March 2025. San Jose Convention Center.
Jensen Huang, CEO of NVIDIA, took the stage. Black leather jacket — the same outfit as always. Numbers flashed on the screen. Fiscal year 2025 revenue: $130.5 billion. Year-over-year growth: 114 percent. Just four years earlier, this company's revenue had been $16.7 billion. An eightfold increase in four years.
There was an even more staggering number. Backlog: over $500 billion. Big Tech companies, startups, and governments around the world were waiting for this company's chips. Demand overwhelmed supply. Huang used the phrase "AI factory" on stage.
The headcount was 26,000. Market capitalization hovered around $3 trillion. That worked out to $115 million in market value created per employee.
Two hundred and fifty-three years earlier, when Richard Arkwright built his mill at Cromford, he employed about 300 workers. Those 300 displaced thousands of handloom weavers. Who are Jensen Huang's 26,000 displacing? To answer that question, we first need to follow the money.
In Chapter 12, we examined the technological foundations of cognitive automation — transformers, scaling laws, emergent capabilities. Technology opened the door. Capital is now organizing those possibilities. Who profits, and who gets pushed aside.
Rome's latifundia concentrated capital in land. The factory system of the Industrial Revolution concentrated it in machinery. In the age of AI, capital concentration is happening on top of things you cannot see.
1. The Magnificent Seven — The Fastest Capital Concentration in History
Start with the numbers.
People assume that investing in the S&P 500 means spreading money evenly across America's top 500 companies. Reality looks different.
In 2015, the Magnificent Seven accounted for 12.3 percent of total S&P 500 market capitalization. Apple, Microsoft, Alphabet, Amazon, Meta, NVIDIA, Tesla. Those seven. By 2023, they had reached 28.6 percent. By October 2025, 37 percent. A threefold increase in a decade.
Invest $100 in the S&P 500, and $37 goes to these seven companies. The remaining 493 split $63. The meaning of "diversified investing" has changed.
For historical context: at the peak of the dot-com bubble in 1999-2000, the top ten companies in the S&P 500 represented 27 percent of total market cap. The market called it a "bubble." In 2025, the top ten account for 41 percent — 1.5 times the dot-com peak.
There is a difference. Dot-com companies sold dreams. Today's Big Tech sells profits. The Magnificent Seven posted combined revenue of $2.02 trillion in 2024. Alphabet, Apple, and Microsoft each exceeded $100 billion in annual net income. The MAGMAN six (the Magnificent Seven excluding Tesla) posted earnings growth of 31.7 percent year-over-year. The remaining 494 companies in the S&P managed 13.0 percent.
This concentration has structural causes. Big Tech dominates three things simultaneously: data, computing power, and AI models. This trinity has created a degree of vertical integration without historical precedent.
Rome's Crassus vertically integrated land, slaves, and a private fire brigade. Arkwright integrated spinning machines, the factory system, and labor discipline. Big Tech integrates the entire stack — from chips to cloud, models to services.
Whoever controls the essential factors of production dominates the economic order. This principle has not changed in 2,000 years.
2. Data — The Land of the Twenty-First Century
In 2017, The Economist declared: "The world's most valuable resource is no longer oil, but data."
They were half right.
Data is closer to land than to oil. Oil is a consumable. Burn it and it is gone. Data is a non-rival good. Using it does not deplete it.
The price of oil is set by markets. The price of data is mostly zero — users pay for it with "free services." The essential similarity lies elsewhere: the question "Who owns it?" determines the entire economic structure. That was true of land. It is true of data.
The numbers tell the story. Google's global advertising revenue per user runs $47 annually. For U.S. users, $454. Meta's global ARPU — average revenue per user — is $49.63. By one estimate, the combined value extracted from a single U.S. user across the top three tech companies comes to $1,274 per year.
In a French consumer survey, respondents asked how much they would accept to sell their personal data gave answers ranging from 1 to 100 euros. A gap of more than tenfold between perceived value and actual value extracted. In second-century BC Rome, smallholders did not understand the true value of their land. They had no idea that a latifundium owner could mass-produce olive oil on that same plot and export it across the Mediterranean. Data users are no different.
The mechanism that sustains this gap is the data flywheel. More users generate more data. More data creates better services. Better services attract more users. A self-reinforcing cycle.
Google's global search market share stands at 89.3 to 89.7 percent. It held near 90 percent for over a decade. In late 2024, it dipped below 90 percent for the first time. On mobile, it still commands 93.88 percent. The nearest competitor, Bing, holds 4.01 percent.
Netflix's recommendation algorithm influences 80 percent of content watched on the platform, generating an estimated cost savings of over $1 billion annually. According to McKinsey, data-driven organizations are 23 times more likely to acquire customers and 6 times more likely to retain them.
This is structurally identical to the network effects of Roman roads. As we saw in Chapter 2, more roads meant more trade, and more trade made building roads more economically justified. But there is a critical difference. Roman roads were public goods. The empire owned them. The owners of the data flywheel are private corporations.
This is where the analogy "data is the new land" earns its force. We can compare three enclosures.
The first enclosure. Beginning in the third century BC, Rome's ager publicus was privatized. This was public land — conquered territory declared state property and granted to citizens for use. The Licinian-Sextian law of 367 BC capped individual holdings at 500 iugera. The law was not enforced. Wealthy families exceeded the cap, and the latifundia spread.
The second enclosure. Between 1604 and 1914, over 5,200 enclosure acts were passed in England. 20 percent of England's total area — 6.8 million acres — was privatized. Tenant farmers and landless laborers who lost access to common land flooded into cities. The labor force of the Industrial Revolution was born.
The third enclosure. Since the 2000s, internet space and user-generated data have been undergoing privatization. Legal scholar James Boyle called the expansion of intellectual property rights "the second enclosure movement." This book extends Boyle's analogy to the privatization of data. Shoshana Zuboff, in The Age of Surveillance Capitalism, analyzed the appropriation of "behavioral surplus." Karl Polanyi's warning — the subordination of society to the market — is repeating itself in the digital domain.
The difference lies in the means. Roman aristocrats seized common land through illegal occupation and political connections. English landlords enclosed commons through parliamentary legislation. Big Tech privatizes data through Terms of Service. It is legal consent, but there is no real alternative. You can use the internet without Google — it is possible, but inconvenient.
A common pattern emerges across all three enclosures. A shared resource exists. A productivity revolution raises that resource's value. Once the resource becomes valuable, the powerful privatize it. The weak are separated from it. Institutional responses are attempted but cannot keep pace with privatization.
Speed is the key. Rome's enclosure took 200 years. England's core enclosure period lasted about 100 years. Digital enclosure has been underway for 20 years. The faster the enclosure, the less time institutions have to respond.
There is also the problem of finitude. The total volume of publicly available text produced by humanity amounts to 300 trillion tokens. This is the sum total of recorded human knowledge. AI training datasets are growing at 2.5 times per year, but compute growth runs at 4 times per year — a gap of 1.6x. Exhaustion of high-quality text data is projected around 2028.
If data is "the new land," then high-quality training data is "fertile land." Fertile land is finite. Big Tech, having already consumed most of it, holds the first-mover advantage. This is structurally the same as the finite arable land on the Italian peninsula in the second century BC.
3. Computing Power — The New Strategic Resource
If coal was the energy source of the Industrial Revolution, the energy of the AI age is computational capacity.
And the supply structure of that capacity is concentrated to a startling degree.
NVIDIA holds 86 to 92 percent of the AI GPU market. Place that number alongside one from 120 years ago. In 1899, Standard Oil controlled 90 percent of the U.S. refining market. It peaked at 91 percent in 1904. Separated by 120 years, in entirely different industries, the same number appears.
Standard Oil was broken into 34 companies under the Sherman Antitrust Act in 1911. Antitrust action against NVIDIA is not even under discussion. Part of the reason: it simply makes a better product, faster.
That is not the only reason. The switching costs created by the CUDA software ecosystem also sustain this dominance. Millions of developers have written code on top of CUDA. Migration to an alternative platform is technically possible but expensive. When the source of monopoly is different, the language of law must be different too.
The training costs of frontier AI models accelerate this concentration. Training GPT-4 cost an estimated $78 million in compute alone. Google's Gemini Ultra ran $191 million. Add labor, infrastructure, data acquisition, and failed experiments, and the total reaches hundreds of millions of dollars. Training costs have increased by a factor of two to three annually over the past eight years. By 2027, the cost of a single training run is projected to exceed $1 billion.
An analysis by Epoch AI captures what this trend means: "If training cost trends continue, only the best-funded organizations will be able to develop frontier AI models."
Big Tech's response is vertical integration. Google has developed its TPU since 2015, now in its seventh generation. Amazon built Trainium, deploying 400,000 units for Anthropic. Microsoft co-designed Maia with OpenAI. A co-design model is emerging in which AI models and chips evolve in tandem.
This vertical integration weakens independent AI labs that rely on general-purpose hardware. Hyperscaler custom chips reduce total cost of ownership by 40 to 65 percent compared to commodity GPUs. More than 50 percent of internal inference workloads already run on custom silicon.
The AI startup ecosystem reflects this structure. OpenAI's valuation stands at $500 billion — the highest for any private company in history. Anthropic is valued at $183 billion, ranking fourth. Together, these two companies account for 14 percent of global venture investment in 2025. Funding for foundation model companies has reached $80 billion, representing 40 percent of total global AI funding. The top ten AI companies capture 76 percent of all AI funding.
The very concept of an independent AI startup is being shaken. Google has invested over $1 billion in Anthropic; Amazon has put in several billion more. Microsoft and SoftBank have made massive investments in OpenAI. Big Tech is investing directly in its potential competitors. The logic of capital absorbs the logic of competition.
Capital expenditure reveals the physical reality of this concentration. Look at combined AI infrastructure spending by the Big Four: Meta, Amazon, Alphabet, and Microsoft. It rose from $107 billion in 2020 to $256 billion in 2024. By 2026, projections reach $630 to $690 billion — more than double in two years.
Put that in context. The CHIPS Act allocates $39 billion in direct subsidies for semiconductor manufacturing. Big Tech's projected 2026 AI infrastructure investment is 17 times that figure. Private corporations are outspending national industrial policy.
Even compared to the fiber-optic infrastructure investment during the dot-com bubble, current AI data center spending significantly exceeds that era's levels. This is why "AI bubble" concerns circulate among investors. There are structural parallels with the Railway Mania of the 1840s during the Industrial Revolution. Massive infrastructure investment precedes revenue realization.
4. Semiconductors — The Invisible Bottleneck
Follow the flow of capital far enough, and you arrive at a single bottleneck.
TSMC. Taiwan Semiconductor Manufacturing Company. It holds 70 to 71 percent of the foundry market by revenue. At advanced nodes — 3nm, 2nm — its share exceeds 90 percent. For sub-10nm leading-edge chips, Taiwan produces 92 percent and South Korea 8 percent. One hundred percent is concentrated in two East Asian countries.
Manufacturing a single chip involves an average of 25 countries. In the semiconductor value chain, more than 50 points exist where a single region holds over 65 percent market share. It is a collection of single points of failure.
There is an even more extreme concentration. ASML, a Dutch company, supplies 100 percent of all extreme ultraviolet lithography equipment. Without EUV, chips below 7nm cannot be manufactured. A single EUV system contains 457,000 components. The critical optical elements were developed by Germany's Zeiss over a span of 15 years.
More than 800 global suppliers make up this ecosystem. Thirty years of cumulative R&D have created what is effectively an insurmountable barrier to entry. This approaches a "natural monopoly." Standard Oil's monopoly was the result of artificially excluding competitors. ASML's monopoly is the natural product of technological complexity.
Eighty-four percent of Taiwanese citizens say "the semiconductor industry protects Taiwan." The silicon shield. This is an unprecedented case of technological monopoly functioning as geopolitical deterrence. Under a full-conflict scenario, a Taiwan Strait crisis would inflict global economic losses of $10 trillion.
Nations are trying to disperse this concentration. The U.S. CHIPS Act totals $280 billion. By January 2026, it had catalyzed over $640 billion in semiconductor supply chain investment. TSMC committed an initial $40 billion to its Arizona facilities and announced an additional $100 billion in 2025. The EU Chips Act has attracted 69 billion euros in investment.
Energy presents another bottleneck. U.S. data center power consumption is projected to rise from 183 TWh in 2024 to 325-580 TWh by 2030. Roughly half of the growth in U.S. electricity demand is driven by data centers.
Microsoft signed a 20-year power purchase agreement with the Three Mile Island nuclear plant. Output: 835 MW. Total contract value: $16 billion. The site of America's worst nuclear accident in 1979 is being resurrected as a power source for AI data centers. It is a paradox: an energy source humanity once feared, embraced again for a technology humanity once could not imagine.
5. The Superstar Economy and the Future of Labor
Where does all this concentration lead? To a transfer from labor — that is, ordinary people's income — to capital.
The U.S. labor share of income fell from 65.4 percent in 1970 to 56.8 percent in 2024. An 8.6 percentage point decline over 54 years. Based on U.S. GDP of $25 trillion, this represents $2.15 trillion per year shifting from labor to capital.
This is not a uniquely American phenomenon. According to Karabarbounis and Neiman's 2014 study, labor shares declined in 42 of 59 countries — 71 percent. It is a global trend.
What is driving it? A 2020 study by Autor and colleagues identified the core mechanism: "The rise of superstar firms." In all six major U.S. industries studied, sales concentration increased. The industries with the largest increases in concentration saw the steepest declines in labor share. A handful of hyper-productive firms dominate the market, and their low labor shares drag down the economy-wide average.
Research by Acemoglu and Restrepo draws a more direct causal line. Automation explains 50 to 70 percent of the widening wage inequality in the United States. Each industrial robot reduces local employment by six workers and depresses wages by 0.42 percent. In 2015, when six robots were introduced to a welding line at a Michigan auto plant, 30 of 36 welders were notified of reassignment. The moment when numbers become reality.
AI extends this pattern into cognitive labor. Korinek and Stiglitz characterize AI as "an extreme case of capital-biased technological change" — replacing labor while simultaneously complementing capital.
Recall the Engels' Pause of the Industrial Revolution. As we saw in Chapter 9, between 1780 and 1840 in Britain, productivity rose 88.6 percent while real wages fell 5.2 percent. The divergence between productivity and wages persisted for 60 years. A similar structure is forming in the AI age.
There is a difference. The Displaced of the Industrial Revolution were unskilled or semi-skilled manual laborers. The Displaced of the AI age are highly skilled knowledge workers — in law, accounting, translation, coding. A class with significant social voice.
Remember what we saw in Chapter 12. According to Eloundou and colleagues, 80 percent of the U.S. workforce has at least 10 percent of its tasks exposed to LLMs. Exposure is higher among high-income, highly educated occupations. "Cognitive proletariat" — this term is an analytical analogy, not an academic definition.
Rome's proletariat was a legal category: citizens who failed to meet the property qualification. The Displaced of the AI age are defined not by legal status but by economic vulnerability. Acknowledging the limits of the analogy, the structural resemblance is clear.
The insight for investors: Declining labor share means rising corporate profits. In the short term, good news for investors. But workers are also consumers. When labor income shrinks, the consumer base contracts.
This is precisely what Rome experienced. As we saw in Chapter 4, when smallholders were reduced to urban poor, the empire had to maintain consumption through grain subsidies. "Bread and circuses" was an approach that managed symptoms without solving the structural problem. It produced short-term stability and long-term crisis simultaneously. Investors in the AI age face the same dilemma.
6. Can Institutions Keep Up with Technology?
The trajectory of Standard Oil serves as a reference point. Monopoly formation: 1870. Sherman Antitrust Act: 1890. Breakup ruling: 1911. 40 years.
Big Tech's trajectory is still unfolding. In August 2024, a federal judge ruled that Google had illegally maintained its monopoly in the search market. The Department of Justice demanded structural remedies, including the divestiture of the Chrome browser and a ban on bundling. Remedy hearings began in April 2025.
Around the same time, a federal judge ruled that Meta "is not a monopoly." The FTC lost. Amazon's antitrust trial is scheduled for October 2026. The Apple case is in its early stages and expected to last years.
The EU has been more aggressive. In 2024, it adopted the EU AI Act, the world's first comprehensive AI regulation, set for full enforcement by August 2026. In 2025, it imposed fines of 2.95 billion euros on Google, 500 million euros on Apple, and 200 million euros on Meta. In December 2025, it opened a formal antitrust investigation into Google's use of content for AI model training.
The comparison between Standard Oil and Big Tech reveals a key difference. Standard Oil dominated a single industry — refining. Big Tech dominates platforms spanning multiple industries. Standard Oil's oil was a rival good: if I use a barrel, you cannot. Data is a non-rival good. Traditional monopoly analysis frameworks do not fit.
Whether a monopoly exists depends on how you define the market. The language of law is failing to keep pace with the structure of the economy.
Institutional response has always lagged technological change by a generation. In Rome, from the Gracchan reforms of 133 BC to Augustus's establishment of the principate in 27 BC took 106 years. In the Industrial Revolution, from the first Factory Act of 1802 to universal suffrage in 1928 took 126 years. How long will the institutional redesign of the AI age take?
7. A Translator's 2025
Behind the numbers and the trends, there are people.
Consider a freelance translator in Seoul. Eleven years of experience. Specializing in English-to-Korean technical translation — IT white papers, medical device manuals, legal contracts. A project took about a week and paid two to four million won.
Starting in the second half of 2024, assignments began to dry up. DeepL, GPT-based translation tools, and domestic AI translation engines entered the market. Small companies started using AI for rough drafts and hired translators only for "review." Rates dropped by more than half. A week-long project shrank to a half-day quality check.
He decided to learn the AI tools himself. He fed a technical document into GPT. A rough translation came back in 30 minutes — work that used to take two days. The quality was impressive but imperfect. Context-dependent terminology, industry conventions, the nuances of legal documents — these still required human judgment.
The problem was whether clients could perceive the difference. Most could not.
Midjourney employs 130 to 163 people. Its estimated annual revenue is $500 million — $3.1 to $3.8 million per employee, about 10 to 13 times the figure at a traditional SaaS company. One hundred and sixty-three people are replacing a significant portion of the work that hundreds of thousands of specialists once performed.
You may recall the story of Marco, the translator we met in Chapter 12. Fourteen years of experience, income down 80 percent in two to three years. The trajectory of this Seoul translator is similar. What handloom weavers experienced over 25 to 30 years during the Industrial Revolution is happening now in two to three.
Survivorship bias must be acknowledged. Some translators who have become proficient with AI tools have seen their productivity increase three- to fivefold, and their income along with it. They are "the Discerning." Exposure to the same technology produces diametrically opposite outcomes.
8. Transition — The Essential Factors of Production, the Speed of Concentration
Three forces are driving capital concentration in the AI age.
First, data. Through digital enclosure, user-generated data is being privatized by a handful of companies. The third enclosure is faster and more invisible than the first (Rome) and second (England).
Second, computing power. NVIDIA at 90 percent. TSMC at 90 percent. ASML at 100 percent. The supply structure of computational capacity is concentrated in a tiny number of firms. Frontier AI model development is increasingly becoming a game reserved for the few with sufficient capital.
Third, AI models. The top ten AI companies capture 76 percent of all AI funding. By investing directly in potential competitors, Big Tech is shaking the very concept of the independent AI startup.
Viewed through a historical lens, the essential factors of production have shifted with each era — from land (Rome) to machinery and capital (Industrial Revolution) to data and computing power (AI age). What does not change is the pattern: the essential factors concentrate in the hands of the few, and that concentration reshapes the social order.
What has changed is the speed. Rome's capital concentration unfolded over 200 years. Standard Oil's rise took about 40 years. The Magnificent Seven's ascent took about 10. The velocity of concentration is accelerating exponentially. Institutional response has historically required more than a century. The gap in speed is widening.
Technology opened the door. Capital is deciding the direction. Institutions have not yet caught up.
How does this structure feel at the level of individual lives? In the next chapter, we meet two people in 2025. A knowledge worker being displaced and an AI-native founder. Same technology, same era, opposite trajectories. Two destinies shaped by the architecture of capital concentration.
End of Chapter 13. Next: Chapter 14 — Two People in 2025: The Displaced Knowledge Worker and the AI-Native Founder