Opening: The Season of $300 Billion
January 29, 2025, 5:30 p.m. New York time. Microsoft's earnings conference call begins. CFO Amy Hood reads the numbers. "Fiscal year 2025 AI-related capital expenditure: $80 billion." The figure lands in Wall Street analysts' spreadsheets. That same week, Meta announces $60 billion to $65 billion. Alphabet puts forward $75 billion. Amazon declares over $100 billion.
The four companies combined: $300 billion in announced projections. That sum equals 17 percent of South Korea's GDP. It approaches 34 percent of the U.S. defense budget. Where does the money go? Most of it buys NVIDIA GPUs. NVIDIA's data-center revenue hit $51.2 billion in a single quarter (FY2026 Q3), up 66 percent year over year. One company's chips have become the oil of the AI era.
Three hundred billion dollars. What does that number tell us? Not a simple investment boom. Technological innovation concentrates capital; concentrated capital destabilizes society; a destabilized society pressures institutions to redesign themselves. The formula that operated once in Rome and once in Britain is now running simultaneously inside two empires — the United States and China. Same input — AI. Different institutions. Different output.
This chapter applies that formula to both empires side by side. Technological capability, capital structure, the profile of the displaced, the speed of institutional adaptation. Four dimensions through which to read the United States and China at the same time.
A. Technology — Design Layer vs. Execution Layer
Architects and Builders
January 20, 2025. DeepSeek releases its R1 model. It scores 79.8 percent on the AIME 2024 math olympiad qualifier, close to OpenAI o1's 83.3 percent. On the Codeforces coding benchmark, it hits the 96.3rd percentile. And the model's weights are open, released to the entire world. Anyone can download and use it.
The person behind DeepSeek is Liang Wenfeng (梁文锋). A mathematician who came out of quantitative hedge funds. The company he founded in Hangzhou in May 2023 brought Silicon Valley to a halt in under two years. Blocked from the latest H100 chips by U.S. export controls, his team responded with the less powerful H800 and an efficient algorithm called MoE — Mixture of Experts. Constraint did not kill innovation; it gave birth to a different kind. In Volume 1's framework, GPUs are the "land" of the AI era, and DeepSeek is the party that reinvented farming itself in a land-scarce environment.
What this scene reveals is the structure of the U.S.-China AI competition. Both nations are competing over the same technology, but they are fighting on different layers.
The United States dominates the design layer. American companies hold the highest-performing foundation models. OpenAI's GPT series, Google DeepMind's Gemini, Anthropic's Claude. According to the Stanford HAI report (2024), U.S. researchers account for 40 percent of the top 1 percent most-cited AI papers. The MacroPolo AI Talent Tracker (2024) shows that 60 percent of the world's top AI researchers are affiliated with American institutions. In semiconductor design, NVIDIA commands 90 percent of the AI chip market and 92 percent of the GPU market.
China dominates the execution layer. It leads in the speed of AI application and deployment. AI features are pushed instantly into WeChat Mini Programs, Alipay, and Douyin (the Chinese version of TikTok) — platforms with user bases exceeding one billion. In the sheer volume of AI papers, China ranks first in the world, accounting for 40 percent of global output. More than 150 foundation models are registered with the Chinese government. China produces approximately 4.7 million STEM graduates per year — more than five times the U.S. figure. It leads the world in annual industrial robot installations at roughly 276,000 units.
Design layer and execution layer. This distinction overlaps with the structure of the Industrial Revolution examined in Volume 1. While Britain dominated finance and trade, Germany seized chemistry and electrical engineering. By 1913, Germany's share of world manufacturing output was 14.8 percent, surpassing Britain's 13.6 percent. The country that invented and the country that exploited were not the same. A similar divergence is unfolding in AI today. The United States designs AI's architecture. China executes AI's deployment.
The Gap Is Closing
According to Epoch AI's analysis, the average performance gap between U.S. and Chinese AI models is seven months. As of 2025, it has narrowed to between three and six months. Google DeepMind CEO Demis Hassabis himself acknowledged a gap of "several months." As of February 2026, China's frontier models (GLM-4.7, Tongyi Qianwen (通义千问, Qwen) 3.5, and Kimi K2.5) are rated on par with Western frontier models.
But reading the gap purely as a "shrinking distance" misses the structure. The key point is that the two sides hold different kinds of advantages. America's edge lies in fundamental research and semiconductor design. China's edge lies in application speed and cost efficiency. DeepSeek R1's API costs $0.55 per million input tokens. OpenAI o1 charges $15 for the same volume. That is 20 to 50 times cheaper. The equation "more money buys better AI" has cracked.
The question is which layer ultimately matters more. Do the architects win, or do the builders? The history surveyed in Volume 1 gives no definitive answer. In Rome, the side that designed institutions prevailed. In the Industrial Revolution, the side that exploited new technology first prevailed. In the AI era, the answer remains open.
B. Capital — The Dollar Moat vs. Innovation Through Scarcity
The Roads of Empire
Combined 2025 capital expenditure for the Big Four tech companies is $400 billion. The 2026 forecast jumps to between $635 billion and $665 billion, a 67 to 74 percent year-over-year increase. Amazon at $200 billion, Alphabet at $175 billion to $185 billion, Microsoft at $145 billion, Meta at $115 billion to $135 billion. Seventy-five percent of this spending goes directly into AI infrastructure: servers, GPUs, data centers.
This scale is possible because of the dollar. American Big Tech can issue dollar-denominated corporate bonds at interest rates of 5 to 6 percent. Chinese companies pay 7 to 10 percent for dollar debt. The U.S. stock market's total capitalization stands at $50 trillion, 45 percent of the global total. American AI companies' access to capital is structurally deeper than their Chinese counterparts'.
The dollar's share of global foreign-exchange reserves is 56 percent. It has fallen from a 2001 peak of 72 percent, yet it remains close to three times the share of the euro, the second-place currency, at 20 percent. The renminbi's share is only 2 percent. This gap gives the United States a structural advantage in the AI investment race.
In Volume 1's framework, the dollar is Rome's road network. Just as Rome's approximately 80,000 kilometers of roads governed the movement of legions and the flow of goods, the dollar governs the flow of capital — and that flow determines technological supremacy. Dollar hegemony produces low-cost capital; low-cost capital expands AI investment; AI investment reinforces technological supremacy; technological supremacy reinforces dollar hegemony. This feedback loop is the foundation of American AI investment.
What Scarcity Produced
China's capital structure is different. Government AI capital expenditure totals 400 billion yuan (a Morgan Stanley estimate). The State Council's 2017 Next-Generation AI Development Plan (新一代人工智能发展规划) set the direction; the private sector executes. Baidu (百度), Alibaba (阿里巴巴), Tencent (腾讯), ByteDance (字节跳动), and Huawei (华为) each pour billions of dollars into AI. Yet the total does not reach a quarter of the Big Four's combined spending.
Here DeepSeek's significance becomes clear. The training cost for DeepSeek V3 was $6 million. Meta's comparable model, Llama 3.1 405B, consumed 30.8 million GPU hours; DeepSeek V3 was trained on 2.78 million, less than one-tenth of the compute cost. If the competition can be won with algorithms rather than money, the meaning of the dollar moat changes.
In June 2024, Sequoia Capital published an analysis: a gap of $500 billion exists between the amount invested in AI infrastructure and the revenue actually realized from AI applications. The path by which Big Tech's $300 billion in spending translates into actual returns remains unclear. The equation "scale equals victory" is in question.
In the United States, a game of scale is underway. In China, a game of efficiency is underway. Which game wins is a verdict this book reserves until its final pages. What is certain is that both games are running simultaneously.
Place $300 billion and $6 million side by side. The former is 50,000 times the latter. Did that 50,000-fold difference in spending produce a 50,000-fold gap in performance? As DeepSeek R1's benchmarks demonstrate, it did not. What this disproportion means (whether it signals a bubble, a strategic first-mover investment, or both) is something Parts 2 and 3 of this book will track.
C. The Displaced — Free Fall in Freedom vs. Stagnation Under Control
Chicago, Lincoln Park
The third stage of the formula. Social instability. What comes after technology concentrates capital. People pushed out.
Chicago. Sarah (pseudonym), 42, a paralegal. For twelve years she did case research and document review at a midsize law firm. According to the Bureau of Labor Statistics (May 2024), median annual pay for paralegals is $61,010. The 2024 Legal Trends Report finds that 69 percent of paralegal tasks fall within the scope of AI automation. AI-integrated administrative work yields time savings of 50 percent. Harvey AI completes case research that used to take her six hours in fifteen minutes. Her annual salary: $62,000. An AI subscription: $3,000 to $6,000 per year. The cost difference: ten to twenty times.
Mustafa Suleyman, Microsoft's head of AI, has predicted "automation of all white-collar work within eighteen months." In 2025 alone, 55,000 layoffs were directly attributed to AI. Sarah is one number inside that figure.
The arithmetic after termination. Unemployment benefits of $586 per week, capped at 26 weeks. COBRA health insurance averaging $584 per month nationwide. Student loan repayments. Chicago rent. The numbers quickly produce an expiration date. The U.S. unemployment-insurance replacement rate is 40 to 50 percent, below the OECD average of over 60 percent. Maximum duration: 26 weeks. After six and a half months, a cliff.
What happens to the displaced in America is "free fall in freedom." Every node of the safety net requires individual action: filing for COBRA, enrolling on the ACA health-insurance exchange, finding a retraining program, adding "AI proficiency" to the resume. Everything is a personal choice and a personal expense.
Shenzhen, Longgang District
The same period. The other side of the planet. Longgang District (龙岗区), Shenzhen (深圳). Wang Lei (pseudonym), 38. For ten years he managed a customer-service team at an e-commerce company. His team shrank from 30 to 12, then to 6. AI customer-service systems had begun handling the bulk of inquiries. Alibaba's Taobao (淘宝) AI automatically resolves 97 percent of first-contact queries. JD.com's JIMI system exceeds a 90 percent automation rate. The cost differential between a CS team's payroll and an AI system: 8 to 17 times.
Wang Lei's problem is not just the layoff. There is an informal age barrier in China's tech industry — the "age-35 crisis" (35岁危机). The average age at ByteDance is 27. Over 80 percent of job postings include an age restriction. Wang Lei is 38. His resume is caught by the search filters on BOSS Zhipin (BOSS直聘), a major recruiting platform.
Alibaba's headcount fell from 254,941 in March 2022 to 124,320 in March 2025 — a 51.2 percent reduction. Baidu's workforce dropped 21.1 percent from its 2021 peak to 35,900 at the end of 2024. China's youth unemployment rate (ages 16 to 24, excluding students) stood at 18.8 percent in August 2025. In surveys of preferred employers among recent graduates, state-owned enterprises command 47.7 percent; private companies have collapsed to 12.5 percent. Young people's trust in the private sector is eroding.
What happens to the displaced in China is "stagnation under control." The state attempts to provide a safety net but faces fiscal limits. The effective unemployment-insurance take-up rate is below 1 percent. More than 10 million delivery riders earn monthly incomes of 6,650 to 9,344 yuan depending on city tier (Meituan 2025 report, high-frequency riders). Field reports consistently note that a significant share of delivery riders hold university degrees, though Meituan itself says it does not collect rider education data, and no verified figure exists. "Former manager, current rider" (前经理, 后骑手) — a self-deprecating phrase circulating on WeChat.
Same Formula, Different Forms
Set the two scenes side by side and the structure emerges. What pushed Sarah and Wang Lei out was the same technology. AI drove the cost of cognitive labor down by a factor of ten or more. But the shape of the fall differs.
In the United States, the fall is fast. Once 26 weeks of unemployment benefits expire, income converges toward zero. Yet the legal and social freedom to rebuild from the bottom exists. One can learn new skills, move to a different industry, relocate to a different city. In China, stagnation lasts longer. The age-35 crisis, the hukou (户口) residency system, and negative housing equity create structural traps. Apartment prices in Shenzhen's Longhua (龙华) and Longgang districts have fallen below 50,000 yuan per square meter — down 30 to 35 percent from peak. The negative wealth effect from the property downturn is structurally suppressing household consumption (IMF). Goldman Sachs estimates that the real-estate slump is shaving roughly 2 percentage points off annual GDP growth.
Volume 1's framework maps onto both scenes. The displaced in America resemble the small farmers of Rome, pushed off their land by the latifundia and flooding into the cities. Freedom without protection. The displaced in China resemble the handloom weavers of Lancashire, their wages collapsing as the power loom advanced. Belonging without exit.
"Which side suffers more" is not this chapter's question. This chapter's question is: does the same formula, passing through different institutions, produce different forms of suffering? The answer is yes. Sarah and Wang Lei's stories will unfold fully in Part 4, Chapter 14. Here, they serve as a preview. The specific numbers these two will face — monthly deficits, savings depletion timelines, the height of reemployment barriers — come then. What matters now is confirming that the formula's third stage is operating in both countries.
D. Institutional Adaptation — Nine Months vs. Still Waiting
The Fastest Response, the Slowest Response
The fourth stage of the formula. Institutional redesign. This stage must function for a society to find a new equilibrium. Volume 1's representative case took 64 years — from 1769, when Richard Arkwright built the first water-powered spinning mill, to 1833, when an effective Factory Act was finally enacted. Throughout subsequent history, a similar lag recurred each time a general-purpose technology appeared. Railways: 14 years. Automobiles: 20 to 30 years. The internet: 27 years. Institutional adaptation to civilian technology moved within a band of 14 to 64 years, with 64 years as the upper bound.
For AI, the clock started on November 2022, the day ChatGPT was released.
China moved first. On August 15, 2023, the Interim Measures for the Management of Generative AI Services (生成式人工智能服务管理暂行办法) took effect — nine months after ChatGPT's debut. It was the world's first binding regulation on generative AI. The rules mandated pre-registration (备案) for generative AI services and required that training data be lawful and accurate. Baidu's ERNIE Bot (文心一言), Alibaba's Tongyi Qianwen (通义千问), and other models could launch publicly only after completing this registration.
Chinese regulation accelerated from there. In November 2024, the Cyberspace Administration of China (国家互联网信息办公室, CAC) launched the "Qinglang" (清朗) algorithm governance campaign. In April 2025, three national standards on generative AI security were issued. In September of the same year, rules on AI-generated content labeling took effect, mandating both explicit and implicit labels.
The United States moved in a different direction. In October 2023, the Biden administration issued an Executive Order on AI Safety, requiring developers of large AI models to report safety-test results to the government. But in January 2025, when the Trump administration took office, the executive order was rescinded. The stated rationale: "removing barriers to AI innovation."
In the meantime, more than 100 AI-related bills were introduced in Congress. The number of comprehensive AI regulatory laws that passed: zero. The only AI-related federal legislation enacted was the TAKE IT DOWN Act, a single bill. Seventy-nine percent of Americans support AI regulation. Among Republican voters, 84 percent; among Democrats, 81 percent. Supporters of both parties overwhelmingly want regulation, yet Congress cannot act.
There is a reason it cannot. The seven largest tech companies spent $50 million lobbying Congress from the first through the third quarter of 2025. That works out to roughly $400,000 per congressional working day. Registered AI lobbyists now number 3,570 (26 percent of all lobbyists), a 168 percent increase from 2022. Meta alone employs 87 lobbyists. That is one for every six members of the House of Representatives.
In September 2024, California Governor Gavin Newsom vetoed SB 1047, a bill that would have mandated safety evaluations for large AI models. "At this point, this level of regulation would stifle innovation." That same month, China released its AI Safety Governance Framework (人工智能安全治理框架). One side says "not yet." The other says "already done."
Speed and Quality Are Different Questions
But is fast necessarily good? Before this question, judgment must be suspended.
China's rapid response has a mechanism. When the Party center decides, the State Council implements and local governments enforce: a unitary chain of command. There is no congressional gridlock. A pre-approval system (备案) is in place. The approach is to regulate first and relax later.
But this speed carries costs. The pre-approval system delays service launches. In the first half of 2023, Chinese companies could not release large language model services not because the technology was lacking but because they were waiting for regulatory clearance. The regulations include a requirement to "adhere to core socialist values." Restrictions on expression unrelated to technical safety are mixed into technology regulation. The ambiguity of regulatory criteria triggers corporate self-censorship. Innovation can be chilled.
America's slow response also has a mechanism. The absence of regulation permits rapid experimentation. Against that backdrop, OpenAI and Anthropic were able to ship models at speed. Stakeholders (companies, civil-society groups, academia) have room to weigh in. Individual states try different approaches, and federalism enables a search for optimal regulation.
But this slowness also carries costs. The negative effects of technology on society spread unchecked. Big Tech lobbying weakens regulation. In Volume 1's framework, this is "the return of Crassus." Wealth converts into political influence, and political influence blocks institutional change. Executive orders are rescinded when administrations change. Trump's revocation of Biden's AI executive order is the evidence.
Does the deliberative slowness of democracy produce better institutions — or does it become a fatal delay in the face of technological speed? The answer to that question is reserved until this book's final chapter. Here, only a conditional framework is offered. When the speed of institutional adaptation fails to keep pace with the speed of technological diffusion, the gap converts into social cost. ChatGPT reached 100 million users in two months. Institutional adaptation to general-purpose technologies has historically taken 14 to 64 years. The gap between technological diffusion and institutional adaptation is the widest in history.
Connection to Volume 1: Same Formula, Different Inputs
Volume 1's core formula runs as follows. Technological innovation concentrates capital; concentrated capital destabilizes society; a destabilized society pressures institutions to redesign themselves. In Rome, this formula passed through the spread of the latifundia, the displacement of small farmers, the failed reforms of the Gracchi brothers, and culminated in the institutional redesign of the Principate. In Britain, it passed through the spread of factories, the collapse of the handloom weavers, the Luddite uprisings, and culminated in the institutional redesign of the Factory Acts.
Volume 1 traced this formula within a single society. Volume 2 asks a different question. What happens when the same technology is injected into two societies at the same time?
This chapter has established four findings. First, in technological capability, the United States dominates the design layer while China dominates the execution layer. Second, in capital structure, the United States plays a game of scale underwritten by the dollar's reserve-currency status, while China plays a game of efficiency. Third, the displaced are emerging in both countries — as "free fall in freedom" in the United States and as "stagnation under control" in China. Fourth, in institutional adaptation, China moved within nine months while the United States has yet to pass comprehensive legislation — though speed and quality are questions of different dimensions.
Tracking how the same formula operates differently inside two empires is the purpose of this book.
"This Time Is Different" — And the One Thing That Truly Is
Every era recycles the same claim. "This time is different." When the latifundia spread across Rome, a belief persisted that "Rome's laws and roads guarantee permanent prosperity." In the early Industrial Revolution, Andrew Ure wrote in 1835 that "machinery will liberate humanity from labor." Reality disagreed. Handloom weavers' weekly wages collapsed from 25 shillings to 4.5 shillings — an 84 percent decline. While labor productivity surged, real wages lagged far behind for decades. The economist Robert Allen named this phenomenon "Engels' Pause" — a prolonged gap in which the fruits of productivity failed to reach labor.*
*Author's note: In Volume 1, I cited specific figures when explaining Engels' Pause. Upon re-examining Robert Allen's 2009 original study, the figures Allen presented for the 1780–1840 period were approximately +46 percent for labor productivity and +12 percent for real wages. The figures used in Volume 1 appear to have resulted from differences in period definition (1760 vs. 1780) and dataset selection. In this volume, I have revised the narrative to align with Allen's original figures. The core argument — that the gains from productivity growth did not adequately reach workers for decades, a structural phenomenon — holds regardless of which dataset is used.
In the AI era, the "this time is different" refrain is everywhere. The optimism that technology will raise everyone's productivity. The reassurance that AI is merely a tool. Yet there is one thing that is genuinely different this time.
The target of displacement is cognitive labor. In every previous technological revolution, cognitive workers occupied a zone of relative safety. Rome's scale revolution displaced small-scale agriculture. Human physical labor and cognitive capacity both remained intact. The Industrial Revolution displaced physical labor — spinning, weaving, hauling. "Machines cannot think" was humanity's remaining high ground. Even after the electric and automotive revolutions, the consensus held: "white-collar workers are safe." For 200 years, the prescription was the same: "Get educated and you'll be fine."
AI renders that prescription obsolete. Translation, legal research, coding, analysis, basic creative work. For certain occupations, higher education correlates with higher AI exposure. According to Eloundou et al., approximately 80 percent of the U.S. workforce is exposed to large language models in at least 10 percent of their tasks. In Volume 1's framework, just as the handloom weaver's physical skill was displaced by the power loom, the translator's cognitive skill is being displaced by AI. What is collapsing is the cognitive version of the execution layer.
The dimension of diffusion speed is also different. It took roughly 150 years for the latifundia to spread across the Italian peninsula. The power loom took about 30 years to displace the handloom weaver. The internet took about 15 years to reach one billion users. ChatGPT reached 100 million users in two months. The speed of capital concentration is accelerating as well. NVIDIA's market capitalization leaped from $1 trillion to $3 trillion in nine months. The dematerialization of leverage described in Volume 1 is unfolding in real time.
Yet the formula itself does not change. Technological innovation concentrates capital; concentrated capital destabilizes society; a destabilized society pressures institutions to redesign themselves. What has changed is the speed at which the formula operates and the scope of what it displaces. Is there time for the fourth stage — institutional redesign — to function? Can the stage that took 64 years arrive on schedule in a world where technological diffusion is hundreds of times faster? This is one of Volume 2's central questions.
The claim that "this time is different" is only partially correct. It is true that the target of displacement and the speed of diffusion are different. But whether the formula operates has not changed. What has changed is the time remaining for institutional redesign. That time may be the shortest in history.
Transition: Two Paths, and a Collision
The same AI is following different paths inside two empires.
In the United States: design-layer dominance → dollar-backed massive investment → white-collar displacement → institutional gridlock. In China: execution-layer dominance → efficiency-based investment → displacement across all strata → fast but opaque institutional response.
Part 2 (Chapters 3–7) traces the American path. Silicon Valley's AI ecosystem, the capital loop driven by the dollar and GPUs, white-collar free fall in freedom, the rise of AI-native companies, and the mechanism by which lobbying locks institutions in place.
Part 3 (Chapters 8–12) dissects the Chinese path. State-led AI strategy, the scale of 1.4 billion data subjects, the age-35 crisis and the delivery riders, the discerning amid sanctions, and the structural constraints of real estate, demographics, and debt.
In Part 4, the two paths collide. The chip war creates a bottleneck. The paralegal in Chicago and the customer-service manager in Shenzhen are displaced by the same AI — yet they fall in entirely different shapes.
Can institutional redesign function in time in both countries? What happens when those institutions collide? Carrying these questions, we go first to the United States.
Investor Lens: For Investors Who Have Read Part 1
Part 1's framework yields three takeaways for investors.
First, the four stages of the formula are reflected in asset prices sequentially. During the technological-innovation stage, innovator equities move first (NVIDIA, TSMC). During the capital-concentration stage, infrastructure assets follow (data-center REITs, energy). During the social-instability stage, defensive-asset premiums rise (healthcare, consumer staples). During the institutional-redesign stage, regulatory beneficiaries emerge (cybersecurity, compliance). The market is currently in the middle of the second stage.
Second, the 14-to-64-year band determines the investment time horizon. If institutional adaptation is fast (the lower bound of the band), early identification of regulation-locked sectors becomes possible. If slow (the upper bound), the uncertainty premium persists longer, and volatility itself becomes opportunity. Knowing where your own investment horizon sits within the band is the first act of positioning.
Third, the diverging paths of two empires are the key variable in regional allocation. The same input — AI — produces different outputs in the United States and China, which means a global portfolio cannot lump U.S. and China weightings under a single "tech" theme. Parts 2 and 3 trace the specific differences in each path; Part 5 translates them into asset allocation.