The UAE’s AI Ambitions

The United Arab Emirates (UAE) is placing enormous bets on artificial intelligence (AI) to diversify its economy and become the world’s next technological hub. As the United States develops its strategy for global AI leadership, the UAE presents a critical test case for engaging with technologically ambitious countries seeking to balance relations with both the United States and China—a challenge that will shape the United States’ broader approach to technological partnerships and export control policies. In late 2024, a group of U.S. technology policy and national security scholars, including the four authors of this paper, traveled to the UAE to understand the country’s AI strategy, its position amid U.S.-China competition, and Microsoft’s $1.5 billion deal with Emirati national AI champion Group 42 Holding Ltd. (G42). The visit combined meetings arranged through the authors’ existing networks with those suggested and facilitated by the UAE embassy in Washington, D.C. This paper presents the authors’ following eight assessments about the opportunities and risks of U.S.-UAE cooperation in AI:

1. The UAE’s commitment to AI represents a genuine national priority with broad leadership support and ambitious economic targets.

2. Emirati officials claim to be “decoupling” from China in AI while acknowledging that they are trying to maintain broader economic ties with both China and the United States.

3. Despite the UAE’s stated commitment to U.S. partnership in AI, China is exerting significant pressure to reinforce its technological influence in the region.

4. While Emirati officials accept the national security justification for U.S. export controls, they express frustration with the United States’ implementation of these controls and the UAE’s country classification.

5. Microsoft’s $1.5 billion investment in G42 reflects its strategy to secure leadership in global AI cloud infrastructure and expand its AI applications portfolio worldwide.

6. G42’s ambitions extend far beyond its Microsoft partnership. The company desires to build 10 to 100 times more data center capacity (independently or with additional partners) than outlined in its current Microsoft agreement.

7. According to G42 and Microsoft representatives, the security controls at the G42 data center the authors toured match typical U.S. commercial standards. While the authors did not find evidence to dispute these claims, the broader question remains whether such commercial-grade controls—whether in the United States or UAE—will prove sufficient as AI capabilities advance and state-sponsored threats evolve.

8. The United States faces two key questions: how to engage with “swing states” in the U.S.-China technology competition, and specifically how to approach the UAE’s ambitious AI aspirations. While the immediate decision centers on AI chip exports, this choice will inform the broader strategy of how to diffuse U.S. technology—from AI chips to cloud services—globally.

Global leadership in AI remains the United States’ to lose, but continued U.S. leadership is by no means assured. The United States can support Microsoft-led projects in the UAE while remaining cautious of endorsing the UAE’s broader AI ambitions. U.S. policymakers and AI executives should maintain a healthy realism about the UAE’s vested interests in hedging its bets between the United States and China. They should ask tough questions—such as what the Emirati government is doing to ensure decoupling from China in AI by tech companies other than G42—and scrutinize the UAE’s answers. For now, however, proceed with caution.

Overview

Rapid advancements in artificial intelligence (AI) in recent years have sparked a global race to develop and control the next generation of AI models, which a large and influential group of senior AI executives and leading researchers believe will be a meaningful step toward human-level intelligence. If they are correct, such technology would have transformative implications for both U.S. national security and the global economy. Leading U.S. companies are making extraordinary and unprecedented investments, reflecting their belief that such a technological revolution is nearly at hand. For instance, according to October 2024 financial disclosures, tech giants Microsoft, Google, Amazon, and Meta are expected to surpass $300 billion in 2025 capital expenditures on AI—a projected 25 percent increase from 2024 and nearly double the 2023 figure. These four companies are hardly alone; in November 2024, Elon Musk’s xAI reportedly raised $6 billion to purchase 100,000 Nvidia H100 graphics processing units (GPUs, also known as AI chips) for the company’s Colossus supercomputer. In December 2024, xAI announced plans to expand the facility to 1 million chips, which could require another $54 billion (assuming consistent chip models and pricing). Meanwhile, the United States and China are locked in a fierce struggle over critical elements of the AI and semiconductor supply chain and are accordingly investing tens of billions annually in building domestic AI capabilities.

Against this backdrop, the wealthy and increasingly influential UAE has inserted itself into the global AI race. The UAE government has identified AI as a critical opportunity to diversify its economy for a post-oil future in line with global climate pledges. At the same time, the country’s strategic advantages—abundant energy, rapid construction, and close connectivity to markets in the rest of the Middle East and Global South—are highly attractive to U.S. AI companies looking to expand AI infrastructure worldwide. U.S. and Emirati companies have already announced billions of dollars in partnerships. On January 21, 2025, OpenAI announced plans for a $500 billion private sector investment vehicle called The Stargate Project to build U.S. AI infrastructure for OpenAI alongside Abu Dhabi investment firm MGX and Japanese investment firm Softbank. In September 2024, MGX and U.S. firms Microsoft, BlackRock, and Global Infrastructure Partners also launched a fund for AI data centers and supporting infrastructure with an investment potential of $100 billion. Additionally, in April 2024, Microsoft announced a $1.5 billion investment deal with Emirati AI national champion Group 42 Holding Ltd. (G42). At a government level, UAE President Sheikh Mohamed bin Zayed Al Nahyan (also known as MBZ) met with U.S. President Joe Biden in Washington in September 2024, culminating in a bilateral AI cooperation agreement between the two nations. Emirati leadership has also participated in investment talks with other U.S. industry leaders, such as OpenAI CEO Sam Altman, to support a $7 trillion project to expand global AI infrastructure, including in the Middle East.

Some experts in Washington view the United States’ growing partnership with the UAE in AI as a critical opportunity to compete with China. The United States’ lead in AI—especially in AI chips—could offer an opportunity to coax the UAE and other “swing states,” which are agnostic to U.S.-China competition, away from China’s sprawling technological reach. The Biden administration’s decisionmaking rested on the assumption that foreign governments will be willing to jump through procedural hoops to access superior U.S. technology as they seek to develop their own indigenous AI capabilities. By offering access to U.S. compute power with strings attached—via export license conditions—the theory goes that the United States can regain geopolitical influence in regions where it is slipping. This bargaining position is particularly strong given that the United States can offer a comprehensive technological package—from advanced AI chips to cutting-edge models and cloud infrastructure—that China currently cannot match at scale. These export conditions could secure commitments not only for enhanced cybersecurity but also for the responsible development and deployment of AI. In the case of the UAE, this strategy has, for now, proven effective. In response to U.S. pressure, G42 has reportedly divested from Chinese technology firms, stripped Chinese tech giant Huawei’s equipment from its data centers, and implemented strict security measures (see points two and seven of this paper) to secure shipments of Nvidia H100 chips to the UAE. The U.S. government reportedly approved the export licenses of these chips to a Microsoft-operated facility in late 2024.

Still, there are substantial risks to exporting large numbers of advanced AI chips to a nondemocratic country with strong ties to China and deploying leading U.S. AI models there. Unlike traditional technology transfers, AI chips can be accessed remotely, meaning their computing power could be diverted even if the physical hardware remains secure. Furthermore, hosting model weights abroad increases the risk of data exfiltration, particularly if data center security and IT infrastructure prove inadequate. Some U.S. lawmakers have raised concerns that such transfers could allow entities with connections to the Chinese Communist Party and People’s Liberation Army (PLA) backdoor access to U.S. AI technology and know-how, thereby evading U.S. export controls.

To address these concerns, the UAE has eagerly voiced its commitment to working with the United States in AI. In June 2024, UAE Minister of State for AI Omar Al Olama stated, “The honest truth is in the AI space today, I think we need to be selective of who we work with. . . . But on the AI front, I think there is going to be complete alignment between the UAE and the U.S.” At the Dubai GITEX Global Expo in October 2024, G42 CEO Peng Xiao, a former U.S. national who renounced his U.S. citizenship for Emirati citizenship, similarly remarked that the UAE’s relationship with the United States “cannot be stronger.” He added, “We’ve shown from [the] UAE side how transparent we are, and how we can guarantee the safety and security of this technology even when [it is] being deployed and used here in [the] UAE.”

The ultimate question is whether the United States can trust the UAE’s claims, especially when the stakes may be nothing less than leadership of a new geopolitical order. On the one hand, the United States must ensure that China cannot access U.S. technology and that the UAE is not simply hedging its bets to secure the best AI chips it can obtain by playing the United States and China against each other. More broadly, if AI reaches human-level intelligence, the United States should be cautious not to offshore its advantage too quickly. On the other hand, without rapid expansion of AI and energy infrastructure beyond current U.S. capabilities—or a plan to entice emerging powers to choose U.S. technology over China’s—the United States risks losing its global technological lead altogether. These are the kinds of tough trade-offs that the second Trump administration, which has vowed to “Make America First in AI,” must now face as it decides whether to continue approving export licenses to the UAE, at what rate, and under what conditions.

Introduction

In late 2024, a delegation of U.S. technology policy and national security scholars traveled to the UAE to better understand the country’s AI ambitions, its position amidst U.S.-China competition, and Microsoft’s deal with G42. Prior to the trip, CSIS experts identified stakeholders and arranged meetings with help and independently from the UAE embassy in Washington, D.C., and solicited input from the other delegation members. The UAE embassy also suggested government and private sector stakeholders that may be of interest for the delegation to meet, which the delegation was at liberty to accept or reject.

This paper presents the authors’ key judgments from the trip based on private interviews with top-level UAE government officials, senior Microsoft and G42 executives, a site visit to a G42/Microsoft data center, and conversations with U.S. and allied government officials. The paper concludes by presenting two strategic questions (point eight) to U.S. policymakers about engaging with the UAE’s AI ambitions in the future.

Key Implications for U.S. AI Leadership

1. The UAE’s commitment to AI represents a genuine national priority with broad leadership support and ambitious economic targets.

While many governments notionally aspire to be global leaders in AI, few are as genuinely serious as the UAE. Total Emirati AI investment is not publicly disclosed, but the figure is clearly in the many tens (to possibly low hundreds) of billions of dollars. In March 2024, Abu Dhabi sovereign wealth fund Mubadala, G42, and the UAE state-backed Artificial Intelligence and Advanced Technology Council announced the creation of MGX, an AI investment vehicle expected to eventually reach $100 billion in assets. Six months later in September, MGX, together with U.S. firms BlackRock, Global Infrastructure Partners, and Microsoft, launched a global AI infrastructure fund with an initial $30 billion in private equity capital and $100 billion in investment potential. More recently, on January 21, 2025, OpenAI announced that MGX will be one of the initial four equity funders for the new “Stargate Project,” which plans to invest $500 billion in building U.S. AI infrastructure for OpenAI over the next four years.

Underpinning these investments is a real conviction that profound economic and social transformation from AI is near. Top UAE government strategists told the authors that the UAE aims to have 20 percent of its non-oil GDP come from AI by 2031, in line with the dates set by the UAE National Strategy for Artificial Intelligence 2031, the government’s principal AI strategy document. The strategy projects that AI will contribute $91 billion (AED 335 billion) to the Emirati economy within six years.

Additionally, one foreign government official told the two CSIS authors that several UAE leaders are “true believers” in artificial general intelligence (AGI), a currently hypothetical scenario in which AI meets or exceeds human levels of intelligence across all or most domains. According to this source, UAE National Security Advisor and G42 Chair Sheikh Tahnoun bin Zayed Al Nahyan believes that AI will be able to perform any human-level job in the near future and that this will be an economic and national security revolution. Though Al Nahyan has not made any public statements on this issue, he has repeatedly met with top AI executives who have been outspoken on AGI—including Microsoft CEO Satya Nadella, OpenAI CEO Sam Altman, and Nvidia CEO Jensen Huang—as part of ongoing partnerships between G42, the UAE government, and the three U.S. companies. OpenAI’s press announcement of the Stargate Project, of which MGX is a key investor, concludes with the statement that “All of us look forward to continuing to build and develop AI – and in particular AGI – for the benefit of all humanity.” Al Nahyan is the head of a reportedly $1.5 trillion sovereign wealth empire and embodies a tech-guru persona that puts him in the same orbit as many of the world’s largest tech CEOs.

2. UAE officials claim to be decoupling from China in AI while acknowledging that they are trying to maintain broader economic ties with both China and the United States.

Multiple UAE government officials told the authors that they see decoupling from China in AI as a critical condition for cooperation with the United States. In a private meeting, one top-level government official made this clear: “Working with the U.S. on sensitive technology—for that, decoupling is the right word. I think we have to do it; in that isolated part of our economy, it must be done.”

Several other officials echoed this sentiment. One high-ranking government strategist described cooperation on AI as a zero-sum game in the context of U.S.-China competition, stating, “We put all our eggs in one basket working with the U.S.” Another stated, “This relationship [with the United States] is very critical to us. We take the trust that the U.S. puts in us very seriously.” They concluded that, to meet its long-term goal of cooperating with the United States in AI, the UAE’s government understands that it must take action to enforce decoupling from Chinese AI technology in order to pursue its larger AI-related goals—even if it means hurting Emirati firms.

Emirati officials principally cited G42’s divestment from Chinese companies as evidence of decoupling from China in AI. Prominent U.S. national security officials have openly criticized G42’s ties to China, including its portfolio companies’ ties to Chinese military and intelligence firms, as well as Xiao’s personal affiliations with companies that contribute to Chinese military-civil fusion and human rights abuses. In January 2025, Wired also published a feature on G42 Chair and National Security Advisor Tahnoun bin Zayed Al Nahyan, which detailed years of covert hacking and mass-surveillance operations orchestrated by the Emirati government and state-controlled firm DarkMatter to monitor journalists, human rights activists, and other members of state opposition. Following a public scandal in 2019, DarkMatter was dissolved and reportedly subsumed in large parts by G42.

To combat these claims, G42 argues that it has demonstrated its commitment to the United States by publicly cutting ties with China. As early as December 2023, Xiao told the Financial Times, “In order for us to further our relationship—which we cherish—with our U.S. partners, we simply cannot do much more with [previous] Chinese partners.” Xiao added, “We cannot work with both sides. We can’t.”

Additionally, one senior G42 representative told the authors that G42 had “willingly” divested its passive investments in Chinese tech firms ByteDance, xFusion, and Honor in November 2023 to “disambiguate” G42’s position vis-à-vis China and demonstrate transparency. The statement supports reporting from the Financial Times in February 2024, in which 42X Fund, G42’s $10 billion investment arm, told the outlet that it had “divested from all its investments in China.” G42 also said that it is taking pains to strengthen security measures in its data centers, including additional security controls (see points four and seven) and scrapping what one of its representatives told the authors was $1.7–$2 billion worth of Chinese hardware, including from Huawei, from its facilities (although one G42 representative admitted that a sizable portion of the Chinese equipment had already significantly depreciated in value by the time it was removed from the data centers). Microsoft officials told the authors that their staff have personally visited G42 facilities and seen Huawei equipment being stripped and discarded.

Still, there are several reasons to be skeptical of G42 and the UAE’s decoupling claims. For one, China has raised no public issue with G42 stripping Chinese hardware from their facilities (a peculiar silence given its retaliation in response to similar actions from other countries in the past), which could suggest some kind of quiet agreement between the two nations. For another, in July 2024, Bloomberg reported that G42’s divestments from Chinese firms were transferred to a new Abu Dhabi investment vehicle, Lunate. Lunate, like G42 and multiple Emirati sovereign wealth funds, is overseen by the UAE’s national security advisor. The transfer of investments raises serious questions as to whether G42—and the wider UAE—is serious about divesting from Chinese companies in AI. Lunate publicly advertises that it has $105 billion of assets under management, including one S&P China/Hong Kong Exchange Traded Fund with top holdings in Chinese tech companies Alibaba (27.7 percent), Meituan (18.9 percent), and Xiaomi (8.10 percent).

For another, it is not obvious to what extent the UAE is expanding decoupling from China in cutting-edge technology, including AI, to other major Emirati tech companies. G42 is undoubtedly viewed by the UAE’s government as a national champion. Through a deeply interwoven relationship between G42 leadership and the government, the company has become a symbol for the UAE’s AI development writ large. There are, however, other Emirati companies that merit scrutiny, particularly those working with U.S. and Chinese companies simultaneously. In March 2022, for instance, Huawei and UAE state-owned telecommunications company e& (formerly Etisalat Group) launched a 5G cloud edge computing platform called 5G Edge Box. The announcement came just months after e& launched a similar 5G edge cloud computing platform with Microsoft. According to Microsoft, the platform will deploy “Microsoft Azure Stack Edge to offer 5G, IoT, and AI applications.”

e& continues to collaborate with both Huawei and Microsoft to provide 5G cloud edge networks. In February 2024, e& signed a new memorandum of understanding with Huawei to build energy-efficient networks in the UAE, which states that “e& will continue to work with Huawei to achieve network decarbonisation across its ICT infrastructure, including . . . data centers.” It is unclear from public statements to what extent Huawei’s “network decarbonization” of e& information and communications technology infrastructure could affect e&’s current or future U.S.-operated data centers in the UAE. For example, in October 2024, e& and Amazon Web Services (AWS) announced a $1 billion cloud agreement, which “combines AWS’s cloud infrastructure and solutions . . . with e&’s network capabilities.”

Furthermore, as of 2021, e& owns 40 percent equity in wholesale data center developer and leasing operator Khazna, the main developer of Microsoft’s AI data centers in the UAE (described in more detail in point five below). G42 owns the remaining 60 percent and claims Khazna as a G42 portfolio company. Figure 1 illustrates G42’s portfolio companies and examples of foreign partnerships, and Figure 2 maps G42’s portfolio companies across the AI value chain.

▲ Figure 1: G42’s Portfolio Companies, Associated Investment Firms, and Examples of Foreign Partnerships. Source: G42 and partner companies’ websites and public press statements; Miriam Gottfried, “Silver Lake Invests About $800 Million for Minority Stage in Abu Dhabi’s G42,” Wall Street Journal, updated April 14, 2021; and CSIS analysis. Note that MGX has been included as an associated investment firm for which G42 is a founding partner and G42 Chair Tahnoun bin Zayed Al Nahyan is also the head.

▲ Figure 2: G42’s Portfolio Companies Across the AI Value Chain. Source: “Our Companies,” G42; G42 portfolio websites; and CSIS analysis.

Outside the realm of technology, the UAE’s relationship with China runs deep. On November 1, 2024, China and the UAE celebrated their 40th year of diplomatic relations, marked by a public statement from Chinese President Xi Jinping in which he described the two countries as “good friends who trust each other and good partners for win-win cooperation.” UAE President MBZ likewise said, “The UAE remains committed to consolidating and developing a comprehensive strategic partnership with China.” Senior Emirati officials speak of a golden era of Chinese-Emirati cooperation, a sentiment echoed by the Chinese ambassador to the UAE who described this as the “best period of time since we established diplomatic relationships in 1984.”

Indeed, China is the UAE’s top trading partner, with UAE government estimates valuing non-oil trade at $82 billion in 2023, a 34 percent increase from 2021. Despite a show of partnership with the United States in AI, the UAE continues to work closely with the Chinese government and technology firms in sensitive areas such as 5G, drones, and other military technologies. For instance, the UAE’s Air Force and China’s PLA Air Force conducted a joint military exercise in July 2024 in Xinjiang province, China, which may have allowed the PLA access to intelligence about Western combat aircraft capabilities. The joint exercise came just months before the Biden administration designated the UAE a “major defense partner” of the United States, expanding bilateral defense and security cooperation through joint military exercises among other measures.

For these reasons, one top-level government strategist privately called the idea of a complete decoupling from China “anathema” to UAE policymakers. Because Emirati policy is driven by geoeconomics rather than geopolitics, the strategist said, the UAE will continue to develop a strong trade relationship with China at the same time as with the United States. UAE Assistant Minister for Advanced Science and Technology Omran Sharaf publicly stated in a November 2024 interview on AI and diplomacy that this delicate navigation between two great powers is what other third-party countries are “trying to learn from the UAE model.” While countries caught between U.S.-China competition are increasingly under pressure to choose between U.S. and Chinese technology and values, the “countries who are able to navigate that the best are the ones who get the most out of the opportunities that are there globally, whether economic or other areas that involve science and tech,” Sharaf said. This thinking suggests a more careful balancing of UAE diplomatic relations between the United States and China (including in technology) than the zero-sum commitment to the United States that other UAE government officials have described.

The UAE’s approach to decoupling from China requires careful distinction between rhetoric and reality. While G42’s actions demonstrate company-level separation from Chinese technology in AI, the broader picture is more nuanced. First, what constitutes the scope of this decoupling remains unclear—whether it applies narrowly to AI or extends to all advanced technology sectors such as 5G. Second, the transfer of Chinese investments from G42 to Lunate suggests more of a reorganization than true divestment. Third, other major Emirati tech companies continue to maintain simultaneous partnerships with both U.S. and Chinese firms. This selective decoupling approach is perhaps unsurprising—even the United States maintains extensive economic ties with China while restricting specific strategic technologies. The critical question is not whether the UAE will completely decouple from China, which would be an unrealistic expectation, but whether it can establish credible boundaries in strategic technology sectors.

3. Despite the UAE’s stated commitment to U.S. partnership in AI, China is exerting significant pressure to reinforce its technological influence in the region.

The Chinese government and Chinese companies, notably Huawei, are reportedly spreading the message that U.S. delays in approving exports of AI chips to the UAE prove that the United States is a selfish and unreliable partner. One UAE government official in the Ministry of Foreign Affairs privately noted that China is putting a great deal of pressure on the UAE government to lure them away from the United States in strategic tech, posing a tough diplomacy challenge for the UAE. Similarly, a senior Microsoft representative said that Huawei is trying to leverage the United States’ slow approval to ship semiconductors to the UAE “to the extreme” in its effort to get UAE officials back on board. The representative pointed to Huawei’s top “diamond” sponsorship of the UAE’s flagship GITEX Global Expo, held in Dubai in October 2024, as evidence that China has not given up on the goal of being the dominant AI supplier to the region.

One government official said that while Chinese technology is inferior to that of the United States, the UAE must get advanced technology from somewhere to boost its competitiveness. In late 2024, the Chinese firm Alibaba released a new open-weight large language model (LLM), Qwen2.5. Senior executives at two top-tier U.S. AI firms told one CSIS author that Qwen2.5’s release led them to dramatically reassess the extent to which the United States has a meaningful technological lead in AI. Specifically, one executive told the CSIS author that “the conventional wisdom that U.S. AI firms have an 18-month lead on Chinese firms is clearly wrong and needs to be updated.” Jack Clark, a cofounder of Anthropic, wrote that “Qwen models outperform rival Chinese [computer code-generating AI] models from companies like Yi and DeepSeek and approach or in some cases exceed the performance of powerful proprietary models like Claude 3.5 Sonnet and OpenAI’s o1 models.” More recently, impressive jumps in the performance of Deepseek’s “V3” model have fueled concerns about China’s advancement of frontier model capabilities. On January 21, 2025, former OpenAI Senior Advisor Miles Brundage tweeted, “Stargate + related efforts could help the U.S. stay ahead of China, but China will still have their own superintelligence(s) no more than a year later than the US, absent e.g. a war.” These statements are important because Emirati officials repeatedly stated in private meetings that their perception of U.S. technological leadership was the key reason for their willingness to decouple from China in AI technology. If the United States were to lose this strategic advantage, the UAE could decide to redirect its attention to Chinese technology firms, much like it did with 5G.

Still, UAE, G42, and Microsoft leadership cited several reasons the UAE would not—or could not—switch so easily. First, while China is making progress in some areas like AI models, the United States maintains critical advantages in the broader AI technology stack, particularly in advanced chips and the scale of production needed for large AI infrastructure. Model performance is just one metric of technological leadership—China continues to lag significantly in AI chip development and production capability due to U.S. export controls. These controls limit both the quality of chips China can produce and, critically, the quantity they can manufacture at scale, making it difficult for China to match U.S. capacity in providing the computing infrastructure needed for advanced AI development. While model performance comparisons may narrow, China’s limited ability to produce advanced AI chips at scale could preserve U.S. strategic advantage.

Second, one senior Microsoft executive argued that even if the UAE wanted to, switching back to Huawei AI infrastructure may soon become prohibitively technically challenging. Microsoft officials claimed that if the United States approves Microsoft’s exports of AI chips, they can “lock” the UAE into a U.S.-led technological ecosystem, allegedly due to high switching costs between different cloud and AI infrastructures as well as network effects for AI-enabled software applications. Creating technological lock-in becomes particularly powerful when combined with China’s limited ability to provide comparable computing infrastructure at scale.

If this claim proves true, it would suggest that the United States could exploit its current technological lead to secure an enduring advantage over China in the next few years. The authors are not in a position to assess the accuracy of the executive’s claim definitively, and it should be regarded with some degree of skepticism: Large AI companies have deep pools of engineering talent that they could use to overcome lock-in, if necessary. Additionally, the fact that companies like Anthropic use multiple cloud providers for their AI models—with Anthropic even beginning to use AWS’s custom chips—suggests that while switching and working across infrastructure platforms may be difficult for many companies, it is not prohibitively difficult for those as well-resourced as Anthropic, OpenAI, or G42.

G42’s own strategy further illustrates this point: the company aims to “abstract the silicon away” through partnerships with multiple chip providers. It is already the largest customer of AI chip startup Cerebras and is running one of its clusters in California, while also recently partnering with Qualcomm to procure AI chips. In several conversations, representatives from Core42, G42’s sovereign cloud and AI infrastructure portfolio company, emphasized their goal to build on top of different types of AI chips and abstract it away for their customers. This diversification of chip suppliers and infrastructure suggests that while AI chips alone may not provide an insurmountable competitive moat (though Qualcomm and Cerebras are both subject to U.S. export controls), offering a comprehensive technology stack—including chips, cloud infrastructure, and AI applications—could create more durable strategic advantages.

4. While Emirati officials accept the national security justification for U.S. export controls, they express frustration with the United States’ implementation of these controls and the UAE’s country classification.

Several UAE government and AI company representatives readily acknowledged and agreed with the underlying reasons for the United States’ concern in exporting AI chips to the UAE in private meetings. Most notably, one UAE government official said that the UAE understands that the United States cannot “shoot itself in the foot” regarding AI. The official stated that the United States has a right to place restrictions given what the official described as the “dual-use nature” of the technology and the boost that it will give to military and intelligence capabilities.

On October 15, 2024, in a public interview at the GITEX Global Expo, G42’s Xiao echoed the UAE government official’s words. Xiao was asked whether he felt that the United States is using then-proposed caps on chip exports to some countries, including in the Middle East, as leverage for achieving its diplomatic goals. Xiao responded:

“I cannot read the mind of the U.S. policymakers, but in many ways, I understand their position. . . . When a nation, or even when a company or individual [has] invented a new technology, it’s only natural [that] you want to hold on to that edge . . . to make sure [that] because this new technology is so cutting edge, before you figure out how to properly use it, it cannot be possibly misused by others. So, in many ways, I understand that position [the] U.S. government has taken. At the same time from our side—we’ve shown from [the] UAE side how transparent we are and how we can guarantee the safety and security of this technology, even when [it is] being deployed and used here in [the] UAE. So, I think that door is opening up for us to do a lot more.”

Another top G42 executive privately agreed with Xiao, saying that he recognizes that G42 is not the target of U.S. export controls but rather “just collateral damage” in U.S. competition with China. The executive expressed support for an “N-1” strategy, in which the United States maintains exclusive access to the most cutting-edge AI chips for a certain period before allowing them to be exported abroad. The executive said that with sole ownership of the latest and most powerful chips for six months, for example, the United States could benefit from enhanced performance first and exercise greater control over AI diffusion.

Still, the executive said that G42 has implemented strict controls to demonstrate as much compliance and transparency to the United States as possible. According to Semafor reporting from September 2024, Microsoft-operated G42 data centers use military-grade encryption and are audited by U.S. Department of Defense contractors to test for potential security breaches that could be exploited by China. The Microsoft deal also reportedly forbids G42 from using Microsoft services or AI chips for surveillance and mandates that it seek permission from Microsoft to share the technology with other governments or militaries. On top of these controls, the G42 executive stated that G42 has applied even stricter security controls on its data centers than is required by the U.S. government and Microsoft (see point seven). According to G42 documents shared with the authors, additional controls include but are not limited to:

• Standard data center information security policies and procedures (such as personnel security, cybersecurity, and campus security measures);

• U.S. export license requirements (such as entity list restrictions, conformity inspections, U.S. Bureau of Industry and Security notification of noncompliance, record retention, and computer security monitoring);

• over 100 controls voluntarily adopted by G42 from the National Institute of Standards and Technology (NIST) Risk Management Framework SP 800-53 security and privacy controls for information systems and organizations; and

• enhanced customer due diligence procedures, including rigorous Know-Your-Customer requirements with particular attention to customers who train frontier models.

Many of these controls appear to be standard security measures commonly implemented by commercial data centers and major cloud providers. Without significantly more detailed information about their implementation—as cybersecurity effectiveness often lies in technical specifics—it is difficult to assess whether these controls represent meaningful improvements or adequately address the unique threats posed by state actors.

A G42 executive indicated that while these controls are necessary to demonstrate compliance and cooperation with the United States, they are placing a heavy financial burden on the company. The executive stated that “our costs are too high at the moment due to compliance,” which is causing G42 to lose customers to other companies without such strict controls. The executive concluded that they would like to see the United States standardize compliance requirements and apply them to companies around the world.

Multiple UAE government officials also voiced frustration that the United States’ tight restrictions are unjust given the UAE’s demonstrated willingness to play by the United States’ rules. One official said that the UAE’s categorization as a “D:4” country on the Department of Commerce Control List Country Chart (alongside countries like Saudi Arabia, Pakistan, and Yemen) is unfair, objecting to being “grouped with our crazy neighbors,” given its demonstrated cooperation with the United States in AI and other sensitive technologies in the past. The official pointed to the UAE’s successful collaboration with the United States on space and nuclear projects as proof of this commitment to technological security. For example, in January 2024, NASA announced that the UAE’s Mohamed bin Rashid Space Centre will provide crew members and an airlock module for its new lunar space station, Gateway. Similarly, in January 2009, the United States and the UAE signed a so-called 123 Agreement for bilateral cooperation on peaceful nuclear energy, which remains in effect today.

More broadly, UAE officials seek clearer guidance on pathways to improved country designation and more consistent, long-term policies that transcend changes in U.S. administrations. One official said that U.S. export controls are clear on an entity-to-entity basis but not at a countrywide level, making the government hesitant to take further action to decouple from China until the United States gives it a clear road map as to what it will take to move up the country designation list. Furthermore, this official said that the United States’ constantly changing export requirements makes real cooperation near impossible, stating, “We respect the U.S. system, we understand your concerns, but as a partner we need stability.” Another top-level official similarly stated that changes in U.S. administrations’ strategies toward China make it unclear how the United States expects the UAE and “countries like us” to navigate their relationship with both powers.

Officials stated that the recently announced Validated End User program for data centers represents a promising step toward establishing a more predictable framework for technology transfers while maintaining necessary security safeguards. The U.S. government has since (in January 2024) announced a Framework for AI Diffusion that builds upon this VEU program, though this development occurred after the authors’ field research.

5. Microsoft’s $1.5 billion investment in G42 reflects a strategic pivot toward securing leadership in global AI cloud inference infrastructure and AI applications.

“It’s all inference, right?”

In the last year, Microsoft has faced considerable scrutiny from the Department of Commerce, Congress, and U.S. national security agencies over its deal with G42 and its efforts to secure shipments of Nvidia AI chips to the UAE. This deal speaks to Microsoft’s evolving strategy for global AI leadership, which involves establishing a strong position in AI inference infrastructure to run AI models rather than an exclusive focus on training them.

Microsoft’s FY 2025 Q1 earnings call on October 30, 2024, with CEO Satya Nadella provides some important clues to Microsoft’s views on the increasing importance of inference (referring to the general process in which an AI model is deployed post-training). In response to an analyst’s question about Microsoft’s strategy for investing in AI models, Nadella responded:

“Capital outlay for training . . . is going to be rate limited by your monetization of inference in a given generation, right? So just like in the past, we would allocate capital to build out cloud based on the demand signal we were seeing and then we would project the demand, and that’s what we would build it for. So, you can think of training essentially as that, right, which is you’re building the next-generation model so that then you have a more capable model that then drives more inference demand.”

Microsoft’s interest in AI models as a way to boost cloud inference demand may help to explain its partial split from OpenAI over the last year; at one time, Microsoft needed to jumpstart AI model demand for its cloud service to run models. However, as October’s earning call suggests, demand is now booming and self-reinforcing such that it might not require continued Microsoft investment and allegiance to an individual model maker. On a December 2024 BG2 podcast episode, Nadella went as far as to say that there will be no one AI model to rule them all in the future. He said, “I don’t think it’s going to be winner-take-all. . . . The world will demand, even ex-China, multiple providers of frontier models distributed all over the world.”

Instead, Microsoft may view inference as the ultimate moneymaker in AI. In the October 2024 earnings call, one analyst asked:

“Analyst: These models are getting bigger, more expensive, but you also pointed out how in the inference phase, we’re likely to get paid. How does that cycle look like in inference for Microsoft?

Satya Nadella: I mean the good news for us is that we’re not waiting for that inference to show up. . . . This is going to be the fastest growth to $10 billion of any business in our history, it’s all inference, right? One of the things that may not be as evident is that we’re not actually selling raw GPUs for other people to train. In fact, that’s sort of a business we turn away because we have so much demand on inference. . . . We kind of really are not even participating in most of that because we are literally going to the real demand, which is in the enterprise space or our own products like GitHub Copilot or M365 Copilot.

So, I feel the quality of our revenue is pretty superior in that context. . . . If this was just . . . a bunch of people training large models and that was all we got, then that would be ultimately still waiting . . . for someone to actually have demand, which is real.”

Nadella reported that Microsoft’s cloud platforms, including those used to offer AI services (and therefore “inferencing” AI models), are already driving total revenue growth. He said, “We are off to a solid start to our fiscal year, driven by continued strength of Microsoft Cloud, which surpassed $38.9 billion in revenue, up 22 percent.” Microsoft’s Executive Vice President and Chief Financial Officer Amy E. Hood concurred that Microsoft “saw continued share gains across many of our businesses. In our commercial business, increased demand, and growth in long-term commitments to our Microsoft Cloud platform drove our results. Commercial bookings were ahead of expectations and increased 30 percent and 23 percent in constant currency.”

On the BG2 podcast episode, Nadella argued that the breadth of Azure’s geographical hubs and capabilities to handle a range of different workloads will be a competitive advantage for Microsoft in the future. He said, “There is going to be fierce competition between the seven, eight, nine, ten of us [hyperscalers] at different layers of the stack,” but that Azure’s structure makes it “slightly different” than other cloud providers. Specifically:

“We built out Azure for enterprise workloads with [a] lot of data residency. . . . We have 60- plus more regions than [other providers] . . . [and] we built [Microsoft] cloud for a lot of heterogeneous enterprise workloads, which I think in the long run is where all the inference demand will be.”

Microsoft’s view that cloud infrastructure for AI inference is where it will see enormous returns on investment—and that the more geographically and computationally diverse it is, the better—helps to explain the company’s push for global AI infrastructure, including in the UAE.

Microsoft’s Inference Bid in G42 and the UAE

Private conversations with Microsoft and G42 executives revealed that G42 and Khazna are building three new data centers in the UAE for Microsoft. The centers will be 30, 30, and 100 megawatts each, the equivalent of nearly 100,000 H100 chips (approximately 15,000, 15,000, and 50,000–60,000 respectively). These data centers will run Microsoft Azure for AI inference, helping to consolidate a regional Azure hub in the Gulf.

Microsoft and other leading hyperscale cloud and AI firms stated that—if U.S. AI chip export licenses continue to be rapidly approved—nearly all of them would consider UAE opportunities as attractive as, or more attractive than, those in the United States, especially in terms of realizing the goal of rapidly developing gigawatt-scale data centers. While the United States grapples with how to domestically build out energy and AI infrastructure on a scale comparable to the Manhattan Project, Microsoft finds partnership with G42 and the wider UAE attractive for four key reasons:

5.1. Availability and Rapid Expansion of (Increasingly Green) Energy Generation Capacity

U.S. AI companies are scrambling to secure long-term sustainable energy sources to fuel unprecedented data center growth, which the U.S. Department of Energy estimates could consume approximately 6.7 to 12 percent of total U.S. electricity by 2028. While bottlenecks in U.S. energy infrastructure currently threaten to place a de facto cap on AI development and deployment in the United States in the near future, the UAE’s capacity to build rapidly and its abundance of energy are incredibly attractive to U.S. AI companies.

The UAE has mobilized impressive capital in recent years to build out energy infrastructure, including for renewable energy, as it strives to meet its ambitious pledge to be carbon-neutral by 2050. In 2023, the UAE government announced a $54 billion investment package to triple its renewable energy supply over the next seven years. Several large-scale renewable energy projects are already underway, such as the monumental Mohammed bin Rashid Al Maktoum Solar Park in Dubai, which is expected to reach 5 gigawatts of installed solar capacity by 2030. Though solar power has historically been problematic for data centers due to fluctuations in supply (between day and night as well as seasons), developments in energy and battery storage capabilities are changing the game. In February 2024, for instance, UAE developer Moro Hub opened the world’s largest Tier III 100+ megawatt data center powered by solar energy in Dubai. U.S. companies Microsoft and Dell, as well as Chinese company Huawei, have reportedly already signed on to be customers.

Furthermore, thanks to a robust engineering and operations base from its oil industry, the UAE has an active labor force with strong technical know-how. As one UAE government official put it, compared to the United States, “we’re part of the world that still remembers how to build.” The official cited the speed at which the UAE built four nuclear reactors for its Barakah Nuclear Energy Plant as a prime example: According to Emirati officials and the World Nuclear Association, the time between breaking ground on the first reactor and connecting the last reactor to the grid took just under 12 years. In contrast, construction of nuclear reactor Units 3 and 4 of the Alvin W. Vogtle nuclear power plant in the United States—the most recent nuclear reactors to enter commercial operation in July 2023 and April 2024, respectively—took 14 years and ran more than $30 billion over budget. The combined net capacity of Vogtle Units 3 and 4 (2,214 megawatts), meanwhile, is only just over half that of the UAE’s four reactors (5,348 megawatts), which it completed in two fewer years. The UAE is already considering building a second nuclear power plant, which government officials told the authors would likely take only seven to 10 years due to efficiencies learned during the construction of the first.

UAE investment in energy infrastructure translates to substantial gains in electricity generation capacity over the next decade. According to a 2024 analysis from PwC Middle East, the UAE’s future electricity generation capacity (from a mix of renewable, nuclear, and fossil fuel sources) will significantly outstrip future domestic demand. In 2025, PwC forecasts, the UAE will have 22.5 gigawatts of total generation capacity, while domestically consuming only 18.3 gigawatts. By 2030, capacity is expected to grow to 26.6 gigawatts while consumption remains a modest 21.2 gigawatts. Finally, by 2035, generation capacity is likely to reach 29.6 gigawatts while consumption rests at 23.2 gigawatts of electricity. If the UAE government decides to build a second nuclear power plant with the same specifications as the Barakah plant by 2035, total generation capacity could soar to 34.4 gigawatts. Figure 3 illustrates this trend:

▲ Figure 3: Forecasted Excess UAE Electricity Generation Capacity to 2035 in Gigawatts (GW). Source: “eMobility Outlook 2024: UAE Edition,” PwC, June 13, 2024; and CSIS analysis. Authors’ note: PwC Middle East’s data reflects proprietary data from Emirati utility companies. Projected generation capacity figures incorporate major planned infrastructure projects by these companies and the UAE government up to 2035. They do not, however, include a second nuclear reactor, which has been added to this graph by the authors with approval from PwC Middle East. PwC Middle East’s projected domestic consumption figures consider factors such as expected population growth and economic expansion over the next decade (including an unspecified amount of growth in installed data center capacity). Figures have been converted from terawatt hours (PwC version) to gigawatts to be more easily comparable to other figures in this paper and are rounded to the nearest decimal.

Though the UAE’s excess electricity generation capacity will not necessarily be directed solely to powering data centers in the UAE, even a fraction of this figure could be a considerable addition to AI infrastructure in the short term. Just 10 percent of excess capacity in 2030, for instance, would be enough to power over 500 megawatts of data centers (approximately equivalent to 300,000 H100 AI chips)—more than sufficient for the additional 420 megawatts of installed data center capacity that Khazna projects will be built in the UAE by 2029 (specifically 850 megawatts in 2029 compared to 429 megawatts in 2024).

Of course, the UAE government could decide to dedicate only a small percentage of its excess electricity generation capacity to powering AI data centers over the next decade. But if AI is indeed the kind of national priority that UAE leadership claims it is (see points one and six), this figure could be a considerable strategic advantage for adding gigawatts of power to AI infrastructure in the next decade. As AI adoption grows, this advantage becomes even more valuable since serving millions of users is substantially more energy-demanding than the occasional training of new AI models.

5.2. Rapid and Cost-Efficient Construction of Data Centers

In October 2024, Khazna CEO Hassan Alnaqbi announced a new Tier-III, 100-megawatt AI-optimized data center in the northern emirate of Ajman. The 100,000 square meter facility will include 20 data halls of AI GPUs using 5 megawatts of capacity each and will be built on an accelerated timeline of 15 to 18 months. In private conversations, Khazna and G42 executives confirmed that the center is being built for Microsoft.

Construction of the 100-megawatt facility is reportedly expected to cost $272 million. According to Alnaqbi, the average cost of construction per megawatt for a group of new Khazna data centers will be between $8 million and $12 million. For comparison, 2024 U.S. market data shows that the average construction cost of U.S. data centers ranges from $10 million to $11 million per megawatt and is projected to increase 5 to 7 percent annually. In terms of total construction costs, one 2023 survey of the global data industry estimates that the average price (excluding hardware) of a 20-megawatt data center typically falls between $100 million and $200 million. Moreover, construction of a 100-megawatt data center can exceed $1 billion. Therefore, if Khazna’s construction estimates are correct, they would be almost four times cheaper than the global average.

5.3. The Opportunity to Lock In High-Demand Customers with Sovereign Cloud Services and Create Network Effects for Their AI Applications

Sovereign Cloud Services

The first part of the Microsoft-G42 deal is to develop a sovereign cloud for the UAE government and public sector using Microsoft Azure and G42-Khazna construction capabilities. In September 2023, Microsoft announced a partnership with G42 to “unlock new opportunities for digital transformation with joint sovereign cloud and AI offering.” The announcement states:

“Microsoft’s sovereign cloud offering will allow UAE public sector and regulated industries to use new platform capabilities for securing sensitive data, providing access to the latest cloud and AI features available on Azure public cloud and helping them comply with local privacy and regulatory requirements. G42’s deep understanding of UAE sovereignty requirements and technical capabilities are central to customizing the offering to help address customer’s specific needs.”

Microsoft reaffirmed this commitment in the more recent $1.5 billion deal with G42 from April 2024. The April press release states that “G42 will run its AI applications and services on Microsoft Azure and partner to deliver advanced AI solutions to global public sector clients and large enterprises,” as well as “to introduce sovereign cloud offerings and collaborate on unlocking the potential of advanced AI capabilities on the Azure public cloud platform.”

In private meetings, G42 stressed that the Azure sovereign public cloud platform remains central to the new G42-Microsoft partnership. This platform is codeveloped between Microsoft and Core42 at the AI compute and cloud platform layers of the AI value chain (see Figures 1 and 2). G42 executives explained that G42 and the UAE government want to expand local data center capacity inside the UAE so that they can store and process sensitive data that, under UAE laws, is not permitted to leave the country. Moreover, Core42 software built on top of Microsoft Azure confidential compute encrypts all data, helping to strengthen the UAE’s data sovereignty. Specifically, any data running on the Core42-Azure stack in the UAE can be accessed—but not read—by foreign governments using extraterritorial authorities over cloud providers overseas, such as the United States’ CLOUD Act or China’s Data Security and Personal Information Protection laws. As a result, one G42 representative said that G42 can technically comply with foreign data laws while simultaneously protecting sensitive “sovereign” data by making it illegible. In return, for Microsoft, UAE companies generating data (including G42’s own companies) will become Microsoft customers, offering Microsoft the opportunity to lock in demand for its cloud services.

G42 executives highlighted G42’s healthcare portfolio company, M42, as a case study of this dynamic. M42 is creating databases of Emiratis’ biometric data as part of a government-sponsored genome sequencing project to improve healthcare diagnostics and drug discovery in the UAE. M42 has collected approximately 700,000 DNA samples thus far—an extraordinary amount of data that requires storage and management and that offers the opportunity for extensive AI analysis. Under the UAE’s Health Data Law, however, in the absence of specific government authorization the storage and processing of health data must take place within the country. Additionally, the sensitive nature of the data means that the UAE wants to host the data on secure encrypted cloud servers that prevent the data from being accessed by foreign governments. As a result, the UAE government and the regulated companies are eager to expand sovereign data center capacity inside the UAE.

G42 executives told the authors that M42’s genomics project is already consuming vast amounts of G42’s local data center capacity and indicated that the firm would be a strong and rapidly growing customer for Microsoft’s new cloud facilities. As the UAE strives to expand its healthcare and pharmaceutical sectors, it will require even greater capacity to store and process data on encrypted platforms. More broadly, G42 and the UAE’s many other companies that handle other forms of sensitive data, such as its space technology company Space42, will likely become strong customers for Microsoft Azure public cloud in the future. This all suggests that Microsoft will secure customers for its planned data centers in the UAE with little trouble.

Creating Network Effects with Microsoft AI Products

In addition to capturing high-demand customers for Azure, Microsoft hopes to become the first and best developer of AI applications that stimulate network effects toward its services. The network effect describes the concept that the more people use a service or product, the more valuable it becomes. It is frequently used in relation to e-commerce or social media platforms, where the quality of the service grows as more users become members. Additionally, the more people use a particular product or service in an individual’s social or professional network, the more likely they are to use it, too, creating a reinforcing cycle.

Over the last 30 years, the network effect has helped to make Microsoft’s Office suite the default operations platform for businesses worldwide. Entrenched norms and efforts to make business software interoperable have ensured that Microsoft products like Word and Excel have become the backbone of the professional world. Through technological diffusion—and not just innovation—Microsoft has become the world’s largest tech companies in terms of market capitalization.

Microsoft would like to reproduce this kind of effect with its consumer and enterprise AI products in the future. Nadella acknowledged this as a Microsoft strategy on the December 2024 BG2 podcast, where he said that “network effects is always going to be at the software layer, at the app layer,” and that there will be “different network effects in the consumer [sector and] the enterprise [sector]” due to different demands for its products, such as Microsoft Copilot for consumers and GitHub Copilot for software developers.

Microsoft executives told the authors that their ultimate hope is to supply emerging markets with their products before Chinese technology firms can get there first. Should Chinese firms overcome U.S. export controls and produce superior AI products (and be able to offer them at scale), they will still face the challenge of competing with an established incumbent across the developing world. Even if a Chinese application is technically better, network effects could ensure that Microsoft customers remain loyal to their products. For this reason, Microsoft is eager to jumpstart a virtuous cycle of demand for its AI products in the UAE and abroad.

5.4. A Gateway to Emerging Markets in the Global South

Another key element of the April Microsoft-G42 deal is the expansion of Microsoft’s AI services to the developing world. Microsoft’s press statement about the deal is specifically titled: “Microsoft invests $1.5 billion in Abu Dhabi’s G42 to accelerate AI development and global expansion.” In private conversations, Microsoft and G42 executives stated that, like the UAE, many governments in the Global South are looking to develop their own public clouds that will store and process data locally in line with national laws (see point 5.3. above). They see this as a critical opportunity for Microsoft to expand its Azure cloud services and for G42 to finance and build out the data center infrastructure abroad. Microsoft’s press statement suggests as much:

“G42 and Microsoft will collaborate to ensure the benefits of secure AI technologies and cloud capabilities are responsibly shared with growing economies. . . . G42 and Microsoft will also work together to bring advanced AI and digital infrastructure to countries in the Middle East, Central Asia, and Africa, providing these nations with equitable access to services to address important governmental and business concerns while ensuring the highest standards of security and privacy.”

Public announcements by G42 and Microsoft have already confirmed that one of these countries is Kenya (see below). Though other countries have not been confirmed under the deal, agreements from 2021 and 2022 between G42 and the government of Kazakhstan, as well as private and state-owned infrastructure companies in Egypt and Indonesia make these likely contenders for expanding Microsoft Azure with G42.

Microsoft and G42’s work in Kenya, which includes developing a 1-gigawatt data center and building a Swahili language AI model, is the first step in this global expansion project. This project began in March 2024, when G42, without Microsoft, announced a partnership with Kenyan firm EcoCloud to build a 1-gigawatt data center in Kenya, “powered by the untapped potential of Kenya’s 10 gigawatts in geothermal energy.” According to the press release, the facility would start at 100 megawatts before scaling to 1 gigawatt to “usher in an era of cloud computing and AI services” in Kenya. Kenyan President William Samoei Ruto reportedly supported the deal personally. It was only later, in May 2024, that Microsoft announced it would join this venture with G42. One top G42 executive privately said that Microsoft’s senior leadership had been “inspired” by G42’s plans in Kenya and “saw the opportunity” to expand Microsoft’s business in the Global South. Microsoft executives likewise confirmed that this was a pivotal moment for the company’s top executives, leading them to give G42 $1.5 billion. G42 executives told the authors that the company was not initially looking for capital investment from Microsoft, only a partnership agreement.

Under the second May 2024 agreement with Kenya, Microsoft and G42 will commit an initial investment of $1 billion in expanding Kenya’s digital ecosystem, in which “one of the Kenyan investment priorities is a state-of-the-art green data center that will be built by G42 and its partners to run Microsoft Azure in a new East Africa Cloud Region.” The public announcement states that Microsoft and G42 will specifically:

“Work with the government of Kenya and will design and operate the new East Africa cloud region as part of a ‘trusted data zone’ based on global standards to protect digital safety, privacy and security. With technical assistance and support from G42 and Microsoft, Kenya will establish the new data center . . . under which data from other countries may be governed by their local laws, even while stored and resident in Kenya.

Kenya will utilize the new data center and cloud services for governmental and citizen services.”

Both G42 and Microsoft executives told the authors that they see this project as the first step to expanding services to many other governments with similar sovereign needs as the UAE.

Microsoft executives expressed significant enthusiasm about what this partnership with G42 could mean for increasing U.S. influence in middle and emerging powers vis-à-vis China. Though they acknowledged the U.S. national security concerns raised by the Microsoft-G42 deal, they argued that in its absence, Chinese technology companies would have snapped up the opportunity to provide AI infrastructure to the UAE. More broadly, the partnership offers an opportunity to rebuff Chinese technology companies’ dominance of critical technology infrastructure in Africa. In recent years, for instance, leading Chinese tech firms like Huawei and ZTE Corporation have developed AI-enabled “smart city” infrastructure and surveillance systems in countries across the African continent, including Kenya. “If the U.S. has the chance to win over Africa [against China], it’s because of what we’re doing here,” one Microsoft executive said.

On January 3, 2025, Microsoft President and Vice Chair Brad Smith called the “promotion of American AI exports” a “critical priority” for U.S. policymakers and companies in the next year. “The rapid development of China’s AI sector has heightened competition between American and Chinese AI, with much of this likely to play out during the next four years in international markets around the world,” Smith wrote. Moreover:

“While the U.S. government rightly has focused on protecting sensitive AI components in secure datacenters through export controls, an even more important element of this competition will involve a race between the United States and China to spread their respective technologies to other countries. Given the nature of technology markets and their potential network effects, this race between the U.S. and China for international influence likely will be won by the fastest first mover. Hence, the United States needs a smart international strategy to rapidly support American AI around the world.”

Multiple Microsoft, G42, and UAE government representatives stressed that the UAE has the diplomatic clout and available capital to become a “gateway to the Global South” for U.S. companies. With closer ties to developing countries than the United States has and enormous capital reserves, the UAE government and companies like G42 could effectively “derisk” U.S. companies’ entry into new markets, they argue.

The UAE committed $97 billion in foreign investment to Africa in 2022 and 2023—three times the amount of Chinese investment in the continent over the same period. It also became a BRICS+ member in 2024, joining a club of emerging economies anticipated to represent 45 percent of global GDP in purchasing power terms by 2040. According to the World Trade Organization, BRICS+ members already represent approximately 25 percent of global exports, with the UAE second only to China (and 14th globally) as the largest exporter among BRICS+ countries at 2.1 percent of the global share.

6. G42’s ambitions extend far beyond its Microsoft partnership. It desires to build 10 to 100 times more data center capacity than currently outlined in its agreement with Microsoft.

At the GITEX Global Expo in October 2024, G42’s Xiao presented the concept of an AI “intelligence grid” similar to today’s electric utility grids. He said:

“As intelligence, through the form of artificial intelligence, becomes a utility, it becomes as important, if not more important than electricity. . . . That means we have to build the intelligence grid—the equivalent of [an] electricity grid; the data centers, connectivity, the computing infrastructure for inferencing, to be able to pipe that intelligence-as-utility to everyone who needs it.”

Xiao estimates that the world today has less than 60 gigawatts of data centers but will need between 300 and 500 gigawatts to spread AI across the globe, with each gigawatt of data center and energy infrastructure costing $45 billion. According to studies of AI training and infrastructure costs and the authors’ calculations, AI chips account for nearly half of this total cost at approximately $20–$24 billion per gigawatt. Xiao further said, “That’s a type of infrastructure we’ve got to build for the world to be able to take advantage of AI and to benefit all humanity. And for me, that’s the greatest opportunity.”

G42’s deal with Microsoft to build two 30-megawatt and one 100-megawatt data center in the UAE is only the beginning of a much bigger project to spread “intelligence-as-utility” worldwide. In a private meeting, one senior G42 executive said the company wants to build gigawatt-scale data centers because “we think this [ability to rapidly and cheaply build large energy and data center infrastructure] could be our competitive moat.” Specifically, G42 and Khazna’s speed and cost-effectiveness of construction and the UAE’s abundance of energy are great advantages in domains where the United States and the rest of the world are struggling to compete.

Khazna reportedly already has 360 megawatts of data center capacity spread over 24 centers in the UAE, with another eight under construction. Khazna’s Alnaqbi has publicly stated that “based on the demand we have . . . we expect the UAE to ramp up and accelerate demand to at least 850 megawatts by 2029” (see point 5.1). However, private discussions with G42 executives indicate that G42’s ambitions to build out AI infrastructure—with or without Microsoft as their partner—far exceed this number:

• By the end of 2025, G42 plans to bring roughly 100 megawatts of data center capacity into operation, including the two 30-megawatt centers built for Microsoft.

• By the end of 2026, G42 plans to bring hundreds of megawatts of capacity into operation, including the AI-optimized 100-megawatt center in Ajman built for Microsoft.

• By the end of 2027, G42 plans to reach multiple gigawatts of data center capacity, including from data centers outside the UAE, such as the 1-gigawatt facility in Kenya.

Top G42 executives told the authors that they think they will be able to generate around 5 gigawatts of data centers within the UAE, most of which will be used for AI deployment. They did not provide a timeframe for building out this capacity but implied that Microsoft may not be their only partner if it cannot fill these data centers quickly enough once they are constructed. They did not suggest which other companies G42 might pick as additional or alternative partners to Microsoft.

7. According to G42 and Microsoft representatives, security controls at the G42 data center align with U.S. commercial standards, though the adequacy of such controls for protecting strategic AI capabilities remains uncertain.

While in the UAE, the authors toured a 21-megawatt data center built by G42 and Khazna for Microsoft. Security controls installed in this center to address cybersecurity, insider threat, and break-in risks include NIST SP 800-53 controls, over 100 SP 800-53 security and privacy controls for information systems and organizations, and Department of Commerce export license requirements. The authors of this paper are not in a position to fully verify G42’s claims about its broader security controls, though the implementation of these commercial security standards appeared consistent with typical industry practice and is plausible.

A key question, however, is whether these measures are wholly sufficient. Standard data center security measures typically address privacy and cybercrime issues rather than full-fledged attacks from nation-states like China. As AI capabilities advance and the stakes increase, U.S. and UAE commercial security standards may need significant enhancement to address evolving threats from state actors. A RAND analysis indicates that securing AI model weights and related intellectual property requires significantly higher security standards than what current AI companies or cloud providers—whether in the United States or the UAE—are currently achieving. While cybersecurity measures are currently a primary focus due to the risks of deploying model weights abroad and potential exploitation by adversaries, compute providers could play a much broader role in ensuring secure AI development.

Furthermore, cybersecurity represents only one dimension of the challenge. Today’s risks extend beyond traditional export control concerns about the physical location and ownership of AI chips. Critical considerations include how compute resources are being used, who can access them remotely, and how resulting AI models and weights might be shared or deployed. Even when AI chips are physically secure within a facility, their computational power can be accessed remotely, and trained models can be distributed widely—potentially to unauthorized users. This requires robust know-your-customer procedures and comprehensive monitoring of compute usage patterns, capabilities that current security frameworks may not adequately address.

If the United States deems the current controls inadequate for protecting U.S. technology in Emirati data centers or elsewhere, it should clarify what additional controls may be required and where.

8. The United States faces two strategic questions regarding UAE AI development: first, whether to continue to support the near-term G42-Microsoft partnership plan, and second, how to approach the UAE’s broader AI ambitions.

The United States can support Microsoft-led projects in the UAE while remaining cautious of endorsing the UAE’s broader AI ambitions. If Microsoft’s claim that cloud infrastructure can create significant lock-in and network effects for its cloud services is true, then supporting the Microsoft-G42 deal could strengthen U.S. technological leadership and make switching to alternative providers (like Huawei) increasingly costly. This “stickiness” of cloud platforms could help cement U.S. dominance in global AI infrastructure while providing a model for future partnerships with other emerging technology powers. The new U.S. Framework for AI Diffusion further reinforces this advantage by allowing U.S. companies like Microsoft to obtain Universal Validated End User status, enabling them to deploy data centers globally with a single license, whereas local companies in the UAE would require an authorization which is country specific and subject to more stringent caps. As Chinese technology firms continue to strengthen their presence in neighboring countries like Saudi Arabia—which is increasing joint ventures with companies like Alibaba and SenseTime, and recently invested $400 million in AI startup Zhipu AI—developing such a model for U.S. foreign engagement on AI is more important than ever.

However, the United States should also bear in mind that today’s cooperation could evolve into tomorrow’s competition. At the GITEX Global Expo, for instance, UAE Adviser to the President and Secretary-General of the Advanced Technology Research Council Faisal Al Bannai said:

“I think the opportunity we have, as UAE, and as the stakeholders here in the country [is]: One, you can leverage all the good technologies that are coming out, let’s say from the U.S., right? The different partnerships that are happening with G42, Microsoft, [and] a number of other entities. But at the same time, we are also building our own AI platforms—something that we control, we own, and we can take progress from.”

The UAE has already made a significant effort to develop indigenous LLMs such as G42’s Arabic-English LLM Jais and the government-backed Technology Innovation Institute’s Falcon 2, which the institute claims outperforms Meta’s Llama 3 model. Private conversations with UAE technology and business professionals revealed that, in reality, these Emirati models perform poorly compared to U.S. models. They said Emirati models are not reliable nor high performing enough for many industry use cases, and that they currently use U.S. options in their work. With increasing access to U.S. compute and talent, however, the United States should think through a scenario in which indigenous Emirati models—controlled exclusively by a nondemocratic government with significant ties to China—become real global competitors. If the next world order is indeed determined by who wields the best AI systems, then the decision to equip other countries with the fundamental ingredients to develop this technology (i.e., chips) should not be taken lightly. Even if the UAE remains faithful to its promise to exclusively partner with the United States in AI, it has said little so far about what a world in which it has outgrown this partnership might look like.

Conclusion: Proceed with Caution

U.S. leadership of the global AI ecosystem is at a critical turning point. Recent developments in China’s AI models have caused several leading AI executives to outright reject the belief that the United States still holds an 18-month lead on China in AI development. At the same time, Chinese technology firms’ investment and infrastructure projects in emerging economies threaten to lock the United States out of economically and geopolitically strategic markets in the Global South. However, China’s ability to produce advanced AI chips at the scale needed to support broad AI deployment could remain significantly constrained by U.S. export controls, limiting their capacity to backfill U.S. technology.

While renewing U.S. industrial policy and onshoring supply chains is a frequent policy priority in Washington, these policies should not be pursued at the expense of expanding a U.S.-led AI ecosystem abroad. Similarly, rigid restrictions on the deployment of U.S. AI technologies exclusively in the United States and its allies will do little to further U.S. influence in the many “swing states” that are unlikely to pick sides in the U.S.-China competition.

Rather, reaffirming the United States’ leadership in AI requires a comprehensive strategy: ensuring emerging powers choose U.S. technology to realize their AI ambitions while establishing robust frameworks for secure and responsible deployment. The traditional approach of simply selling technology without ongoing oversight is insufficient given the unique characteristics of AI systems—their remote accessibility, the sensitivity of model weights, and the potential for unauthorized sharing or deployment.

There will always be inherent risks to offshoring AI to countries outside the United States, let alone to nondemocratic ones with strong ties to China. Thus far, the United States’ approval of relatively modest clusters of advanced AI chips to G42 indicates that the potential benefits—both commercial and strategic—outweigh the risks. However, as the U.S.-UAE partnership evolves, U.S. policymakers and AI executives should maintain a healthy realism about the UAE’s vested interests in hedging its bets between the United States and China. They should have no illusions about the nature of the Emirati political regime, nor about some of the uses to which the UAE is likely to apply its AI technologies. They should ask tough questions, such as what the UAE government is doing to ensure decoupling from China in AI by tech companies other than G42, and scrutinize the UAE’s claims about these and other security measures. If they do not think that the current security controls for AI chips are sufficient to protect the United States’ advantage, they should articulate clearly and consistently what those controls should be. For now, however, proceed with caution.

---

Gregory C. Allen is the director of the Wadhwani AI Center at the Center for Strategic and International Studies in Washington, D.C.

Georgia Adamson is a research associate for the Wadhwani AI Center.

Lennart Heim is an associate information scientist at RAND and professor of policy analysis at the Pardee RAND Graduate School.

Sam Winter-Levy is a fellow for technology and international affairs at the Carnegie Endowment for International Peace’s Technology and International Affairs Program.