Energy constraints and burdensome regulations are increasingly pushing AI companies to build data centres abroad. But while such deals are a boon for tech companies and countries in the Middle East, they could cause big problems back home.
In recent months, a flurry of deals has sent US officials scrambling, experts told Transformer. Abu Dhabi-based G42 is building data centres as part of a $1.5b investment deal with Microsoft. AWS has committed $5.3 billion to build at least 3 data centres in Saudi Arabia, which will be used for AI, among other things. A new $100 billion alliance to build data centres counts Abu Dhabi’s MGX as a member — with plans for facilities both stateside and overseas.
The list goes on. In September, American data centre company Equinix opened its fourth UAE facility. Last week, Google announced a partnership with Saudi Arabia’s Public Investment Fund, which will reportedly serve as the starting point for a $100b AI hub in the country. Sam Altman, meanwhile, has reportedly discussed building data centres with the UAE, only pivoting to domestic facilities when officials in Washington expressed concern about building this vital technology in the Middle East.
This interest in building data centres abroad is driven by a “power crunch” in the US, according to Konstantin Pilz, a researcher at the Special Competitive Studies Project specialising in semiconductors. As AI tech giants race to build the large data centres needed for ever more powerful AI, global power requirements for AI data centres are predicted to grow by 130 GW by 2030, according to a report from the Institute for Progress. But total American power generation is only likely to grow by 30 GW.
Tim Fist, one of the report’s authors, told Transformer that the US is likely able to build only a few 5 GW data centres by 2030, even if the government steps in to help. A Trump presidency may further exacerbate the problem: the proposed tariffs on foreign imports would sharply bump up the cost of renewable energy, which relies heavily on Chinese components.
Complicated regulations and environmental protections compound the problem, with recent attempts to secure power for data centres vetoed by regulators and special interest groups. Just this week, regulators vetoed an Amazon deal to expand a data centre with power from a nearby nuclear power plant, while Meta reportedly gave up on building a new facility because a rare species of bee was discovered living nearby.
Overseas, these problems disappear. Unlike the US, Gulf states can build power capacity quickly. Countries such as the UAE and Saudi Arabia, Fist said, have a track record of building new plants fast. “It’s likely that they're going to be very good at building the next ten plants [too],” he said, thanks to their expertise, supply chains, abundant natural resources, and lighter permitting regimes. Adding to the region’s appeal is easy access to capital, with governments eager to invest in new industries to reduce their dependence on oil revenues. “There's a lot of money just waiting there,” Pilz said.
But though the deals might help satisfy AI companies’ ever-growing demand for compute, they could cause big problems for the US government — particularly when it comes to securing America’s lead in the AI race. Experts fear that if models are trained abroad, it will be much easier for foreign adversaries such as China to steal them. “We should expect it to be significantly harder to adequately secure model weights if they build abroad,” Yusuf Mahmood, a research fellow at RAND, told Transformer. Pilz concurred, saying that it would be “a lot easier” for China to steal models if they’re trained abroad. Among other reasons, he noted that the US would have more limited intelligence capabilities abroad to investigate and prevent theft.
Theft might not even be necessary. The recipient countries could just smuggle chips on to China. And “even if the chips remain in these Middle Eastern countries, they could just provide either cloud access to China,” Pilz said, “or even just train AI models and send them to China.”
Beyond just securing the models and chips, the country housing compute has more influence on AI policy, including copyright policy, safety regulations, and controlling who can use the chips. “The more compute you have outside of the country, the less impact you have on if people are using this compute responsibly,” Pilz said.
The US has one tool to maintain its authority: export controls. Almost all AI chips are designed and sold by US companies, so the US can block their export. It has shown a willingness to do so: as of mid-September, the Financial Times reported that Microsoft still had not received licenses to ship some of the necessary components for its deal with G42. Bloomberg, meanwhile, reported that the Biden administration was considering capping AI chip exports to the Middle East in the interest of national security, though it’s unclear if the Trump administration will take the same stance.
But preventing Middle Eastern countries from buying AI chips altogether could be a mistake. As China vies for influence in the Middle East and elsewhere, the US risks alienating its allies: blocking exports might push Arab countries into closer alliances with China, weakening US interests in the region. And as AI companies need compute, stopping them from building it in the Middle East might just slow down US AI progress altogether.
One way to square the circle is to offer compute — but with conditions. Informally, that already seems to be happening. Seemingly in an attempt to assuage US concerns, G42 has said it is cutting ties to Chinese suppliers, and that it will not do business with any company on the US export controls list. Saudi Arabia, meanwhile, is reportedly taking steps to limit its involvement with Chinese firms in order to satisfy US security concerns.
Going forward, the US may expand this “conditional compute sharing” into a formal program. The Commerce Department recently announced an expansion of the Validated End User program to include data centres. The program should make it easier for approved foreign data centres to get access to American chips — but it requires them to comply with certain security standards upfront. The US government may also require things like on-site reviews, new reporting, or security commitments from the host country's government.
According to Fist, such an approach could allow “the United States to pursue a kind of like chips-for-peace-style, multilateral governance regime for AI”. The devil, though, is in the details: whether the program will actually mitigate the risks of offshore data centres depends on what conditions are implemented and how well those conditions are enforced.
Disclosure: The author’s partner works at Google DeepMind.