The administration of U.S. President Donald Trump has signed multiple agreements giving Middle Eastern countries greater access to cutting-edge U.S. semiconductors and allowing them to build state-of-the-art AI data centers in the region. Bringing these countries closer to the U.S. technology ecosystem has been a long-standing goal of U.S. foreign policy, and access to chips gives Washington leverage to make it happen. Smart deals have the potential to protect U.S. technology, expand foreign markets for U.S. companies, and build economic and political partnerships with geopolitical swing states.

But there are also significant risks to these deals if they’re rushed through. Ill-judged agreements that offer up huge quantities of chips could offshore a growing U.S. industry and hand control over a strategic technology to countries whose interests only partially align with those of the United States. The administration needs to promote U.S. industry without giving away control over frontier AI development or enabling technology diversion to China.

So, what should observers look for to assess the merits of an AI deal? Many of the specifics will likely not be disclosed, but deals—good or bad—will carry key indicators. Some of these will likely be revealed by public reporting, while others will be available to those in Congress and the executive branch. Some may even be left undecided in initial deals and left to later implementation decisions. Important indicators include chip numbers, key beneficiaries, security measures, and concessions. 

What Are the Chip Export Limits?

AI companies are planning dramatic increases in the computing power used to train their most advanced systems over the next few years. The largest AI data center in the world, built by xAI, reportedly has around 200,000 Nvidia H100 GPUs, a powerful AI chip, and the company plans to amass as many as 1 million of Nvidia’s next generation GB200 chips at the site. (Each GB200 has the performance of about 2.5 H100s.) Other companies, including OpenAI’s Stargate joint venture, have similar plans to scale up to millions of chips. Meanwhile, the world’s fastest traditional supercomputer, El Capitan, at the U.S. government’s Lawrence Livermore National Laboratory, runs advanced nuclear simulations on 44,000 of the latest generation AMD chips.

The AI diffusion rule proposed by former president Joe Biden’s administration and recently rescinded by the Trump administration would have allowed a foreign company that met stringent security standards to buy a level of computing power equivalent to 320,000 H100 chips by 2027. This quantity would likely have been about twelve months, or one generation, behind the frontier by that time. In addition, trusted U.S. companies could have each installed up to 7 percent of their computing base in any one of a broad group of countries, including the Gulf states. (For the largest companies, this would likely have been more than 320,000 H100-equivalents by 2027.) In total, that could have allowed many hundreds of thousands of chips to flow to the Gulf, although they would have been split between several companies.

With those benchmarks in mind, a deal that greenlights the export of tens of thousands of chips over the next two to three years would present a manageable level of risk, if the primary concern is keeping frontier AI development in the United States. This number of chips would still be enough for some concerning activities, such as nuclear weapons modeling and sophisticated intelligence analysis, but restricting chips at this level is unlikely to be practical, given the widespread ability to remotely access large clusters of this computing power. The first phase of the reported deal between the United States and Saudi Arabia will allow the new state-backed Saudi AI company Humain to buy 18,000 of the latest Nvidia chips. That level is enough to allow significant commercial AI development and deployment, but not enough to threaten the U.S. lead in frontier AI. And many of the commercial benefits of AI development will come not from possession of the physical chips, but from the ability to develop consumer apps and business products using cloud services elsewhere. 

On the other hand, a deal that allows Saudi or Emirati companies to amass more than around 400,000 H100-equivalent chips over the next two years would put those companies on track to build frontier-scale data centers in the near future. And a deal that allows the export of the equivalent of more than a million H100s by 2027 to a single company—like the deal the administration is reportedly considering with the Emirati company G42—could be enough to build the world’s largest AI training cluster. (Reporting on the Saudi deal is less clear, but company announcements suggest chip sales in the hundreds of thousands over the next five years.)

Who Are the Main Beneficiaries?

Whether the chips go to a single—and likely state-backed—Gulf company or are split among several competitors also matters. The latter is less likely to result in the offshoring of frontier AI development and more likely to promote a healthy global AI ecosystem that uses U.S. technology to bring commercial products to businesses and consumers around the world. Of course, there is always a risk that states could pool their chips if they wanted to build one massive model, which is why the overall limits matter. But nationalizing and aggregating existing data centers would not be straightforward, and dividing chips across companies will mitigate the risk.

If a few companies are singled out for big shipments, the administration should consider making them joint ventures between American companies or subsidiaries of the big U.S. players, as a better bet for U.S. security. If U.S. officials know that a U.S. company is ultimately on the hook, they should be less worried that sensitive technology will be diverted to adversaries or used to surveil American citizens. Major governance issues will remain, including which country has ultimate authority over the development and use of AI systems in these data centers, and the risks of misuse won’t go away entirely, since U.S. companies have limited visibility into how their customers use their chips. But the U.S. government will have more tools available to manage them. Giving sole control over large numbers of chips to a Gulf company, by contrast, would limit the ability of the U.S. government to ensure that the technology isn’t used in malicious ways or diverted to adversaries down the line.

What Security Measures Are in Place?

The core of any deal should include measures to protect U.S. technology and intellectual property from being stolen or misused. Strong protections, which should be backed up by legal guarantees from the host government and robust verification measures, could include:

• Requirements that companies receiving large quantities of chips cut any financial or ownership connections to people or entities in China, Russia, and other U.S. adversaries, and strip out any technology they currently use that is sourced from those countries. For example, companies building major data centers with U.S. chips shouldn’t be part-owned by Chinese investment firms or using networking technology from blacklisted companies such as Huawei or ZTE.
• Requirements that large data centers developing or deploying advanced AI models follow the National Security Agency’s cybersecurity guidance on deploying AI systems securely and meet standards equivalent to the security level the federal government uses for contractors handling unclassified data.
• Restrictions on Chinese customers using exported chips remotely to train advanced AI models, as well as for military or intelligence purposes. Some defense hawks might push to bar all access by arms-embargoed countries, but keeping Chinese and other customers dependent on foreign cloud services has benefits for the United States. Shutting them off from AI services would risk further splintering the digital world and undermine long-standing U.S. defenses of the open internet.
• Requirements that companies receiving chips obtain authorization from the U.S. government before re-exporting them or transferring them to another entity.
• Regular reporting to the U.S. government on chip location and security measures, in order to limit the risks of diversion to countries of concern, alongside cooperation with U.S. law enforcement in any investigation into potential violations.

The U.S. government’s strongest tool to ensure compliance is the threat of cutting off future export licenses, as AI chips last only three to five years before they need to be replaced. Robust information sharing and verification measures will give U.S. officials confidence that partners are complying with the security guarantees. The U.S. government shouldn’t make a deal unless the receiving country will reliably furnish Washington with the information and ongoing leverage necessary to curtail or terminate chip access if security requirements are violated. Of course, this kind of contingency is exactly what foreign governments will want to avoid. But U.S. national interests demand it, and the Trump administration has the leverage—likely the greatest the country will have for years to come—to get it.

What Concessions Did the Other Side Agree To?

The right mix of asks will vary by country, but in general, the United States should ask partners to codify robust controls on the export of advanced technology to countries of concern, including China; restrict outbound investment in the Chinese technology industry; and limit inbound investment from Chinese funds that could siphon away sensitive technology.

In the case of the Gulf states, a top priority should be restricting Gulf investments in China’s AI and semiconductor sectors. Billions of dollars from sovereign wealth funds would do a lot to accelerate China’s technological progress. Moreover, U.S. investors face limits on their ability to finance these sectors in China, so allowing state-backed Gulf funds to do so while offering them U.S. technology would give them an unfair advantage over their American counterparts. For countries like India and those in Southeast Asia that lack sovereign wealth funds but are hot spots for chip smuggling, improved export control laws and enforcement will matter most.

Security guardrails and guarantees of continued U.S. control over AI governance should be at the heart of any deal. But the United States could also use its leverage to push for additional concessions on unrelated issues. It could demand an end to joint military exercises with China (both the UAE and Saudi Arabia have conducted these exercises in the past two years) and verifiable cancellation of any suspected Chinese military facilities in the UAE. It could push for limits on support for armed proxies that threaten regional stability, restrictions on the use of Huawei in 5G networks across the Gulf, and support for U.S. positions in the UN (most of the Gulf states have abstained on numerous U.S.-backed UN votes over the past few years). Better U.S. military access, assistance in U.S. operations against transnational criminal organizations, and reductions in tariffs and other trade barriers that keep U.S. products out of foreign markets could also be on the table.

These are just a few examples of additional requests U.S. policymakers could make of states in exchange for access to U.S. computing power. Analysts and policymakers across the government will rank these priorities differently, and there are benefits and costs to making sweeping issue linkages in international negotiations. The administration seems to be putting heavy weight on promises of hundreds of billions of dollars of Gulf investment in the United States, but these headline numbers are often flimsier than they appear. They frequently contain projects that were already underway, and optimistic projections have a habit of dissolving under scrutiny. Trading large quantities of chips for illusory investment commitments would be a mistake. For these and any future deals, policymakers and observers should ask hard questions about the full range of concessions Washington receives.

Access to chips alone does not guarantee that a state will be able to train competitive frontier models. Many of these countries face talent shortages, and the United States and China are likely to remain the world’s AI centers for the foreseeable future. But the Gulf states have the energy and the capital to be serious players in the AI race, and chip orders on the scale the administration is contemplating raise the risk that they will emerge as serious competitors. U.S. policymakers should think hard about whether this is an outcome they wish to encourage—and what they’re getting in return.