24th Place

In January 2026, Microsoft released data that should have reframed every conversation about AI competition. The company had been tracking what it calls "AI diffusion"—the percentage of working-age populations actually using generative AI tools—across nations worldwide. The results were striking, but not in the way the American tech industry expected.

The United States, which leads the world in AI infrastructure, frontier model development, and venture capital investment, ranked 24th in AI usage among its working-age population. Twenty-fourth. Behind Ireland, behind Spain, behind Belgium. The usage rate: 28.3 percent.

The United Arab Emirates ranked first, at 64 percent adoption. Singapore second, at 60.9 percent. These are not countries known for frontier AI research. They have no models competing with GPT-5 or Claude. They didn't pioneer transformer architectures or invent reinforcement learning from human feedback. They did something different: they diffused capability throughout their populations while America was busy inventing it.

I've been thinking about this data for weeks now, because it crystallizes something I've suspected for a long time about how technological advantage actually works. The conventional wisdom—that the nations and organizations leading AI research will dominate the AI era—is almost certainly wrong. And the historical evidence for why it's wrong is overwhelming, if you know where to look.

In 2022, Jeffrey Ding, a political scientist at George Washington University, published research that should have upended how we think about technology competition. His book, Technology and the Rise of Great Powers, won the 2025 Lepgold Book Prize, but its central finding remains underappreciated: across three industrial revolutions, what determined economic advantage wasn't who invented breakthrough technologies. It was who diffused them.

"Where innovations are adopted more effectively is more important than where they are first introduced," Ding argues. The insight seems counterintuitive until you examine the evidence.

Consider the Second Industrial Revolution. Germany dominated the frontier technologies of the era—chemicals and electrical equipment. By 1900, Germany produced 90 percent of global synthetic dyes and controlled roughly 50 percent of world electrical equipment exports. These were the AI equivalents of their time: transformative, general-purpose technologies that would reshape every industry they touched.

Yet the United States became the preeminent economic power. Not Germany. How?

The answer lies in diffusion infrastructure. In 1862, the Morrill Act financed land-grant colleges dedicated to mechanical arts. Engineering schools multiplied from 6 in 1862 to 126 by 1917. By 1900, 88 percent of mechanical engineering students were in land-grant colleges emphasizing practical, applied training—not elite research. The result: U.S. machine intensity was more than twice that of Britain and Germany by 1907.

American advantage came not from breakthrough inventions but from what British observers called "adaptation of special apparatus to a single operation in almost all branches of industry." Americans didn't invent interchangeable manufacturing. They diffused it—across sewing machines, bicycles, typewriters, and eventually automobiles. The resistance to diffusion wasn't technical; it was cultural. British manufacturers saw American methods as crude, unsophisticated, lacking the craft wisdom that "real" engineering required.

The pattern repeats with eerie consistency. Britain's First Industrial Revolution succeeded not because of heroic inventors like James Watt, but because of what economic historians call "tweakers" and "implementers"—ordinary engineers who spread mechanization throughout the economy. Britain's advantage wasn't elite scientific genius. It was "the average level of technical literacy" among machinists. Widening the base of capability, not deepening the peak.

The most instructive case, though, is the Soviet Union—because it demonstrates what happens when innovation capacity and diffusion capacity diverge completely.

By 1970, the Soviet Union led the world in R&D spending as a percentage of GNP. They pioneered technologies across multiple domains. A 1977 cross-national assessment ranked the Soviet Union most successful in pioneering innovations and least successful in diffusing them throughout the economy.

The continuous casting of steel provides the clearest example. The Soviet Union pioneered this technology in the mid-1950s—five years before Japan adopted it. By 1980, only 10.7 percent of Soviet steel used continuous casting. Japan, which adopted the technology later, produced 59 percent of its steel with the technique.

General Secretary Brezhnev acknowledged the problem explicitly in 1971: "The links connected with the practical realization of scientific achievements and their adoption in mass production are the weakest." He was describing, with unusual precision, a diffusion deficit—a structural gap between the capacity to invent and the capacity to spread.

The Soviet system had every incentive to innovate. Research institutes competed for prestige and funding. Scientists received recognition for breakthroughs. What the system lacked was the infrastructure to diffuse—the training programs, the industry-academia linkages, the decentralized experimentation that allows ordinary workers to adopt and adapt technologies for local conditions.

Innovation capacity without diffusion capacity equals stagnation. The Soviets won every innovation metric and lost the Cold War.

What Ding's research reveals is that general-purpose technologies—steam power, electricity, information technology, and now AI—share a specific property that makes diffusion decisive. They improve continuously over time, apply broadly across economic sectors, and generate synergies with complementary innovations. But they only deliver transformative value after decades of diffusion through education systems, organizational restructuring, and complementary innovations.

The electricity parallel is especially instructive. The first commercial electric dynamo appeared in the 1880s. Electrification's boost to manufacturing productivity materialized about five decades later—occurring only after factories had restructured their layouts entirely. For decades, factories wired for electricity mimicked the layouts designed for steam power. They put electric motors where steam engines had been, ran central shafts through the building, and saw minimal improvement.

The transformation came when factory organization reconstituted around electricity's unique properties—when power could be distributed to individual machines, when layouts could optimize for workflow rather than power transmission, when the entire production system reorganized around what electricity made possible rather than substituting electricity into a steam-era architecture.

The pattern for AI looks similar. Organizations are installing AI where analysts used to sit. They're adding AI copilots to existing workflows. They're substituting AI into industrial-era organizational structures and wondering why the productivity revolution hasn't materialized. They're electrifying the steam factory.

The enterprise data tells the same story the national data does, just at a different scale.

As of early 2026, 88 percent of organizations report using AI in at least one business function. That sounds like success—until you look at what "using AI" actually means. According to research across multiple surveys, only 8.6 percent of companies have AI agents deployed in production. Nearly two-thirds remain stuck in the pilot stage. A March 2025 survey of 120,000 enterprise respondents found 63.7 percent report no formalized AI initiative at all.

MIT's research found that 95 percent of generative AI pilots fail to deliver measurable return on investment. Not disappointing returns—zero returns. Organizations are pioneering AI adoption and failing to diffuse the capability to where value gets created. They're the Soviet Union of the AI era.

The mechanism is identical to what Ding identified at the national level. Organizations invest heavily in AI innovation—data science teams, frontier model access, research partnerships—while underinvesting in diffusion infrastructure: training programs, process redesign, organizational restructuring. They're measuring the wrong thing. They count pilots launched and models deployed when they should be counting decisions informed and capabilities spread.

One data point captures the disparity precisely. MIT and Erik Brynjolfsson's research found that organizations invest ten times as much in intangible assets associated with IT—training, reorganization, new processes—as in the IT itself to see returns. The technology is the smaller investment. The diffusion infrastructure is the real cost. Most organizations have inverted this ratio entirely.

This framework reframes the US-China AI competition in uncomfortable ways.

The dominant narrative focuses on frontier capability: Who has the best models? Who leads in chip production? Who attracts the most AI researchers? By these metrics, the United States maintains significant advantages. American companies lead in frontier model development. Export controls have constrained Chinese access to cutting-edge AI chips. Top AI talent continues to favor American institutions.

But Ding's research suggests these metrics may be measuring the wrong competition.

China has pursued a distinctly different strategy: open-source diffusion. DeepSeek's emergence in January 2025 demonstrated that Chinese AI labs could achieve frontier-competitive performance with dramatically fewer resources—and more importantly, could release those capabilities as open-source models that anyone could deploy. Microsoft's diffusion data shows the result: "DeepSeek's success reflects growing Chinese momentum across Africa, a trend that may continue to accelerate in 2026."

China isn't trying to aggregate users at centralized platforms the way American AI companies do. It's spreading capability to markets American frontier labs don't serve. This is a diffusion strategy, not an innovation strategy. And if Ding's historical research holds, diffusion strategies win.

The research assessment from Recorded Future states the implication directly: "AI diffusion rather than innovation will very likely determine the 'winner' in the competition to economically and geopolitically benefit from the technology." China may be investing in diffusion infrastructure—the equivalent of land-grant colleges for AI—while America optimizes for frontier research that generates prestige but not spread.

For business leaders, the diffusion thesis suggests a fundamental reorientation of AI strategy.

First, measure diffusion, not capability. The relevant question isn't "do we have AI?" or "how good is our AI?" It's "what percentage of decisions in our organization are informed by AI?" Organizations should track adoption breadth—the percentage of employees using AI tools, the number of business functions with AI integration, time-to-adoption for new capabilities—rather than pilot counts or model quality metrics.

Second, invert the investment ratio. If the 10:1 ratio holds—and the evidence suggests it does—then most organizations have their AI budgets backwards. They're spending on models and data science teams while underinvesting in training, process redesign, and organizational restructuring. The diffusion infrastructure costs more than the technology, and it's where the return actually materializes.

Third, recognize that diffusion is structural, not promotional. You cannot communicate your way to AI adoption. The Soviet Union didn't fail at diffusion because of insufficient propaganda about continuous casting. They failed because their institutional structure couldn't support spread—centralized control blocked local adaptation, rigid hierarchies prevented experimentation, and the "softer stuff" (as Ding calls it) of culture and incentives worked against adoption.

Organizations face the same structural barriers. Data science teams centralized in corporate functions can't diffuse capability to the edges where decisions happen. Governance frameworks designed to control AI use prevent the experimentation that drives adoption. Training programs that upskill a small cadre of "AI champions" replicate the Soviet error of deepening elite capability rather than widening the base.

Fourth, understand that you may be winning the wrong race. The organization with the most sophisticated AI models, the best-credentialed data science team, and the most impressive pilot results may be losing to competitors who deploy simpler tools more broadly. A "good enough" system used by 70 percent of employees beats a brilliant system used by 5 percent. The land-grant colleges won against elite research universities. The tweakers and implementers won against the heroic inventors.

One complication deserves acknowledgment: diffusion doesn't happen linearly. There's substantial evidence that AI adoption involves threshold effects—that value emerges discontinuously at critical levels of diffusion rather than accumulating proportionally.

Electrification showed this pattern. Factories didn't see proportional gains from adding more electric motors. They saw minimal gains for decades, then transformative gains once organizational restructuring crossed a threshold. The phase transition came from reorganization, not electrification per se.

AI may work similarly. Organizations below a threshold of AI diffusion—call it 30 or 40 percent of decisions informed by AI—may see negligible returns regardless of the quality of their models or the sophistication of their pilots. Organizations above the threshold may see qualitatively different returns as system-level intelligence emerges.

This creates a strategic trap. Incremental investment in AI diffusion yields negligible returns until the threshold is crossed—at which point organizations that didn't invest are too far behind to catch up. The McKinsey data showing that high performers set growth and innovation objectives rather than efficiency objectives may reflect this: organizations pursuing transformative AI adoption cross the threshold, while those pursuing incremental efficiency gains never do.

The practical implication is sobering: AI investment below the diffusion threshold may be wasted investment, regardless of how sophisticated the technology. This argues for concentrated, committed diffusion efforts rather than distributed pilot programs—the opposite of how most organizations approach AI adoption.

The diffusion paradox—that nations and organizations leading AI innovation often lag in AI adoption—isn't actually paradoxical once you understand the mechanism. Innovation and diffusion require different capabilities, different investments, different institutional structures. Excellence at one doesn't produce excellence at the other.

The Soviet Union excelled at innovation and failed at diffusion because their system was structured to reward breakthroughs and structured to prevent spread. Centralized control, elite institutions, and rigid hierarchies optimize for concentration, not distribution. The American system that overtook Britain in the Second Industrial Revolution worked in the opposite direction: land-grant colleges, practical engineering training, and decentralized experimentation that let capability spread throughout the economy.

The AI era is replaying this pattern in real time. The United States leads in frontier capability and lags in adoption. The UAE leads in adoption with negligible frontier capability. Enterprise organizations run impressive pilots that never reach production. China pursues open-source diffusion while America pursues proprietary model excellence.

The historical evidence is unambiguous about which strategy wins. Diffusion determines advantage. Invention without spread is economically irrelevant. The nation or organization that diffuses pattern-recognition capability throughout their system first—not the one that builds the best models—will dominate the AI era.

Microsoft's January 2026 data isn't a curiosity. It's a leading indicator. The question for every organization is whether they're tracking innovation metrics that generate prestige or diffusion metrics that predict success. And for most, I suspect, the answer is the former—which means they're optimizing for the Soviet outcome, not the American one.

The race isn't to build better AI. It's to spread AI further. And that race has barely been acknowledged, let alone won.

Microsoft Corporation

Academic Research

MIT & Productivity Research

Enterprise AI Adoption Data

Technology Competition Analysis

US Industrial Development

German Second Industrial Revolution

Soviet Technology Diffusion

Electrification Timeline

DeepSeek

McKinsey Research

About the Author

Sravan Ankaraju is a researcher and writer focused on technology adoption, organizational transformation, and the intersection of innovation and diffusion. Based in Dallas, Texas, he develops frameworks for understanding how technological capability spreads through organizations and economies.