The reemerging concentration problem

When we first measured ecosystem concentration in the earliest editions of the Global Startup Ecosystem Report, a select few startup ecosystems captured the overwhelming share of global exit value. The shape of innovation, by every meaningful metric, was a power law, with Silicon Valley sitting on top, way ahead of even the immediate runners-up.

Then something changed. Over the following decade, that concentration began to ease. My hometown Berlin, Tel Aviv, Bengaluru, São Paulo, Shenzhen, Stockholm, Toronto, Riyadh, Lagos: new hubs climbed the rankings. Capital pools deepened locally. Founder networks recycled successes into new firms. 

Ecosystem Value shows how far that diffusion went. Between GSER 2020 and GSER 2024, the share of global Ecosystem Value held by the Top 20 ecosystems fell from 75.4% to 69.7%. The Top 10's share fell from 63.8% to 58%, and Silicon Valley's own share slipped from 27.3% to 24.5%.

That broadening had a name in the academic literature: regional diffusion. Through 2023, it looked like a one-way trend – the inevitable result of factors like improved global connectivity, smarter institutions, and better governance. But the trend is now reversing: the Top 20's share has risen in each of the last two editions, returning to 75.4% in GSER 2026 and erasing four years of diffusion. The Top 10 now hold 64.4% of global Ecosystem Value, slightly more than in GSER 2020, and Silicon Valley's share has climbed back to 28.3%. 

This reversal is being driven by AI. As we shared before, AI-Native Ecosystem Value crossed $1 trillion globally in 2025, growing 507% since 2020. This is twelve times faster growth than the rest of tech, and twenty times faster than the EU economy. North America captured 78% of global AI-Native venture capital between H2 2023 and 2025. The U.S. alone produced 43.2% of all AI-Native startups created in that window; the U.K., the closest European competitor, produced 9%.

Our dear friend Josh Lerner of Harvard Business School has spent decades studying the geography of innovation. The pattern he describes for new and disruptive technologies is rarely cited in startup policy circles, but it should be. In his framing, the geography of Frontier Tech obeys five laws:

By Lerner’s framework, the AI head start is structurally durable. Diffusion will happen, but it will mostly be the diffusion of use, not creation. The question for ecosystems outside the originating handful is whether the AI activity that arrives will be the kind that compounds – such as research, models, and frontier engineering – or the kind that consumes.

One of the major questions facing us today is whether AI will have similar effects to previous general purpose technologies. Previous technological waves like mechanization, electrification, and computing clearly eliminated many jobs (such as farm laborers, weavers, and switchboard operators), but the broad consensus among historians is that they also created new types of jobs that more than compensated for these losses. The short-term, localized impact was often devastating, but the longer-term, wider economic gains were substantial.

A slightly different question is whether technology changes the distribution of who benefits. Here there is much less consensus, particularly over the split of gains between labor and capital – that is, whether benefits accrue more to workers, through their ability to be more productive, or instead to those who own the machines. 

How will AI fit these historical patterns? Recent work by Acemoglu suggests that even when AI raises productivity, the gains may be captured more by capital than by labor. If we combine that with Lerner’s finding that the high-skill creation of new technologies can remain geographically concentrated for decades, as well as the rapid diffusion and global nature of many digital services, it suggests that the benefits of AI could be concentrated both in capital's hands and in a few places, while its costs (like technological unemployment) are felt much more widely. 

In previous technological revolutions, local disruption was often compensated at the national level by other changes; with AI, such national-level compensation mechanisms may not appear. The risk for ecosystems outside the pioneer hubs is therefore not that they fail to adopt AI. It is that they become highly productive consumers of a technology whose most valuable economic assets are owned somewhere else.

There is another, less visible force reinforcing this divergence: budgets are moving far more slowly than the technology. In a recent Startup Genome roundtable with ecosystem leaders, 54% of respondents who provided a clear figure said they were allocating less than 5% of their 2026 budget to new AI-focused policies or initiatives. At the same time, an average of 62% of ecosystem budgets was simply carried over from the previous year. The technology stack may be changing at extraordinary speed, but the institutions charged with responding to it are still largely funding yesterday’s priorities. 

Governments outside the pioneer hubs cannot afford to layer a few AI programs on top of a portfolio designed for the previous era. They need to start from a blank sheet: identify the capabilities that will determine where AI creation happens — entrepreneurial talent, specialized communities, frontier programs, early-stage capital and commercialization — and then ask which existing expenditures still deserve to survive. In a power-law technology transition, incremental budgeting produces an incremental response to an exponential shift.

None of this means the geography of AI is predetermined. But it does mean that the window for changing it may be shorter than policymakers assume. If Lerner is right, the locations that establish an early advantage in creating a frontier technology can retain it for decades.

For ecosystems outside the pioneer hubs, the objective cannot therefore be AI adoption alone. It must be to capture a meaningful share of AI creation: the founders, researchers, specialized investors, communities, and companies that produce the technology and retain its economic upside locally. That will require choices, including moving money away from policies designed for the last technological era.

The last decade showed that the geography of innovation can broaden. The next decade will determine whether AI continues that diffusion or reverses it.