As the generative AI gold rush continues, entrepreneurs, investors and technology thinkers predict who will win. Some believe this time it will be smaller, geographically diverse companies and startups that triumph, rather than the big tech companies and their coastal hubs.
And indeed, a new generation of AI applications powered by large language models (LLMs) are rapidly proliferating on the internet, fueled by licensing agreements with companies like OpenAI and the use of open source software. Meta, for example, has long open-sourced its key algorithms, including its latest generative AI system known as LLaMa 2, for developers to use freely. Meanwhile, British research firm Google DeepMind has also historically open-sourced many of its key algorithms, allowing programmers to freely download and build upon them.
All of this forms a compelling story. As Cade Metz of The New York Times recently summarized, “Many in the field believe that this kind of freely available software would allow anyone to compete.”
Yet despite this beautiful picture of tech inclusion, research from the Brookings Institution and others suggests that the winner-take-all dynamics of digital technologies, and the mechanics of AI itself, work against such decentralization, at least across cities. Moreover, our latest research confirms that over the past year, more than half of generative AI job openings have been concentrated in just 10 metropolitan areas—many of which are the same coastal metropolitan areas where AI optimists believe the technology should spread.
Thus, rather than democratizing the technology, generative AI may further centralize AI activity in the absence of deliberate intervention.
Technological innovation tends to be concentrated in key hubs, and AI is following the same path.
As we detailed in our July report on building AI cities, digital industries tend to concentrate in a few major hubs, particularly the Bay Area, for specific reasons, including the innovation benefits of local clusters, the need for deep pools of specialized talent, and the benefits of “winner takes most” network effects deployed across huge platforms.
To be sure, some job diffusion occurs over time as moderate-wage jobs in a field begin to spread to new regions. But even so, Stanford economist Nicholas Bloom and his team have shown that over the past two decades, most disruptive technologies have remained highly concentrated in core regions, giving these “pioneer” regions a long-term advantage.
For AI, the predictions don’t seem to be all that different, even though the technology is still in its early stages. According to a 2021 Brookings Institution study, the Bay Area and 13 “early adopter” metro areas account for more than half of the country’s AI activity in federal contracts, conference papers, patents, job ads, job profiles, and startups. More recently, a July report found that nearly half of all job ads for generative AI roles over the past 11 months were concentrated in just six coastal metro areas: San Francisco, San Jose, California, New York, Los Angeles, Boston, and Seattle.
Now, our new analysis has found an even greater concentration of AI posts.
Nationwide, more than 60% of generative AI jobs posted in the year ending July 2023 were concentrated in just 10 metropolitan areas. Nearly a quarter of those were in the Bay Area, with the rest concentrated in a few large “superstar” cities.
The nature of AI may lead to further geographic concentration
Looking at the new job advertisement figures, the story of generative AI, and of AI more broadly, looks a lot like that of social media, the earlier internet boom, and even the PC boom before that, in terms of geographic concentration. As the latest digital technology, AI appears to be developing along the same highly clustered path as earlier digital services, as it requires deep pools of existing expertise and talent.
Moreover, key features of today’s AI models may well exacerbate the tendency toward concentration. Indeed, the projected increase in accessibility and affordability of computing costs, and the relatively modest costs of tweaking and training new versions of existing “base” or “foundation” models, suggest that in the near future, more AI work will be able to be done anywhere, decentralizing the industry to more locations.
But even so, policymakers and technologists need to consider the costs, technical talent, and computing hurdles that tend to place foundational model building and training in frontier hubs. These, combined with first-mover advantages of all kinds, have helped to centralize AI development and lock it into core management and design hubs dominated by the most resource-rich companies and clusters.
So, despite its potential to drive productivity gains for businesses, exciting new use cases in new sectors, and driving economic benefits for local economies, AI will likely continue to dominate in the same regions.
Significant investments could help AI activity spread to more places
So what should be done? Going forward, countries, states, and industry will need to step in proactively to drive a more inclusive AI geography.
At the heart of these efforts should be locally oriented measures to increase the availability of key inputs to AI innovation and commercial adoption not just in a few but in many promising regions. To this end, a July report from the Brookings Institution cited a range of potential federal, state, and local initiatives, some of which are already being piloted and are worth reiterating here:
To spread AI research and development more broadly, the country needs to expand the National Science Foundation’s National Artificial Intelligence Institute program, through which 19 universities are pursuing diverse, locally-focused research programs serving nearly 40 states. Such investments are critical because universities provide a widely dispersed network of hubs for technology growth.
Additionally, the federal government should establish and build out the proposed National AI Research Resource (NAIRR), which would democratize access to a range of critical data and computational resources: For example, NAIRR could be used to train foundational models for public use, thereby reducing the prohibitive costs of such work in non-superstar hubs.
As a result, states and localities, in partnership with federal agencies, must urgently expand digital education and training efforts, with a special focus on ensuring underrepresented groups have access to AI skills pathways in areas with growing AI activity.
States also now have a unique opportunity to leverage multiple “region-based” industrial policy programs to turn emerging regional AI activity into true growth clusters. Coincidentally, AI has been designated a “major industry” by both the Department of Defense and the representative CHIPS and SCIENCE Acts, meaning that AI development in new locations is a priority for the U.S. government. States and regions should explore new ways to access and deploy resources to build emerging regional AI ecosystems.
Decentralizing AI activities is possible – if you put in the effort
In short, the growth of AI may be different from the growth of any digital technology to date: The proliferation of open source development and cheaper, more accessible cloud-based computing may free up AI activity from the concentration of resources, talent, and research institutions in early mover hubs.
But as this new data on generative AI job ads shows, the winner-take-all dynamics of the digital sector’s past may also be a harbinger for AI. If that’s the case, and indeed either way, federal, state, and local governments will need to act now to broaden the reach of AI developments and ensure their benefits for more people in more places.
