Gulf AI Compute: How Washington Trades Chips for Alignment

On 10 July 2026 the US Commerce Department’s Bureau of Industry and Security (BIS) gave the United Arab Emirates a substantial upgrade under American export controls. The UAE left Country Groups D:3 and D:4 for Country Group A:5, and its government agencies, along with two local AI companies, G42 and Core42, may now receive advanced AI chips from Nvidia and others without applying for a licence each time. A week later G42 welcomed the decision as an important milestone in the US–UAE technology partnership.

The rule’s new supplement also carries a deadline. Unless BIS says otherwise, the authorisation for G42 and Core42 expires automatically on 6 April 2027. The rule’s preamble is blunter: if the two companies have not become US companies by then, they will have to reapply to keep their approved status. The eight American firms on the same list (Amazon, Apple, Google, Meta, Microsoft, OpenAI, Oracle and xAI) and their subsidiaries carry no expiry date.

Saudi Arabia is not in the rule at all. Its state-backed AI company, Humain, was cleared in November 2025 to buy the equivalent of up to 35,000 Nvidia GB300 chips, subject to rigorous security and reporting requirements. The kingdom remains in Country Groups D:3 and D:4, so advanced chips still need a licence.

The gap between the two shows how Washington now uses chip controls. Export controls were built to keep technology away from adversaries; in the Gulf they have become a bargaining chip. Washington trades access to chips for a partner’s position between America and China: cutting AI ties with China, leaving American firms in charge of the compute, pledging matching investment in the United States, and accepting that access can be withdrawn.

Clearing the chips is only the first step. They have to go into data centres measured in gigawatts, and power, grids and this year’s war with Iran may prove harder to overcome than licensing. For Taiwan, the deal is also a mirror: TSMC makes the chips, but Washington decides where they go.

Sovereign AI, built on someone else’s chips

When Gulf governments talk about sovereign AI, they mean keeping compute, data and models at home and under their own control. The UAE started earlier. G42 is chaired by Sheikh Tahnoon bin Zayed Al Nahyan, the president’s brother and the country’s national security adviser. In May 2025 G42 announced Stargate UAE with OpenAI, Oracle, Nvidia, Cisco and SoftBank: a 1 GW cluster in Abu Dhabi that G42 is building and OpenAI and Oracle will operate, with the first 200 MW expected to go live in 2026.

Stargate UAE is only the first phase of the UAE–US AI Campus, a 10-square-mile (26 km²) site planned for 5 GW, the largest of its kind outside the United States. Separately, Microsoft is adding 200 MW of data-centre capacity in the UAE through Khazna Data Centers, a G42 subsidiary, due to start coming online before the end of 2026.

Saudi Arabia’s champion is Humain, launched in May 2025 by the Public Investment Fund and chaired by Crown Prince Mohammed bin Salman. It plans AI factories of up to 500 MW, running on several hundred thousand of Nvidia’s most advanced GPUs over five years, starting with 18,000 GB300s, and it is aiming for 1.9 GW of data-centre capacity by 2030 and 6 GW by 2034. Tareq Amin, its chief executive, has been explicit about the ambition: to be “the third-largest AI provider in the world, behind the United States and China”.

All of this is done in the name of sovereignty, yet the chips, the software and the export licences come from America. What Gulf states control is where the data centres sit, where the data is stored and who runs them; whether the chips arrive, and who may use them once they do, is settled by rules written in Washington. Under the 2025 US–UAE agreement, the campus’s compute is reserved for US hyperscalers and approved cloud providers, and the two governments are to tighten know-your-customer checks on who gets access.

Chips work as leverage because they are the easiest layer of AI to control: physical goods that cross customs and come from a handful of producers. How the layers differ is covered in the explainer What Do AI Controls Actually Control — Chips, Models, or Uses?

From blocking rivals to setting terms

Washington’s doubts about G42 go back years. According to The New York Times, when Sheikh Tahnoon visited the White House in June 2023, Jake Sullivan, then the US national security adviser, raised Washington’s concern that G42 had concealed the extent of its work with China. That December G42’s chief executive, Peng Xiao, told the Financial Times that the company would phase out hardware purchases from Huawei. “We cannot work with both sides,” he said.

In April 2024 Microsoft invested $1.5 billion in G42, its president, Brad Smith, joined G42’s board, and the two companies signed an intergovernmental assurance agreement developed with both governments. Tarun Chhabra, a White House technology adviser, later described the deal as, on balance, a positive development, because it gave G42 an alternative to Huawei.

The turning point came in May 2025. The Trump administration rescinded the Biden administration’s AI Diffusion Rule, whose compliance requirements were due to take effect on 15 May. That rule sorted countries into tiers: 18 destinations, Taiwan among them, could receive advanced chips without a licence, while most other countries were placed in a second tier with allocations. Neither the UAE nor Saudi Arabia was in the top tier. In its place came country-by-country deals.

Donald Trump toured the Gulf that same week. Humain was launched and signed with Nvidia; the US and the UAE agreed a US–UAE AI Acceleration Partnership and unveiled the 5 GW campus. Commerce Secretary Howard Lutnick said the agreement “contains strong security guarantees to prevent diversion of U.S. technology.” BIS also issued guidance warning industry about the risks of using Chinese advanced computing chips, naming specific Huawei Ascend chips.

Approvals then came in batches. In September 2025 Microsoft secured licences to ship the equivalent of 60,400 more A100 chips, in the form of GB300s. On 19 November the Commerce Department cleared G42 and Humain to buy the equivalent of up to 35,000 GB300s each, conditioned on “rigorous security and reporting requirements” that BIS would monitor. An analysis by the Arab Gulf States Institute in Washington asked at the time whether the allotment was the start of a longer export schedule, or a smaller gesture meant to preserve American leverage.

The UAE signed the US-led Pax Silica declaration in January 2026, the ninth country to do so (on the alliance itself, see The Price of the Silicon Alliance). In February G42 announced plans for an assured-compute framework that would continuously verify the location and physical control of its chips and track their use cryptographically. How the chips are used once they arrive had become part of the deal.

The July rule wrote these terms into regulation. BIS said the upgrade recognised the UAE’s status as a Major Defense Partner and its support for US national security interests, including Operation Epic Fury, the American campaign against Iran that began in February 2026, and that the UAE had reaffirmed its commitment to matching investment in American AI infrastructure. Reuters has reported that, in return for chip access, the UAE agreed to cut AI ties with China. BIS also said it would review licence applications involving MGX, an Abu Dhabi AI investment firm, favourably.

A narrower upgrade than the label suggests

Country Group A:5 is the group eligible for the Strategic Trade Authorization licence exception; most of its members are NATO and EU states, along with partners such as Japan, South Korea, Australia and India. As a member, the UAE’s government and approved companies can now receive certain military items, commercial satellites, and dual-use items for oil and gas, desalination and civil nuclear power without individual licences.

For chips, the list still rules. Advanced computing items bound for the UAE still need a licence unless the recipient is an approved end user named in the new supplement. Government agencies qualify, but not state-owned companies or government contractors; among local firms only G42 and Core42 are listed, and their approval expires. Washington is extending trust to named companies, not to the country. How end users are judged under export controls is covered in the explainer If You Do Not Operate in America, Why Do Its Export Controls Still Reach You?

That explains a Bloomberg report in September that G42 executives had held exploratory talks about selling a majority stake to American companies, or setting up a new vehicle in the US, so that the company could keep buying chips beyond next year; no decision had been made. G42’s investors already include Microsoft, the private-equity firm Silver Lake and Mubadala, an Abu Dhabi sovereign fund. If it comes to that, the UAE would gain a secure supply of chips and give up control of G42.

Saudi Arabia’s position is different. It has not signed Pax Silica, and its advanced chips still need licences. In September, at the LEAP 2026 technology conference, Saudi Energy signed three agreements. In one, its transmission arm, National Grid SA, will supply power to Humain’s AI data-centre project in Riyadh; in another, its project development subsidiary signed a memorandum of understanding with Huawei on data and AI centre projects.

That does not mean Riyadh has chosen Beijing. In August Saudi Arabia signed a joint defence pact with Turkey and Pakistan, another way of adding relationships beyond the United States (see Saudi Arabia, Turkey, and Pakistan Align). It is classic middle-power hedging. On chips, the price appears to be access that comes later, and in smaller amounts, than Abu Dhabi’s.

Chips can be cleared by a rule; power plants cannot

An export licence can change in a day; power stations and grid connections take years. That gap may be the Gulf’s real constraint.

Start with the UAE. The UAE–US AI Campus is planned at 5 GW. For comparison, Barakah, the country’s only nuclear power station, has four 1.4 GW reactors, 5.6 GW in all; since all four entered commercial operation in 2024 it has generated about 40 TWh a year, which its owner, the Emirates Nuclear Energy Corporation (ENEC), says meets a quarter of the UAE’s electricity needs. Wood Mackenzie, an energy consultancy, estimates that all the UAE’s data centres together used about 3 TWh in 2025, roughly 2% of national demand of 173 TWh, and will pass 6 TWh by 2030.

By our rough estimate, a 5 GW campus running flat out all year would draw about 43.8 TWh, more than Barakah generates and a quarter of the UAE’s 2025 demand. Stargate UAE’s 1 GW alone would use about 8.8 TWh, nearly three times what every data centre in the country consumed in 2025. Real loads will be lower than full capacity, but the gap is too wide to ignore. Either the campus will be built far more slowly than announced, or the power supply will have to grow much faster than current forecasts.

The UAE is not standing still. In December 2025 TAQA, Abu Dhabi’s national energy company, and the Emirates Water and Electricity Company (EWEC) reached financial close on a 1 GW gas-fired plant at Al Dhafra that will supply data centres and is already under construction. Masdar is building a solar-and-battery system designed to deliver 1 GW around the clock, also for AI data centres. As AFP has noted, oil, gas and year-round sun let the UAE build power plants quickly, which elsewhere is often the biggest bottleneck for data centres.

Saudi Arabia’s gap is wider. According to the Ministry of Communications and Information Technology, the kingdom’s operating data-centre capacity rose from 68 MW in 2021 to 440 MW in 2025, and stood at 467 MW in the first quarter of 2026. Humain’s target of 1.9 GW by 2030 alone is about four times today’s national total; its 6 GW target for 2034 is nearly 13 times.

Humain is making progress. In August AMD, Cisco and Humain said that infrastructure built on AMD’s MI355X GPUs was live in Saudi Arabia, with a next phase of up to 250 MW to begin deployment in 2027. In September National Grid SA agreed to supply Humain’s Riyadh data-centre project, though the announcement gave no capacity or timetable.

The war has made data centres targets

Beyond power there is security. On 28 February 2026 the United States and Israel launched strikes on Iran (the American operation was named Epic Fury), and Iran retaliated against Gulf states that host American forces. On 1 March drones struck two Amazon Web Services facilities in the UAE and damaged a third site in Bahrain, knocking two of the three availability zones in AWS’s UAE region offline. By September AWS said it could not restore data held only in one of those zones, and that most customers had re-established their operations in other regions.

In April Iran’s armed forces released a video naming Stargate UAE as a potential target, with a map purporting to show where it is being built. In September Reuters, citing six people familiar with the matter, reported that the UAE was revising its 5 GW plans: the single campus in Abu Dhabi will likely become a network of data centres spread across the country, possibly with air defences and some facilities underground. Two of the sources placed the site near Al Dhafra Air Base, which hosts US troops. G42 said work was progressing as planned and that “the project’s specifics are subject to continuous review in line with the security, resilience and operational standards expected of critical infrastructure at this scale.”

There is a less obvious tension here. BIS cited the UAE’s support during Operation Epic Fury as a reason for the upgrade, and Iran’s retaliation was aimed at Gulf states hosting US forces and at the facilities of American tech companies. The alignment that won the UAE its chips has also made the places that house them targets. Dispersal, hardening and air defence all cost money and time. Daniel Benaim, who was US deputy assistant secretary of state for Arabian Peninsula affairs under President Biden, has called the conflict an “economic earthquake” for Gulf states that long marketed themselves as safe havens.

When Microsoft announced in September that it would invest more than $10 billion in the region by 2030, it wrote: “Conflict in the Middle East has reinforced the connection between digital resilience and digital sovereignty.” In the Gulf, the cost of sovereign AI now includes the cost of protecting it.

Taiwan makes the chips but does not decide where they go

Much of the hardware in this bargain runs through Taiwan. Nvidia’s Blackwell GPUs are made on TSMC’s custom 4NP process, an advanced node, and more than half of the system partners Nvidia lists for GB300 are Taiwanese: Foxconn, Quanta’s QCT, Wistron, Wiwynn, Inventec, Pegatron, Gigabyte, Asus and ASRock Rack. When Nvidia announced a supply-chain expansion in October, the substrate maker Unimicron and the packaging and testing firms SPIL and KYEC were on the list. “Demand is extraordinary. Capacity is the constraint,” said Jensen Huang, Nvidia’s chief executive.

Whether those chips can go to Abu Dhabi or Riyadh, however, is decided in Washington. Nvidia’s chips embody American technology and fall under the extraterritorial reach of US export controls, so a Taiwanese firm shipping servers to the Gulf has to check whether the buyer is approved and whether a licence is needed. How much it sells depends, in the end, on how much Washington clears.

Taiwan sits inside the same system. In September 2025 the US told TSMC that the Validated End-User status of its Nanjing fab would be revoked at the end of the year, after which suppliers would need individual licences to ship controlled equipment there; Taiwan’s economics ministry said the move would affect the predictability of the plant’s operations. What Washington took back from TSMC is the same kind of company-level privilege it has now granted G42.

Taiwan cooperates, and it pays. In June 2025 its economics ministry added Huawei and SMIC to its entity list for strategic high-tech exports. Under the US–Taiwan deal of January 2026, Taiwanese companies pledged at least $250 billion of new direct investment in the US, alongside at least $250 billion of credit guarantees from Taiwan. The Gulf trades alignment and investment for chips; Taiwan pays in the same currency.

Formal labels can mislead. In the US country groups, Taiwan sits in A:6 with Israel and Singapore while the UAE is now in A:5, yet buyers in Taiwan need no licence for these chips at all unless they are headquartered in China or Macau, whereas recipients in the UAE must be named one by one. Pax Silica is similar: the UAE is a signatory, while Taiwan takes part as a non-signatory participant, having endorsed the declaration’s principles through a joint statement with the US in January. What a label says and what it gets you are different things.

Taiwan has its own sovereign-AI effort. In 2023 the National Science and Technology Council launched TAIDE, which retrains Meta’s open Llama models on traditional-Chinese and Taiwan-specific data. The National Center for High-performance Computing opened a cloud computing centre in the Southern Taiwan Science Park in December 2025 with 15 MW of power, and plans a 120 MW centre at Shalun for 2029. At the opening, President Lai Ching-te said that computing power is national power, and that Taiwan cannot remain merely a supplier of hardware.

The scale is of a different order. Shalun’s 120 MW is less than a fortieth of the UAE’s planned 5 GW campus and less than two-thirds of Stargate UAE’s first block. Power is tight in Taiwan too: since 2023 Taipower has declined to supply new data centres of more than 5 MW north of Taoyuan, and only in July 2026 did it say the limit could be relaxed for dedicated zones near power plants.

For Taiwan, “chips for alignment” means that leverage does not automatically sit with whoever makes the chips. Taiwan’s silicon shield rests on manufacturing that is hard to replace, but the flow of chips is decided by whoever writes the rules. What Taiwan can do is be the most trustworthy manufacturing and compliance partner, checking Gulf customers thoroughly so that Washington has no reason to tighten, and spend its limited power and money on what must stay under its own control (language models, data and critical applications) rather than chase gigawatt-scale compute.

Three things to watch

First, whether G42 and Core42 restructure as US companies before 6 April 2027, or BIS extends their approval. Second, whether Saudi Arabia wins list-based treatment of its own or keeps applying licence by licence. Third, whether Stargate UAE’s first 200 MW and Humain’s Riyadh campus come online on schedule despite the limits set by power and war.

Chips have already bought Washington alignment in the Gulf. Whether that alignment turns into working compute will be decided by power plants, grids and air defences.