specific lesson that Mistral had already been demonstrating: efficiency-led model development is a viable path for actors who cannot or will not spend at hyperscale, which is the position Europe is in. The continent’s best hope at the model layer looks more like DeepSeek and Mistral than like a hundred-billion-dollar American training run.

Generative video opened a new creative front

A year after the language-model race, a parallel race in generative video has emerged, and Chinese companies are among its clearest early leaders. Kuaishou’s Kling, launched publicly in June 2024, has become the most widely used video-generation platform in the world by its own and partners’ accounts. By the 2025 World AI Conference, Kling reported supporting a global community of over 45 million content creators, powering the generation of more than 200 million videos and 400 million images, and serving over 20,000 enterprise clients across marketing, animation, film, and game production. In December 2025, Kling’s 2.6 model added simultaneous audio-visual generation, producing voiceovers, sound effects, and ambient audio in the same pass as the visuals, collapsing a production workflow that previously required separate dubbing.

The competitive field is crowded and almost entirely Chinese at the leading edge alongside the American players. ByteDance’s Seedance, Kuaishou’s Kling, and Shengshu’s Vidu are pushing on the hardest problem in the field: consistency across shots, keeping characters, props, and styles stable as a scene changes. Kling’s O1 model, launched in December 2025, was described as the first unified multimodal video model integrating generation, editing, and understanding in a single architecture, positioned directly against OpenAI’s Sora and Runway. Shengshu’s Vidu reached an estimated annual revenue of 20 million dollars on subscriptions within about a year of global launch.

The structural advantages echo the rest of the Chinese AI story. China sits on the densest short-video ecosystems on earth; Kuaishou alone reported around 409 million daily active users in mid-2025, spending over two hours a day in the app. That is an unmatched source of training data and a built-in distribution channel. Chinese firms also tend to identify a commercial pain point first, areas where businesses will pay, which is why Kling and its rivals are aimed squarely at advertising, e-commerce product videos, and film production rather than at open-ended novelty.

There is a geopolitical reading that European policymakers should not ignore. Because most leading Chinese video and image models are open-weight and open-license, while many American tools are restricted by API access and geographic licensing, Chinese creative AI is easier to adopt globally, especially in countries American vendors restrict. Analysts have described this as a form of cinematic soft power: the tools that shape how the world makes video may increasingly be Chinese, with the cultural and influence implications that follow. For Europe, which has a genuine asset in Black Forest Labs on image generation, the creative-AI front is one more arena where it has a foothold but no scale, and where the open-weight Chinese stack is setting the global default.

Open-source became geopolitical leverage

The decision by Chinese labs to release their strongest models openly is often read as generosity or as a workaround for distribution constraints. It is better understood as strategy. An open-weight model that becomes the foundation for thousands of downstream products embeds Chinese technology into the global AI supply chain in a way that no closed API can, and it does so without requiring Chinese firms to win the consumer-facing competition directly. When an estimated 80 percent of US startups building derivative AI products use Chinese base models, China has achieved a kind of infrastructure influence that survives even export controls and app-store bans.

The breadth of the open ecosystem reinforces the point. DeepSeek, Qwen, GLM, Kimi, and others ship under MIT, Apache 2.0, or equivalent licenses, giving any team the freedom to fine-tune and deploy commercially without vendor lock-in. More than 90 percent of Chinese enterprises use open-source technologies, and the AI Plus initiative explicitly prioritizes open-source development as a national goal. This is not incidental; it is policy aligned with commercial behavior, and the alignment is what makes it powerful.

Premier Li Qiang’s framing at Davos, that China’s innovation is open and that the country is willing to share its technologies with the world, is a deliberate contrast to the more closed, dominance-oriented posture associated with leading American labs. Open-source becomes a tool of positioning: China presents itself as the provider of AI as a public good, particularly to developing countries, while the United States is cast as the gatekeeper restricting access through export controls and licensing. Whether or not that framing is fair, it is effective diplomacy, and it lands with audiences in the Global South that resent being locked out of frontier tools.

The strategy carries real tensions that a complete account must name. Monetization pressure is pushing some Chinese firms back toward closed releases and higher prices; both Alibaba and Zhipu have released recent models in closed form, at least initially, and Baidu and others have raised model and cloud prices. The open-weight commitment may prove partly tactical, useful while building market share and influence, and less durable once the businesses need returns. There is also the censorship dimension: a Chinese open-weight model carries content restrictions wherever it is deployed, so a developer in another country using a Chinese base model may inherit constraints on politically sensitive topics. For Europe, the lesson cuts two ways. Open weights lower the barrier for European firms to build on capable models cheaply, which is a genuine opportunity. But building European products on Chinese foundations also imports dependencies and constraints, which is exactly the kind of strategic vulnerability the EU’s sovereignty rhetoric is meant to avoid.

Beijing’s bid to write the global rules

China’s ambitions extend past deploying AI at home to shaping how the rest of the world governs it, and 2025 was the year this became explicit. At the World AI Conference in Shanghai on July 26, 2025, Premier Li Qiang unveiled a Global AI Governance Action Plan, a thirteen-point roadmap, and proposed a World AI Cooperation Organization, potentially headquartered in Shanghai, to coordinate global AI development and standards. Li warned delegates, many from developing nations, against AI becoming the “exclusive game” of a few countries, an unmistakable reference to American export controls.

The proposal is positioned as an alternative to Western-led initiatives like the Bletchley Declaration and the G7 Hiroshima Process, which Beijing characterizes as partial or exclusionary. The Chinese framing emphasizes multilateralism, inclusion, and a development-oriented approach aimed at the Global South, in deliberate contrast to the United States’ deregulation-and-dominance strategy and the EU’s rights-based regulation. The action plan calls for working through the International Telecommunication Union, the International Organization for Standardization, and the International Electrotechnical Commission to set technical standards on security, industrial application, and ethics, and it prioritizes digital infrastructure support, including compute, data centers, and clean electricity, for developing countries.

The capacity-building apparatus is already taking shape. China has established a China-BRICS AI Development and Cooperation Centre in Shanghai and a China-Laos AI Innovation Cooperation Center, designed to build AI infrastructure and cultivate talent in partner countries. This echoes the Digital Silk Road, through which Huawei and other state-backed firms built telecommunications networks across Asia, Africa, and Latin America, and it serves a similar purpose: laying groundwork that favors Chinese firms and standards as global AI demand grows.

The honest assessment includes the catch. The Carnegie Endowment noted that even as Beijing offers capacity-building, it would likely retain control over the technologies underpinning the AI it exports, and that a Chinese-made model might automatically censor content embracing liberal norms regardless of where it is deployed. The governance bid is genuine diplomacy, but it is also an attempt to set rules that advantage Chinese technology and embed Chinese norms. For Europe, this is a direct contest. The EU has been the world’s standard-setter in technology regulation, through the AI Act and the broader Brussels effect, and China’s governance offensive is a challenge to that role precisely in the arena where Europe believed it had a durable advantage. The competition over who writes the rules may matter as much as the competition over who builds the models.

The United States remains the pole both must measure against

Any China-versus-Europe comparison that ignores the United States distorts the picture, because both are competing in a field the United States currently leads. The numbers are not close at the top. US private AI investment in 2024 reached around 109 billion dollars, nearly twelve times China’s and roughly twenty-four times the United Kingdom’s. American institutions produced 40 notable models in 2024 against China’s 15 and Europe’s three. The United States holds roughly nine times China’s compute capacity and seventeen times Europe’s. The frontier of capability, the CUDA software moat, the concentration of top talent, and the deepest capital markets all sit in the United States, and that lead is self-reinforcing.

This reframes the question the topic poses. China being technologically ahead of Europe does not mean China leads the world. It means China occupies a clear second place that is closing on the United States in some dimensions while pulling away from Europe in most. Europe, by the cold metrics, is competing for third. The American advantage in foundational research and global platform distribution coexists with China’s advantage in rapid implementation and scale deployment, a division that the Brookings Institution and others have described as structural rather than temporary.

The United States also shapes the China-Europe contest directly through policy. Export controls on advanced chips are the single biggest external constraint on Chinese AI, and they are an American instrument. Washington’s 2025 AI Action Plan pivoted toward deregulation and aggressive promotion of AI exports to allies, an approach that sits between China’s deployment-first model and Europe’s rights-first model and competes with both. Europe is squeezed: it cannot match American compute and capital, and it cannot match Chinese deployment speed and state coordination, which is why its strategy emphasizes the things neither superpower prioritizes, namely trust, rights, and sovereignty over sensitive workloads.

The constructive reading for Europe is that a three-pole world is not a two-horse race it has already lost. The United States leads at the frontier, China leads in deployment, and Europe can occupy defensible ground in regulation, scientific AI, defense, and efficiency-led specialization. The destructive reading is that both other poles are outbuilding Europe so decisively that its niches will be absorbed over time. The evidence supports a sober middle position: Europe is genuinely behind both, more behind China on deployment than on research, and its window to establish durable advantages in its chosen niches is open but narrowing. Measuring China against Europe without this American backdrop produces a misleading sense that the contest is bilateral. It is not. It is a three-way race in which Europe currently runs third, and the gap to China specifically is widening fastest in the deployment dimension this article has documented.

Banking and insurance absorbed AI earliest

Among the sectors deploying AI in China, banking, financial services, and insurance came first and run deepest, which is why this category holds the largest single share of China’s AI market. Financial institutions adopted AI early for fraud detection, algorithmic trading, customer-service automation, credit scoring, and risk management, and the maturity shows. Fraud-detection systems that flag anomalous transactions in real time, underwriting models that price risk from richer data, and chatbots that handle a large share of routine customer interactions are standard rather than experimental across major Chinese banks and insurers.

The structural reason finance led is that it combines high transaction volume, clear monetary stakes, and abundant structured data, the ideal conditions for machine learning to deliver measurable returns. A model that reduces fraud losses or improves credit decisions pays for itself quickly, which is why financial firms everywhere adopt AI early, and why Chinese institutions, operating in the most digitized payments economy on earth, could move faster than peers elsewhere. Ant Group’s rollout of AI-powered payment systems and the integration of AI across Alipay’s risk and service layers illustrate how deeply the technology now sits in the financial plumbing.

The agentic-commerce wave is extending finance AI into new territory. When Alibaba’s Qwen lets users compare products and complete payments within a chat, and when ByteDance’s Doubao books tickets autonomously, AI is moving from analyzing transactions to executing them. This blurs the line between a financial service and an AI assistant, and it raises questions about authorization, liability, and consumer protection that regulators are only beginning to address. The same speed that makes Chinese finance AI impressive also concentrates new risks, from automated decisions that are hard to contest to the systemic implications of many institutions relying on similar models.

For the European comparison, finance is a sector where the gap is narrower than in deployment-heavy areas like robotics or autonomy, because European banks and insurers are sophisticated AI adopters with strong data-science capabilities, operating under a regulatory regime that mandates explainability and fairness in automated financial decisions. Europe’s financial AI may be more cautious and more auditable, which has genuine value in a sector where errors and bias carry legal and social consequences. The trade-off is familiar: China deploys faster and learns faster, Europe deploys more carefully and protects consumers more thoroughly, and which approach proves superior depends on whether the speed advantage compounds faster than the trust advantage matures.

The scale of the financial-AI base in China is easy to underestimate from the outside. Hundreds of millions of people run their entire financial lives through two super-apps, Alipay and WeChat Pay, which means a single AI improvement to fraud scoring or credit assessment reaches a user base larger than the population of any Western market the moment it ships. That concentration is why the banking, financial services, and insurance category consistently registers as the largest slice of China’s applied-AI spending. The data these platforms sit on is not only large but unusually complete, linking payments, lending, investment, insurance, and daily consumption in one graph, which lets models see patterns that a fragmented Western financial system, split across banks, card networks, and separate apps, struggles to assemble.

The newest move is the embedding of large language models directly into the financial interface rather than running them quietly in the back office. Ant Group has built AI assistants that handle wealth-management questions, insurance comparisons, and health-claim guidance in natural language, turning what used to be a forms-and-menus experience into a conversation. When the assistant that answers a financial question is also the system that can execute the resulting transaction, the distance between advice and action collapses, and that collapse is exactly what makes agentic finance both powerful and hard to govern. Europe’s banks are testing similar assistants, but compliance regimes that require a clear audit trail for every automated recommendation slow the leap from chatbot to autonomous agent, which is the difference the deployment numbers keep capturing.

Government offices fold AI into the administrative state

One of the less-discussed but most consequential deployment fronts is government itself. China is integrating AI into public administration, judicial work, and urban management at a scale that turns the state into one of the largest AI users in the country. The City Brain platforms pioneered in Hangzhou are the visible face of this, fusing traffic, public-security, and municipal data into systems that manage urban operations in real time, with City Brain 3.0 built on DeepSeek-R1 extending into patrolling and grid coordination.

The judicial system is a striking case. China’s Supreme Court has pushed all courts to develop competent AI systems by 2025 for use in trials and administrative work, and reporting has described systems that recommend whether prosecutors should seek arrest or suspended sentences. Embedding AI into prosecutorial and judicial decisions is a deployment frontier most legal systems approach with extreme caution, and China’s willingness to move quickly here reflects the same deployment-first logic that runs through its economy, applied to the machinery of the state. The efficiency case is real; the due-process risks are equally real, and they are being run at scale.

Routine administration is being reshaped too. Government service portals increasingly use AI to handle citizen queries, process applications, and route requests, reducing the friction of dealing with the bureaucracy. Population forecasting and industrial-trend analysis are being used to plan everything from school placements to resource distribution. The state is using AI not only to surveil and control but to deliver services more efficiently, and for many citizens the latter is the more visible effect in daily life.

For Europe, government AI is an area where the contrast is sharpest in philosophy. The EU AI Act treats AI in law enforcement, justice, and essential public services as high-risk, subject to strict oversight, human-in-the-loop requirements, and transparency obligations, precisely to prevent the kind of automated judicial recommendation systems China is deploying. European governments are adopting AI for administrative efficiency, but within guardrails designed to protect citizens against automated decisions affecting their rights. This is the clearest example of a gap that is a choice rather than a capability deficit. Europe could deploy AI in courts and policing as aggressively as China; it has decided not to, on grounds of rights and due process. Reading that restraint as falling behind misunderstands what the European system is optimizing for, even as it concedes that China’s administrative state is becoming more AI-saturated than any European counterpart.

The speed of adoption inside the state accelerated sharply once a cheap domestic model existed to run on. After DeepSeek’s reasoning models arrived in early 2025, government bodies, state-owned enterprises, and local administrations across the country moved to deploy them, partly because the cost structure made wide rollout affordable and partly because a homegrown model carried none of the security objections attached to foreign systems. Provincial governments stood up their own AI service platforms, courts and hospitals plugged the same models into their workflows, and municipal data centers were retooled to host them. A capable open-weight model produced domestically turned out to be the missing piece that let the deployment-first instinct reach every level of administration at once, because it removed both the licensing cost and the political risk that a proprietary foreign system would have carried.

That diffusion has a quieter governance consequence worth naming. When the same handful of models underpin services, courts, hospitals, and city-management systems nationwide, the model layer becomes a kind of shared infrastructure whose assumptions, blind spots, and content rules propagate everywhere it is installed. A flaw or a built-in constraint is no longer contained to one application; it is replicated across the institutions citizens depend on. The efficiency gains are real and visible, but the concentration of so much public function on a narrow base of state-aligned models is a structural feature European systems, with their insistence on auditability and human review, are deliberately designed to avoid, and it is the part of China’s administrative AI story that the deployment statistics never capture.

The economics of cheap intelligence

Underneath the sector-by-sector story sits an economic shift that explains why deployment has accelerated: the cost of useful AI has collapsed, and China has both driven and benefited from that collapse. DeepSeek’s roughly six-million-dollar training run was the dramatic signal, but the more important trend is in inference, the cost of actually running a model to serve a request. DeepSeek’s pricing, on the order of cents per million input tokens, and the mixture-of-experts architectures that let firms run complex queries at a fraction of the GPU cost of dense models, have pushed the price of capable AI down to the point where deploying it across an entire economy is affordable.

This is the precondition for everything else in this article. Mass deployment only makes economic sense when the per-use cost of AI is low enough that the value it adds exceeds what it costs to run, and that threshold has dropped sharply. A hospital can run a strong model locally on modest hardware. A logistics firm can apply demand forecasting across thousands of facilities. A manufacturer can put computer-vision inspection on every line. The “token economy” that Chinese commentators describe, where intelligence is metered and cheap, is what turns the AI Plus penetration targets from aspiration into arithmetic.

The economics also reveal a strategic tension within China’s own AI sector. The capital expenditure is enormous, Alibaba’s 123 billion yuan in 2025, ByteDance’s reported plans to spend around 23 billion dollars on AI infrastructure, even as these figures remain well below American hyperscaler spending of 75 to 94 billion dollars each. Monetization has not kept pace with investment, which is why Alibaba’s net income plunged 66 percent in a year of heavy AI capex, and why firms are raising prices and releasing closed models to improve returns. The era of free, cheap, open Chinese AI may be a phase rather than a permanent condition.

For Europe, the economics of cheap intelligence is mostly good news, and an underused opportunity. The collapse in the cost of capable AI, driven substantially by Chinese open-weight efficiency, means European firms can access strong models cheaply without building their own. The constraint is not the cost of the model anymore; it is the cost and difficulty of deployment, integration, and the organizational change required to use AI well. Europe’s adoption gap is therefore less about technology cost and more about diffusion capacity, the ability of European firms, especially the credit-constrained SMEs that dominate the economy, to actually integrate AI into operations. That is a different problem from the one the investment figures suggest, and it points to a different kind of policy response than simply spending more on compute.

The risks China is deploying straight through

A deployment-first strategy means running risks that a more cautious approach would slow down to address, and China is running several at once. The clearest is in healthcare, where the speed of DeepSeek’s adoption across hospitals outran the legal framework for accountability. Chinese researchers have warned explicitly about automation bias, the danger that overworked clinicians defer to AI recommendations, and about the absence of clear liability when an AI-assisted diagnosis causes harm. Deploying medical AI into 90 hospitals before resolving who is responsible for its errors is a calculated bet that the benefits arrive faster than the harms, and it is a bet whose downside is borne by patients.

Surveillance carries a different category of risk. Facial-recognition systems produce false matches, and false matches in a system tied to social management and policing can mean wrongful detention. The expansion of AI into judicial recommendations raises the prospect of automated decisions affecting liberty without adequate human oversight or appeal. These are not hypothetical concerns; they are the predictable consequences of deploying probabilistic systems into high-stakes state decisions at scale, and the June 2025 facial-recognition rules address commercial misuse far more than state use.

The autonomy front showed its risks plainly when a Wuhan robotaxi malfunction stranded passengers and disrupted traffic, prompting a reported pause on new permits. Humanoid robotics faces the gap Wang Xingxing identified, hardware racing ahead of the general intelligence needed to make it reliable in unstructured settings. The data-center glut demonstrated that state-directed investment produces waste alongside scale, with billions sunk into facilities that may not find productive use. And the open-weight strategy that spreads Chinese influence also spreads Chinese content restrictions, embedding censorship into systems used far beyond China’s borders.

There is also a systemic risk in the model layer itself. Hard-coded content restrictions mean Chinese models operate with constrained information in politically sensitive domains, which limits their reliability for some uses and reflects a deeper subordination of the technology to political control. Monetization pressure and the partial retreat from open weights suggest the commercial foundations are less settled than the deployment numbers imply. The honest summary is that China is not deploying AI carefully; it is deploying it fast and managing the consequences as they arrive. For some applications, that trade favors China decisively, because the learning and economic value compound. For others, particularly those touching liberty, health, and accountability, the deferred costs are real and may surface later. A reader trying to judge whether China is “ahead” should weigh not only what it has deployed but the risks it has chosen to carry forward unresolved, because those risks are part of the true cost of the speed.

AI deployment reshapes work for ordinary professionals

The deployment story is ultimately about people, and the effect of pervasive AI on ordinary Chinese workers and professionals is one of the most important and least settled aspects of the whole picture. With roughly 34 percent of job functions at Chinese companies already integrated with AI tools, the technology is changing daily work across white-collar and increasingly blue-collar roles. Clinicians use AI for diagnosis support and documentation. Factory workers collaborate with humanoid and industrial robots. Customer-service staff are augmented or replaced by chatbots. Teachers use AI lesson companions. Creative professionals use Kling and similar tools to produce in hours what once took days.

The optimistic framing, which Chinese policy promotes, is augmentation: AI handles routine cognitive and physical tasks, freeing workers for higher-value work, and the education system is being retooled to produce a workforce comfortable with these tools from primary school. The bet is that pervasive AI raises national productivity enough to offset demographic decline, which is a genuine strategic concern for an aging China with a shrinking working-age population. Humanoid robots are explicitly framed as a response to labor shortages rather than purely as a cost-cutting measure.

The harder reality is displacement, and it is uneven. Roles built on routine tasks, in customer service, basic analysis, content production, and repetitive manufacturing, face the most pressure, and the speed of Chinese deployment means the adjustment is happening fast. The education push is partly an attempt to manage this transition by preparing the next generation for AI-complementary work, but it does little for workers already in displaced roles. The social management of AI-driven labor disruption is a problem China is encountering before most other countries, simply because it is deploying first and at scale.

For European workers, the comparison is double-edged. Slower deployment means slower disruption, which buys time for adjustment, retraining, and the social-protection systems Europe is known for to respond, an advantage that rarely appears in the technology-race framing but matters enormously to the people involved. The cost is that slower deployment also means slower productivity gains at a time when Europe’s aging workforce faces the same demographic pressures as China’s. The choice between fast disruption with fast productivity gains and slow disruption with slower gains is not obviously settled in either direction, and it is a choice about social values as much as about technology. China has chosen speed and is absorbing the disruption; Europe has chosen caution and is forgoing some of the gains. Which looks wiser in a decade depends on how each society handles the human consequences of its choice.

From fast follower to deployment leader

The position China occupies in 2026 was built in identifiable stages, and the history clarifies why deployment, not the frontier, became its strength. The foundational stage ran from the 2017 New Generation AI Development Plan, which set the 2030 leadership goal and, unusually, proposed security supervision early. Through 2020, China built research capacity and absorbed global advances, often described, fairly at the time, as a fast follower. The industrialization stage from roughly 2020 onward saw the ecosystem expand to more than 4,500 enterprises spanning chips, algorithms, data, platforms, and applications, with the domestic market reaching close to 600 billion yuan. The deployment stage, accelerated by DeepSeek and the AI Plus initiative in 2025, is where China converted accumulated capacity into pervasive use.

The backbone enabling this is infrastructure most consumers never see. Huawei is not only a chip maker but China’s second-largest cloud provider, with its own open-source Pangu model family, and it supplies the networking, computing, and cloud layers that hospitals, factories, and city governments build on. Alibaba Cloud and Tencent Cloud provide the model-hosting and inference capacity that turns the AI Plus targets into running systems. The combination of domestic cloud, a dense hardware supply chain, ultra-high-voltage power transmission, and the world’s most built-out mobile-payment and digital-identity systems gives China an integrated stack that few countries can match end to end.

The telecommunications layer matters more than it appears. China’s national networks, the Smart Education portal serving the entire basic-education system, the digital-yuan platform, and the data systems coordinated through the National Data Administration form a substrate on which AI applications scale rapidly. When a model is released, it does not need new infrastructure to reach scale; the rails already exist, and hundreds of millions of users are already on them.

The historical lesson for Europe is sobering and specific. China’s deployment lead did not come primarily from superior models; it came from a decade of building the infrastructure, the supply chain, the digital-payment rails, the cloud capacity, the state coordination, into which AI could be poured. Europe’s deployment gap is partly a model gap but mostly an infrastructure and coordination gap, and that is harder to close with a single funding initiative than the headline investment numbers suggest. The InvestAI billions can buy compute, but they cannot quickly assemble the integrated, coordinated stack that China spent a decade constructing. The history explains why the gap is structural, and why closing it requires more than money.

Retail and consumer life reorganized around AI

Retail is where AI touches the most Chinese citizens directly, and the reorganization runs deeper than recommendation engines. China’s e-commerce platforms, Taobao, Tmall, JD.com, Pinduoduo, Douyin’s commerce features, use AI across the entire purchase journey: personalized discovery, dynamic pricing, demand forecasting, inventory optimization, fraud prevention, and increasingly autonomous fulfillment. The mobile-first character of Chinese commerce, with around 84 percent e-commerce AI adoption leading the world, means these systems operate at a density of usage that produces continuous feedback and improvement.

The agentic-commerce shift is changing the consumer relationship itself. With Alibaba’s Qwen completing transactions inside a chat and ByteDance’s Doubao handling bookings autonomously, the act of shopping is collapsing into a conversation with an assistant that compares options and executes the purchase. This is a genuinely new consumer interface, and Chinese platforms are deploying it at scale before Western counterparts have moved past pilots. Livestream commerce, already a Chinese innovation, is being augmented with AI-generated product videos from tools like Kling, letting merchants produce marketing content at near-zero marginal cost.

Physical retail is being reshaped too. Computer-vision systems power cashier-less stores, inventory robots, and footfall analytics. Facial-recognition payment, where permitted under the 2025 rules, lets customers pay by looking at a camera. Smart logistics connects online and offline inventory so that a purchase can be fulfilled from whichever location is closest. The consumer experience in urban China is more AI-mediated than almost anywhere else, from the moment a product is discovered to the moment it arrives.

For the European comparison, retail illustrates the diffusion gap clearly. European retailers use AI for recommendations, pricing, and supply-chain optimization, and the leading ones are sophisticated, but the depth and speed of integration lag China’s, constrained by more fragmented markets, stronger consumer-privacy protections that limit data collection, and slower adoption of mobile payment and super-app models. The European consumer is protected from the data-intensive personalization that powers Chinese retail AI, which many Europeans would count as a benefit rather than a deficit. The gap in retail AI is therefore another instance of the central pattern: China deploys deeper and faster, Europe protects consumers more and deploys less, and the question of which is better depends entirely on whether the value being weighed is economic efficiency or individual privacy. What is not in dispute is that the Chinese retail economy runs on AI to a degree the European one does not, and that the gap is widening as agentic commerce scales.

The content layer of Chinese retail has been transformed almost as quickly as the transaction layer. Generative video tools have collapsed the cost of producing the marketing that drives online sales, and the scale is already large: Kuaishou’s Kling platform reported more than 45 million creators and over 200 million generated videos, with more than 20,000 enterprise clients using it to turn product photos into polished promotional clips in minutes rather than days. For a merchant on a livestreaming platform, the ability to generate an endless stream of product videos at near-zero marginal cost changes the economics of selling, because the bottleneck was never the product but the content needed to surface it in a feed. A retail ecosystem where both the recommendation engine and the advertising creative are AI-generated is operating on a different cost curve from one where humans still produce the marketing, and that curve is what lets the smallest Chinese sellers compete with national brands.

The consumer-facing assistants are beginning to close the loop between content and purchase. When the same platform can generate the video that catches a shopper’s attention, answer questions about the product in natural language, and complete the payment without leaving the conversation, the entire funnel collapses into a single interface controlled by one company’s models. Western retail is moving in the same direction, but it is doing so across separate advertising, search, and payment systems owned by different firms, which slows the integration. China’s super-app structure, where one app already holds messaging, payments, identity, and commerce, gave its platforms a head start on agentic retail that European fragmentation cannot easily match, and it is why the consumer-facing edge of the deployment gap is most visible precisely where ordinary people shop.

The verticals where Europe still holds an edge

A fair account has to name where Europe is not behind, because the deployment-gap story can slide into a caricature of European decline that the evidence does not support. Scientific AI is the clearest case. Google DeepMind’s AlphaFold, developed in London, solved protein-structure prediction at a level that earned a share of the 2024 Nobel Prize in Chemistry and transformed drug discovery and structural biology. No Chinese lab has produced a scientific AI contribution of comparable significance, and Europe’s combination of world-class universities, strong basic research, and deep domain expertise in fields like chemistry, physics, and the life sciences is a genuine and defensible advantage. AI that advances science rewards exactly the strengths Europe has and the scale-and-speed approach does not prioritize.

Defense AI is more contested but increasingly a European strength. Germany’s Helsing has built a serious business in autonomous military systems, with its Centaur agent piloting a Saab Gripen fighter on test flights and the company extending into AI-driven drones and submarines. Its partnership with Mistral brings frontier model capability into defense applications. Europe’s defense-AI firms are competitive at the leading edge, and the strategic imperative created by the security environment in Europe is driving investment and urgency that did not exist a few years ago. This is one arena where European deployment is accelerating rather than lagging.

Efficiency-led model development is a third niche, and it suits Europe’s constraints. Mistral built competitive models with far less compute than American rivals, demonstrating the same lesson DeepSeek taught: capability can come from architectural cleverness rather than brute scale. For a continent that cannot match American or Chinese compute, efficiency is not a consolation prize but a viable competitive strategy. The companies most likely to keep Europe relevant at the model layer are the ones that do more with less, and Mistral is the proof of concept.

Specialized application leaders round out the picture. DeepL dominates a valuable translation niche where quality and enterprise trust matter. Black Forest Labs leads in image generation. Wayve is a credible player in autonomous-driving AI. Industrial software, where European firms like SAP and Siemens are embedding AI into enterprise and manufacturing systems, plays to a traditional European strength in industrial technology. Europe’s edge is concentrated, not broad: it leads in scientific AI, holds strong positions in defense and efficient models, and has defensible niches in translation, image generation, and industrial software. The risk, as noted earlier, is that niches can be absorbed by general-purpose models and that several of these firms depend on non-European compute and foundations. But the existence of these strengths means the honest comparison is not China ahead everywhere; it is China ahead in deployment and most aggregate metrics, with Europe holding real advantages in a handful of high-value verticals where depth beats scale. Those verticals are where European strategy should concentrate, because they are the ground Europe can actually defend.

The open questions the evidence cannot settle yet

Several questions central to judging this contest cannot be answered with the evidence available in 2026, and intellectual honesty requires naming them rather than resolving them by assertion. The first is compute. China leads in deployment while depending on a hardware base it does not fully control, constrained by export controls, stuck at 7-nanometer process technology, and reliant on foreign high-bandwidth memory. Whether China can achieve genuine semiconductor self-sufficiency, or whether the compute ceiling eventually caps its AI ambitions, is genuinely unresolved, and it is the single biggest uncertainty in any forecast. Huawei’s production ramp suggests progress; the persistent process-node gap suggests limits.

The second is whether deployment depth converts into durable capability advantage or merely reflects current conditions. The argument that deployment compounds, through data, operational experience, and feedback, is plausible but unproven at the level of national AI capability. It is possible that frontier model quality, where the United States leads, matters more in the long run than deployment breadth, in which case China’s lead over Europe in deployment would be less decisive than it currently appears. The relationship between deploying AI widely and developing AI capability is an open empirical question, not a settled fact.

The third is whether China’s risk tolerance proves wise or costly. Deploying medical AI ahead of liability frameworks, judicial AI ahead of due-process safeguards, and surveillance AI at population scale are bets that the benefits arrive before the harms compound. If those deferred risks surface as scandals, failures, or social backlash, the speed advantage could reverse into a liability, and the more cautious European approach could look prescient. The Wuhan robotaxi pause is a small early signal that even China hits limits.

The fourth concerns Europe’s strategy directly. Whether a continent can remain a relevant AI actor through regulation, scientific excellence, and selective niches, without competing on deployment or compute, is the untested premise of European policy. The bet that trust and rights are competitive assets rather than just costs has not been validated by results, and the investment figures suggest it is under strain. It is equally untested whether the InvestAI program and the gigafactory plans can close enough of the gap to matter, or whether they socialize an expensive race Europe cannot win. These four uncertainties mean that confident predictions about the China-Europe technology balance in 2030 are unjustified. What can be said is what the current evidence shows, and what the genuine unknowns are, which is more useful than a false forecast dressed up as analysis.

Lessons European businesses can take from this

The China-Europe comparison is usually framed for policymakers, but it carries concrete, actionable lessons for European businesses, and these are worth setting out directly because they are where individual firms have agency that governments do not. The first lesson is that the cost of capable AI has collapsed, and waiting for a perfect model or a fully clear regulatory picture is now the expensive option. Chinese firms deployed good-enough models across operations and accumulated the operational learning that compounds. A European company that delays adoption until the EU AI Act’s every provision is litigated will fall behind competitors, including Chinese-influenced ones, that are already learning from deployment. The practical takeaway is to start deploying AI on real operational problems now, within the guardrails the law provides, rather than treating regulation as a reason to wait.

The second lesson is to exploit open weights rather than building from scratch. The collapse in model cost means European firms can build on capable open-weight models, including the Chinese ones that now dominate the open ecosystem, without the capital expenditure of training their own. The strategic caution is dependency: building core products on Chinese foundations imports both technical dependencies and content constraints, so firms in sensitive sectors should weigh European or open alternatives like Mistral’s models against the convenience of the cheapest option. The choice of foundation model is now a strategic decision about sovereignty and risk, not just a technical one about benchmarks.

The third lesson concerns where to compete. China’s deployment advantage is broad but its frontier and scientific-AI advantages are narrower, and Europe’s defensible ground is in depth, scientific AI, defense, efficient specialized models, industrial software, and applications where trust and auditability are themselves the product. A European business is unlikely to win by trying to match Chinese deployment scale in consumer applications. It is far more likely to win by occupying a vertical where domain expertise, regulatory compliance, and trust create a moat that scale alone cannot breach. The right competitive posture for most European firms is specialization and trust, not scale.

The fourth lesson is about diffusion capacity inside the organization. China’s adoption is fast partly because the surrounding infrastructure and the workforce are oriented toward AI, down to mandatory AI education in schools. European firms cannot wait for the school system, but they can invest now in the organizational capability to deploy AI well: data infrastructure, integration skills, and workforce training. The binding constraint on European AI adoption is increasingly not the cost of the model but the capacity to integrate it, and that capacity is built deliberately. Firms that invest in deployment capability, not just in models, will capture the value; those that buy AI tools without the capacity to integrate them will not.

The fifth lesson is to treat compliance as a feature rather than only a cost. Europe’s regulatory environment is real and binding, but the predictability it creates, and the trust it can confer, are assets in markets where customers and regulators value safety and accountability. A European firm that builds AI products which are demonstrably compliant, auditable, and rights-respecting has something Chinese deployment-first products often lack, and there are global markets that will pay for it. The European regulatory burden can be turned into a differentiated value proposition for buyers who need AI they can trust and defend, particularly in regulated industries like finance, healthcare, and the public sector.

The synthesis for European businesses is that the lesson of China is not to imitate China. It is to recognize that the deployment race is being lost on speed and scale, that those are not games most European firms can win, and that the response is to deploy faster than European caution instinctively allows while concentrating on the verticals, trust, specialization, and scientific depth, where Europe’s structural strengths actually create advantage. The Chinese example is most useful as a demonstration of what happens when an economy treats AI as infrastructure to be used now rather than a technology to be perfected first. European firms that internalize the urgency without abandoning their genuine advantages are the ones most likely to remain competitive in a market where both larger poles are moving faster.

The competitive map by domain shows no single winner

Stepping back from the sector-by-sector detail, the most accurate summary of the China-Europe contest is a map that varies by domain rather than a single verdict, because “ahead” resolves differently depending on what is being measured. On frontier model capability, the United States leads, China follows closely and is gaining, and Europe trails both with the notable exception of efficient specialized models from Mistral and scientific models from DeepMind. On open-weight models, China clearly leads the world, having made its systems the default substrate for a large share of global development, while Europe contributes through Mistral and a few others.

On deployment depth, China is comprehensively ahead of Europe and arguably ahead of everyone, embedding AI across manufacturing, healthcare, surveillance, commerce, autonomy, education, and government at a scale and speed no other economy matches. This is the domain where the gap with Europe is widest and where it is widening fastest, and it is the strongest support for the claim that China is technologically ahead of Europe. On compute and semiconductors, the United States leads decisively, China is constrained but building, and Europe is the weakest of the three, dependent on imports and lacking the coordinated demand that funds domestic capacity.

On robotics and embodied AI, China dominates, taking over 80 percent of humanoid installations and the majority of industrial-robot deployments, supported by an unmatched hardware supply chain. Europe has pockets of strength in industrial automation but nothing approaching China’s scale. On autonomous driving, China leads in commercial deployment through Apollo Go and others, with Chinese services poised to reach European streets before homegrown European ones. On generative video and creative AI, China holds an early lead with Kling and its rivals, while Europe has a foothold in image generation through Black Forest Labs.

On scientific AI, Europe leads through DeepMind’s AlphaFold and its research base, a genuine and significant advantage China has not matched. On defense AI, Europe is increasingly competitive through Helsing and others, driven by the security environment, while China pursues civil-military fusion at greater scale. On AI governance and standards, the contest is direct and unresolved: Europe set the global template through the AI Act and the Brussels effect, and China is mounting a serious challenge through its Global AI Governance Action Plan and the proposed World AI Cooperation Organization, courting the Global South with a development-oriented, open-source-friendly alternative.

The map shows China ahead in the largest number of domains and decisively ahead in deployment, the United States ahead at the frontier and in compute, and Europe leading only in scientific AI and contesting in defense and governance. For the specific question of whether China is technologically ahead of Europe, the map gives a clear answer in most domains: yes, and the lead is largest precisely in the deployment dimension that converts technology into economic and strategic power. The honest qualifications, Europe’s scientific edge, its defense competitiveness, its governance role, and the unresolved questions about compute and whether deployment compounds, prevent this from being a total verdict. But they do not overturn the central finding. Across the domains that can be measured, China is ahead of Europe in artificial intelligence, and the gap is growing in the areas that matter most for the next decade.

What Europe would need to change to close the gap

If the diagnosis is that Europe trails China most in deployment, the question that follows is what would actually narrow that gap, and the honest answer is that the required changes are structural and uncomfortable rather than a matter of one more funding round. The first change is coordination. China’s deployment advantage rests heavily on the ability to coordinate demand across the state, directing procurement, mandating adoption, and creating captive markets that fund domestic capacity. Europe’s demand is fragmented across 27 member states, each procuring separately, which is why analysts at Bruegel argued that coordinated public procurement across the EU, for AI compute used in public administration, healthcare, defense, and research, is one of the few levers Europe actually holds. Without a mechanism to aggregate European demand, the continent cannot generate the captive market that funds a hardware-improvement cycle or accelerates deployment at scale.

The second change is capital. Europe’s investment gap is rooted in fragmented capital markets and limited access to equity, especially for the small and medium enterprises that make up most of its economy and cite high upfront costs and binding credit constraints as barriers to AI adoption. Deepening the internal market and advancing the savings-and-investment union are the structural reforms that would let European firms fund AI adoption, and they are slow, politically difficult, and only indirectly about AI. The European AI gap is partly a capital-markets problem wearing a technology costume, and treating it purely as a technology-funding question misses the deeper constraint.

The third change is energy. Europe’s high electricity prices make AI infrastructure more expensive to operate than in China or much of the United States, and the gigafactory ambitions run directly into this wall. Closing the deployment gap requires either cheaper energy, which collides with environmental and security commitments Europe has made for good reasons, or a strategic decision to subsidize the energy cost of strategic compute, which is expensive and contestable. There is no painless way around the energy constraint, and pretending otherwise produces plans that look good on paper and fail in operation.

The fourth change is regulatory calibration, and this is the most contested. The EU AI Act’s defenders argue it creates predictability that accelerates compliant deployment; its critics, including the June 2025 European Parliament committee motion, argue that too much regulation combined with weak investment is causing Europe to fall further behind. The realistic position is that the law’s intent is sound but its implementation must avoid becoming a deployment brake, which means clear guidance, proportionate obligations for lower-risk uses, and regulatory bodies resourced to give fast answers rather than slow uncertainty. Regulation that protects rights without paralyzing deployment is achievable, but only if implementation prioritizes clarity and speed, which has not been Europe’s strong suit.

The fifth change is cultural and organizational, and it may be the hardest. China deploys fast partly because its firms, workforce, and state share an orientation toward using AI now, reinforced from the school system upward. Europe’s more cautious culture around technology, rooted in legitimate concerns about privacy and unchecked power, slows adoption. Changing that does not mean abandoning the caution; it means building the organizational capacity to deploy AI well within European values, investing in data infrastructure, integration skills, and workforce training so that the binding constraint, which is increasingly diffusion capacity rather than model cost, is addressed directly. The uncomfortable conclusion is that closing the deployment gap requires changes to European capital markets, energy policy, regulatory implementation, and organizational culture, none of which a single AI strategy can deliver, and most of which take years. That is why the gap is likely to persist even if Europe executes its current plans well, and why a realistic European strategy focuses less on catching China in deployment and more on dominating the verticals where Europe’s structural strengths create durable advantage.

A realistic read on the China-Europe technology balance

Pulling the threads together, the most defensible read on where China and Europe stand in artificial intelligence in 2026 is neither triumphalist nor dismissive of either side. China is technologically ahead of Europe across most measurable dimensions of AI, and the lead is largest and growing fastest in deployment, the dimension that converts the technology into economic output, social organization, and strategic power. From manufacturing floors and tertiary hospitals to robotaxi fleets, surveillance networks, classrooms, payment apps, and government offices, AI is woven into the fabric of the Chinese economy at a depth and speed Europe has not approached. The State Council’s AI Plus penetration targets, 70 percent of key sectors by 2027 and 90 percent by 2030, are the explicit statement of a deployment-first strategy that is already further advanced than any comparable European effort.

This lead is real but conditional. It rests on a compute base China does not fully control, constrained by export controls, a stuck process node, and import dependence for advanced memory, which is the single largest threat to its trajectory. It involves running risks, in medical accountability, judicial automation, surveillance overreach, and unresolved liability, that a more cautious approach would slow down to address, and some of those deferred costs may surface later. And it reflects a political system that enables deployments Europe has deliberately chosen to prohibit, which means part of China’s “lead” in areas like population-scale biometric surveillance is a difference in values rather than a deficit in European capability.

Europe is genuinely behind, more so than its strong research base and handful of excellent companies suggest, and the structural causes, fragmented capital markets, fragmented demand, high energy costs, and slow diffusion, are deep and resistant to quick fixes. But Europe is not without standing. It leads the world in scientific AI through DeepMind’s Nobel-recognized work, it is increasingly competitive in defense AI through firms like Helsing, it has proven that efficiency-led model development works through Mistral, and it holds the global standard-setting role in AI governance that China is now actively contesting. These are real assets, concentrated in high-value verticals where depth beats scale, and they are the ground a realistic European strategy should defend and extend rather than abandoning them to chase a deployment race it cannot win.

The forecast that the evidence supports is modest and bounded. Through the rest of the decade, China is likely to retain and probably widen its deployment lead over Europe, because the structural advantages behind it, state coordination, supply-chain integration, cheap energy, and an oriented workforce, are durable and the European constraints are slow to change. Whether that deployment lead translates into a decisive long-term capability advantage depends on unresolved questions, above all whether China overcomes its compute ceiling and whether deployment depth compounds into capability the way its proponents claim. If China solves compute and deployment compounds, the gap with Europe becomes structural and possibly permanent. If the compute ceiling holds and frontier capability matters more than breadth, China’s lead over Europe in deployment will coexist with a shared position behind the United States at the frontier, and Europe’s niche strengths will matter more than they appear to today.

For the original question, whether China is technologically ahead of Europe and where AI is used across China, the answer is clear and specific. AI is used across essentially every sector of the Chinese economy, deployed faster and deeper than anywhere else, from factories and hospitals to streets, schools, shops, and government. And yes, China is ahead of Europe in artificial intelligence, decisively in deployment and in most aggregate metrics, with Europe retaining real but narrow advantages in scientific AI, defense, efficient models, and governance. The gap is not a verdict on European capability so much as a consequence of different choices about speed, control, rights, and the role of the state, made by two systems optimizing for different things. China chose to deploy AI everywhere, fast, and is living with both the gains and the risks of that choice. Europe chose to build the rules first and protect its citizens, and is living with the slower diffusion that follows. Which choice looks wiser in 2030 is the question neither the data nor anyone honest can yet answer, but the direction of travel, for now, runs in China’s favor.

Common questions about China’s AI deployment and the gap with Europe

Author: Jan Bielik
CEO & Founder of Webiano Digital & Marketing Agency

This article is an original analysis supported by the sources cited below

State Council approves AI Plus initiative to boost industry adoption Trivium China’s breakdown of the August 2025 AI Plus policy, including the sector adoption targets for 2027, 2030, and 2035 that anchor China’s national deployment strategy.

China’s approach to developing its AI industry A RAND Corporation assessment of how China funds and directs its AI sector, including the state investment funds and the expectation that waste accompanies scale at the application layer.

The great AI race: China’s approach to developing its AI industry APCO Worldwide’s overview of China’s national funds, enterprise base, and the policy logic linking the 2017 plan to the current build-out.

China artificial intelligence market size and forecast Market research tracking the size of China’s AI market and its projected growth trajectory into the early 2030s.

Chinese open-source LLM models A survey of the leading Chinese open-weight models, including DeepSeek, Qwen, GLM, Kimi, and others, with notes on their global download growth.

The best Chinese open-weight models An analysis of how far China’s best open-weight models trail the leading proprietary systems, and why the gap is narrower than often assumed.

USA vs China AI and LLM statistics Comparative figures on AI adoption, enterprise open-source use, manufacturing penetration, and the deployment metrics that distinguish the two economies.

Transforming industries with AI: lessons from China A World Economic Forum account of how digital twins, predictive maintenance, and generative AI are reshaping Chinese manufacturing and other sectors.

Humanoid robot installations and the Chinese lead Reporting on the scale of China’s humanoid robot sector, the leading firms, and the national action plan targeting widespread deployment by 2027.

China’s Baidu says weekly robotaxi rides hit 250,000 CNBC’s report on Apollo Go reaching driverless ride volumes comparable to Waymo, with detail on its cities, safety record, and global expansion.

Baidu’s Apollo Go leads global autonomous driving with 17 million orders Coverage of Apollo Go’s cumulative ride totals, cost trajectory toward profitability, and the partnerships extending its model abroad.

China’s censorship and surveillance were already intense, and AI is turbocharging them CNN’s report, drawing on Australian Strategic Policy Institute research, on AI-driven surveillance, judicial recommendation systems, and content control at population scale.

Evaluating DeepSeek and large language models in clinical settings A research paper examining the performance and limitations of DeepSeek-based models in medical reasoning and clinical decision support.

Artificial intelligence in Chinese hospital systems A peer-reviewed study on the integration of AI diagnosis and decision-support tools across Chinese tertiary hospitals and the regulatory questions it raises.

AI-assisted diagnosis and automation bias in healthcare Research examining the risks of liability gaps and automation bias as hospitals adopt AI diagnostic tools faster than oversight frameworks mature.

The limits of chip export controls A CSIS analysis of how US semiconductor export restrictions have shaped, rather than stopped, China’s AI development and the compute gap that remains.

DeepSeek, Huawei, and export controls A CSIS examination of China’s domestic chip efforts, the persistence of foreign hardware dependence, and the strategic implications of DeepSeek’s efficiency work.

China’s bid for smart cities An Observer Research Foundation analysis of city-brain platforms, smart-grid coordination, and AI-managed urban infrastructure in China.

China’s AI education experiment Reporting on Beijing’s compulsory AI curriculum, the national smart-education platform, and the scale of the school-level rollout.

Inside China’s AI education push An in-depth look at the tiered AI curriculum, the smart-education infrastructure, and the ideological dimension of the program.

China’s pivot on global AI governance A Carnegie Endowment analysis of China’s diplomatic strategy on AI rules, including the World AI Cooperation Organization and outreach to the Global South.

Global AI Governance Action Plan The Chinese foreign ministry’s statement on its proposed global AI governance framework, presented at the World Artificial Intelligence Conference.

Europe needs a strategy to close the artificial intelligence compute gap A Bruegel analysis of Europe’s compute and demand-aggregation problems and the two-track strategy it recommends for staying relevant.

USA, Europe, and China: the best AI models compared A comparison of model output and investment across the three regions, including the figures on notable models and private AI investment.