Add to Google Preferred Sources
AI is everywhere — in headlines, earnings calls, and government budgets — yet the economist who literally wrote the book on how progress ends is now warning that the productivity payoff will disappoint. Carl Benedikt Frey, Oxford professor and author of *How Progress Ends*, argues on the Trumponomics podcast that AI is still waiting for its “separate condenser moment,” the enabling innovation that transforms a clever prototype into an economy-wide force. Without it, AI automates production but leaves humans holding the bill for verification, a problem the computer revolution never had. The deeper argument is institutional: progress depends on a cycle of decentralized exploration and centralized scaling, and both the US and China are now building the incumbent structures that historically kill growth. Frey sees killer acquisitions, patent-office capture, and political capitalism hardening the American system, while China’s 15-year pivot toward security priorities is loading weight onto state-owned enterprises — its least innovative sector. The AI race between the two countries, he says, matters far less than what everyone else does: export controls are pushing the rest of the world toward Chinese and European open-weight alternatives, fragmenting the AI stack just as it should be diffusing. For investors betting that AI will deliver a decade-long productivity miracle, Frey’s message is blunt: “If you believe that we are entering a new renaissance for economic growth, you’re likely to be mistaken.”
Key Elements
The most startling thing Carl Benedikt Frey says on the Trumponomics podcast is not that AI will disappoint — plenty of skeptics have said that — but that the computer revolution was actually the bigger deal. Speaking to Bloomberg’s Stephanie Flanders, the Oxford economist and author of How Progress Ends: Technology, Innovation, and the Fate of Nations draws a line through 200,000 years of human history and arrives at a conclusion that should sober anyone building a portfolio on AI-driven growth: “If you believe that we are entering a new renaissance for economic growth, you’re likely to be mistaken, because AI is actually likely to give less of a productivity boost than the computer revolution.”
That is not what the market is pricing. Tech giants are pouring hundreds of billions into data centers; the “Magnificent Seven” have added trillions in market cap; AI-related job postings in the UK just hit a record 9.4% of all listings. But Frey’s argument is not that AI does nothing — it is that the gap between invention and transformation is institutional, and institutions are currently moving in the wrong direction.
Why progress is unnatural
Frey’s core thesis is that progress is not the default state of any economy. It took 200,000 years to produce an industrial revolution. Most countries are still not rich. Britain, where he lives, has endured two decades of productivity stagnation.
“If progress was inevitable, it would not have taken 200,000 years to have an industrial revolution,” he noted.
The mechanism that produces progress is an institutional cycle that must re-decentralize each time a technological wave matures. In the early, uncertain phase of any new technology, you need many competing explorations — different approaches, different funders, different bets. Once a prototype emerges, centralized scaling takes over. As diminishing returns set in, the system must de-centralize again, or growth stops.
The Soviet Union proves both halves. It grew for more than four decades by scaling technology invented elsewhere — exploiting Ford’s open-door policy to transfer assembly knowledge and build a domestic vehicles industry. In the age of mass production, the command system worked “fairly well,” because standardized production lets factory managers be benchmarked against one another. The system broke when global returns to mass production petered out in the 1970s. Two things failed at once: you cannot benchmark novelty, and the Soviet funding funnel was a single pipeline. An engineer could approach the Red Army, perhaps two or three other options, and if all declined, the idea died.
| Soviet exploration system | American exploration system | |
|---|---|---|
| Funding routes | Monolithic — engineer goes to the Red Army, ~2–3 options, then the idea dies | Decentralized finance — many independent backers, each a shot on goal |
| Benchmarking | Works when production is standardized | Impossible for novel technology; requires bets, not benchmarks |
| Fate | ~Zero contribution to the computer revolution | Funded the computing era, including Google |
The Google example is revealing. Bessemer Venture Partners famously declined to invest in Google in 1999 and probably regrets it, but in a decentralized system the refusal did not kill the company. Others stepped in — necessary because search looked settled, with AltaVista and Yahoo dominating, and Google’s viability had to be demonstrated by someone with capital and conviction.
More invention, less transformation
The cycle helps explain today‘s central disconnect. Every indicator of inventive output — patents, scientific publications — is up, while the economy points in the opposite direction. Research productivity and breakthrough innovation are down.
“We’re getting more in terms of scientific and inventive output, but it seems that we’re getting less transformational output,” Frey told Flanders.
Read against history, the computer revolution was more transformative than AI appears today. Computers and the internet connected the world’s best scientists, streamlined the research process, and put the world’s knowledge into people’s pockets. The payoff: a decade-long productivity upsurge, mostly confined to the United States.
AI is different in one critical dimension. Computers automated downtime — the weeks or months researchers spent waiting for materials to arrive. AI automates production — drafting, analysis, code — but the output must still be verified by humans. The net gain is production time saved minus verification time. The computer revolution had no such subtraction.
There is also a behavioral problem. Since the computer era, inventors and scientists have taken on more projects — the data shows it. A powerful new tool offers two choices: drill deeper on a few problems, or drill more holes across many. The aggregate outcome of “drill more holes” is attention spread thinly, making breakthroughs less likely at any given time. AI is reproducing the pattern: people use it to produce more output rather than spend ten years producing nothing and then emerge with a real breakthrough. Academic incentives — publish or perish — actively reward that choice.
“Adoption rates will tell you little,” Frey said. “What matters is adoption for what.” If AI is adopted for email, it will not drive growth in any meaningful way. If it is adopted for scientific research that develops new products and technologies, that is a different matter entirely.
China’s real innovation machine — and what’s throttling it
The same institutional lens upends the standard “China just copies” story. China is a country of 1.4 billion people, Frey noted, so it would be extraordinary if there were no innovation at all. But crucially, the system is not Soviet central planning.
“China is not as centralized as is commonly believed,” he said. “What you have in China is a system where provincial governors are competing against each other based on growth targets for promotion inside the one party system.”
This creates a tournament of political competition that is genuinely hard to replicate. European industrial policy tends to be anti-competitive — German rearmament means plowing funds into Rheinmetall. Chinese industrial policy can be pro-competitive, because provinces seed new firms that compete against one another. The firms actually leading Chinese innovation are mostly startups, privately funded, and often foreign-funded — structurally similar to the West.
The real difference: China has no rule of law, so political connections determine who gets a seat at the table when priorities change. And priorities have changed. Over the past 15 years, the Chinese Communist Party has shifted from economic targets toward political objectives — self-sufficiency, national security, common prosperity. That pushes reliance onto state-owned enterprises, because private firms are less keen to pursue non-economic national goals. By any measure, Chinese SOEs are less innovative and less productive.
“And that I think is unlikely to change,” Flanders added, reinforcing Frey’s assessment.
The inversion between the two systems is stark. In the 2000s, the expectation was that China would become more like the US. The opposite is happening: the US now resembles the political capitalism once associated with China. Being on the right side of the current administration matters, which is one reason Frey pointed to OpenAI’s openness to letting the government take a five percent stake in the company. If the state is on your side, you are on the right side.
How the US builds barriers to its own progress
Seen through the institutional cycle, the American system’s historical advantage is its capacity to fund new entrants over incumbents — and the US is now busy taxing that advantage. Frey’s pattern is unambiguous across every technology wave: leaders of one wave do not lead the next.
| Incumbent leader | Successor wave | Outcome |
|---|---|---|
| Bicycle manufacturers | Automotive | Did not lead |
| Legacy media companies | Social media | Did not lead |
| Legacy car companies | Electric vehicles | Did not lead |
| Legacy retailers | E-commerce | Did not lead |
“It’s new firms that tend to develop new technology and new kinds of industries,” Frey said. “And old industries have every incentive to prevent that sort of competition.”
The US has developed specific mechanisms for letting them do so. “Killer acquisitions” let incumbents buy promising startups purely to shut them down. A revolving door with the US Patent and Trademark Office lets patent examiners grant incumbents low-quality patents, then take jobs at those same firms. The aggregate result: business dynamism has declined even though every major technology of the past four decades — the personal computer, the internet, the cloud, and AI — has made setting up and operating a company cheaper. Less entry, more barriers, and an incumbent class with every incentive to keep it that way.
Why scale alone will not make AI the exception
Flanders pushed the strongest counter-case: maybe AI is genuinely different. The recent trajectory suggests returns come from scale — throwing resources at large language models — and the gains accrue to whoever has the most energy and chips. If fixed costs are prohibitive, new entrants cannot matter, and the need for decentralized exploration has been superseded by a brute-force race.
Frey’s rebuttal has three parts.
First, brute force only works against a static distribution of events. The world changes constantly, and resilience matters. An AI managing supply chains can perform well until something unprecedented happens — and no amount of training data covers the unprecedented.
The Go case is the sharpest evidence. In 2016, AlphaGo beat Lee Sedol 4-1, achieving superhuman performance in a game long considered the benchmark of machine intelligence. Yet around 2024, human amateurs using standard computers beat the best available Go programs quite easily, simply by exposing them to new positions and concepts absent from their training. Superhuman performance against today’s distribution says nothing about tomorrow’s.
Second, AI is still awaiting its enabling breakthrough. Early steam engines were so energy-inefficient they could only drain coal mines; the separate condenser made them efficient enough for railroads and steamships. Frey believes something comparable must arrive before AI’s economic payoff generalizes.
“I think AI is still waiting for what I call the separate condenser moment,” he said.
Third, humans are dramatically more data-efficient than current AI, learning from just a few examples. Getting machines to that level requires innovation, not more scale. The path is genuinely uncertain: it may be large language models, small language models, world models, or something entirely different. What is knowable is that greater data efficiency “is clearly not something we’re going to get through just through scaling.”
The fragmentation game
For countries behind the frontier — which is almost everyone — the historically proven route to growth is adoption of technology invented elsewhere. The postwar US made this explicit policy: Marshall Aid shared American technology with allies, contributing to the postwar miracles in Europe, Japan, and Korea. The reason the entire world is not rich is not that the technology is unavailable, but that institutional and cultural constraints block adoption.
AI is changing this equation because national security has inserted itself into diffusion. In 2026, the Trump administration imposed restrictions on foreign use of Anthropic’s latest model, and Frey expects similar measures at greater scale. Dependence on the technology leader is now a strategic vulnerability.
Frey doubts the “who’s three months ahead” question — whether OpenAI, Anthropic, Google, or a Chinese lab leads — matters much at all, unless one firm gets onto a curve where it genuinely pulls away. And even then, the pull-away might come from outside the large-language-model paradigm entirely.
The more consequential question is what everyone else builds on. Export controls are pushing non-frontier countries toward Chinese or European open-weight alternatives, a dynamic Frey expects to accelerate precisely because the US has made reliance on its stack look like a geopolitical risk. The sharpest irony is that this is exactly how China closed the gap — Beijing embraced an open-weight ecosystem in large part because US chip export controls forced it to.
| Strategic option | Dynamics |
|---|---|
| Remain on US frontier technology | Historically easiest; now restricted by national-security controls |
| Pivot toward Chinese or European open-weight technology | Triggered by perceived US unreliability as a trading partner |
| Build domestic open-weight capacity | Heavy lift, but full independence |
The AI stack is fragmenting just when it should be diffusing, and the countries that lose access to the frontier are the ones most likely to build alternatives that eventually compete with it.
What investors should watch is not the weekly model-release calendar but the institutional signals: whether production AI systems exhibit the same resilience failures the Go amateurs exploited, whether US enforcers move against killer acquisitions, whether China’s SOE pivot costs it the frontier, and which non-US ecosystem becomes the default home for the majority of the world’s AI developers. Frey’s central point is that technology does not determine its own fate — institutions do. And on current evidence, the institutions of 2026 are building walls, not bridges.
Full content available at:Why AI Isn’t Actually Boosting Productivity | Trumponomics
- BIS Warns: AI Boom Fuels Both Supply and Demand, Clouding Central Bank Rate Decisions
- Over 1,100 AI Insiders From OpenAI, Anthropic Demand Global Brakes on ‘Runaway’ AI Race
- UK Job Postings Fall 11% in First Half, Graduate Roles at 20-Year Low While AI-Related Demand Hits Record High
- China Vows Timely Incremental Policies as Growth Shift Tests Stability
Once added, BigGo Finance appears first in Google Search Top Stories, so you get the broadest, most up-to-the-minute, and most comprehensive global financial news first.
