Regional Disparities

In May 2026, the President of Kenya went on television to explain why a data center was not going to be built. The facility, a billion-dollar project backed by Microsoft and the Emirati firm G42, had been announced with fanfare as proof that the AI economy was coming to East Africa. Then the engineers did the arithmetic. Running the data center at full capacity would have drawn roughly a third of Kenya's entire installed electricity supply—about 3,000 megawatts for the whole country, and this single building wanted a thousand of them. President Ruto suspended the project. There was simply not enough power.

Hold that image next to another one from the same season. In the first quarter of 2026, global venture funding crossed three hundred billion dollars in three months, and roughly eighty percent of it went to artificial intelligence. A single company, OpenAI, closed a financing round larger than the annual economic output of most nations on Earth. Anthropic raised thirty billion dollars, xAI twenty, Waymo sixteen. Every one of those companies sits within a few miles of the same stretch of the San Francisco Bay.

This is the geography of AI in a sentence: a building in Nairobi that cannot get electricity, and a neighborhood in California where more capital arrives in a quarter than some continents will see in a decade. AI is not spreading evenly across the map. It is pooling—in particular cities, particular countries, particular grids—and almost everywhere the pool does not reach is becoming, in relative terms, poorer, slower, and more peripheral to the century's defining technology. Understanding how that concentration works, why it feeds on itself, and what it costs the places left out is the subject of this chapter.

The Pull of the Superstar Cities

Start with where the money goes, because money is the least ambiguous signal. In 2025, AI companies attracted $211 billion in venture capital worldwide—an 85 percent jump over the previous year, and roughly one out of every two venture dollars invested anywhere on the planet (The AI Economy, 2026). Of that global total, the San Francisco Bay Area alone captured 60 percent—$126 billion—despite hosting only about 22 percent of the world's AI deals. The region is not just winning; it is winning disproportionately to its own already-outsized share of activity. And within the Bay Area, the concentration folds in on itself again: $113 billion of that $126 billion went to just 92 companies raising rounds of $100 million or more. AI now absorbs 81 percent of all startup capital in the region, up eleven points in a single year.

Talent tracks the money. California posted 170,881 AI job openings in 2025, more than 17 percent of the United States total, according to the Stanford AI Index. Texas and New York followed at some distance; together the top three states accounted for roughly a third of all AI hiring in the country (Stanford HAI, 2026). The very highest concentrations show up in a handful of dense metros—Washington, D.C., Seattle, the Bay Area—where the share of jobs touching AI runs several times the national average. Globally the same clustering repeats at the level of cities: Singapore, Hong Kong, and Luxembourg top the charts for AI's share of local hiring, tiny high-income enclaves pulling in work far out of proportion to their size.

None of this is unique to AI. It is the latest and most extreme expression of a pattern that has been tightening for two decades. More than 90 percent of United States innovation-sector job growth between 2005 and 2017 landed in just five coastal metros. What AI has done is take an already lopsided map and press down harder on the winners. Even wealthy, well-educated cities that are not in the club feel the squeeze. San Diego—home to world-class universities and a serious biotech industry—watched its venture funding fall to a six-year low in late 2025, as capital consolidated into the metros building AI and left the rest to compete for what remained.

Why the Advantage Compounds

The reason a place like San Francisco keeps winning is not mysterious, and it is not mainly about weather or taxes. It is agglomeration: the self-reinforcing productivity of putting the right people, firms, and money close together.

When leading AI labs cluster in one region, they create a thick local market for a specific kind of labor—researchers who understand large-model training, engineers who have shipped inference systems at scale, product managers fluent in what the models can and cannot do. That thick market is itself the draw. An engineer choosing where to live knows that in the Bay Area a single bad job can be replaced by ten others within a bike ride, which lowers the risk of moving there, which brings more engineers, which makes the region more attractive to the next lab. Capital behaves the same way: investors want to be where the deals are, deals happen where the founders are, founders go where the investors and engineers already sit. Specialized suppliers appear—the law firms that know how to paper an AI licensing deal, the recruiters with the right Rolodex, the cloud brokers who can find spare GPU capacity. Knowledge itself leaks usefully between people who bump into each other, in ways that video calls still do not fully replicate. Each of these forces would matter on its own. Together they lock in.

The crucial property of agglomeration is that it is far easier to start than to copy. A region that missed the first wave can lure an individual company with tax breaks, or fund a research center, or subsidize an office. What it cannot easily manufacture is the whole ecosystem at once—the suppliers and the labor pool and the investor density and the culture of people who have done this before. You can buy a building. You cannot buy the fact that the person in the next building has already solved the problem you are stuck on.

The Fifty-Year Rule

History suggests how long this kind of lock-in tends to last, and the answer is sobering. Economist Aakash Kalyani and colleagues, tracking how new technologies spread through the American economy, found that a major novel technology takes on the order of fifty years to disperse geographically—and that the dispersion is slowest, and least complete, for exactly the high-skill, high-value jobs that matter most. The places where a technology was pioneered remain its high-skill center for decades after the routine work has scattered. In their data, 56 percent of the most economically impactful technologies of the era traced back to just two American locations: Silicon Valley and the Northeast Corridor (Kalyani et al., Quarterly Journal of Economics, 2025).

The lived evidence is all around us. Detroit still anchors American automotive engineering more than a century after the Model T. Silicon Valley still designs the world's most advanced chips fifty years after the integrated circuit. What disperses first is the low end—the assembly, the call centers, the off-the-shelf deployment. What stays put is the frontier: the research, the strategic decisions, the highest-paid roles. Over time the average skill level of a technology's jobs falls as the work commoditizes and spreads, but the commanding heights barely move.

Apply that clock to AI and the implication is stark. If the frontier is concentrating in San Francisco, Seattle, and a few peer cities in the 2020s, the fifty-year rule would push meaningful dispersion of the highest-value activity toward the 2070s. Ordinary people in second-tier cities and lower-income countries might get the tools far sooner—a small business in Nairobi or Nashville can already rent a frontier model by the token—but the labs, the capital, and the power to shape where the technology goes next would stay clustered for a working lifetime.

There is a serious counterargument, and honesty requires stating it, because it cuts to whether the historical clock even applies. AI is not a factory. Using a model requires no local plant, no physical supply chain at the point of use—only a browser and a connection. Knowledge, measured by patent citations, already crosses borders roughly twice as fast as it did in the 1970s. On that view, the use layer of AI could diffuse in years, not decades. But the fifty-year rule was never mainly about the tools; it was about the high-value production of the technology—the research, the founding, the investing—and those activities run on agglomeration economics that a fast internet connection does nothing to dissolve. The likeliest outcome is a split-screen: rapid global spread of AI consumption, and stubborn, decades-long concentration of AI creation and the wealth it generates. The gap between using AI and owning the companies that make it may be the defining inequality of the next fifty years.

The Divides, Measured

Concentration shows up as two great gaps: within countries, between city and countryside, and between countries, between the Global North and the Global South. Both are wide, and by the best current measures both are getting wider.

Inside the wealthy world, the urban-rural gap is already concrete. The OECD estimates that 45 percent of jobs in its member countries' urban regions are exposed to AI-driven automation, against just 13 percent in the most rural regions; measured per worker, urban employees face about 32 percent exposure to generative AI in their tasks versus 21 percent for rural workers (OECD, 2024). Exposure is double-edged—it means both disruption and opportunity—but the workers positioned to gain from AI, to bolt it onto their output and become more productive, are overwhelmingly the urban ones. The infrastructure gap underneath this is widening independently: across the OECD, the absolute broadband-speed gap between metro and remote regions grew from 22 to 58 megabits per second, with metropolitan download speeds now running 44 percent faster than in areas far from cities (OECD, 2025). AI applications that lean on the cloud need not just any connection but a fast, stable one—precisely the thing rural areas are falling further behind on.

At the global scale the numbers are starker still. Microsoft's AI Diffusion Report, one of the first attempts to measure worldwide generative-AI use directly, found that 16.3 percent of the global population used these tools in the second half of 2025. But the aggregate hides the split: 24.7 percent of working-age people in the Global North, against 14.1 percent in the Global South—and the gap between them widened from 9.8 points in the first half of the year to 10.6 points in the second, as Northern adoption grew nearly twice as fast. Every one of the ten countries that gained AI users fastest was a high-income economy (Microsoft AI Economy Institute, 2026). The trajectory, at least for now, is divergence, not catch-up.

Income tier Population with internet access Generative-AI adoption trend
Low-income countries ~27% Lowest; growing slowest
Lower-middle-income ~53% Lagging, uneven
Upper-middle-income ~80% Rising, concentrated in cities
High-income countries ~93% Highest; fastest recent gains

The bottleneck beneath the adoption gap is physical, and Africa shows it most vividly. A continent of 1.4 billion people holds less than 1 percent of the world's data-center capacity. Capacity in its five largest markets is projected to grow from roughly 400 megawatts today to somewhere between 1.5 and 2.2 gigawatts by 2030—real growth, but from a rounding-error base, and the continent would need on the order of 700 new data centers to give its citizens a genuinely inclusive digital future (McKinsey, 2025). The binding constraint is not code or ambition; it is electricity. The suspended Kenyan project that opened this chapter was not an outlier but an illustration of a rule: energy supply has become the single most critical issue facing digital-infrastructure investors across Africa, and foreign funding for the grid expansion that would fix it fell sharply through 2025, with African foreign direct investment down 42 percent in the first half of the year. Without power, the data centers sit dark. Without data centers, latency makes real-time AI impractical. Without either, no amount of policy ambition at the software layer can compensate. These are problems that take decades and staggering capital to solve.

The Global South is not one place. India, Brazil, the Gulf states, and parts of Southeast Asia are building real domestic AI capacity and belong to a different story. But for the large majority of low- and lower-middle-income countries, the current path points toward the margins of the AI economy rather than toward its center.

How the Forces Feed Each Other

The gaps described so far are not independent. Three mechanisms—talent migration, underrepresentation in data, and exclusion from governance—each widen regional disparity on its own, and, worse, they interlock, each one deepening the conditions that produce the others.

graph TD
    A[AI activity concentrates in hubs] --> B[Skilled workers migrate to hubs]
    B --> C[Lagging regions lose capacity to build AI]
    C --> D[Little local data collected]
    D --> E[Populations underrepresented in training data]
    E --> F[AI serves those populations poorly]
    F --> G[Weak local demand and benefit]
    G --> C
    C --> H[No delegation, staff, or funds for governance forums]
    H --> I[Rules written without them]
    I --> J[Standards fit hub conditions, not theirs]
    J --> G
    B --> A

Take talent first. A computer-science graduate in rural Iowa, or in Lagos, faces a labor market where the best-paid and most technically ambitious AI jobs sit in a few coastal metros in a few rich countries. The rational individual move—to pack up and go—drains human capital from the very places that can least afford to lose it. AI intensifies a brain drain that long predates it, because it has pushed the salary premium on frontier technical skill to extraordinary heights and concentrated the top jobs in an unusually small number of postal codes. Each departure makes local capacity-building harder and signals to the next student that the future is elsewhere.

The thinning of local talent has a second-order effect that is easy to miss: it means little local data gets created in a usable form, which brings us to underrepresentation. AI models are built from datasets that overwhelmingly reflect urban, high-income, English-speaking, Global North life—because that is where the infrastructure to generate and capture data exists. Regions outside those zones become blind spots in the training corpus, and their blind-spot status lowers the incentive to invest in collecting data there, which keeps them blind. The exclusion is structural, not malicious, which is exactly what makes it durable.

The third force, governance exclusion, closes the loop. Fewer than a third of developing countries have a national AI strategy, and until recently 118 nations—most of the Global South—sat outside the major forums where AI's rules were being written, the processes run by the G7, the G20, the OECD, the EU (UNCTAD, 2025). The barrier is rarely a lack of will; it is a lack of resources. A government without spare technical staff cannot send experts to engage complex regulatory proposals, cannot fund a standing delegation, cannot implement a standard even where it agrees with one. So the standards get set by and for high-income conditions—assumptions about bandwidth, institutional capacity, and use cases that do not travel—and then those standards are deployed back into contexts they were never designed for. The August 2025 UN resolution establishing a Global Dialogue on AI Governance and an Independent International Scientific Panel, the first bodies to seat all 193 member states, is a genuine step, and its explicit provision for funding developing-country participation is an admission that a seat at the table is worthless without the means to occupy it. Whether it changes the underlying arithmetic remains to be seen.

What Underrepresentation Does to Real People

It is worth making the data problem concrete, because abstractly "underrepresentation in training data" sounds like a technicality, and in practice it decides who a system works for.

In medicine, the skew is measurable. A review of publicly available clinical datasets used to build health AI found that more than 70 percent over-represented White, higher-income populations; Hispanic patients made up just 2.8 percent of the data despite being 18 percent of the United States population, and Black patients 7.3 percent against a 13 percent share (Nature Digital Medicine, 2025). A model learns the patients it is shown. A skin-cancer classifier trained mostly on light skin misses cancers on dark skin. A diagnostic tool tuned to the disease prevalence and health profiles of Boston and San Francisco can misfire for a rural population, or a Sub-Saharan one, whose bodies and illnesses it never learned. When such systems are deployed anyway—often because they are the only option, or because their limits are invisible to whoever bought them—they produce systematically worse outcomes for the people least able to contest the result.

In language, the exclusion is even more total. More than 90 percent of the text used to train large language models is Standard American English; less than 7 percent is every other language on Earth combined. Roughly 92 percent of GPT-3's training was English. For the thousands of languages spoken by the majority of humanity, the models simply perform worse—one audit found that when a leading model handled community-level Nigerian Pidgin, its quality score dropped 36 points against its English performance. A tool that is a marvel in California is a stutter in Lagos, and the people it fails are precisely those with the fewest alternatives.

The same logic reaches into finance. Credit-scoring and lending models trained on populations with thick formal financial histories tend to misjudge people who transact in cash, or through informal networks, or in economies where the recorded data is sparse—the effect is not neutral, it systematically rates the underrepresented as riskier or renders them invisible to the system entirely. Across all three domains the pattern is the same: the model is confident, the model is wrong, and the cost lands on the community that was never in the data.

The Honest Uncertainty

Before drawing conclusions, a caution about the numbers themselves, because this is a domain where the measurement is genuinely shaky and false precision would be its own kind of dishonesty.

The global adoption figures—16.3 percent here, a 10.6-point gap there—rest on methodologies that differ sharply by country and are only a few years old. "AI use" means different things in different surveys: typing a prompt into a chatbot, using a phone feature that quietly runs a model, deploying an enterprise system. Data-collection capacity is itself unequally distributed—the same infrastructure deficits that suppress AI adoption in poor countries also suppress our ability to measure AI adoption there, which means the worst-off regions are the ones we know least about, and the true gaps may be wider or narrower than the headline numbers suggest. We should treat the direction of these findings as robust and the decimal points as provisional.

The harm from underrepresentation carries a subtler uncertainty. That clinical and language datasets are skewed is established fact. That the skew degrades real-world outcomes at scale is strongly suggested by controlled studies—the skin-cancer and Pidgin results are real—but we do not yet have systematic, population-level measurement of how much worse off underserved communities actually end up because of it, in part because studying a harm requires data infrastructure in exactly the places that lack it. The mechanism is clear; the magnitude, at scale, is not yet pinned down. Building the capacity to measure it is itself one of the things the affected regions cannot currently afford to do.

And the fifty-year clock, as already noted, is an inference from past technologies whose applicability to a purely digital one is exactly what is in dispute. It may prove too pessimistic for AI's use layer and roughly right for its production layer. We are, in an important sense, running the experiment in real time.

What Concentration Costs

Grant the uncertainty, and the costs are still large and pointed in one direction. The most direct is unrealized productivity: a region that cannot adopt AI forgoes the efficiency gains that would have raised its output, wages, and living standards, and that gap compounds year over year as the adopters pull ahead. There are innovation costs too—AI applications never invented because the problems they would solve are invisible to researchers clustered in a few rich cities, problems in tropical agriculture or minority-language education that no one in the room has any reason to see. And there are human-capital costs, the people in lagging regions who could have contributed to building AI and never got the education, mentorship, or opening to try.

Brookings researchers describe the danger as a "next great divergence": pioneering regions compounding their advantages while lagging ones fall into a self-deepening trap, where slow adoption yields weak growth, weak growth yields less infrastructure investment, and thin infrastructure makes adoption harder still. This is not a hypothesis about AI; it is a description of what the digital economy already did over the past twenty years, and there is little in AI's trajectory to suggest a gentler outcome without deliberate intervention.

The political costs may in the end be the sharpest. When whole regions watch prosperity flow to distant cities while their own communities stall, the reliable result is resentment and instability. The deindustrialization of manufacturing regions across the United States and Europe in the 2000s and 2010s is the recent, unignorable rehearsal: the economic displacement was real, the political backlash was real, and the policy response was inadequate to contain either. AI is setting up the conditions for a similar dynamic—potentially faster, and reaching a wider swath of workers than the factory closures ever did. A country can absorb a great deal of inequality of outcome. It absorbs far less easily the sense that the future is being decided somewhere else, for someone else.

Whether the Window Is Still Open

If the fifty-year rule holds for AI's high-value activities, then the window to shape where they land is open now and will not stay open. Agglomeration hardens with each funding round and each cohort of engineers who move to the same few cities. The clusters forming in the 2020s are not yet as entrenched as Detroit's or Silicon Valley's took decades to become—which is precisely why the present moment is the point of maximum leverage and diminishing returns. Intervention in the next several years, before the ecosystems fully set, can plausibly still bend the map. Intervention in the 2040s, against clusters that have compounded for twenty years, would be fighting the same uphill battle that every latecomer region has lost. The uncomfortable truth is that the most effective time to act is now, when the concentration is most visible and its beneficiaries most powerful and least inclined to share.

What would acting actually look like? The policy toolkit is not a mystery; the constraint is political will and money, not knowledge. Domestically, the levers are broadband investment reaching genuinely rural areas, public funding for AI research centers in second- and third-tier cities, incentives for firms to locate AI operations outside the established hubs, and AI education and workforce training in the places being left behind. The United States has done deliberate geographic redistribution before and knows it can work: rural electrification in the 1930s, the interstate highways, the land-grant universities all spread opportunity across a far wider map than the market alone would have. The catch is scale and duration—these worked because they were sustained for decades, and a digital-era equivalent would have to be too.

Internationally the agenda is more ambitious and the record thinner. Bridging the divide means real funding for digital infrastructure and, above all, electricity in low-income countries; technology-transfer mechanisms that let developing nations build their own AI capacity rather than merely rent AI services from foreign firms; and governance inclusion that is resourced, not just announced. The World Economic Forum estimates that closing Africa's "compute paradox" could unlock over a trillion dollars in economic value by 2030—a figure that only underscores how far current commitments fall short of the opportunity, let alone the need.

This raises the hardest question, which is one of obligation rather than economics. Do the countries and companies growing rich on AI owe anything to the regions being structurally left behind? A case can be made that they do—that much of the data these systems learned from was scraped from the whole of humanity, that the talent powering the hubs was often educated at other nations' expense before it emigrated, that a technology sold as a lever for global development ought to reach beyond the places already at the frontier. But an obligation without a mechanism is a sentiment. The mechanisms that could give it teeth—binding technology-transfer commitments, funded governance participation, infrastructure financing at a scale that matches the gap—all require the actors who currently benefit most from concentration to accept limits on advantages they enjoy. History suggests that rarely happens from goodwill alone. It happens under sustained pressure from those left out, or when the winners come to see that runaway divergence is a systemic risk that eventually reaches them too. Which of those forces arrives first, and whether either arrives while the window is still open, is the open question on which a great deal of the coming decades depends.

Summary

AI's geography is one of extreme and compounding concentration. In 2025, AI companies drew $211 billion in venture capital—half of all venture money on Earth—and the San Francisco Bay Area alone captured 60 percent of it, most of that flowing to fewer than a hundred companies. Talent, research, and decision-making power cluster in the same handful of metros, and the pattern is tightening, not loosening.

The concentration persists because agglomeration is self-reinforcing—thick labor markets, dense capital, specialized suppliers, and leaked knowledge each attract more of the same—and because historical evidence suggests the high-value activities of a new technology take roughly fifty years to disperse. AI may spread faster at the use layer, where no physical infrastructure is required, but its production—the labs, the capital, the founding—runs on the same agglomeration economics that have kept prior clusters dominant for lifetimes.

The gaps are wide and widening. Urban OECD workers face far higher AI exposure than rural ones, atop a broadband gap that is growing. Globally, Global North adoption (24.7 percent) outruns the Global South (14.1 percent) and pulls further ahead each half-year, while Africa—holding under 1 percent of the world's data-center capacity, and unable in places to power the facilities it is offered—faces an infrastructure deficit measured in decades and trillions. Talent migration, underrepresentation in training data, and exclusion from governance forums each deepen disparity and feed one another in a loop. Underrepresentation is not abstract: it means worse medical AI for non-White and rural patients, near-useless language tools for most of humanity, and financial systems that misjudge the undocumented.

The data on all of this is directionally solid but numerically shaky, and the scale of the harm from data underrepresentation is not yet systematically measured—partly because measuring it requires infrastructure the affected regions lack. What is not in doubt is the cost: forgone productivity and innovation, and a political backlash of the kind deindustrialization already produced. The tools to intervene are known and the window is open now, while the clusters are still soft. It will not stay open, and whether the countries and companies with the power to act do so—out of obligation, pressure, or self-interest—is the question that will decide whether AI narrows the world's divides or carves them permanently into the map.

Key Takeaways

  1. Concentration is extreme and intensifying. AI drew $211B in venture capital in 2025—half of all VC globally—and the Bay Area captured 60 percent of it, with $113B going to just 92 companies. The top three US states hold a third of all AI hiring.

  2. The advantage compounds and resists copying. Agglomeration—dense talent, capital, suppliers, and knowledge spillovers—is easy to start and hard to replicate, which is why historical technologies take ~50 years to disperse and their high-value work barely moves at all.

  3. AI may split use from creation. The consumption layer could spread fast because it needs no local infrastructure; the production layer—labs, capital, wealth—likely stays concentrated for decades. The gap between using AI and owning it may be the defining inequality.

  4. Both gaps are widening. Global North AI adoption (24.7%) outpaces the Global South (14.1%) and the gap grew within 2025; urban-rural exposure and broadband gaps in rich countries are widening too.

  5. The Global South bottleneck is physical. Africa holds under 1% of world data-center capacity; Kenya suspended a $1B facility that would have needed a third of the national grid. Electricity, not software, is the binding constraint.

  6. Three forces interlock. Talent migration, underrepresentation in training data, and governance exclusion each widen disparity and reinforce one another in a self-perpetuating loop.

  7. Underrepresentation has concrete victims. Over 70% of clinical AI datasets over-represent White, higher-income patients; 90%+ of LLM training text is Standard American English. The result is worse healthcare, language tools, and financial access for those already underserved.

  8. The data is directionally solid but numerically uncertain, and the scale of data-underrepresentation harm is not yet systematically measured—because doing so requires infrastructure the affected places lack.

  9. The window is open now. Clusters are still soft enough to shape; the tools (rural broadband, regional research funding, technology transfer, funded governance inclusion) are known. What is missing is the political will and the enforceable obligation to deploy them at scale.

Sources

Last updated: 2026-07-14

V2 (in progress) Previous: V1