The US is home to some 60% of the top 2,000 AI scholars in the world and attracted roughly one-quarter of all AI specialists that relocated globally between 2022 and 2024; its total AI talent pool has grown to nearly half a million people, the largest in the world. And US-based AI scholars have contributed 35% of the world’s most influential papers in the field of AI since its inception.³ 3 A.E. Ezugwu, J. Greeff, and Y. Ho. “A Comprehensive Study of Groundbreaking Machine Learning Research: Analyzing Highly Cited and Impactful Publications Across Six Decades,” Journal of Engineering Research (October 2023). Notes: 3 A.E. Ezugwu, J. Greeff, and Y. Ho. “A Comprehensive Study of Groundbreaking Machine Learning Research: Analyzing Highly Cited and Impactful Publications Across Six Decades,” Journal of Engineering Research (October 2023).

US-based GenAI startups have also received unparalleled private investment: a total of $65 billion since 2019. In addition, AI-directed capital expenditures within established tech firms like Alphabet, Amazon, Meta, and Microsoft are expected to exceed $200 billion for 2024. With massive computing capacity, a full-stack approach extending to hardware codesign, deep pockets, global reach, access to the world’s top talent, and in-house model developers or deep partnership with outside developers, these companies are singularly positioned to support US leadership in GenAI in the foreseeable future.

Beyond talent and capital, the US also has the infrastructure to support inference at scale—that is, to supply GenAI on an ongoing basis to model users. It is the largest data center hub in the world, with an estimated capacity of approximately 45 gigawatts in 2024. The US has reliable access to cutting-edge hardware (through US-based Nvidia and strong geopolitical ties to Taiwan-based TSMC) to support data center expansion, though long lead times to connect to the electrical grid may create challenges.

Made-in-China Models Are Catching Up

Some signs indicate that China is catching up to the US on some key enablers and making rapid progress in the production of top LLMs. Two Chinese companies, Alibaba and the GenAI startup 01.AI, contribute over one-quarter of the world’s top open-source models. Established tech giants Baidu and Tencent have also released high-performing models, as have a new generation of GenAI startups, the so-called AI tigers: Zhipu AI, Baichuan AI, Moonshot AI, and MiniMax AI. Top Chinese models have substantially reduced the gap compared to state-of-the-art alternatives in the last year. (See Exhibit 2.) And top Chinese models have effectively closed the gap in Chinese language benchmarks.

The Chinese GenAI ecosystem has benefitted from access to foreign open-source or open-weights foundation models: 01.AI’s Yi model family, for example, is based on Meta’s Llama. (See the sidebar “A Landscape of Open-Source and Closed Models.”) The Chinese government has signaled its intent to ensure an ample supply of domestic open-source options that play a similar role in the future, supporting a growing number of smaller companies as they ramp up their technical capabilities.

Although it trails the US, China still has a strong AI talent bench—and potent national research champions like Tsinghua University and Shanghai Jiao Tong University. The four AI tigers are all founded by Tsinghua faculty or alumni. China’s strength in AI talent is apparent in its leadership on patents, with more than 76,000 AI patents filed with the World Intellectual Property Organization between 2019 and 2023, about four times as many as the US.

China also has ample data center infrastructure, with approximately 20 gigawatts of capacity, and significant capital through its public R&D budget, which has averaged $50 billion per year between 2016 and 2023. In fact, government venture capital funds demonstrate China’s commitment to boosting AI innovation through public investment. Government VCs have invested nearly a quarter of their total funds in AI firms, exceeding $180 billion in total since 2000, of which roughly $100 billion was invested between 2019 and 2023.⁴ 4 Martin Beraja, Wenwei Peng, David Y. Yang, and Noam Yuchtman. “Government as Venture Capitalist in AI,” National Bureau of Economic Research Working Paper Series (July 2024). Notes: 4 Martin Beraja, Wenwei Peng, David Y. Yang, and Noam Yuchtman. “Government as Venture Capitalist in AI,” National Bureau of Economic Research Working Paper Series (July 2024). And private investment, domestic as well as foreign, has actively driven progress among the AI tigers, which together with 01.AI have raised more than $6 billion in the last few years—including from Chinese hyperscalers, notably Alibaba.

Current limitations in China’s access to state-of-the-art chips for AI model training and inference are likely to delay rather than impede further progress. Many Chinese companies reportedly retain access to chips, and Nvidia has designed models not covered by US trade restrictions; Chinese companies acquired $5 billion worth of these chips in 2023.

Furthermore, China is investing aggressively in its domestic chip manufacturing capacity: the government has pledged approximately $40 billion, and Huawei recently released its Ascend 910 chips, optimized for AI workloads. Regardless of whether it can compete with Nvidia at the cutting edge of computing hardware—and especially the software layer on top of it—Huawei has reached a milestone in recent months with the training of a high-performing LLM, iFlytek’s Xinghuo 4, entirely on its Ascend platform.

Chinese companies also have ambitions outside their home market. Alibaba, for example, is expanding its data center footprint in Malaysia, Philippines, Thailand, South Korea, and Mexico—extending the reach of its Qwen family of generative AI models.

GenAI Middle Powers on the Rise

This US–China story has fueled the “two superpowers” narrative—and convincingly so. But momentum is building in other parts of the world. The EU, for example, has strong complementarity among member states with established strengths. In the Gulf, highly concentrated and agile capital, coupled with affordable energy, is accelerating progress even in the absence of a robust, established tech sector. South Korea and Japan both have strong technology sectors, with the capital to scale.

Made in the European Union

For some, what may be out of reach at the individual country level becomes feasible at the bloc level. The EU stands to benefit from the complementarity of enablers among member states and the scale it can achieve as a unified market.

Building on Distributed Strength. The EU is already home to a nascent GenAI startup ecosystem. France-based Mistral has contributed seven of the world’s top LLMs and has received $1.2 billion in funding to date. The German startup Aleph Alpha has also developed powerful foundation models (two of the world’s top LLMs), though it has most recently pivoted to industry-specific applications.

Smaller GenAI startups are also attracting investment. Kyutai, founded in 2023, has received around $350 million and already released its Moshi AI model, specialized in advanced speech; with $220 million in seed capital, H is pushing to develop productivity-enhancing AI agents; Poolside, meanwhile, has received investments worth $626 million to build a leading model for code generation; and Black Forest Labs has obtained around $150 million to further its text-to-image foundation model.

While these companies are modest in size and funding compared to US and Chinese counterparts, they have a strong foundation of talent on which to build. The EU has an extensive and growing pool of AI specialists, of which there are more than 100,000 in France and Germany alone. (See Exhibit 3.)

The EU also has the advantage of being a massive market, with a combined GDP of $18 trillion. It’s so large that its comparatively demanding regulatory framework for AI, including the General Data Protection Regulation (GDPR) and the recently enacted EU AI Act, is unlikely to dissuade GenAI suppliers from operating there.⁵ 5 Meta, which in July 2024 announced its intent not to make certain models available in the EU for regulatory reasons, is not a supplier of GenAI as defined in this article, as it develops foundation LLMs but does not itself supply (and monetize) inference for end users. Notes: 5 Meta, which in July 2024 announced its intent not to make certain models available in the EU for regulatory reasons, is not a supplier of GenAI as defined in this article, as it develops foundation LLMs but does not itself supply (and monetize) inference for end users. In fact, its regulations could create demand for EU-developed and -hosted intelligence, which may be viewed as more trustworthy and protective of users’ data.

The EU has the world’s third-largest data center capacity, after the US and China, with 8 gigawatts in 2024. While electricity prices in the region are high, considerable variability can be exploited: industrial electricity can cost 20% less in France, and close to 60% less in Sweden and Finland, than in Germany. Moreover, companies looking to supply GenAI models to the EU market may have no alternative to scaling data centers there for the aforementioned regulatory reasons.

Building on the above strengths, we already see complementarity in action: Mistral’s models, for example, are designed in France, trained on Italian supercomputers, and served to clients largely out of Swedish data centers—powered by the chips manufactured using EUV lithography equipment produced exclusively by Netherlands-based ASML.

Bridging the Funding Gap. The EU’s greatest challenge lies in securing the investments necessary to keep up with model scaling and expand data center infrastructure (which includes upgrading the power grid). The EU has a history of lagging investment in the tech sector, and therefore it has been unable to scale tech champions: while US GDP is 1.5 times greater than that of the EU, the market cap of the US share of the world’s 1,000 largest public tech companies is close to 18 times greater than the EU share.⁶ 6 As of late October 2024, the market capitalization of the US share of the world’s largest 1,000 tech companies stood at $24.7 trillion, compared to $1.4 trillion for the EU share. Notes: 6 As of late October 2024, the market capitalization of the US share of the world’s largest 1,000 tech companies stood at $24.7 trillion, compared to $1.4 trillion for the EU share.

As a result, EU-based technology firms have only a fraction of the investment muscle of their US counterparts: The 20 largest EU tech companies spent a combined $40 billion on R&D in 2022, compared to $200 billion for the 20 largest tech companies in the US. Unsurprisingly, GenAI startups in the EU have only received $3.5 billion in investments since 2019, or 5% of the private investment received by US-based GenAI startups.

There are signs of positive momentum, however. The Draghi report on European competitiveness and recent announcements by the European Investment Bank and the European Commission all stress the need for a stronger EU tech ecosystem—with the funding necessary for startups to scale within the bloc. There are pools of resources the EU could tap into, notably public R&D spending across member states, which averaged about $40 billion per year between 2016 and 2022—but doing so will require a concerted effort in a complex environment of distributed decision making.

Precedents for such collaboration exist, particularly in joint efforts like the creation of the European airplane consortium Airbus. But orchestration across national boundaries has also at times been challenging, as with the Gaia X initiative to develop trustworthy and sovereign cloud infrastructure for the EU. It remains to be seen whether the present policy momentum translates into timely action.

Made in Saudi Arabia or the UAE

Because of the high cost of building a top-performing GenAI ecosystem, some countries can use publicly available IP and invest strategically in talent and computing power. Countries may also rely on strong access to capital to outsource the development of models which they can nevertheless own and serve at scale. This approach is best suited to economies such as Saudi Arabia and the UAE, which have centralized access to significant capital resources like sovereign wealth funds.

Investing in Growth. In the Gulf region, Saudi Arabia and the UAE are each making significant investments to diversify their economies and contribute to the region’s broader technological acceleration.

The commitment to direct investments toward AI is clear. The UAE has launched a $10 billion AI VC fund—more than eight times the total funding received to date by Mistral, for example. Saudi Arabia, meanwhile, plans to invest $40 billion in AI development, on top of a recently announced plan to invest $100 billion in data center expansion (Project Transcendence) and its earlier $100 billion tech fund (Project Alat). These committed investments draw on much larger pools of capital: of the 20 largest sovereign wealth funds worldwide, UAE is home to five, totaling roughly $2 trillion in assets under management, while Saudi Arabia’s Public Investment Fund (PIF) manages $920 billion.

Both countries have invested heavily in building an AI R&D ecosystem through universities and sovereign-wealth-backed startups like G42 in the UAE. Since 2022, the AI talent pool in the UAE and Saudi Arabia has grown by 36% and 17%, respectively, and incoming net AI talent migration has grown by 40% and 70%. (See Exhibit 4.)

Data center infrastructure in the Gulf remains small in absolute terms, with Saudi Arabia and the UAE each around 0.4 gigawatts in capacity. But it is growing rapidly, in part enabled by the region’s low energy costs, which can be 30% to 50% lower than in the US on average. The recent easing of cutting-edge chip export restrictions to the UAE—which the Saudi government expects will eventually extend to Saudi Arabia as well—will contribute to sustained access to hardware optimized for AI workloads. Still, the pace of actual data center buildout is an open question.

Progress to date has been notable: Saudi Arabia’s Aramco has produced what is reportedly the largest industrial GenAI model in the world, and the Saudi Data and AI Authority released the leading Arabic LLM family, ALLaM, the largest versions of which are built on the basis of Meta’s Llama-2. Two of the Emirati Technology Innovation Institute’s Falcon models continue to be counted among the world’s top LLMs, and the UAE’s G42 has developed a high-performing Arabic model family, Jais—which speaks to the ability of developers in the region to access a robust corpus of Arabic data for model training.⁷ 7 330 billion Arabic tokens were reportedly used to train Jais; for comparison, Meta’s Llama 3.1 was trained on 15 trillion tokens. Notes: 7 330 billion Arabic tokens were reportedly used to train Jais; for comparison, Meta’s Llama 3.1 was trained on 15 trillion tokens.

Reaching Critical Mass. It remains to be seen whether Saudi Arabia and the UAE can sustain the AI talent pipeline needed to foster a self-sustaining model-development ecosystem. For all the impressive growth in this regard—with roughly 7,000 and 5,000 AI specialists now residing in the UAE and Saudi Arabia, respectively—the Gulf’s talent pool remains small compared to countries such as Germany (55,000 AI specialists) or France (50,000).

More important, for Saudi Arabia and the UAE to become profitable suppliers of GenAI, they have to reach sizeable technology export markets to justify the growing capital demands of GenAI model development. The UAE’s G42 appears to be doing precisely this with its Nanda LLM, designed to cater to Hindi-speaking consumers in India.

Made in Japan or South Korea

Some countries have capabilities across critical GenAI enablers but lack scale. For these nations, which include South Korea and Japan, investing to achieve this scale can create greater competitiveness, at least at a domestic or regional level. South Korea and Japan also have large internal markets and sizeable capital pools to provide the investment required.

The path for these countries is less certain given the challenges they face—but their strength in some key enablers suggests that it’s too soon to count them out.

Capitalizing on a Legacy of Tech Strength. Both South Korea and Japan have strong, heavily concentrated tech ecosystems, dominated by conglomerates that are large R&D spenders. The annual R&D spending of South Korea and Japan’s 20 largest tech companies is greater than that of other GenAI middle powers, at $28 and $26 billion, respectively, compared to $14 billion in Germany and $4 billion in France. (See Exhibit 5.) This focus on R&D translates into strong performance on patents: South Korea and Japan rank first and third among GenAI middle powers in AI patents since 2013.

South Korea and Japan also benefit from their important positions along the hardware value chain, which gives them reliable access to top-quality chips. South Korea holds an approximately 30% market share in high-end (sub-10 nanometer) chip manufacturing, although it trails Nvidia on cutting-edge chips and is expected to lose share in the coming years. Both South Korea and Japan are leading exporters of critical inputs in chip making: indium for South Korea and photoresist processing, material removal and cleaning, manufacturing automation equipment, and arsenic for Japan.

While neither country has produced world-class LLMs to date, there are visible efforts around foundation model development. In South Korea, Samsung and Naver have built their own LLMs—as has telecom giant KT Corp, using AI chips manufactured domestically. Naver, as the largest internet search engine in the country, benefits from its access to data (South Korea is one of only four countries where Google doesn’t dominate the Internet search market).

In Japan, a partnership between universities and private companies recently released the open-source Fugaku-LLM, a model that has strong performance in Japanese-language benchmarks and trained using the Riken supercomputer Fugaku. Others, like tech conglomerate Rakuten, have used foreign open-source models like Llama and Mistral, which are then optimized for the Japanese language.

Finding Sufficient Market Demand and Computing Power. To justify the requisite investments, South Korea and Japan would need to secure large enough target markets for their generative intelligence. South Korea is less advantaged here than Japan, whose economy is roughly two and a half times as large. But South Korean developers appear to be more aggressively moving to secure export markets for GenAI: Naver, for instance, is partnering with Saudi Arabia to jointly develop an Arabic language LLM, while other South Korean companies are building specialized models targeting the Thai, Vietnamese, and Malaysian markets, among others.

Japan has a larger market and greater data center infrastructure, at 1.9 gigawatts of capacity for 2024. This is on par with Germany and well ahead of South Korea’s 0.9 gigawatts. (Planned construction could double capacity in South Korea in the coming years, capitalizing on comparatively low energy prices.) It is not clear yet, however, whether Japanese model developers will be able to compete as suppliers of GenAI with US hyperscalers, many of which are increasing investments to expand their footprint in Japan; to date, no obvious champion for domestic GenAI model development has emerged there.

In South Korea, where Naver and Samsung are clearly making inroads, it remains to be seen if funding will suffice to allow these companies to keep pace with the world’s leading models, which continue to scale. This will largely depend on Samsung, which accounts for 70% of total R&D spending among the 20 largest tech companies in the country.

In both South Korea and Japan, the odds of securing a place as competitive domestic and possibly regional suppliers of GenAI may depend on a concerted effort that capitalizes on the history of strong partnership between the private and public sectors around strategic national priorities. Both governments have voiced strong government-backed AI ambitions; it remains to be seen how those ambitions translate into action, and how quickly.

Betting on Research Breakthroughs

LLMs are largely developed using open IP; the architecture on which they are built was first published by Google in 2017. Increasingly, however, companies are improving LLMs through engineering innovations that are layered onto that open IP. Yet even those improvements fall within the current paradigm of scale, which makes capital intensity and computing-power capacity critical sources of advantage.

This paradigm is not the end of fundamental innovation in AI. It is possible that new approaches that don’t rely on the original architecture of today’s LLMs can deliver superior performance at lower computational cost. Such innovation, if kept proprietary, would reshuffle today’s sources of advantage—privileging whoever is in possession of it.

Breakthrough innovations of this kind could come from the US, China, or the GenAI middle powers—but also from countries with a strong legacy of AI R&D but limited strength across today’s key enablers. Notably, the UK, Canada, and (to a lesser extent) Israel are strong centers of AI research and have made important contributions to the field. (See Exhibit 6.)

Canada and the UK are home to a sizable share of the world’s top AI scholars; they have produced many notable AI models and have published some of the most important AI research papers. Both countries also rank highly in terms of the number of AI specialists residing there—the UK is fourth in the world and Canada is seventh.

This strong bench of talent has driven economic benefit for both countries. Canadian-based GenAI startup Cohere has produced four of the world’s top LLMs (all from the Command model family). The UK-produced DeepMind, one of the world’s top AI research labs, was acquired by Google in 2014 and has been at the heart of the company’s AI innovations, including AlphaGo, AlphaFold, and Gemini—one of the world’s top LLM families.

Israel, with a significantly smaller population than the UK and Canada, is still among the 15 countries with the most top AI researchers and has one of the highest numbers of AI specialists on a per capita basis. Despite its small scale, Israel’s strong concentration of AI talent has enabled it to develop five of the top LLMs (from the Jamba and Jurassic model families) through AI21 Labs.

A Call to Action for Leaders

Both company and country leaders need to be able to navigate the new geopolitics of AI by building their geopolitical muscle: the ability to sense coming shifts and adapt their operating model.

On the one hand, companies have a vested interest in a diversified supply of GenAI. The lessons of Covid are stark: a massive disruption can create severe supply-chain chokepoints and destroy value, which underlines the importance of optionality.

Executives will also need to be able to ensure smooth operations across the geographies in which they do business, despite differences in regulation, language, and legacy tech infrastructure. Certain models may be available in some countries but not in others, which argues for building regional GenAI supply chains and localizing part of a multinational business’s tech operations.

If all options originate in just two countries, it could lead to disruptions in the technology’s availability due to geopolitical shocks. CEOs should therefore consider a portfolio approach to generative AI. Utilizing a variety of models—even a combination of open source and closed—can enable companies to exploit individual models’ specific strengths while increasing options for model access across jurisdictions.

On the other hand, because AI sovereignty may soon become a critical source of national security, economic value, and soft power, governments of countries or in regions with potential to become GenAI middle powers should consider what it would take to claim the space. Ensuring robustness across enablers will involve different immediate priorities for each: some will need to focus on attracting and retaining the right talent, others will need to emphasize investment to boost domestic AI champions, and still others may need to prioritize expanding their data center infrastructure.

The choices that both private and public sector leaders make will be shaped by regulatory and policy action that may create drastic shifts, for example, in tariffs, the restriction of international talent flows, or data regulations. All the while, the tech landscape is quickly evolving, and the cost of entry is rising just as rapidly. In this environment, the ability to understand the full GenAI landscape will be critical.