Opening: A Classroom in Bangkok, a Hospital in Nairobi, a Government Office in Jakarta
The Classroom's Choice
March 2028. A public school on the outskirts of Bangkok.
Teacher Supachai turns on the screen at the front of his classroom. An AI tutor interface loads. Until last month, an American-made service occupied this screen. This month, it changed. DeepSeek.
The reason is simple. The cost is one-twentieth.
The school's management committee discussed the matter two months earlier. The vice principal arrived with a performance comparison table. Thai-language support was actually better on DeepSeek. It read student names accurately in Thai script and delivered feedback in Thai. Faced with the calculation that a single budget line item would shrink to one-twentieth, the meeting did not last long.
Supachai does not think about which worldview he is teaching. He picked what fit the budget. His forty-one students do not think about it either. They type questions and receive answers. They do not know which country's algorithm generated those answers.
In a world where algorithms replace textbooks, this choice is a technology choice. It is also, simultaneously, an institutional choice. An AI trained on different data, aligned to a different value system, is teaching forty-one students in Supachai's classroom how the world works.
This classroom in Bangkok is a microcosm of an enormous fork in the road.
Nairobi and Jakarta
That same hour, in the radiology department of a Nairobi hospital.
Dr. Amina signs the last page of a contract. It is an agreement to switch the hospital's diagnostic AI to a Chinese-made system. She read the performance comparison report three times. It matches the American product. The price is one-third. The training data includes East African imaging. Internal evaluation concluded it fits Kenyan patients better.
While Amina signs, the hospital's IT manager is coordinating the server migration schedule. From now on, the hospital's diagnostic AI connects to a data center in Beijing.
In Jakarta, civil servant Hadi processes paperwork. It is an approval document to migrate the administrative processing system from an American cloud to Alibaba Cloud. In a year when the Indonesian government's digital transformation budget was slashed, Alibaba offered a package deal: localization support and technical training bundled together. Hadi says the decision was not difficult. "The cost is half and the service is better. Nationality has no reason to be an issue."
But for some, nationality is the issue.
The Great Fork
These events are not isolated.
The teacher in Bangkok, the doctor in Nairobi, the civil servant in Jakarta. Each chose according to their own budget and needs. But when those choices accumulate, they become a structure. The question: which country's AI ecosystem becomes the default technology base for the next generation. Once that default is set, every service, dataset, and relationship built on top of it is determined alongside it.
This is not a question of trade routes. It is a question of standards.
If you built wheels to Roman road specifications, you could access the entire Roman Empire's market. Deviate from those specifications, and you could not travel the empire's roads. In the AI era, the roads are API compatibility, data formats, and governance norms.
The question shifts. "Over the next twenty years, in which direction does the world split?"
This chapter does not predict the future. Predictions speak in probabilities, but we do not know those probabilities. Instead, it speaks in conditions. Which conditions, when met, strengthen which path. Three paths, the conditions that open them, and the variables that determine them.
The Contours of Three Paths
There are three paths.
The first: a path where the United States maintains AI technological hegemony. When frontier model leadership, the dollar's reserve currency status, and the alliance network are all sustained simultaneously, this path is strengthened. The United States controls the standard for AI infrastructure and extends its existing financial hegemony through digital dollar dominance.
The second: a path where China achieves a practical overtaking. In a world where "good enough AI" displaces "the best AI," this path is strengthened when energy advantages and Global South proliferation combine. A world where cost and accessibility become more important variables than performance.
The third: a path where the world splits into two technology blocs. Neither the United States nor China secures a decisive advantage, and hardware stacks, software ecosystems, and governance norms each diverge. A world where Supachai in his Bangkok classroom and a fintech CTO in Singapore both bear the cost of duplication.
The core formula identified in Volume 1 — technology explodes, capital concentrates, the displaced are created, institutions scramble to catch up — is operating in both empires simultaneously. The outcomes of that formula are not singular. Three roads diverge.
It took sixty-four years from Richard Arkwright's water frame in 1769 to the Factory Acts of 1833. Add sixty-four years to 2026 and you reach 2090. The twenty years in between — 2026 to 2045 — are the years this book's readers will be investing, choosing careers, and raising children. In the middle of the Engels' Pause, they must act without knowing which of the three roads will materialize.
Section A: Tools for Reading Scenarios — The Six-Variable Framework
What London Missed in 1893
1893 was a remarkable year.
That year, Germany's steel production surpassed Britain's for the first time. London's The Economist ran it on the front page. The data was clear. The reaction of British readers was equally clear. It split two ways. One was anxiety. The other was reassurance.
Reassurance won.
"We have finance. We have the pound. We have the Royal Navy. What does it matter if steel has been overtaken?"
This reassurance was half right. By 1913, Germany's share of world manufacturing output stood at 14.8 percent. Britain's had fallen to 13.6 percent (Bairoch statistics). The manufacturing rankings had reversed. German organic chemicals commanded 90 percent of the global market. Yet London remained the center of world finance. The pound was the reserve currency. The Royal Navy upheld the principle that it "must be stronger than the second and third greatest naval powers combined."
Then, in 1914, the First World War erupted.
Germany chose a military gamble. The pressure came from failing to convert its manufacturing overtaking into hegemonic transition. The war ended in defeat. In the process, Britain accumulated massive debts to the United States. The pound's strength drained slowly. The dollar overtook the pound as the dominant trade credit currency in the 1920s. By foreign exchange reserves, the dollar surpassed the pound in the late 1920s. The Bretton Woods system formalized dollar hegemony in 1945.
From steel reversal to financial hegemony transfer: thirty to fifty years.
Panic over a single indicator's reversal is an error. So is complacency. Historical hegemonic transitions occurred only when multiple variables shifted simultaneously.
In 2025, the MMLU benchmark gap between US and Chinese AI performance narrowed from 17.5 percentage points (2023) to 0.3 percentage points (Epoch AI). Panic over this number is the 1893 error. So is the reassurance that "we still lead at the frontier."
Six Variables Extracted from History
Examining four historical cases where productivity revolutions and hegemonic competition ran in parallel reveals a common set of operating variables.
Rome versus Carthage. The technology variable was "protocol standardization": roads, aqueducts, legal systems. The institutional variable was the graduated expansion of citizenship. Carthage was economically resilient (it minted coins even during the Punic Wars). But a system dependent on mercenaries had no institutional scalability. Rome won by extending citizenship across the Italian peninsula, broadening its human base.
Britain versus Germany (1870-1914). Manufacturing overtaking did not equal hegemonic transition. The inertia of financial hegemony, alliance networks, and Germany's strategic blunder of military gamble — despite seizing "the next wave" (chemistry and electricity of the Second Industrial Revolution) — proved decisive.
The United States versus the Soviet Union (1947-1991). Soviet defeat was not a failure of technology. The Soviets developed nuclear weapons, semiconductors, and space technology. The problem was the absence of institutional flexibility to diffuse those technologies into the civilian sector. The civil-military conversion whereby DARPA gave birth to the internet and the internet gave birth to Silicon Valley did not exist in the Soviet Union.
The transition from Britain to America (1914-1945). The only "peaceful transition." Mass production, financial innovation, and above all, the exhaustion of British capital through two world wars completed the transfer.
Six variables emerge from these four cases.
First, the technology frontier. Who seizes the next wave. Not a single benchmark, but sustained leadership in system-level innovation: agentic AI, scientific discovery applications. The United States currently leads, but the gap is closing fast. Chinese open-source AI's global market share jumped from 1.2 percent in late 2024 to 30 percent by August 2025 (LLM Stats).
Second, capital and financial structure. Financial hegemony transitions far more slowly than manufacturing hegemony. The dollar currently accounts for 56.32 percent of global foreign exchange reserves (IMF COFER). The renminbi sits at 2 percent, and this structure will not change as long as capital controls persist.
Third, alliances and networks. Technology alliances are alignments of interest. The United States uses the CHIPS Act to anchor fab investments by TSMC, Samsung, and Intel, and coordinates EUV export controls on ASML with the Netherlands, Japan, and South Korea. China embeds infrastructure across the Global South through the Belt and Road Initiative (BRI) and the Digital Silk Road.
Fourth, institutional flexibility. The variable that Soviet failure proved. The presence or absence of institutional pathways for diffusing technology into the civilian sector. The United States moves slowly, but fifty states running decentralized experiments create feedback loops. China moves fast, but errors risk amplification rather than correction. Over the long term, this variable is decisive.
Fifth, human capital. China produces 47 percent of the world's AI researchers. Yet most train at American graduate schools and work at American companies. A $100,000 increase in H-1B visa fees threatens this pipeline, even as China has launched the K-visa to attract talent in the opposite direction.
Sixth, energy and infrastructure. Energy in the AI era is coal in the Industrial Revolution. China added 430 GW of net power capacity in 2024, fourteen times the US figure of 30 GW (National Energy Administration / Jefferies). The United States, meanwhile, controls 92 percent of the NVIDIA GPU market. Energy favors China; advanced semiconductors favor the United States.
Arrange these six variables in a single table and the current position of US-China competition becomes visible. The United States holds advantages in the technology frontier, capital and finance, and alliances and networks. China holds a structural advantage in energy and infrastructure. Human capital and institutional flexibility carry vulnerabilities on both sides.
Six Variables and Three Scenarios
A reversal in any one of these six variables does not trigger a hegemonic transition. The steel reversal of 1893 proved that. But when all six begin shifting simultaneously — that carries a different meaning.
In the US-China AI competition, the three scenarios are defined by combinations of these six variables.
Scenario A gains ground when the United States maintains three variables: the technology frontier, capital and finance, and alliances and networks. Scenario B hardens when China converts its energy and infrastructure advantage into energy-intensive AI applications, compensates for human capital constraints with robotics, and expands alliances and networks into the Global South. Scenario C is strengthened when both sides maintain their respective technology frontiers but alliance networks split into two blocs and institutional flexibility fails to find common ground.
The three scenarios are sets of conditions. Not predictions.
Given that, where do the six variables currently point?
As of 2025, the United States holds an advantage in three variables: the technology frontier, capital and finance, and alliances and networks. China leads structurally in one: energy and infrastructure. The remaining two (human capital, institutional flexibility) carry vulnerabilities on both sides. In this configuration, Scenario A rests on the broadest foundation.
But the direction of movement matters more than a static advantage. If key allies defect from export control coordination or adopt Chinese AI infrastructure, the alliance and network variable destabilizes, strengthening Scenario C. If China's energy advantage translates into cost competitiveness in AI applications and Global South adoption accelerates, Scenario B gains ground. If populism within the United States leads to immigration restrictions and technology regulation, Scenario A erodes not from the outside but from within.
This is why this book assigns the highest weight to Scenario A while ranking Scenario C above B. The structural change most actively underway is not overtaking but fracturing. Chapter 16 converts these weights into specific numbers and a monitoring framework.
Section B: Scenario A — The United States Maintains Technological Hegemony
Silence in the Hearing Room
Washington, D.C. A committee hearing room in the Russell Senate Office Building.
Spring 2025. A Senate AI subcommittee hearing. In the witness chairs sit a tech executive, an economist, and a labor specialist. Mustafa Suleyman, Microsoft's head of AI, had stated publicly earlier that year: "Within eighteen months, all white-collar work will be automated."
In the hearing room, senators nod. They ask questions. They ask more questions. Two hours pass. The hearing ends.
No legislation is produced.
In all of 2025, a single federal AI bill passed the United States Congress. The TAKE IT DOWN Act, a law mandating the removal of nonconsensual fake sexual images. Comprehensive federal legislation addressing the economic impact of generative AI, white-collar job displacement, or algorithmic accountability: zero.
That same year, the seven largest tech companies spent $50 million on lobbying. That amounts to $400,000 per day Congress was in session. AI lobbyists represented 26 percent of all lobbyists, a 168 percent increase from 2022.
That same day in Beijing. A joint AI ethics management measure from ten ministries was officially announced. From draft publication to enforcement: nine months.
The two countries' clocks run at different speeds. That gap is the internal risk of Scenario A.
2035, If Scenario A Materializes
Spring 2035. A community hospital in Pittsburgh.
Denise, an administrative coordinator with a nursing background, studies her screen. The AI assistant system has organized today's case list for her. Thirty-four patients. Insurance claims, discharge coordination, post-care alerts. Eighty percent has been processed automatically. Only the remaining 20 percent requires her judgment.
As recently as 2028, three people did this work.
Yet Denise's hourly wage has actually risen. The hospital chose not to cut coordinators but to transform the role. Repetitive processing goes to the AI; difficult conversations with patients' families go to her. Pittsburgh's "AI Transition Stipend" program covered retraining costs. The program is funded by federal AI tax revenue.
The system, running on AWS AI infrastructure, offers its menu in English, Spanish, and Arabic. Pittsburgh's hospital began supporting Arabic in 2031. In a world where the American AI standard became the global standard, that standard has reached this hospital screen.
Denise does not think about the backdrop. She processes today's cases.
This is one facet of a world where Scenario A has materialized. Hegemony arrives without a declaration. Through the menu language on a hospital screen, through the funding source of a retraining program, through the API standard of an AI system.
Five Conditions Met Simultaneously
A path in which the United States maintains technological, financial, and institutional hegemony in the AI era requires five conditions to hold simultaneously. If any one falls away, Scenario A weakens.
Condition 1: Sustaining the technology lead.
Not merely ranking first on benchmarks. Sustained leadership in system-level innovation: agentic AI, multimodal reasoning, scientific discovery applications. Current US frontier models (GPT-5, Claude Opus, Gemini 3 Pro) maintain benchmark leads. But the gap is narrowing.
MMLU gap: 17.5 percentage points (2023) to 0.3 percentage points (2025). HumanEval gap: 31.6 percentage points to 3.7 percentage points. Time lag: averaging seven months, now compressed to three to six months (Recorded Future analysis). Chinese open-source AI's global market share jumped from 1.2 percent in late 2024 to 30 percent by August 2025. Nine of the top ten open-weight models are Chinese-made.
These numbers point in one direction. The effort required to maintain the technology lead is growing exponentially.
Condition 2: Stabilization of immigration policy.
Silicon Valley does not function without immigration. A significant share of core researchers at OpenAI, Google DeepMind, and Anthropic hail from India, China, and Eastern Europe. The Trump administration's $100,000 increase in H-1B visa fees exerts reverse pressure on this pipeline.
FAANG companies expanded Indian local hiring to 33,000 (an 18 percent year-over-year increase). Thirty to 40 percent of those who study in the US now prefer employment in India afterward. If the flow of talent toward the United States reverses, Condition 1 (sustaining the technology lead) destabilizes. The conditions are interconnected.
Condition 3: Continuation of the dollar reserve currency system.
The dollar currently accounts for 56.32 percent of global foreign exchange reserves. Down from its 2001 peak of 72 percent. Yet some analyses argue that after removing exchange rate effects, the real decline is a mere 0.12 percentage points (St. Louis Fed). The renminbi sits at 2 percent. The euro at 20 percent. With no viable alternative, dollar hegemony persists on inertia alone.
But this is not a permanent condition. Just as the transition from sterling to the dollar took thirty to fifty years, dollar erosion is proceeding gradually. If paying for AI services in dollars becomes the global standard, dollar hegemony is reinforced. If AI services paid in renminbi proliferate, dollar hegemony erodes.
Condition 4: Alliance network cohesion.
CHIPS Act implementation is underway. Subsidies of $7.86 billion for Intel, $6.6 billion for TSMC, and $4.75 billion for Samsung have been committed. As long as ASML's EUV equipment monopoly holds, China's advanced semiconductor production faces a structural bottleneck. ASML's 2024 China DUV revenue stood at 36.1 percent annually (peaking at 49 percent in a single quarter) before tightening export controls compressed it to roughly 33 percent annually in 2025.
The risk to this condition lies within the alliance. Trump's tariff policies generate friction with allied nations. Instability in the Taiwan Strait threatens the TSMC supply chain. An alliance is cohesion, not a contract. It holds only when incentives are aligned.
Condition 5: Sustained expansion of R&D investment.
AI capital expenditure by the Big Four tech companies expands from $400 billion in 2025 to an estimated $635-665 billion in 2026. Combining five hyperscalers (including Oracle) brings the total to $660-690 billion. This investment is the largest private expenditure ever directed at a single technology sector.
But investment must convert to revenue for the virtuous cycle to continue. The lag between AI investment and monetization is fueling "AI bubble" debates. Amazon's free cash flow (FCF) is projected to turn negative. If investment fails to generate returns, the next round of investment is imperiled.
How Scenario A Operates
When all five conditions hold, Scenario A operates as follows.
American frontier AI models maintain their lead. The standards for global AI cloud infrastructure (AWS, Azure, GCP) calcify around a US center. Subscribing to AI services in dollars becomes the world standard. Digital dollar hegemony replaces and reinforces physical dollar hegemony. Allied nations depend on the American ecosystem, and China converges into a domestic AI application powerhouse. Its challenge for global hegemony is thwarted.
This is Scenario A's logical chain.
The Threat from Within
The greatest threat to Scenario A is not China.
Direct AI layoffs in 2025 reached 55,000. Ford's CEO publicly stated that "AI will replace half of white-collar workers." Salesforce's CEO said "AI already handles 50 percent of the workload." In January and February of 2026 alone, 32,000 tech workers were laid off.
Seventy-nine percent of Americans favor AI regulation. Eighty-four percent of Republican voters, 81 percent of Democratic voters. In American history, few issues command such bipartisan consensus. Yet no legislation emerges from the hearing room.
Big Tech lobbying fills that gap. The $50 million spent by the seven largest tech companies constitutes a structural force blocking congressional action. Marcus Licinius Crassus privatized Rome's fire brigades. When the displaced saw their homes burning, he negotiated the price. The Crassuses of the twenty-first century operate the same way, preventing the displaced's suffering from translating into legislation.
This is the internal contradiction of Scenario A. The democratic institutions that created technological hegemony cannot absorb the costs that technological hegemony produces. The displaced's anger accumulates into populism. When populism becomes policy, immigration is blocked, alliances destabilize, and Scenario A's conditions unravel in sequence.
Scenario A collapses not from the outside but from within.
Conditions That Invalidate Scenario A
Scenario A requires all five conditions to hold simultaneously. Conversely, if any of the following occurs, the scenario is structurally weakened.
First, sustained technology reversal. If Chinese AI models lead US models on major benchmarks (MMLU, HumanEval, MATH) for three consecutive cycles, and that lead holds for six months or more. A "narrowing gap" and a "reversed gap" are qualitatively different events.
Second, a critical decline in dollar share. If the dollar's share of foreign exchange reserves (IMF COFER basis) falls below 50 percent and the renminbi's share breaks through 5 percent. Given the precedent that the sterling-to-dollar transition took thirty to fifty years, the acceleration of movement matters more than the threshold itself.
Third, alliance defection. If a key technology ally (the Netherlands, Japan, or South Korea) unilaterally resumes advanced equipment exports to China, or if major fab investment commitments within the US bloc are delayed or canceled on a large scale. An alliance is an alignment of interests. When interests diverge, cohesion dissolves.
These conditions are not predictions. They are monitoring criteria.
Section C: Scenario B — China Overtakes
2035, If Scenario B Materializes
August 2035. A logistics warehouse on the outskirts of Lagos.
Operations manager Chidi stands at the warehouse entrance holding a tablet. The screen powers on, a Chinese-language interface flickers briefly, then switches automatically to English and Yoruba. It is an Alibaba Cloud-based inventory management AI. Twenty-seven warehouses across Nigeria are connected through this system.
Chidi knows where this system came from. Had he not known the cost, he would not have chosen it.
In 2031, Nigeria's logistics industry evaluated an American-made system. The annual license cost was four times higher. Yoruba support was "under review." Alibaba's local partner already had a technical team in Lagos. The contract did not take long.
A camera is mounted above the warehouse entrance. It automatically identifies incoming shipments. The moment Chidi taps his confirmation on the tablet, the data flows to a server in Shenzhen. Chidi knows this. He does not care.
This is the logic of Scenario B. Technological hegemony derives not from peak performance but from the default installed in the most places. The warehouse in Lagos is that default.
The Training Center in Astana
Astana, the capital of Kazakhstan. Autumn 2025.
A vocational training center in the city. A sign on the concrete exterior is written in both Chinese and Kazakh. The facility was established as part of China's Digital Silk Road program.
At the front of the classroom stands Wang Tao, an engineer dispatched by Huawei. Forty trainees are taking an introductory course on Huawei's cloud platform. The textbook is written in both Chinese and Kazakh. The AI hardware was provided by the Chinese government.
In six months, the trainees will receive Huawei-certified AI technician credentials. Jobs are already arranged upon graduation. Twelve identical training centers operate across Kazakhstan. Nine in Uzbekistan. Five in Kyrgyzstan.
No one calls this "hegemonic expansion." They call it "technology cooperation."
Regardless of the label, the default technology worldview of these trainees becomes the Chinese ecosystem. The systems they install, maintain, and teach run on Huawei's operating system. When the Kazakh government next procures AI infrastructure, these trainees will be the advisors.
This is Scenario B on the ground.
The Core Logic of Scenario B
The core logic of Scenario B (where China overtakes the United States or comes to share technological hegemony) rests on a single proposition.
As the performance gap in frontier AI narrows, cost and accessibility become more important variables than performance.
Supachai's choice in Bangkok illustrates this. Even if American AI is 0.3 percentage points more accurate than Chinese AI, if it costs twenty times more, Supachai chooses the Chinese AI. Dr. Amina in Nairobi chose Chinese diagnostic AI over the American alternative not because it was superior in performance but because of cost and local fit.
"Good Enough AI" displacing "the best AI" in the market. This is the heart of Scenario B.
Five Conditions
Condition 1: Achieving semiconductor self-sufficiency, or a practical workaround.
Made in China 2025 set a semiconductor self-sufficiency target of 70 percent. By total production within China, the figure is close to 50 percent, but Chinese companies' own design-and-manufacture share stands at 19 to 23 percent. Semiconductor equipment self-sufficiency sits at 13.6 percent. There is no EUV capability.
Yet Huawei's self-sufficiency push is underway. The Ascend 910C is produced on SMIC's 7-nanometer process; despite a 40 percent yield ceiling, targets are 600,000 units in 2025 and 1.6 million dies in 2026. SMIC's advanced-node monthly production capacity is projected to expand from 45,000 wafers in 2025 to 80,000 by 2027.
The critical question: how far can DUV multi-patterning go without EUV? Structural disadvantages in cost and yield persist, but if the goal is "sufficient quantity at sufficient performance," full self-sufficiency is not required.
Condition 2: Strategic exploitation of energy advantage.
In the Industrial Revolution, Britain's abundant coal powered the steam engine era. In the AI era, coal is electricity.
Chinese data center electricity costs are less than half of US rates (Fortune). In 2024, China added 430 GW of net power capacity, more than fourteen times the US figure of 30 GW (National Energy Administration / Jefferies). Projected surplus power by 2030: 400 GW. The US power grid upgrade is mired in permitting processes that take decades.
What does this mean? DeepSeek built a competitive model with 2,048 H800 GPUs at a training cost of $6 million — less than one-tenth the resources of US Big Tech. If energy costs are half, inference costs fall accordingly. With economies of scale, the marginal cost of "good enough AI" drops dramatically.
DeepSeek's API price is $0.55 per million input tokens. OpenAI's reasoning model o1 costs $15. Model tier and performance differ, so direct comparison has limits, but the price gap itself sends an unmistakable signal to the market.
Condition 3: Default adoption across the Global South.
In Africa, DeepSeek usage is two to four times higher than US AI services (Microsoft AI Diffusion Report 2025). Temu holds 24 percent of global cross-border e-commerce, effectively tied with Amazon at 25 percent (IPC 2025). China's digital services trade surplus reached a record $33 billion in 2025.
The Digital Silk Road is not only hardware. Like the training center in Astana, it cultivates technical talent and seeds ecosystems. If five billion people in the Global South adopt the Chinese AI ecosystem as their default, data and users accumulate, feeding a virtuous cycle of model improvement.
Condition 4: Technological compensation for the demographic crisis.
China's demographic cliff is undeniable. Births in 2025: 7.92 million. The share of the population aged sixty and over rises from 23 percent in 2025 to above 30 percent by 2035. Total population has declined for four consecutive years.
The Chinese government aims to offset this decline with "Physical AI." Humanoid robots, autonomous vehicles, and smart factories are the pillars. "Physical AI" emerged as a core theme in the 2026 semiconductor supercycle (SEMICON Korea 2026). If this strategy succeeds, the demographic cliff's impact is mitigated. Whether robots can replace newborns, however, is an economic and social question before it is a technological one.
Condition 5: A soft landing on debt.
Macro leverage ratio: 302.4 percent (2025, up 12.4 percentage points year-over-year). Government debt under the IMF's expanded definition: 127 percent of GDP (2025 estimate). Net FDI outflows hit a record $168 billion in 2024. Real estate investment fell 13.9 percent year-over-year in the first three quarters of 2025.
Goldman Sachs estimates the property downturn drags approximately 2 percentage points off annual GDP growth. Morningstar, however, projects property market stabilization by late 2026 to 2027. If China can complete restructuring while avoiding a Japan-style prolonged stagnation, this condition is met.
How Scenario B Operates
When all five conditions hold, Scenario B operates as follows.
DeepSeek-style efficient AI combines with energy advantages. The marginal cost of AI services drops dramatically. Lower costs drive default adoption of the Chinese AI ecosystem across the Global South. The choices made in Bangkok, Nairobi, and Jakarta are replicated billions of times over. Data and users accumulate. Models improve. Improved models drive further adoption.
If this virtuous cycle takes hold, American frontier AI's technology lead fails to convert into a market lead. The best models exist, but they lose market share to a cheaper, more accessible competitor. The result is a practical sharing — or reversal — of technological hegemony.
The Double Constraint
Two structural constraints stand in Scenario B's way.
The first is the demographic cliff. Consumer contraction, domestic market shrinkage, labor force decline. These are structural problems that short-term policy cannot solve.
The second is the institutional ceiling on innovation. Alibaba's permanent headcount fell from 254,941 in 2022 to 124,320 in 2025 — a 51.2 percent reduction. Baidu stood at 35,900 employees by the end of 2024, down 21.1 percent from its peak in 2021. Is this AI-driven efficiency or regulatory suppression? Either way, in a society where 73.1 percent of college graduates aspire to civil service or state-owned enterprises — and only 12.5 percent prefer private companies — creative destruction in the private sector is stifled.
Tang ping (lying flat) is both economic pressure and a silent protest against the system. When the Cyberspace Administration of China (CAC) censors tang ping-related social media posts, it is severing an early-warning system for social discontent. When institutions operate without feedback, errors go uncorrected.
Scenario B materializes only when all five conditions are met and the double constraint is overcome. Those conditions are not easy to satisfy. But they are not impossible.
Conditions That Invalidate Scenario B
If Scenario B's double constraint (the demographic cliff plus the institutional ceiling) worsens beyond the point of recovery, the scenario is invalidated.
First, semiconductor self-sufficiency stalls. If China's narrowly defined semiconductor self-sufficiency rate (own design and manufacture) fails to exceed 30 percent by 2030, the hardware bottleneck structurally blocks the expansion of the "good enough AI" strategy.
Second, debt crisis materializes. If the macro leverage ratio exceeds 350 percent while GDP growth falls below 3 percent, fiscal capacity to sustain both AI infrastructure investment and social safety net expansion is exhausted.
Third, Global South defection. If India builds an independent AI ecosystem and Chinese AI service market share in the Global South stagnates or declines, the premise of the "default adoption" strategy collapses. India's choice is the largest unknown in this variable.
Fourth, accelerating demographic cliff. If the total fertility rate falls below 0.8 and robotic labor substitution fails to extend beyond manufacturing into services and caregiving, domestic market contraction breaks the virtuous cycle.
Section D: Scenario C — A Fractured World
2035, If Scenario C Materializes
November 2035. The transit zone at Istanbul Ataturk Airport.
Ercan, a digital policy advisor to Turkey's Ministry of Health, pulls two smartphones from his bag. One is for the US bloc. Accessing Google's healthcare AI system requires this device. It is compliant with EU medical data regulations. The other is a Huawei device. The system connecting to the Central Asian health network supports only this one.
He is on his way from a conference in Geneva to Baku. Which phone he uses depends on his destination.
Ercan's job is to advise on which AI standard Turkey should adopt. But his bag already gives the answer. Turkey adopted both. It spent the past two years developing a middleware layer compatible with both standards.
Someone once asked how much it cost. Ercan paused for a moment, then answered that it was cheaper than the cost of giving one up. For now.
This is the daily reality of Scenario C. Fracturing proceeds without a declaration. Through two smartphones in an airport transit zone, through the overtime of a middleware development team, through the qualifier "for now."
Two Server Racks
2029 (hypothetical). Singapore.
Fintech startup Paybridge. CTO Amirah Ahmad stands at the entrance to the server room, looking at two racks.
The left rack. AWS cloud connection, NVIDIA H200 accelerators, OpenAI API integration. Services for US-bloc clients run here. Half of the customers in Singapore, Malaysia, and Thailand connect through it.
The right rack. Alibaba Cloud connection, Huawei Ascend accelerators, DeepSeek API integration. Services for China-bloc clients. The other half (customers in Indonesia, Cambodia, Myanmar) connect here.
The same service runs on two stacks. Two codebases. Two authentication systems. Two payment pipelines. Forty percent of the development team is devoted to this dual maintenance.
Amirah says that companies like hers bear the cost of the technology competition. Not the United States. Not China. Companies like hers.
This is Scenario C on the ground.
Fracturing Already Underway
Scenario C is not a future possibility. It is already underway.
Hardware stack fracturing. The US bloc consists of NVIDIA GPUs + TSMC foundry + ASML EUV + SK hynix and Samsung HBM. The China bloc consists of Huawei Ascend + SMIC foundry + DUV-based alternatives + domestic HBM development. The two stacks operate on different technical standards. Huawei's CANN (Compute Architecture for Neural Networks) framework is incompatible with NVIDIA's CUDA ecosystem. Choosing one stack closes off the other's entire software ecosystem.
By 2026, this hardware bifurcation becomes visible (Pantheon Insights analysis). Which hardware you choose determines every software decision built on top of it.
Software and model ecosystem fracturing. The US bloc has OpenAI API, Claude, Gemini. The China bloc has DeepSeek, Qwen (Tongyi Qianwen), GLM, and Kimi.
The role of open source deserves attention. Nine of the top ten open-weight models are Chinese-made (ChinaTalk). Open source could serve as a bridge across the fracture — or as a pathway for Chinese ecosystem proliferation worldwide. Which it becomes depends on regulatory choices by individual governments.
The "Decoupling America's AI from China Act" — a bill introduced in the wake of the DeepSeek shock — seeks to restrict open-source adoption as well. If passed, fracturing accelerates.
Governance and norms fracturing. The United States favors minimal regulation. One federal AI bill in 2025. Trump revoked Biden's AI safety executive order on his first day in office. China implemented the world's first binding generative AI regulation (August 2023, nine months from conception). The EU applies the AI Act in full starting August 2026. Forty-four countries participate in the Global Partnership on AI (GPAI), but no shared framework exists between the United States and China.
Governance fracturing runs deeper than hardware fracturing. Which AI is permitted and which is prohibited, how data is processed, who bears liability. When these norms do not align, system integration itself becomes impossible.
Eric Schmidt put it this way: "A China-led internet and a non-China internet are diverging." In a September 2025 interview, Schmidt acknowledged his earlier judgment was wrong. "I thought the United States and China were competing on equal footing in AI. But they were doing something very different from what I thought." His warning: with China competing through open-source models and the United States through closed models, "most of the world that corresponds to the Belt and Road will end up using Chinese models, not American ones." Kai-Fu Lee concurs: "The world is splitting into countries that adopt Chinese apps and countries that adopt American apps."
The Three Axes of Scenario C
Fracturing proceeds simultaneously along three axes: hardware, software, and governance.
When all three axes diverge simultaneously, Scenario C calcifies. If only one diverges, the other two can maintain connections. When all three diverge, exchange between the two ecosystems becomes technically impossible.
Where do things stand now? Hardware divergence is underway. Software divergence has begun. Governance divergence is already reality. All three axes show active divergence, but the calcification stage has not yet been reached.
The Distribution of Fracturing's Costs
The costs of Scenario C are not distributed equally.
The United States and China each build and innovate within their respective blocs. A significant share of that cost is borne by economic actors within each bloc.
But third-country companies like Amirah's startup bear a double cost. Firms serving clients in both blocs must maintain two infrastructures. Forty percent of a development team going to dual maintenance means 40 percent of the resources available for innovation disappear.
Global innovation also slows. Knowledge, data, and talent exchange between the two ecosystems declines. The overall pace of AI advancement decelerates.
Consumer costs rise. Redundant infrastructure, incompatible devices, segregated services. The bill is passed to the end user.
The "coexistence illusion" examined in Volume 1 formally ends in this scenario. The belief that economic interdependence would prevent technological decoupling — this belief fails in the face of export controls and standards fragmentation.
Yet Fracturing Creates Opportunity for Some
Scenario C does not impose costs on everyone equally.
Countries possessing technology needed by both sides see their bargaining power grow as fracturing deepens. SK hynix and Samsung supply HBM to the US bloc (NVIDIA). Simultaneously, China pours massive resources into domestic HBM development. With Korean memory companies deeply embedded in the US bloc's supply chain, South Korea's leverage in Scenario C may actually strengthen.
This point is developed in detail in the Epilogue, which examines four "third players": South Korea, India, the EU, and the Middle East.
Conditions That Reverse Scenario C
Scenario C materializes when divergence along all three axes (hardware, software, governance) calcifies. Conversely, the following conditions could reverse the fracturing.
First, an AI safety accord. If the United States and China agree on a joint framework for AGI safety. Precedent exists: the Nuclear Non-Proliferation Treaty (NPT) was signed in 1968, in the depths of the Cold War. Just as the fear of mutually assured destruction (MAD) compelled cooperation, the existential risk of AGI could play an analogous role.
Second, reintegration of technical standards. If both sides agree on a common AI interface standard, or if a de facto compatibility layer forms organically in the market. Precedent: TCP/IP unified what could have been a fragmented network into a single internet.
Meanwhile, one sub-variant of Scenario C requires consideration. Not US-China bipolar fracturing, but multipolar dispersion. If the EU constructs a third AI governance bloc built on Mistral and the AI Act, and if India leverages 1.3 billion people's data and its Digital Public Infrastructure (DPI) to form an independent technology bloc, the world disperses into three or four poles rather than two. In this variant, neither the United States nor China can impose its bloc's default. The costs of fracturing rise, but the space for negotiation widens. The "third players" discussed in the Epilogue are the critical variables in this variant.
Section E: Scenario Intersections and the Time Axis — Reality Is a Hybrid
From 1769 to 2090
- Richard Arkwright built a water-frame factory in Cromford, Derbyshire. No one knew how it would reshape the world of Lancashire's handloom weavers.
Around 1785, a Lancashire handloom weaver's weekly wage was twenty-five shillings. Skilled weaving was a source of pride.
By 1835, the same weaver's weekly wage had fallen to four and a half shillings. Productivity soared. Real wages lagged for decades. The long gap in which the fruits of productivity failed to reach labor — that was the Engels' Pause.
In 1833, the Factory Acts passed. From Arkwright's factory to the Factory Acts: sixty-four years.
Those sixty-four years represent the upper bound of the historical band for institutional adaptation — fourteen to sixty-four years. Railways took fourteen years. The internet took twenty-seven. Sixty-four years is the longest case in the band.
Add sixty-four years to 2026 and you reach 2090.
The AI era is different, of course. Technology cycles have compressed to three to five years. Whether institutional cycles compress at the same rate is unknown. If the next wave arrives before institutions adapt to the current one, the structural gap may accumulate rather than close.
In the middle of that uncertainty, this book's readers must make decisions. The twenty years from 2026 to 2045 are the window.
The Scenario Matrix by Time Horizon
The three scenarios unfold differently by sector, time horizon, and region.
Short term (2026-2030).
In AI model development, Scenario A dominates. American companies lead in frontier models. With Big Four tech CapEx expanding to $635-665 billion, overwhelming advantage in capital deployment is sustained.
In AI applications and deployment, A and B are mixed. Developed-market defaults are American services, but DeepSeek and Chinese AI service adoption spreads across the Global South. Choices like Supachai's accumulate.
In semiconductors, A dominates but the seeds of C are planted. While the US bloc's NVIDIA + TSMC + ASML supply chain holds, China expands Huawei Ascend production and builds an independent stack.
In finance, dollar inertia maintains A. The structure of renminbi at 2 percent and the dollar at 56 percent does not change in the short term.
In governance, C has already begun. The full application of the EU AI Act (August 2026), US minimal regulation, and Chinese comprehensive regulation — a three-way divergence calcifies.
Medium term (2030-2035).
This is the true inflection point.
As the AI model gap narrows further, A and B compete. If the performance of "good enough AI" reaches a level practically indistinguishable from frontier models, cost and accessibility become the decisive variables.
In semiconductors, technological divergence deepens. Whether C calcifies depends on how far China's self-sufficiency push reaches. If China succeeds in developing independent EUV or next-generation lithography technology, Scenario B hardens. If it fails, the China bloc under Scenario C compensates for performance disadvantage with cost and scale.
Governance fracturing calcifies into C. If the GPAI fails to produce a joint US-China framework, international AI governance fragments into three pieces: the "Brussels Effect" spreading the EU standard to third countries, voluntary standards from US corporations, and China's state-imposed standards.
Long term (2035-2045).
Uncertainty in this period is extreme.
Three variables prove decisive. Whether AGI (artificial general intelligence) is achieved. Stability in the Taiwan Strait. How China's debt problem resolves.
If an American company achieves AGI, Scenario A is irreversibly strengthened. The technology gap reopens, and its velocity makes catch-up impossible.
If a military conflict erupts in the Taiwan Strait, Scenario C shifts abruptly. A severed TSMC supply chain confronts the entire US bloc with a semiconductor crisis. Paradoxically, the greatest cost falls not on China but on the world at large.
If China achieves a soft landing on debt and continues leveraging its energy advantage, Scenario B can materialize.
The conditions for any single outcome to materialize with certainty cannot be known today.
What Goldman Sachs Left Unsaid
Goldman Sachs projects that GenAI will lift global GDP by 7 percent ($7 trillion). That translates to 1.5 percentage points of annual productivity growth (based on Epoch AI analysis).
McKinsey projects that AI will generate $13 trillion in additional economic activity by 2030.
These numbers say "AI will deliver growth." They do not say who captures that growth.
Whether Goldman Sachs's $7 trillion concentrates in one bloc or another under one scenario or another — that is what determines the next twenty years.
Under Scenario A, this $7 trillion concentrates in US Big Tech and the dollar ecosystem. Under Scenario B, it disperses into Chinese AI services and the Global South data ecosystem. Under Scenario C, a portion is lost to duplication costs, and the remaining growth splits between two blocs.
How investors and policymakers position themselves under each scenario — that is the subject of the next chapter.
Section F: Democracy versus Authoritarianism — The Gap Between Two Clocks
What the New Deal Teaches
October 24, 1929. Stock prices on Wall Street collapsed.
Black Thursday. The year after, the year after that, the year after that — the American economy kept contracting. By 1932, unemployment reached 25 percent. Factories stopped. Banks failed. Farmers burned their harvests.
Franklin D. Roosevelt took office in March 1933. Four years after the onset of the Great Depression. The New Deal's policies did not produce tangible effects until the external shock of the Second World War.
Democratic institutions repeat this pattern. They respond only after a crisis grows large enough.
Abolition of slavery came ninety years after the founding, after the external shock of the Civil War. Women's suffrage came after the external shock of the First World War. New Deal labor law came after the external shock of the Great Depression.
Has the AI crisis "not yet grown large enough"? If 55,000 white-collar layoffs are not enough, how many are? How many 2026 graduates must fail to find jobs before Congress moves?
The Speed of Two Clocks
At the federal level, AI legislation in the United States is gridlocked. Meanwhile, thirty-eight states have adopted around 100 AI-related measures. Decentralized experimentation at the state level is underway. Colorado mandated AI impact assessments. Illinois enacted regulation of AI hiring algorithms. California is debating an AI transparency bill.
Federal gridlock does not block state-level innovation. This is democracy's mechanism of decentralized experimentation. Fifty states become fifty laboratories. What works rises to the federal level.
But the speed is slow. Seventy-nine percent of the public favoring AI regulation does not translate into federal legislation. This is democracy's "democratic deficit."
China's clock is different. From generative AI regulatory concept to enforcement: nine months. The world's first binding generative AI regulation. Government AI capital expenditure: 400 billion yuan in 2025. "Monday's decision becomes nationwide deployment by Friday," as Oxford's Blavatnik School of Government documented.
On speed alone, authoritarianism appears advantageous. But speed does not guarantee accuracy.
Speed and Error Correction
Authoritarian systems have two structural weaknesses.
The first is information distortion. When tang ping-related social media posts are censored by CAC directive, the signal for social discontent is severed. Policymakers make decisions without knowing actual conditions. Errors are not corrected but amplified.
The second is the suppression of entrepreneurship. After the Jack Ma affair, the behavioral patterns of Chinese private entrepreneurs shifted. Seventy-three point one percent of college graduates aspire to civil service or state-owned enterprises. Creative destruction in the private sector, the wellspring of innovation, is suppressed.
Yuval Noah Harari has argued that in the AI era, centralized data processing may be more efficient than distributed systems. This argument carries a precondition. "Efficiency" and "adaptability" are not the same. An efficient system achieves its goals rapidly. An adaptive system detects when its goals are wrong and changes course.
The problem with authoritarian systems is that they are very fast when the goal is correct. When the goal is wrong, the mechanism for recognizing and correcting that error is weak. The Soviet Union did not collapse from military failure. It collapsed because it failed to correct economic planning errors for seventy years.
China differs from the Soviet Union for a reason. "Fragmented authoritarianism" — maintaining central control while permitting flexible policy experimentation at the local level — provides a degree of error correction. Just as Shenzhen served as a space for market economy experimentation, certain regions function as policy laboratories.
Whether this mechanism is sufficient for the speed of the AI era has not been tested.
AI's Effect on Institutional Speed
Two hypotheses compete.
Hypothesis A: AI accelerates institutional adaptation.
AI becomes a tool for policy analysis, public opinion simulation, and impact assessment. A thousand-page regulatory proposal is analyzed and simulated by AI in twenty-four hours. Both the quality and speed of policy decisions improve. The historical band's upper bound of sixty-four years could compress to thirty or forty.
If this hypothesis is correct, democracy's "crisis-then-response" pattern breaks. The signals of AI crisis are detected early, enabling preemptive response.
Hypothesis B: Accelerated destruction and accelerated adaptation cancel each other out.
AI also accelerates the speed of destruction. If technology cycles compress to three to five years, the next wave arrives before institutions adapt to the current one. When the Factory Acts passed in 1833, the next innovation after the water frame was decades away. In the AI era, GPT-5 arrives before adaptation to GPT-4 is complete; agentic AI arrives before adaptation to GPT-5.
If this hypothesis is correct, the Engels' Pause recurs repeatedly, each pause shorter in duration. The structural gap persists or accumulates.
Current evidence can be read as supporting or refuting either hypothesis.
Is Democracy Slow but Right?
Institutions are slower than technology. The historical base rate suggests that democratic institutions have achieved more durable adaptation — but that is a past record, not a guaranteed future.
In the AI era, this proposition is under challenge.
"Was democracy slow but right, or does speed become justice in the AI era?"
The answer will be rendered by history over the next twenty years, depending on which of the three scenarios is strengthened.
What we can do now is read, in real time, which of the three paths is being reinforced — while we wait for that answer.
Connection to Volume 1: Same Formula, Three Outcomes
Volume 1 showed a history in which technological explosions followed a single thread. Roman roads reshaped Mediterranean trade. Arkwright's factory reshaped Britain's labor structure. AI is reshaping cognitive labor.
But the AI era is not a single thread.
The same explosion — AI — is proceeding in two empires simultaneously. In 1769 Lancashire, the explosion occurred within a single institutional context. In 2026, two fundamentally different institutional systems — democracy and authoritarianism — are absorbing the same technology at the same time. The result diverges into three paths.
The core formula presented in Volume 1 — technology leads to capital concentration leads to social unrest leads to institutional redesign — operates in all three scenarios. The difference is the speed and direction of each stage.
In Scenario A: US-led frontier AI produces hyperconcentration in the Big Five tech companies. Fifty-five thousand white-collar workers are displaced. Populist backlash forms. But democratic institutions — slowly — find a new equilibrium through adaptation.
In Scenario B: Chinese state capital and private AI produce hybrid concentration. Tangping and the age-35 crisis manifest as forms of social unrest. Top-down institutional redesign proceeds rapidly, but the limits of error-correction capacity threaten long-term sustainability.
In Scenario C: The formula operates in both blocs. Third countries bear the institutional redesign costs of both blocs simultaneously. Supachai in Bangkok, Amina in Nairobi, Amirah in Singapore — each navigates a choice between two algorithms in their own way, bearing the cost of that choice.
The formula is identical. Only the outcomes differ.
In Volume 1, it took thirty years for the Lancashire handloom weaver's weekly wage to collapse from twenty-five shillings to four and a half. The clock of the AI era runs faster. Yet the institutional clock remains slow. This gap — between the speed of technological change and the speed of institutional adaptation — is the source of pain for Volume 1's readers, for this chapter's readers, and across all three scenarios simultaneously.
In that space of pain, the discerning move differently.
Bridge: From Map to Action
Three paths have been drawn. A: the United States maintains frontier AI, dollar reserve status, and the alliance network, controlling the world standard. B: China leverages energy advantages and efficient AI to become the Global South's default. C: simultaneous divergence along the hardware, software, and governance axes hardens into two technology blocs.
The three paths are not mutually exclusive. Different scenarios operate simultaneously across sectors, time horizons, and regions. In the short term, A dominates. In the medium term, fracturing deepens. The long term is uncertain.
What matters is not predicting which path is right. It is reading in real time which path is being strengthened. DeepSeek's adoption in a Bangkok classroom is a signal for B. The absence of legislation from a Washington hearing room is a signal of A's internal vulnerability. Two server racks in a Singapore fintech is a signal that C has already arrived on the ground.
Reading signals and acting on signals are different capabilities.
Add sixty-four years to 2026 and you reach 2090. In the middle of the Engels' Pause, we must act now.
On December 29, 1989, the Nikkei index hit an all-time high of 38,915 yen. The title of a Tokyo brokerage report read: "Japan, Hegemon of the Twenty-First Century." What that investor needed was not a more accurate prediction. It was a better positioning framework — one that would not collapse under any scenario.
Key data sources: Epoch AI (US-China AI benchmark gap, time lag), IMF COFER / St. Louis Fed (dollar foreign exchange reserves), LLM Stats (open-source model market share), CNBC (Big Tech CapEx, white-collar layoffs), Fortune / Brookings (China energy additions), White & Case (China AI regulation timeline), Drata (US AI legislation status), Issue One (Big Tech lobbying), Goldman Sachs / McKinsey (GenAI GDP impact), WEF Global Risks 2026 (geoeconomic risk ranked #1), Oxford Blavatnik School (democracy vs. authoritarianism AI policy speed), Bairoch (1913 Germany-Britain manufacturing share), Microsoft AI Diffusion Report 2025 (DeepSeek Africa usage), Pantheon Insights / Yale SOM (technology bloc divergence), ChinaTalk (share of Chinese-made open-source models), SCMP (Alibaba and Baidu employment), Zhaopin (graduate job preferences), CISSM / UMD (AI regulation approval rate)