The AI Race Beyond Chips

By David Yu, PhD, CFA, NYU Stern and Shanghai, DRI-WISC Affiliate

The AI race is framed as one for chips, but access to them only grants permission to build. Financing contracts determine whether the actual capacity gets built and who really controls it. Shaping the global diffusion of AI requires treating capital formation alongside export policy as dual levers of strategic influence, rather than viewing finance as an afterthought.

On July 10, 2026, the U.S. Department of Commerce moved the United Arab Emirates into a category called Country Group A:5 under the rules governing exports of sensitive American technology. Specifically, it allowed the UAE government and approved Emirati companies to buy AI chips and servers each time without an individual license, rather than opening the market to everyone.

However, the terms of that decision matter more than the headline. Commerce did not simply relax a restriction, it prominently linked easier technology access to the UAE’s commitments to invest in American AI infrastructure. Ultimately, the parties negotiated technology access and capital together.

The arrangement’s physical evidence was clear too, as under a bilateral framework (announced in May 2025), Stargate UAE paired a planned one-gigawatt computing cluster in Abu Dhabi with Emirati capital directed into Stargate facilities in the United States. Chips flowed one way, money the other, and each commitment depended on the other.

Yet public scrutiny remains focused on only half of the transaction. Export licensing generates congressional hearings, legal analyses, and sustained press attention, while financing is reduced to a dollar-figure headline. That overlooks two crucial distinctions: permission is not capacity, and ownership is not control. Both are ultimately settled through contracts, not export rules.

What a chip license actually does

After more than two decades of underwriting capital-intensive assets, legal permission almost never surfaces as a critical bottleneck. These pitfalls can be rewired across impediments to get around bottlenecks through some combination of legal, tax, risk, and jurisdictional maneuvers by legal and advisory “magicians.”

A license simply answers a binary yes or no—may this buyer purchase this equipment? But it leaves untouched building the electrical substation feeding the building (currently a severely constrained supply chain), moving a project up the years-long queue to connect to the power grid, or securing water for cooling. And it does nothing about the problem that keeps project financiers awake—debt remaining outstanding longer than the customer contracts supporting it.

Why financing has become the bottleneck

The scale of AI infrastructure explains why this now concerns a matter of national policy. The International Energy Agency expects data center electricity consumption worldwide to rise from some 485 terawatt hours in 2025 to about 950 by 2030, with demand from AI-focused facilities roughly tripling over that period.

Once a buildout exceeds what even the largest technology companies can fund, the decisive institution is no longer the only one that designs the chip. It is also the one willing to price the surrounding risks, from power supply and construction delay to currency exposure and obsolescence, and to carry them until the project is finished.

Why sovereign capital behaves differently

Gulf states have been building toward that capability. MGX, an Abu Dhabi investment vehicle, closed its first fund in July 2026 at $49 billion in commitments, above its $45 billion target, and weeks later joined the AI Infrastructure Partnership and BlackRock’s Global Infrastructure Partners in acquiring Aligned Data Centers at an enterprise value of roughly $40 billion.

The numbers are the least interesting part because private funds write checks that large too. What differs is the motive. A private fund answers to investors who expect capital returned on a defined schedule, which narrows its question to whether returns arrive quickly enough. A sovereign fund can pursue a financial return, a domestic industry, and a strategic relationship through one transaction, and can wait far longer for all three to mature. In infrastructure, where risks arrive early and returns late, that patience is the genuinely distinguished advantage.

Two projects—the same demand but different results

Indonesia shows what happens when financing is structured carefully. DBS and UOB, two Singaporean banks, arranged a 6.7 trillion rupiah ($378 million USD) financing facility with the Indonesia Investment Authority for a three-data center campus developed by DayOne, a Singapore-headquartered digital infrastructure platform. Denominating financing in rupiah rather than dollars reduced the project’s exposure to currency mismatch, the risk that a shift in exchange rates makes debt payments unaffordable relative to what a project earns. Choices of that kind often decide whether a project reaches financial close, when funding is legally committed and construction begins.

However, Kenya showed what happens when those conditions are not met. It announced that a $1 billion data center project involving Microsoft and G42 (an Abu Dhabi-based AI and technology conglomerate) stalled over capacity payment guarantees and the scale of power the facility would require. Officials described it as delayed rather than canceled, though nothing has been built.

It’s important to note that the gap isn’t in wealth or savings. The Africa Finance Corporation reported in 2026 that African institutions hold more than $2 trillion in domestic capital. Much of it is concentrated in low-risk assets, including government bonds, as relatively few infrastructure projects achieve the structure and credit quality that institutional investors require—across dependable power supply, credible customers, workable currency structures, and contracts that specify who bears each risk. At today’s levels of readiness, the IMF estimates that AI would add only about 0.2 percent to sub-Saharan African GDP over the coming decade, in contrast to closer to 4 percent if power, connectivity, skills, and institutions improve.

What the loan documents decide

The distinction between ownership and control applies directly to these facilities.

A data center’s building and electrical plant may last for decades, while the processors inside become economically obsolete much sooner as technology layers develop unpredictably. Drawing on my experience in aviation finance—where long-lived airframes routinely outlast rapidly evolving engine options and avionics packages—this mismatch between physical asset lifespans, technology cycles, customer contracts, and debt maturities is a familiar trap. If customer contracts expire before the debt matures, lenders inherit significant refinancing and re-leasing risk. More importantly, paying for a facility does not mean deciding what happens inside it. A 2026 paper now under academic review mapped 46 announced African AI projects worth $12.7 billion and found a recurring pattern of expanding physical investments, while control of the computing stays concentrated among a few global technology firms.

Hosting hardware is not the same as governing it. The relevant question about cross-border AI transactions isn’t who paid for the building, but who controls access to the capacity, who sets prices, who governs the data, whose operating standards apply, and what happens to those rights when the interests of host government, investor, operator, and technology supplier diverge? Those terms are settled in financing and operating agreements that are rarely public. Yet they decide who can afford to use the capacity, whose rules govern the data inside it, and whether a country that hosts a data center becomes a participant in the AI economy or is merely a landlord.

The three tests

If governments could apply the rigor they bring to export licensing to infrastructure finance, three questions would address most of what matters.

Who can use the compute, and can that use be checked? A government that cannot audit what runs inside a facility in its own territory has no way to enforce any commitment it was promised.

Can the project survive being wrong? Demand may arrive lower than the model assumed, power may cost more, and the processors may age faster than the repayment schedule. A structure that only works in the best case pushes the losses onto whichever party is least flexible to exit, usually the host country.

Who governs the data, the capacity, the operating standards, and the exit? Answer these badly and a country supplies the land, power, and water while the pricing, customer list, and data rules are set elsewhere.

Signing is only half of the process, as governments need to track what actually happens—the financing closing, how construction progresses, how many megawatts came online, and which customers receive the capacity. Announced investment totals only blur aspirations with money genuinely well spent. The gap is where the disappointment usually shows up.

Export policy determines where advanced computing may go. Capital formation determines whether it becomes durable capacity, who controls it, and what obligations stay. Having read the first closely, it’s time to instead scrutinize the loan documents.