10 Questions for AI Leaders: Mehran Gul

Mindstream interviews a lot of people building AI. Mehran Gul is one of the few people mapping where it's all heading.

His newly published book, The New Geography of Innovation, just landed on the FT's Best Books list - and his views on the US-China race, the myth of a global AI ecosystem, and who gets left behind are critical knowledge for anybody following AI.

Mehran Gul: The current consensus is that the United States leads at the frontier while China is better at scaling and diffusion, or that the United States dominates closed models while China excels at open models.

That view underestimates how much progress China has made at the scientific layer, including in fundamental research. According to the Stanford AI Index, 41 of the 100 most-cited AI papers are from China, including the most-cited paper. Models such as Kimi can look surprising if you focus only on products; they are much less surprising if you have been tracking the underlying progress in Chinese science.

The other misconception is that there is a genuinely global AI race. In practice, it is overwhelmingly a contest between two countries. Other economies have carved out important niches, but very few are broad competitors across research, models, capital, talent and hardware.

The great majority of investment, frontier capability and research talent is concentrated in the United States and China. That does not look very global to me.

Silicon Valley's dominance remains remarkably strong. In my book, I argue that the Valley is not becoming less relevant; if anything, its importance is increasing. Many of the defining companies of the AI era - Nvidia, OpenAI, Anthropic and Databricks - are still based there.

More than $200 billion in private venture capital flowed into AI companies in the Valley last year, orders of magnitude more than in almost any other ecosystem. The ingredients that made Silicon Valley successful are clearly still working.

But that does not mean every successful ecosystem needs to reproduce the Silicon Valley formula. Some of the Valley's defining advantages - especially immigrant talent and deep pools of private capital - are far less prominent in China.

Beijing and Shanghai are not highly international cities in the way the Bay Area is, and China's technology industry is overwhelmingly domestic. Elon Musk, Satya Nadella, Sundar Pichai, they all came to the US as students and then stayed.

The Chinese tech industry is not led by people like that. Different ecosystems can succeed through different combinations of talent, industrial capacity, state support, market scale and institutional structure.

That concentration is already happening. AI is often presented, like earlier technological waves, as a democratising force, but its economic structure is highly centralising.

Many of the biggest winners are incumbents from the internet era or even before it: Microsoft, Alphabet, Meta, Nvidia. Unlike earlier start-ups that could emerge from a garage, frontier AI requires enormous expenditure on chips, data centres and technical talent. OpenAI and Anthropic are losing extraordinary sums of money as they build at industrial scale, which makes it exceptionally difficult for ordinary start-ups to compete.

Both trends can be true at once: innovation is becoming more geographically dispersed, while economic power becomes more concentrated.

Companies such as Revolut in the United Kingdom, SK Hynix in South Korea, Spotify in Sweden and a growing group of Chinese technology firms are becoming larger and more important. But the biggest US companies are growing even faster. The result is not a narrowing gap, but a widening one between a handful of technology giants and almost everyone else.

Its position is not as unassailable as it can appear. There is a striking gap between inputs and outputs. The United States outspends China by an enormous margin: the Stanford AI Index reported that US AI companies attracted 23 times as much private investment as their Chinese counterparts in 2025.

Yet the performance gap between leading US and Chinese models has narrowed to roughly 2.7 per cent. In other words, a huge difference in capital is producing only a marginal difference in model performance.

Competition is also moving beyond benchmark performance towards cost, reliability, openness and trust. Even if closed US frontier models remain slightly ahead, Chinese models may still win on adoption.

Recent OpenRouter data showed China overtaking the United States in token consumption. It is entirely possible for the United States to retain a narrow technical lead while a large share of consumers and businesses use Chinese models.

The US advantage in hardware is probably more durable, but it is often misunderstood. It does not rest on Nvidia alone; it is an allied-system advantage. The stack includes advanced lithography equipment from the Netherlands, memory chips from South Korea and fabrication capacity in Taiwan.

Export controls prevent many of those suppliers from providing their best technology to China. China needs to recreate domestically an international supply chain from which the United States benefits. That is a much harder challenge.

It is difficult to argue that no part of the industry is in a bubble. The strongest case is not that AI is useless, but that spending and valuations have run far ahead of proven economic returns.

The largest technology companies are committing hundreds of billions of dollars to chips and data centres even as free cash flow comes under pressure, while many businesses remain stuck in pilot projects and only a minority report meaningful profit gains from AI.

Demand is also unusually concentrated among a handful of hyperscalers and frontier laboratories. Leveraged infrastructure companies may depend on one or two major customers, while chipmakers increasingly invest in or finance the companies purchasing their products, creating a circular flow of capital.

Rapid improvements in hardware could also make today's expensive equipment economically obsolete long before it is fully depreciated, while venture funding and stock-market gains have become heavily concentrated in a small number of AI companies.

AI may be a genuine technological revolution wrapped in a financial bubble. The technology can be transformative while investors still dramatically overestimate how quickly, broadly and profitably it will be commercialised.

My instinct is that you should build before you regulate. Regulation introduced before an industry has properly developed can be premature. But there is also a misconception that regulation is either the principal reason Europe cannot innovate or a substitute for innovation. I do not think either claim is quite right.

Europe's central problem is that it does not have enough leading AI companies. Whether a particular regulation is enacted or withdrawn does not by itself solve that problem. The EU has a more robust regulatory framework than the United States across many areas, not only AI. The two issues - industrial weakness and regulatory ambition - operate partly on separate planes.

Indeed, Europe's toughest regulatory actions are often directed at large US technology companies, which should theoretically create room for local competitors. The problem is that, in many cases, those local competitors do not exist at sufficient scale. Regulation may shape the market, but it cannot substitute for building companies.

I am not an AI doomer. Much of the current anxiety - both about mass job destruction and about a sudden Skynet scenario in which humanity becomes subordinate to superintelligent machines - is overblown.

I am more persuaded by the argument advanced in the Princeton paper 'AI as Normal Technology': AI is likely to be profoundly consequential, but it will diffuse through institutions, companies and society rather than arriving as a single, instantaneous rupture.

The internet, smartphones and computers changed almost everything, yet they now feel ordinary. I expect AI to follow a similar path. We will undoubtedly make mistakes in how it is deployed, and there will be serious questions about power, labour and governance. But the kind of general superintelligence imagined by the most alarmist forecasts is not as close as many people think. I suspect we will still be debating the same things a decade from now.

It is already widening the gap. Even under the optimistic scenario in which AI successfully automates repetitive work, many of the first jobs affected will be tasks previously outsourced to countries such as India or Malaysia - call centres, support functions and routine coding work. Those activities have served as entry points into the global economy. If they are automated, several rungs may disappear from the development ladder.

The second problem is concentration. We speak of a 'global' AI race, but the core capabilities remain overwhelmingly concentrated in the United States and China. Latin America and Africa are largely absent, and even Europe is struggling to reach the starting line.

The ten largest US technology companies are now worth more than the GDP of every country except the United States itself.

More troublingly, there is no obvious path for most countries to join the frontier. Are we likely to see globally competitive foundation models emerge from Eastern Europe, or Nvidia-class hardware from India, in the near future? Probably not.

The most important question is not whether the United States or China wins, but what happens when two countries run away with the technological foundations of the next economy.

The biggest surprise has not been a particular product, whether GLM-5.2, Kimi or another model. It has been the rise of Chinese science. ResNet, developed by researchers at a Beijing laboratory, became the most-cited paper in artificial intelligence and is the most-cited paper in any scientific field published in the twenty-first century.

The products that attract headlines are downstream of that scientific output. ResNet in particular - and China's broader progress in frontier computing research - changed my view of what the country could contribute at the deepest layer of technological development.

The country most likely to surprise us in the next decade is the same one that surprised us in the last: China. I think it is only getting started. Many people still have a mental block when imagining China moving ahead in AI. They may fear the possibility, but they often treat it as a distant dystopian scenario rather than a plausible competitive outcome.

Yet China is already at or near the frontier in electric vehicles, solar power, batteries and drones. Its strength is not limited to manufacturing at scale; it also produces important scientific discoveries and some of the world's best products in these fields. It has achieved this despite export controls, blacklisting and tighter restrictions on scientific collaboration, and those pressures are encouraging it to build more capabilities at home.

The Huawei story is instructive. The company initially buckled under sanctions, then returned with a far more self-reliant technological stack. Many of the external levers once capable of constraining it are now less effective. The question is whether something similar could happen across the Chinese technology system as a whole. That is what I would watch most closely.

Mehran Gul is the author of The New Geography of Innovation (William Collins / Simon & Schuster), recipient of the Financial Times/McKinsey Bracken Bower Prize, and an FT Best Book of the Year selection. He was previously at the World Economic Forum and attended Yale, where he was a Fulbright Scholar, Fox International Fellow, and Teaching Fellow. 

You can get Mehran’s book, The New Geography of Innovation, here!