Opening: The Monday Morning Shockwave
January 27, 2025. Monday, 9:30 a.m.
The opening bell rang at the New York Stock Exchange. Within three minutes, NVIDIA's stock began to fall. Gently at first. Five percent. Seven percent. Then it accelerated.
The closing bell at 4 p.m. The final decline: -17%.
The market capitalization that vanished in a single day was $600 billion. An amount rivaling Saudi Arabia's entire gross domestic product, evaporated in twenty-four hours. It was the largest single-day market cap loss for any individual stock in history.
The cause was not an American company. It was a single open-source AI model, quietly released over the weekend by a startup in Hangzhou, Zhejiang Province. DeepSeek-R1.
The numbers were straightforward. Benchmark performance matched OpenAI's latest reasoning model, o1. But the API cost was twenty to fifty times cheaper. DeepSeek used less than one-tenth the GPUs of America's leading AI labs.
Wall Street traders asked themselves: "Was our strategy of pouring hundreds of billions of dollars into GPU purchases the right one?"
The combined AI capital expenditure forecast for the Big Four tech companies in 2026 ranged between $635 billion and $665 billion. NVIDIA sat at the center of that supply chain. DeepSeek had shaken the premise underneath it.
But reading the DeepSeek shock as a single company's success story misses the point.
A more fundamental question lurked beneath it. The United States had tried to contain Chinese AI through export controls, but had those controls instead forced a different kind of innovation? Do constraints suppress innovation, or catalyze it? In the language of Volume 1: what kind of system were China's "discerning" building from within the walls of sanctions?
In this chapter, the reader meets four types of Chinese discerning.
DeepSeek's Liang Wenfeng turned GPU scarcity into fuel for efficiency innovation. Huawei is building its own AI chip ecosystem as an answer to American sanctions. ByteDance's Zhang Yiming designed an algorithm that captures the attention of the entire world. Pinduoduo's Colin Huang built a system that connects Chinese manufacturers directly to global consumers.
We examine how each of these systems works, and what limits are built into them.
Section A: DeepSeek — Innovation Designed by Constraint
Liang Wenfeng's 2021 Decision
Fall 2021. An office in downtown Hangzhou. Liang Wenfeng was reading a document.
A graduate of Zhejiang University's electrical engineering program, Liang had co-founded and was running the quantitative hedge fund High-Flyer. Building financial algorithms was his day job. But in that process, he had accumulated deep expertise in operating large-scale GPU clusters. He was a rare figure who understood both algorithms and hardware simultaneously.
The document he was reading was a draft of U.S. Department of Commerce export regulations. It outlined a coming ban on exports of NVIDIA's high-performance AI chips to China.
Liang moved before the regulations took effect. Using High-Flyer's capital, he bulk-purchased NVIDIA GPUs. The exact quantity was never disclosed, but DeepSeek's subsequent holdings were estimated at roughly 20,000 units combining H100 and H800 chips (SemiAnalysis).
In May 2023, he founded DeepSeek. He took no venture capital. The company ran on High-Flyer's own funds. The staff numbered a few dozen.
A caveat: "No VC funding" does not mean "zero state support." In December 2023, DeepSeek received designation as a "National High-Tech Enterprise" from Zhejiang Province authorities. The designation carries tax incentives, government subsidies, and research grants (Lawfare, 2025). The Hangzhou West Science and Technology Innovation Corridor, where DeepSeek is based, is a zone targeted to become China's Silicon Valley, into which massive state subsidies flow. Zhejiang Province distributed computing vouchers worth up to approximately $300,000 per company to AI startups, offsetting cloud costs.
By early 2024, Liang had been invited to provide policy counsel to Premier Li Qiang. DeepSeek's core R&D was funded by private capital, but the soil in which that private capital could operate was cultivated by the state. A classic triangular arrangement.
Then, in January 2025, DeepSeek released R1.
Something stands out about Liang's trajectory. He was not someone who had built his reputation in the AI research community. He was neither an academic star nor a Google Brain alumnus. He was a quant who built financial algorithms. That background became an advantage.
In the world of quantitative funds, cost efficiency is the condition of survival. The competitive edge lies in producing more precise predictions with less computation. Unlimited resources are not an option. Liang carried this mindset directly into AI training. While Silicon Valley raced toward "bigger," he raced toward "more efficient."
In a 36Kr interview (second half of 2024), he explained why he refused VC investment: "Money has never been the problem for us; bans on shipments of advanced chips are the problem." On the relationship between investment scale and innovation, he was blunter: "More investment doesn't produce more innovation. If it did, large corporations would have monopolized innovation." His hiring criterion was not experience but passion and curiosity: "There are no wizards here. Most of us are fresh graduates or doctoral students from top universities." The organizational structure was not hierarchical: "DeepSeek is entirely bottom-up. We don't pre-assign roles; the division of labor emerges naturally." — Liang Wenfeng, 36Kr interview, 2024 (China Talk English translation, January 2025)
Then came his diagnosis of the fundamental limit of Chinese AI innovation: "The real gap is the difference between originality and imitation. If this doesn't change, China will remain a follower forever." — Liang Wenfeng, 36Kr interview, 2024. This self-diagnosis is what distinguishes Liang from other Chinese tech executives. Instead of declaring victory, he acknowledged the gap.
Because the starting point was different, the destination was different.
The Engineering of Scarcity: MoE, MLA, FP8
Understanding DeepSeek's innovation requires going back to the problem formulation.
American AI labs had been competing under the formula: "more GPUs for bigger models." Meta used 30.8 million GPU-hours to train its Llama models. OpenAI and Google do not disclose their figures, but their consumption is significantly higher.
That path was not available to DeepSeek. They did not have enough GPUs.
Constraint reshaped design.
The GPU compute cost for training DeepSeek-V3 was $5.58 million (2.78 million GPU-hours). The additional training cost for DeepSeek-R1, which applied reinforcement learning on top of V3, was considerably less. Even V3's training cost alone was one-eleventh of Meta's comparable model (30.8 million GPU-hours).
Framing this as "they built GPT-4-class AI for $6 million" is inaccurate. The $6 million covers only the pretraining GPU compute cost. DeepSeek's total server CapEx is estimated at approximately $1.6 billion by semiconductor research firm SemiAnalysis. But even $1.6 billion is one four-hundredth of the Big Four's combined 2026 CapEx of $635 billion to $665 billion.
The efficiency gap is real. Its core was three technical innovations.
The first is Mixture of Experts (MoE) architecture. Conventional large language models activate their entire set of billions of parameters for every input. MoE creates multiple specialized smaller networks (experts) and selectively activates only the experts relevant to each input. Energy consumption and compute costs drop dramatically.
The second is Multi-head Latent Attention (MLA). Conventional attention mechanisms must maintain massive Key-Value caches in memory at each processing step. MLA compresses these caches into condensed latent vectors. GPU memory usage falls and processing speed increases.
The third is FP8 training. The standard for AI model training had been 32-bit (FP32) or 16-bit (BF16) floating-point arithmetic. DeepSeek made 8-bit (FP8) training practical. Lowering precision speeds up computation, but errors accumulate. That was the known problem. The DeepSeek team developed fine-grained numerical stabilization techniques that made FP8 viable.
These three innovations are not simple cost reductions. They represent the design of a new training architecture. Resource constraints reshaped design principles.
What the Cost Gap Means
DeepSeek-R1's API pricing is $0.55 per million input tokens and $2.19 per million output tokens. OpenAI's o1 charges $15 for input and $60 for output. A 27x difference on both input and output. In the thinnest use cases, the gap approaches 50x.
This price gap is not simply one company's margin strategy. It is a question aimed at the entire cost structure of the American AI industry.
The Big Four tech companies will spend $635 billion to $665 billion on AI infrastructure in 2026 alone. The premise of this investment is the conviction that "more compute produces better AI." If that conviction is correct, DeepSeek is merely a temporary efficiency optimization. If that conviction is wrong, the Big Four are standing in the middle of the largest overinvestment in history.
NVIDIA CEO Jensen Huang responded quickly. DeepSeek used NVIDIA GPUs, he noted, and deployed the logic of the Jevons Paradox — that more efficient algorithms actually increase AI demand. When price falls, usage rises, and GPU demand rises with it.
OpenAI CEO Sam Altman acknowledged DeepSeek as "an impressive model." He added that "the U.S. will continue to build better models." Google DeepMind CEO Demis Hassabis assessed the gap between the two companies at "a few months."
A few months. A gap once measured in years had become a gap measured in months.
Which side is right remains unresolved. DeepSeek reshaped the competitive terrain. That much is certain.
The Historical Pattern: Design Born of Scarcity
History has repeated this pattern.
The onboard computer of Apollo 11 in 1969 had four kilobytes of memory. Under that constraint, software engineers invented techniques to compress the lunar landing algorithm into four kilobytes. In the 1980s, RISC (Reduced Instruction Set Computer) architecture simplified processor instructions. The simplification itself improved performance.
Constraint simplifies design. Simplification improves efficiency. Efficiency opens new possibilities.
Liang Wenfeng reproduced this pattern within the GPU drought. The constraints America imposed forced Chinese AI teams to confront a question the American AI industry had never bothered to ask: "Can we build the same thing using less?"
The answer to that question was DeepSeek-R1.
Section B: Huawei's Counteroffensive — The Ascend Chip Ecosystem
A Night at the SMIC Shanghai Fab
On the outskirts of Shanghai, inside the SMIC foundry. Vacuum chambers spin through the night.
To produce the same circuits, this factory repeats the lithography and etching process four to eight times. What ASML's extreme ultraviolet (EUV) lithography machines engrave in a single pass, DUV (deep ultraviolet) equipment carves in multiple passes. The process is called multi-patterning.
The cost is three to five times higher. Yield is half. Producing the same chip requires more than twice the wafers TSMC needs.
They do not stop.
Using this process, SMIC achieves 7-nanometer circuitry. It is the workaround China developed after the United States blocked Huawei's access to EUV equipment in 2019. Building 7nm without EUV was considered impossible. SMIC chose to make the impossible slowly.
The Architecture of the Ascend Ecosystem
What Huawei is building is not a single chip. It is an ecosystem.
NVIDIA's AI dominance does not stem from GPU hardware alone. It stems from CUDA, a software platform. AI researchers have been writing code on CUDA for more than a decade. As of 2025, roughly 50% of the world's AI developers are inside China, but the code they write runs on CUDA. Even when America blocks the hardware, the inertia of the software ecosystem remains.
Huawei's objective is to replace that ecosystem.
It is structured in three layers. The hardware is the Ascend 910C chip. The software framework is MindSpore. The computing architecture layer is CANN (Compute Architecture for Neural Networks). The structure maps precisely onto NVIDIA's GPU + CUDA + cuDNN stack.
The production figures reveal the scale of the ambition. The 2025 production target for the Ascend 910C is 600,000 units. For 2026, the target rises to 1.6 million dies. Combined output of Chinese domestic AI accelerators from Huawei and Cambricon is estimated to exceed one million units in 2026.
The Hard Limit: What 40% Yield Means
The limits, however, do not hide.
Ascend chip yield on SMIC's 7nm process runs at about 40%. TSMC's yield at the same node exceeds 90%. Out of every 100 wafers produced, only 40 yield usable chips. The remaining 60 are defective. The cost structure is fundamentally unfavorable.
The developer ecosystem problem is even larger. There is no compatibility with CUDA. Running existing AI code on Ascend requires large-scale rewrites. The CUDA-based libraries and codebases built up by AI researchers worldwide do not function without NVIDIA GPUs. Huawei is fighting against this software inertia. But the trajectory of that fight has begun to shift. As of late 2024, the Kunpeng-Ascend ecosystem had attracted over 8,500 partners and 6.65 million developers (Huawei 2024 Annual Report). That same year, Huawei released CANN 8.0 for Ascend heterogeneous compute architecture and launched the openMind application support kit, accelerating ecosystem expansion.
Two forces continue to push the Ascend ecosystem forward.
One is the Chinese government's "domestic AI chip procurement priority" policy. The state guarantees a market. Major internet companies like Baidu, Alibaba, and Tencent have begun adopting Ascend to meet government procurement requirements. This is a market created by institutional mandate, not technological superiority.
The other force is DeepSeek V4. DeepSeek's next-generation model was released optimized not only for NVIDIA chips but also for Huawei and Cambricon hardware. Chinese AI algorithms have begun optimizing for Chinese-made hardware. Software has started pulling hardware toward it, a gravitational effect.
An Accelerating Roadmap
Huawei's Ascend roadmap continues.
In Q1 2026, the Ascend 950PR launched, designed specifically for prefill and recommendation workloads. In Q4 2026, the Ascend 950DT arrives, targeting decoding and training. In Q4 2027, the Ascend 960 is planned, with the goal of doubling compute, memory, and interconnect performance over the 950 series.
Reaching TSMC's 90% yield is not possible in the near future. But building a functioning ecosystem without TSMC is a different matter. An ecosystem that operates within China, among Chinese companies — for that, 40% yield is a tolerable condition.
The trajectory becomes visible in SMIC's production ramp-up figures. Monthly advanced-node production capacity was 45,000 wafers in 2025. The plan calls for 60,000 in 2026 and 80,000 in 2027. Slow, but the direction is set.
When the United States sanctioned Huawei in 2019, the objective was to contain China's advanced semiconductor capabilities. Six years later, China is manufacturing 7nm chips without EUV. Have the sanctions failed, or are they succeeding more slowly than expected? Even within the United States, the answer to this question is contested.
America's goal was to stop Chinese AI. China's goal is to build AI that works without America. The two goals do not collide.
Section C: ByteDance and Pinduoduo — The Triumph of Global AI Applications
Congressional Hearing, 2025
- A congressional hearing in Washington, D.C.
American lawmakers posed their questions. ByteDance, the company operating TikTok, is a Chinese enterprise. Could the Chinese government access American user data? Was the algorithm manipulating the information environment of American citizens?
ByteDance's legal team came prepared. They proposed an ownership separation of the algorithm. A joint venture with the American corporation Oracle would process U.S. user data exclusively on American servers. The proposal included technical measures to sever the Chinese parent company's access to the algorithm.
Negotiations dragged on.
When the U.S. government tried to expel the "Chinese algorithm," it discovered one thing. The habits of 170 million American users could not be expelled. Users proved stronger than algorithms.
Mining Attention as a Resource
Apply Volume 1's "discerning" framework to Zhang Yiming.
In Volume 1, the discerning were those who built systems others "had no choice but to use." When Richard Arkwright built the water frame, hand spinners had no choice but to stand before his machine. When Jensen Huang built CUDA, AI researchers could no longer work without that platform.
What Zhang Yiming built is TikTok's recommendation algorithm. It is the most sophisticated AI application ever designed to exploit the human dopamine system. The algorithm knows what users want to see before they know it themselves. Eye movements during 0.5 seconds of watching a video, scroll speed, whether they replay — these signals determine the next video in real time.
The results show in the numbers. In 2025, TikTok ranked second in U.S. App Store downloads. Even under the government's "sell or be banned" pressure, its user base grew. An app used by 170 million people had become a system in itself.
In the language of Volume 1, what Zhang Yiming built resembles the factory system of the eighteenth century. Just as Arkwright's factory standardized yarn production, TikTok's algorithm standardized the consumption of human attention. Just as hand spinners had to synchronize their bodies to the rhythm of the machine, TikTok users synchronize their brains to the rhythm of the algorithm. There is one difference. The hand spinners had to go to the factory. TikTok enters the user's bedroom through a smartphone.
ByteDance's valuation is estimated at over $300 billion. It is an unlisted company, yet the largest private tech company in the world by scale. And it keeps growing. The daily active users of its domestic AI chatbot Doubao exceeded 100 million in 2025 (QuestMobile). Its 2025 AI-related capital expenditure reached 160 billion yuan (approximately $22 billion), the largest single outlay by any private Chinese company (The Information/Reuters). Zhang Yiming stepped down as CEO in 2021. He was thirty-six. The company had grown too large, and managing the political pressure that scale generates was not the work he wanted to do, he said.
Pinduoduo: Hacking the Supply Chain
Colin Huang's system operates differently.
Temu, Pinduoduo's international platform, launched in the United States in September 2022. By 2025, its global e-commerce market share stood at 24% — tied with Amazon. In three years, it had drawn level with the world's largest e-commerce company.
Temu's operating principle is simple. It connects Chinese manufacturers directly to global consumers, eliminating every intermediary distribution layer. To understand why prices are so low, look at the structure.
The path of a product sold on Amazon: Chinese manufacturer, exporter, Amazon warehouse, consumer. At each step, a margin is extracted.
Temu's path: Chinese manufacturer to consumer. No middle. An AI price-optimization algorithm handles demand forecasting, inventory adjustment, and shipping route selection in real time.
Consumers ask: "Why is it so cheap?" The answer is that the distribution structure has been removed and replaced with AI.
A $33 Billion Digital Services Surplus
The combined effect of these two systems shows up in the data.
In 2025, China's digital services trade surplus reached $33 billion — an all-time high. A decade earlier, China ran a large deficit in digital services. Advertising fees and commissions flowed to the United States as Chinese businesses used Google, Amazon, and Meta's platforms. The flow has reversed. Americans watch TikTok and buy on Temu, paying Chinese companies.
In the United States, foundational technologies and platforms are built. Jensen Huang's CUDA, Anthropic's Claude, OpenAI's GPT. In China, systems are built through application optimization and user scale. Zhang Yiming's recommendation algorithm, Colin Huang's supply chain AI.
Which strategy is more valuable in the long run? In the terminology of Volume 1, the United States commands the design layer (architecture layer), while China builds scale at the execution layer.
Where those two layers meet remains unknown.
Section D: Patent Quantity Versus Quality — The Boundaries of Innovation
The Hugging Face Leaderboard
Late 2025. An American AI researcher opened the Hugging Face global model leaderboard.
Nine of the top ten open-weight AI models were Chinese. DeepSeek-R1, Alibaba's Qwen series, ByteDance's Doubao, and Beijing-based Zhipu AI's GLM series.
The researcher asked: is this real code and real weights? These models were being downloaded, and they were producing real performance on real benchmarks.
It was evidence far more powerful than any patent statistic.
Quantitative Dominance, Qualitative Gap
Start with the numbers.
China holds 38,210 generative AI patents, over 60% of the global total. The United States holds 6,276 AI patents. In raw volume, a 6x difference.
Two other figures complicate the picture. Only 7.3% of Chinese AI patents are filed internationally. Citation counts compared to U.S. AI patents are seven times lower. China files prolifically but is rarely referenced in international markets.
Three factors explain the quantitative excess.
First, the government subsidy structure. Chinese universities and research institutes are evaluated on patent counts as a measure of research output. More patents mean more subsidies. The criterion became filing potential rather than use potential.
Second, defensive patent practices. Patents are filed not for actual deployment but as litigation shields. They become bargaining chips when disputes arise.
Third, a concentration on applied over basic research. Patents emerging from foundational papers attract high citation counts. Patents specialized in narrow applications attract few. The majority of Chinese AI patents fall into the latter category.
DeepSeek Blurs the Boundary
The assertion that "China lacks foundational innovation capacity" is no longer tenable.
DeepSeek's MoE + MLA + FP8 innovations do not sit at the application layer. They are a redesign of the training architecture itself. This is foundational innovation. The DeepSeek R1 paper became one of the most rapidly cited papers in the AI research community after its release.
By early 2026, the state of China's frontier AI models showed the gap narrowing further. Alibaba's Qwen 3.5 handled 3D game creation, browser automation, and website generation — at 60% lower cost than its predecessor. ByteDance's Doubao 2.0 received peer-level evaluations against ChatGPT and Gemini on complex reasoning and multi-step tasks.
According to analysis by Epoch AI, a leading AI research institution, the performance gap between Chinese AI models and American frontier models stood at three to six months as of 2025. Just two years earlier, the dominant view had placed the gap at one to two years.
China's share of the global open-source AI market jumped from 1.2% at the end of 2024 to 30% by August 2025.
The Patent Paradox
A paradox emerges.
Patent statistics measure the past. Working model code and weights measure the present. China's patent figures reflect a legacy of defensive accumulation, but the nine models on the Hugging Face leaderboard reflect current capability.
By the time the measurement metrics need updating, reality has already moved to the next stage.
What DeepSeek proved is not merely an efficiency innovation. It is a signal that China's foundational AI research capability is entering maturity. The gap between quantitative patent metrics and qualitative capability is narrowing.
Chinese-origin researchers account for 47% of the global AI research community. A significant share of them trained at American universities before returning to China, or remain at research institutes within the country. The researcher pool is larger than America's. As resource constraints ease and research capacity accumulates, the conditions are in place for the gaps in patent quality and citation counts to continue closing.
This is the limitation of framing Chinese AI simply as a "follower." By some metrics, it already leads. By others, it still trails. Leading in patent volume, trailing in citations. Leading in open-source models, trailing in advanced semiconductors. The boundary between follower and leader shifts with every metric.
Connection to Volume 1: The System of the Discerning, and Its Ceiling
In Volume 1, we defined the discerning. They are those who read the grammar of a new technology and build systems that others have no choice but to use. Marcus Licinius Crassus systematized Rome's fires and real estate. Arkwright systematized factory labor. Jensen Huang systematized AI development through CUDA.
China has its own discerning. Liang Wenfeng created an algorithm that converts scarcity into fuel for efficiency. Zhang Yiming created a system that converts human attention into revenue. Colin Huang created an AI that eliminates supply chain friction.
The systems built by China's discerning have a ceiling.
The case of Jack Ma illustrates it most clearly.
Alibaba's founder, Jack Ma, was China's greatest discerning of the 2000s and 2010s. E-commerce, payment systems, financial services — the Chinese internet economy appeared unable to function without the systems he built. Ant Group was on the verge of the world's largest initial public offering. The projected valuation exceeded $300 billion.
In October 2020, Ma gave a speech in Shanghai. He compared China's financial regulatory system to a pawnshop.
Three days later, Ant Group's IPO was forcibly halted. Ma disappeared from public view. He spent time abroad before quietly returning months later. A massive antitrust investigation into Alibaba followed. The fine was 18.2 billion yuan. Between March 2022 and March 2025, Alibaba's headcount was cut in half — from 254,941 to 124,320.
Did Jack Ma make a mistake? Or did the state refuse to tolerate a system-builder becoming a bigger system than the state itself?
China's discerning operate within two conditions.
The first condition: they must align with the state's strategic objectives. When DeepSeek fit the narrative of semiconductor self-reliance, it received state support. When Huawei's Ascend fit the narrative of ending dependence on NVIDIA, government procurement followed.
The second condition: they must not become a system larger than the state. Jack Ma crossed that line.
America's discerning operate under different conditions. Jensen Huang built the CUDA ecosystem, and AI developers became unable to work without it. That dependency created NVIDIA's 92% GPU market share. The U.S. government initiated antitrust discussions, but those discussions concerned competition within the market — not threats to the founder's physical safety.
The scope of freedom to build a system differs. What that difference produces over the long term is addressed in the latter half of this book.
Transition: The Equation of the Pursuer
Take stock of what China's discerning have built.
Liang Wenfeng converted GPU scarcity into algorithmic efficiency. Huawei is constructing a chip ecosystem that is slow but does not stop. Zhang Yiming and Colin Huang completed systems that pull global users into AI applications. A $33 billion digital services surplus is the scorecard.
But two limits overlay this system. One is dependence on foundational technology. DeepSeek innovated in training architecture, but the GPUs were still NVIDIA products. China's semiconductor equipment self-sufficiency rate remained at 13.6%. The other is the institutional ceiling. As Jack Ma's case demonstrates, building a system larger than the state is not permitted.
Beneath the ground on which the discerning's innovations operate, four walls are rising simultaneously. The semiconductor blockade. The debt wall. The paradox of capital controls. The demographic cliff. The pursuer does not face a single wall to clear.
We examine those walls, one by one.
Next chapter: Ch. 12 — The Four Walls: Semiconductors, Debt, Capital, and Demographics Pressing In at Once