Opening: A Signature in 2017, a Shock in 2025
July 20, 2017. The State Council offices in Beijing. That afternoon, officials distributed a thirty-two-page document titled Xin Yidai Rengong Zhineng Fazhan Guihua — the New Generation AI Development Plan. A red state seal adorned the cover. The Premier's signature was affixed inside.
One number mattered above all others: 10 trillion yuan. Roughly $1.4 trillion. The target date: 2030. At the time, Silicon Valley was still basking in the afterglow of AlphaGo. Attention was fixed on autonomous vehicles and the iPhone X. Few people in Washington read the document carefully. Fewer still outside Beijing bothered to ask what it actually meant.
Eight years later, in January 2025, the answer arrived.
A Chinese AI company called DeepSeek published the specifications for its new models, V3 and R1. The numbers were strange. Where American Big Tech AI labs were pouring tens of billions of dollars into comparable models with training costs running into the hundreds of millions, DeepSeek V3's GPU compute cost was $6 million. Its API pricing was one-twenty-seventh that of OpenAI's o1. On performance benchmarks, it matched or exceeded GPT-4 in several categories. And then the company released the model weights to the world for free.
NVIDIA's stock price plunged 17% immediately after the announcement. Roughly $600 billion in market capitalization evaporated. A question rippled through Silicon Valley: "Was it right to invest hundreds of billions of dollars in GPU infrastructure?"
Trace the epicenter of that shock back far enough, and you arrive at the same thirty-two-page document.
A question presents itself. How did China get here? The simple explanation ("the state poured money in") is insufficient. The Soviet Union concentrated state resources on military technology too, and that system ultimately lost the competition. Something different is at work in China's AI strategy.
A more fundamental question follows. Can a state design the direction of innovation? What is the structural difference between America's model, where the market generates innovation on its own, and China's model, where the state draws the blueprint and corporations execute? And who benefits from that difference, and under what conditions does the advantage reverse?
In this chapter, the reader follows the evolution of China's technology strategy from 2006 to 2025, from the slogan of indigenous innovation (zizhu chuangxin) to its transformation into concrete industrial policy, and then into an AI-specific national strategy. We examine the triangular structure formed by the state, state-owned enterprises, and private tech companies, and analyze how this system made an innovation like DeepSeek possible, or whether it happened in spite of it.
Section A: Twenty Years of Design — From Indigenous Innovation to an AI-Architected State
A Journey That Started at 1.4%
In the spring of 2006, China's State Council issued the National Medium- and Long-Term Plan for the Development of Science and Technology (Guojia Zhongchangqi Kexue he Jishu Fazhan Guihua Gangyao), a document mapping the direction of technological development over the next fifteen years, from 2006 to 2020. That year, China's R&D intensity — research and development spending as a share of GDP — stood at 1.4%. The United States was at 2.7%. Japan was at 3.2%. The gap was wider than the numbers suggested: Chinese technology firms were trapped in a structure of licensing foreign technology, importing foreign machinery, and copying foreign software.
The document's central term was indigenous innovation (zizhu chuangxin), a declaration that China would create its own technology rather than importing it from abroad. It looked like a slogan. But this concept would become the underlying grammar of Chinese technology policy for the next twenty years.
At the time, virtually no foreign company took the declaration seriously. China was still "the world's factory." Cupertino designed the iPhone; Shenzhen assembled it. The division of labor between design and execution seemed natural and durable.
The planners in Beijing did not accept this division as permanent.
The Real Legacy of Made in China 2025
In 2015, China's State Council issued Made in China 2025 (Zhongguo Zhizao 2025), a strategy for upgrading the manufacturing sector. Ten priority industries were identified, with AI, semiconductors, and robotics among them. The targets were specific: domestic sourcing of 70% of core components and materials by 2025.
That target was not met. China's semiconductor self-sufficiency rate appeared to reach 50% by total domestic production in 2025, but when measured solely by Chinese companies' own design and manufacturing, the figure was 19-23% (IC Insights/TechInsights). Equipment self-sufficiency stood at 13.6%. American and Western media routinely concluded that Made in China 2025 had failed.
That conclusion was wrong. It missed what mattered.
The real legacy of Made in China 2025 was not target achievement but direction-setting. From the moment the policy was announced, hundreds of Chinese companies began treating domestic sourcing as a strategic priority. The perception that dependence on foreign technology constituted a risk was etched into the minds of corporate leadership. And that perception drove supply chain internalization before American export controls ever became reality.
The irony is that the most effective outcome of Made in China 2025 may have been convincing the United States to take Chinese technology seriously as a threat. The policy was one of the direct triggers for the Trump administration's first-term trade war against Huawei. Alarm led to sanctions, and sanctions accelerated the very domestic sourcing the policy had envisioned.
The plan drew energy from its own backlash.
AIDP: The State Declares the Direction of AI
The 2017 New Generation AI Development Plan (hereafter AIDP) was not merely an industrial policy. It was a historic document in which the State Council explicitly designated AI as a "national strategic technology."
Targets were set in three stages. By 2020: world-leading levels in key AI technologies. By 2025: global best in AI theory, technology, and applications across the board. By 2030: world number one in AI industry scale and technological capability, a "global hub of AI innovation." The 2030 target was the 10 trillion yuan figure.
There are two ways to read these targets.
The first: take the numbers at face value. This produces the question, "Were they achievable?" Did China meet its 2020 targets? Its 2025 targets? Read this way, AIDP is a scorecard to be graded.
The second: read it as a declaration. When a state announces "this is the direction we are heading," what effect does that have on companies, investors, researchers, and students? Read this way, AIDP is a signal that redirects resource allocation.
The first reading deserves its answer on the record. The 2020 target — "world-leading levels in key AI technologies" — was partially achieved. China's share of highly cited AI papers reached 23.2%, first in the world (Stanford HAI). AI patent filings were also the world's highest. But in foundation models and cutting-edge chip design, the country trailed the United States. As for the 2025 target, DeepSeek demonstrated a technological leap, but whether "global best across all applications" has been achieved remains contested.
The second reading, however, captures AIDP's real impact more accurately. From its publication in 2017, China's universities, companies, and local governments began realigning around AI as a priority. AI-related enrollment at Peking University and Tsinghua University expanded. Local governments competed to establish AI industrial parks. Entrepreneurs founded AI startups. The results are embedded in the workforce numbers. According to Tsinghua University's China AI Development Report 2025, China's core AI workforce grew from fewer than 10,000 in 2015 to approximately 52,000 by 2024. The rate of increase did not flatten around the time AIDP was issued — it steepened. The state's direction-setting pulled private-sector resource allocation into its orbit.
The difference from America's VC-driven innovation model is structural. In the United States, the market decides which technology becomes "the next big thing." In China, the state sets the direction first, and resources flow toward it. Both approaches can work. The real question is which one proves more effective, and under what conditions.
AI+: The Latest Edition of the Strategy
In August 2025, China's State Council issued the Opinions on Accelerating the Promotion of AI+ Actions (Guanyu Jiakuai Tuijin Rengong Zhineng+ Xingdong de Yijian, State Council Document No. 11 of 2025). It was the direct successor to the 2015 "Internet+" (Hulianwang+) strategy. Where Internet+ had grafted the internet onto traditional industries, AI+ aimed to deeply integrate AI into them: manufacturing, agriculture, education, healthcare, energy, and transportation alike.
Summarize the policy evolution from 2006 to 2025 in a single line: from catching up to leading, from importing to self-reliance, from execution to design. And at every stage, the state set the direction.
The key point is that this evolution was not improvised. The DeepSeek shock was not a sudden event. It was the accumulated result of twenty years of policy design, surfacing at a specific moment in time.
Section B: The Triangular Structure — State, State-Owned Enterprises, and Private Tech
The Architecture of Three Actors
Calling China's AI innovation "state-led" is half right. The more precise term is a triangular structure.
The state and the Party occupy one vertex. They set direction, allocate resources, design regulations, and grant access to data. State-owned enterprises (SOEs) occupy the second vertex. They serve as stable procurement clients, collection channels for public data, and conduits through which state capital flows. Private tech companies occupy the third. They execute the actual innovation, compete commercially, and open overseas markets.
This triangular structure differs from both America's purely private VC model and the Soviet Union's central planning model. Each vertex depends on the other two.
How does the United States compare? OpenAI reached a $500 billion valuation on the strength of purely private capital. Anthropic recorded a $380 billion valuation. There were no state-designated platforms, no government subsidies. Yet as of 2026, the combined AI capital expenditure of the Big Four — Amazon, Alphabet, Microsoft, and Meta — reached $635-665 billion, demonstrating that private capital has attained state-level scale. The boundary between state and private is blurring in terms of resource magnitude.
But the structure remains different. American Big Tech answers to shareholder value. Chinese SOEs answer to national strategy. This difference determines the direction of resource allocation.
Consider the financial scale of the Chinese triangle in data. Huawei invested 179.7 billion yuan in R&D in 2024, 20.8% of its 862.1 billion yuan in revenue (Huawei Annual Report). Its cumulative R&D investment over the past decade exceeded 1.249 trillion yuan. Alibaba announced in February 2025 that it would invest 380 billion yuan (approximately $53 billion) in AI and cloud infrastructure over the following three years. Baidu's AI cloud revenue reached 30 billion yuan in 2025, with AI-related revenue posting triple-digit growth for nine consecutive quarters (Baidu FY2025 earnings). In absolute terms, these figures trail U.S. Big Tech CapEx. But the triangular structure aggregates state capital and private investment, so comparing individual companies alone does not capture the full picture.
The scale of state capital shifts the frame. As of 2025, more than 2,100 government guidance funds (zhengfu yindao jijin) had been established in China, with a combined target size of 13.5 trillion yuan ($1.86 trillion) (Zero2IPO/Qingke). Government money serves as seed capital designed to attract private investment. The financial sector has joined in as well. In March 2025, Bank of China announced it would extend 1 trillion yuan in lending to AI and science-technology sectors over the next five years. State-owned banks supply capital aligned with national strategy. In the United States, venture capital and public markets serve this function. In China, government guidance funds and state-owned banks do. The sources and flows of capital differ; the scale is becoming comparable.
How the National AI Open Innovation Platforms Work
One of the key instruments the Chinese government chose is the National AI Open Innovation Platform (Guojia Rengong Zhineng Kaifang Chuangxin Pingtai) designation system, which identifies leading companies in specific domains and provides them with concentrated support.
The roster of designated companies maps the terrain of China's AI ecosystem. Baidu was designated for autonomous driving AI. Alibaba for smart cities. Tencent for medical imaging AI. iFlytek (Keda Xunfei) for voice AI. Huawei for AI computing infrastructure.
What do designated companies receive? Access to public data, priority in government procurement, and regulatory sandbox benefits.
Baidu's experience illustrates how this structure works in practice. In 2025, Baidu's AI cloud infrastructure revenue reached approximately 20 billion yuan (up 34% year-over-year), while AI application revenue surpassed 10 billion yuan. Total revenue (129.1 billion yuan) declined 3% as the legacy advertising business contracted, but AI cloud had emerged as the new growth engine. After being designated as the autonomous driving AI platform, Baidu gained access to city-wide traffic data across twenty-two cities, including Wuhan, Chongqing, Beijing, Shanghai, Shenzhen, and Guangzhou. Hundreds of thousands of vehicles, thousands of traffic signals, and the movement patterns of millions of people fed into Baidu's autonomous driving training data.
What would this process have looked like in the United States? Individual agreements with city governments. Private land transit permits. Navigating personal data protection laws. Overlapping regulation at the federal, state, and municipal levels. Compare the time it took Tesla to collect autonomous driving data across the United States with the time it took Baidu to complete the same task across twenty-two Chinese cities, and the advantage of the triangular structure becomes clear.
The Paradox of the State-Owned Client
The triangular structure has its weaknesses. The most serious emerges when SOEs serve as clients.
For Chinese AI companies, SOEs are essential customers. When large state-owned enterprises in the power, telecommunications, financial, and transportation sectors commission AI systems, Chinese AI firms receive effectively non-competitive contracts. Foreign companies face disadvantages in bidding, and competition among domestic private firms is limited.
This structure guarantees revenue in the short term. Over the long term, it suppresses innovation.
State-owned clients value relationships over performance. "Is this AI 10% more accurate than the competitor's product?" matters less than "Has this company built a trusted relationship with us?" For SOE procurement officers who are not exposed to global competition, performance differentials are abstract. If the AI works well enough, it is good enough.
In this environment, AI companies are incentivized not to push performance to its limits, but to build products that are compatible with existing systems and easy to maintain. The pace of innovation is governed by relationships, not competition.
The cost of failure operates differently as well. In the American VC model, failure is part of the portfolio. One success out of ten is sufficient. With Chinese state funding, failure carries political liability. When a platform designated by the State Council fails, it directly damages the careers of the officials responsible. Administrators therefore prefer a slightly less successful success over a safe failure. This narrows the radius of innovation.
There is a case where this structural vulnerability shows up in the numbers. China's estimated government AI capital expenditure in 2025 was approximately 400 billion yuan — a substantial portion of total AI-related public spending of 600-700 billion yuan. Much of this funding flowed into state-owned cloud infrastructure and SOE AI adoption projects. Yet the innovation that stunned the world — DeepSeek — was built entirely with the private capital of a hedge fund, untouched by state money. The paradox: the most innovative outcome was not produced by state funding but emerged precisely where state funding did not reach.
One distinction is necessary here. The 400 billion yuan in government AI capital expenditure is a different figure from the "400 billion yuan in core AI industry scale" that the 2017 New Generation AI Development Plan set as its 2025 target. The former is direct government spending; the latter is a scale target for the entire industry, including the private sector. What was the result? According to a State Council press conference in January 2026, China's core AI industry scale surpassed 1.2 trillion yuan in 2025 (State Council Information Office, January 21, 2026). The number of AI companies exceeded 6,000, accounting for 16% of the global total. Smart computing power reached 1,590 EFLOPS. The target was exceeded threefold. The state set the direction; the private sector overshot the goal in execution.
Stanford HAI's AI Index Report 2025 also documents the output of this structure. As of 2023, China ranked first in AI-related academic publications (23.2% of the global total), with a citation share of 22.6%, also first. In AI patents, China held more than 60% of global filings. However, in notable AI models in 2024, the United States had forty compared to China's fifteen. A quantity-quality asymmetry.
This asymmetry extends to infrastructure. According to the same State Council press conference, China's smart computing power (zhineng suanli) reached 1,590 EFLOPS in 2025. CAICT (Zhongguo Xintong Yuan) and IDC project this figure will expand to 2,781.9 EFLOPS by 2028, roughly 40% annual growth. The pace of quantitative expansion is rapid, but the gap in qualitative capability to design frontier models from scratch has not yet closed.
The state-led system is powerful under specific conditions: where long-term investment is required, the direction is clear, and economies of scale are decisive. AI infrastructure, autonomous driving data collection, and smart city construction all fit those conditions. Huawei founder Ren Zhengfei crystallized this distinction in a November 2025 conversation with ICPC participants: "AI invention creates at most one IT company, but application makes a nation strong." His point was that America pursues artificial general intelligence while China creates value by dissolving AI into the efficiency of existing industries. A 0.1% improvement in coal washing precision, multiplied across 4 billion tons of annual output; a 1% gain in blast furnace efficiency, multiplied across 1 billion tons of steel. Not innovation measured by a single company's market capitalization, but innovation dissolved into the productivity of entire sectors (Caixin Global, December 6, 2025).
Yet in domains where direction is uncertain, where rapid experimentation and failure are essential, and where global competition makes you stronger the more you are exposed to it — different results emerge. Foundational AI research, creative applications, and the discovery of new markets fit these conditions. The core strength of the American VC model reveals itself here. Within a structure where failure is tolerated as part of the portfolio, companies like OpenAI, Anthropic, and Mistral compete and push each other forward. When one of the three achieves a breakthrough, the other two immediately sprint in the same direction. Matching this speed with the tempo of state planning is difficult.
Section C: The Regulatory Paradox — How the World's First AI Regulation Protected Innovation
Nine Months
In April 2023, the Cyberspace Administration of China (CAC) published a draft regulation on generative AI services. ChatGPT had launched fewer than nine months earlier, in November 2022. Four months after the draft's release, on August 15, 2023, the Interim Measures for the Management of Generative AI Services (Shengchengshi Rengong Zhineng Fuwu Guanli Zanxing Banfa) took effect. China became the first country in the world to enforce binding regulation on generative AI.
What was happening in the United States at the same time? Through 2025, not a single comprehensive federal AI bill had been passed. The only AI-related federal legislation to clear Congress was the TAKE IT DOWN Act, a narrowly targeted deepfake law. More than 100 AI-related bills had been introduced and were pending. 79% of Americans supported AI regulation — 84% of Republicans, 81% of Democrats. And still, no bill passed. In a system where Big Tech spent $92 million per year lobbying Congress on AI-related issues, the gap between public opinion and legislative outcome refused to close.
Western observers drew an immediate conclusion from this contrast. "China is fast, America is slow. China controls innovation, America protects it." This conclusion misses two things.
The Two Faces of Shield and Censorship
Read the core provisions of China's generative AI regulation, and it becomes clear that this is not a simple case of "innovation suppression." The regulation mandates pre-launch security assessments (anquan pinggu), legality evaluations of training data, and obligations to prevent the generation of false information. And then there is one more clause: generated content must reflect "core socialist values" (shehui zhuyi hexin jiazhiguan).
On August 15, 2023 — the day the regulation took effect — Baidu's ERNIE Bot (Wenxin Yiyan) and Alibaba's Tongyi Qianwen (Qwen) had already passed pre-launch security reviews and launched their services. Tencent's Hunyuan and ByteDance's Doubao followed in quick succession. By September 2025, 538 generative AI services had registered filings (bei'an) with the CAC. According to CNNIC's 56th report, as of June 2025 China had 515 million generative AI users — 47.4% of all internet users had tried generative AI.
On that same day, ChatGPT remained blocked on the Chinese mainland by the Great Firewall. The regulation's effective date was also the date the domestic protection shield was completed.
The logic of this structure is crisp. Foreign AI services cannot (or will not) satisfy the "core socialist values" clause. Chinese companies understand that satisfying this clause is both the qualification for operating in the Chinese market and the pathway to earning government trust. Regulation doubles as a barrier to entry.
A scene from the ground illustrates how this structure operates in practice. In the predawn hours of August 15, 2023, the engineering team at Baidu's Beijing headquarters had been on standby since the previous night. Approval of ERNIE Bot's public pre-launch security review had come through days earlier. The plan was to officially launch the service that morning. On the same day, ChatGPT remained blocked from Chinese access. The timing was no accident — Baidu had spent months in close coordination with the authorities in preparation. The day the regulation took effect was the day market dominance began.
It is also true that America's slow regulation serves a function: protecting the space for innovation. In the United States, startups can launch services without waiting for regulatory approval. When GPT-4 was released, American startups had API-integrated products within two days. In China, the security review comes first.
But the key point is that Chinese AI companies did not grow despite regulation; they grew with it. The process of meeting regulatory requirements was itself a method of securing government trust, and that trust translated into SOE contracts, data access, and government procurement.
The Evolution of Regulation: The Global Standards Game
China's AI regulation did not end in 2023. In April 2025, three national standards for generative AI security were issued. In July, a Global AI Governance Action Plan was released. In August, ten ministries jointly published a draft of AI technology ethics management measures. In September, rules for labeling AI-generated content took effect.
These regulations share a common thread: the ambition to convert "Chinese standards" into global standards. The Global AI Governance Action Plan was a declaration of intent to proactively propose China's AI governance principles in international forums. With Chinese-made AI surveillance technology exported to more than seventy-five countries, China is now proposing the operating standards for that technology on the international stage.
If you view regulation only as a "barrier to innovation," you miss the real game. Whoever proposes international standards first writes the rules of the market. Just as Europe raised the bar for global data governance with the GDPR, China is attempting the same play in AI governance.
Section D: From BAT to BATH+D — The Meaning of Generational Shift
From Platform Economy to AI-Native
The Chinese tech ecosystem of the 2010s was dominated by BAT, an acronym for Baidu, Alibaba, and Tencent. These three companies divided the Chinese internet along the axes of search, e-commerce, and social networking.
By 2025, the landscape had shifted. BAT still existed, but the nature of their dominance had changed. A new generation had arrived.
A new acronym emerged: BATH+D. Huawei commanded hardware, 5G, and AI chips. ByteDance set a new standard for algorithmic recommendation. Pinduoduo (and its global brand, Temu) became a new axis of consumer behavior data. And DeepSeek sent shockwaves through the open-source AI model ecosystem.
This generational shift is not a simple reshuffling of companies. If the first-generation BAT companies were protagonists of the "platform economy," the second generation is evolving into AI-native enterprises. Platforms created marketplaces for transactions. AI-native companies predict, optimize, and replace individual behavior.
Liang Wenfeng's Choice
The DeepSeek story begins in the summer of 2021.
Liang Wenfeng graduated from Zhejiang University's electrical engineering department in 2008 and co-founded High-Flyer (Huanfang), an AI-driven quantitative hedge fund. High-Flyer became one of the leading firms in China's quantitative investment world. The process of running financial algorithms naturally accumulated experience with large-scale GPU clusters and AI model training.
In 2021, Liang Wenfeng stood at a decisive turning point. As American semiconductor export controls tightened on the horizon, he purchased NVIDIA GPUs in bulk before the restrictions took hold. The total number of GPUs secured (including H100 and H800 units) was estimated at 20,000 (SemiAnalysis). In May 2023, he established DeepSeek in Hangzhou.
The structure was different from the start. No VC funding. No obligation to report quarterly results to investors. No pressure to rush a product to market to meet an IPO timeline. Operating on High-Flyer's own capital, the team could focus exclusively on technical problems.
This structure paradoxically became DeepSeek's strength. OpenAI's Sam Altman, upon seeing DeepSeek R1, acknowledged it as "an impressive model, especially for the price." R1's API pricing was one-twenty-seventh that of OpenAI o1 for both input and output.
DeepSeek's efficiency stemmed from three technical innovations. Its Mixture of Experts (MoE) architecture activated only 37 billion of its total 67.1 billion parameters per token, dramatically reducing computational load. Multi-head Latent Attention (MLA) compressed GPU memory usage. FP8 mixed-precision training achieved double the efficiency of conventional BF16 training.
These innovations share a common origin: they emerged from constraint. An environment in which sufficient H100 GPUs could not be purchased paradoxically forced the maximization of efficiency. Sanctions did not kill innovation. They gave birth to a different kind.
A caveat is necessary. The frequently cited "$6 million training cost" refers only to GPU pre-training costs. According to SemiAnalysis, DeepSeek's total server infrastructure capital expenditure was $1.6 billion, of which $944 million went to cluster operations. The simplification "$6 million = total cost" is a misunderstanding. Nevertheless, the training efficiency relative to comparably performing American models was demonstrably superior.
The Strategic Logic of Open Source
DeepSeek's decision to release its model weights represents a tactical pivot in China's AI strategy.
Western observers were puzzled. While the United States was trying to prevent technology leakage through export controls, why were Chinese AI companies giving their models away for free? The reasons operate on multiple levels.
Ecosystem dominance is the first. Android overwhelmed iOS in smartphone market share through an open-source strategy. If DeepSeek weights become the default tool for developers worldwide, a Chinese company effectively controls the AI infrastructure layer.
Cost pressure on American models is the second. If DeepSeek R1 delivers performance equivalent to OpenAI o1 at one-twenty-seventh the price, companies around the world will choose DeepSeek on cost alone. It is the most effective way to attack OpenAI's revenue model.
Regulatory circumvention is the third. Open-weight models are not subject to U.S. AI export controls. Weight files are data; code is software. As of early 2026, no regulation exists to block them.
The results of this strategy are visible in the numbers. As of early 2026, nine of the top ten global open-weight models were Chinese-made. Beyond DeepSeek, Alibaba's Tongyi Qianwen (Qwen) and ByteDance's Doubao series were widely used on global platforms such as HuggingFace. Chinese open-source AI's global market share surged from 1.2% at the end of 2024 to 30% by August 2025.
AI Patents: The Light and Shadow of 60%
China's AI patent holdings also boast impressive numbers. China accounts for more than 60% of global generative AI patents, with 38,210 registrations, more than six times the United States' 6,276.
But reading these numbers at face value misses something important.
Of the 38,210 patents, only 7.3% were filed internationally; the remaining 92.7% are Chinese domestic patents. Domestic Chinese patents have lower registration thresholds and carry no legal protection outside China. A single American patent is often filed simultaneously in the United States, Europe, and Japan. Most Chinese patents are valid only in China.
A more revealing metric is citation count. The average citation count for Chinese AI patents is one-seventh that of American AI patents. This indicates that Chinese patents are concentrated in applications and improvements of existing technology, with relatively few foundational research patents that invent new principles on the scale of the Transformer architecture.
DeepSeek has cracked this picture. DeepSeek's MoE refinements, MLA, and FP8 training techniques were innovations at the foundational architecture level, not mere applications. The era in which one could flatly assert that China lacked the capacity for foundational innovation has begun to close.
Connection to Volume 1: The State Version of the Design Layer
In Volume 1, the reader met Richard Arkwright. He did not invent the spinning machine. James Hargreaves's Spinning Jenny and the roller spinning technology of Lewis Paul and John Wyatt already existed. Arkwright's true innovation was not a machine. It was the design of a new production architecture: the factory system. Organizing workers by the clock, standardizing the process from raw material input to finished goods output, combining capital and labor in a new way.
Extend Volume 1's central dichotomy (the execution layer and the design layer) to the level of nations, and the picture changes.
The United States is currently a design-layer nation in AI. The AI model architectures of GPT, Claude, and Gemini. NVIDIA's GPU designs. The CUDA software ecosystem. The AI training frameworks of PyTorch and TensorFlow. The entire world uses tools designed in America.
China is currently, in substantial part, an execution-layer nation in AI. It applies, deploys, and distributes the tools America designed. Fastest, cheapest, and widest. AI customer service, smart cities, e-commerce recommendation algorithms, facial recognition systems.
But an important qualification attaches to this dichotomy.
Arkwright himself was initially a more efficient executor of existing technology. His real innovation emerged when he moved from execution to design. DeepSeek may be the signal that China has begun the same migration.
There is another difference. Arkwright was exposed to market competition. If a more efficient competitor appeared, he could be displaced. In the nineteenth-century textile industry, more efficient firms did in fact emerge and push aside early leaders. The Chinese state's control of the design layer is relatively insulated from market competition.
This is simultaneously a strength and a vulnerability.
The tension between state design and private execution surfaces in the statements of business leaders as well. CATL founder Robin Zeng (Zeng Yuqun) declared at Davos in 2024: "We are fighting climate change. Whatever the geopolitical issues, we need to find a way." But ahead of APEC in 2025, he struck a different tone — emphasizing that Chinese corporate technology innovation was still in its early stages and that substantial investment in AI hardware and software was needed. Optimism facing the global stage coexists with sober diagnosis facing the domestic audience. The executing actors within the triangular structure accept the state's blueprint even as they internally recognize its limits.
The strength first. It enables long-term investment. Strategies spanning ten or twenty years can be executed without market volatility. AIDP was issued in 2017 and pursued consistently through 2025 precisely because of this stability.
The vulnerability lies on the other side. Protection without competition can lower the bar for innovation. When the state sets a direction, innovations outside that direction struggle to attract resources. When the direction proves wrong, the cost of correction is high. And most fundamentally, the state does not always choose the right direction.
In the language of Volume 1: the Chinese state is attempting to play Arkwright's role, not at the factory level, but at the national level. Not inventing the technology itself, but designing the system in which technology operates. The problem is that Arkwright had the market to validate his designs. The mechanism that validates the state's designs is different, slower, and sometimes does not work at all.
Transition: The Next Question
China's AI strategy can be stated in a single sentence: not developing the technology, but designing the system in which the technology operates.
The triangular structure of state, SOEs, and private tech companies. A regulatory framework that doubles as a barrier to entry. An open-source strategy for ecosystem capture. Every one of these elements accumulated over twenty years to produce the DeepSeek shock.
But this system requires fuel. Data.
The payment flows of WeChat Pay. The movement patterns of Didi Chuxing. The viewing records of Douyin. The facial data collected by 700 million surveillance cameras. This is the fuel of the Chinese AI system.
A central question remains. Does the scale and density of this data actually translate into AI performance advantage? What pitfalls lie along that path? Examining the reality of that fuel is the next step.
Next chapter: Ch. 9 — The Data of 1.4 Billion: Why China Dominates AI Application