From Racing China to Pacing the AI Frontier

The latest regulatory developments on economic security & geoeconomics

By Paul Nadeau, Visiting Research Fellow, Institute of Geoeconomics (IOG)

Strait of Hormuz Update: As of Sunday, September 6, the seven-day moving average of ships passing through the Strait is 3 according to the International Monetary Fund’s Portwatch. The seven-day moving average at this time last year was 97.

OFAC Issues License for Venezuela Telecom Sector: On August 21, the U.S. Treasury Department’s Office of Foreign Assets Control (OFAC) issued a general license authorizing the supply of certain items and services to Venezuela related to telecommunications, as well as a general license authorizing negotiations of and entry into contingent contracts for investment in Venezuela’s telecommunications sector.

U.S. Removes Syria as State Sponsor of Terrorism: On August 24, the U.S. State Department rescinded Syria’s designation as a State Sponsor of Terrorism and delisted Hay’at Tahrir al-Sham (HTS) as a terrorist group in recognition of actions taken and further commitments made by the Syrian government to fully distance Syria from acts of international terrorism.

U.S. Targets Far-Left Groups with Sanctions: On August 26, the U.S. Treasury Department’s Office of Foreign Assets Control (OFAC) sanctioned three entities and two individuals to counter violent far-left terrorist groups, including Autistici Inventati, an Italy-based entity that supplies specialized digital architecture, tools, and services for Antifa and other activists, Palestine Action, a UK-based organization proscribed as a terrorist group by the UK government in July 2025, and the transnational group Masar Badil which allegedly operates as a front for the Popular Front for the Liberation of Palestine, a designated Foreign Terrorist Organization.

Trump Uses Emergency Powers to Protect Electrical Equipment: On August 26, Donald Trump signed an executive order declaring a national emergency under the International Economic Emergency Powers Act (IEEPA) in order to protect the security, integrity, and reliability of bulk-power system electric equipment used in the United States by generally prohibiting certain foreign-produced equipment, including associated critical software and digital capabilities that could pose cybersecurity or operational risks, from being purchased or installed in the United States.

Trump Announces U.S. Control of Venezuela Oil Reserves: On August 29, Donald Trump announced that the United States had reached an agreement with Venezuela that gives the United States majority control (55 percent) of Venezuela’s proven oil reserves through a new private joint venture. U.S. Secretary of State Marco Rubio said that the agreement will bring nearly $100 billion in private investment into Venezuela.

U.S. Customs & Border Protection Announces Supply Chain Visibility Effort: On September 2, U.S. Customs and Border Protection (CBP) issued an Advance Notice of Proposed Rulemaking designed to improve visibility into the supply chains of goods imported into the United States, to more effectively detect, stop, and prevent illicit imports that evade U.S. customs and trade laws.

U.S. Sanctions Grandson of Cuba’s Raul Castro: On September 3, the U.S. State Department sanctioned, Fidel Castro Calis, the grandson of former Cuban president Raul Castro along with Banco Exterior de Cuba, a Cuban state-owned bank, and four entities associated with Cuba’s metals and mining sector.

Canada Retaliates on U.S. Tariffs; U.S. Counter-Retaliates: On September 8, Canada announced retaliatory tariffs against the United States on roughly $20 billion in U.S. exports in response to tariffs imposed by Donald Trump following the collapse of trade negotiations on August 21. Canada’s tariffs are designed to match U.S. tariffs dollar-for-dollar and target sectors including steel, dairy, appliances, agricultural equipment, pulp and paper, and electronics. In response to Canada’s announcement, the United States announced import bans on certain Canadian alcoholic beverages, motorcycles and dairy products (some cheese products are tariffed at 50 percent rather than banned) and are set to go into effect on September 29.

UK Expands Sanctions on Iran: On September 8, the United Kingdom’s Foreign, Commonwealth, and Development Office introduced sanctions amendments that expand and strengthen the UK’s regime against Iran, including new sectoral trade, financial, and transport restrictions. In a separate announcement on September 8, the Foreign, Commonwealth, and Development Office sanctioned several Israeli nationals for their involvement in violence against Palestinian individuals.

U.S. Targets SE Asia Cybercriminals: On September 9, the U.S. Treasury Department’s Office of Foreign Assets Control (OFAC) targeted Xinbi Guarantee, an illicit online platform used by Chinese cybercriminals for scam operations, as well as two Southeast Asian companies that help enable its activities. In a concurrent announcement, U.S. Justice Department announced that it had seized infrastructure and digital asset wallets used by the platform.

Canada Announces New Russia Sanctions: On September 10, Canadian Prime Minister Mark Carney announced a series of initiatives to support Ukraine amidst the ongoing invasion by Russia, including sanctions on eight individuals who have engaged in the violation of Ukrainian children’s rights, including their unlawful deportation and forcible transfer, as well as their exposure to indoctrination and militarization.

Operation Economic Outcast Updates:

・On August 28, as part of Operation Economic Outcast, the U.S. Treasury Department’s Financial Crimes Enforcement Network (FinCEN) sanctioned the general manager of an Iranian bank’s Dubai branch as well as a Hong Kong-based trading company tied to a sanctioned Iranian exchange house as part of its push to further counter Iranian shadow banking activity.

・On August 28, the U.S. Treasury Department’s Financial Crimes Enforcement Network (FinCEN) proposed a rule, as part of Operation Economic Outcast and pursuant to Section 311 of the USA PATRIOT Act, that would revoke Banque Misr UAE’s correspondent banking access to U.S. financial institutions after assessing that its customers include apparent front companies used by Iran to evade U.S. sanctions and launder money. The proposed ⁠punishment is expected to come into effect in less than 30 days after a public ​comment period.

・On September 4 the U.S. Treasury Department’s Office of Foreign Assets Control (OFAC) designated, Golden Global Yatirim Bankasi Anonim Sirketi, a Turkey-based financial institution and its subsidiaries for facilitating tens of millions of dollars’ worth of transactions for the Islamic Revolutionary Guard’s Quds Force and providing the Iranian regime with key correspondent banking access.

・On September 8 the U.S. Treasury Department’s Office of Foreign Assets Control (OFAC) announced sanctions on 36 targets for supporting Iran’s aviation sector, as well as covert front companies, foreign intermediaries, and deceptive transshipment routes that Iran relies on to obtain U.S.-origin aircraft and sensitive technology and to move weapons, personnel, and illicit cargo.

・On September 10, the U.S. Treasury Department’s Office of Foreign Assets Control (OFAC) sanctioned entities and individuals supporting Kata’ib Hezbollah and Lebanese Hezbollah, groups that serve as Iran’s proxies in the Middle East.

・On September 10, the U.S. Treasury Department issued a policy of presumed denial for licensing requests to engage in activities prohibited by the Iranian Transactions and Sanctions Regulations.

From Racing China to Pacing the AI Frontier

By Andrew Capistrano, Visiting Research Fellow, Institute of Geoeconomics (IOG)

Ever since the “DeepSeek moment” in January 2025, the US government and the top AI labs seemed to have aligned interests: to keep pushing the AI frontier forward in order to “win the race” against China. That shared imperative gained urgency in response to cheaper Chinese open-weight models, which undercut the economics of cutting-edge American proprietary models while setting off alarm bells within Washington’s national security establishment. The resulting alignment also helped justify “hyperscaling”, an enormous expansion of the US compute base supporting frontier labs such as Anthropic and OpenAI.

And the size of the hyperscaling wager has been extraordinary. The four largest US hyperscalers spent approximately $410 billion in 2025 and plan to spend around $725 billion in 2026—more than $1.1 trillion across two years, although not all of that expenditure is specific to AI. Crucially, this spending served both corporate and national purposes. Superior access to advanced chips, data centers, and electricity promised the frontier labs and their infrastructure partners a defense against the commoditization of the model layer, while the same investment expanded the American compute base, accelerated capability development, and was expected to widen the US lead over China. In this way, private competition mobilized capital toward a national strategic objective.

But on 12 September, Anthropic CEO Dario Amodei published a major essay arguing that frontier labs must instead “slow the pace” of AI capability improvements. OpenAI CEO Sam Altman has stated he agrees. This reversal is striking, and—at least for the frontier labs—its timing is convenient. Moonshot AI’s July 2026 release of Kimi K3 showed that a Chinese open-weight model could approach frontier performance at a far lower API price than Anthropic’s leading model, eroding the durable commercial advantages that hyperscaling was supposed to create. It is therefore difficult to avoid the conclusion that, while “pacing” may dovetail with legitimate AI safety concerns, it also serves the frontier labs’ self-interest by offering a way to manage the weakening commercial logic of hyperscaling: committing ever more capital to the compute required for each new generation of frontier models, even as open competitors erode the proprietary returns.

The national security imperative to “win the race”, however, remains—which is why Amodei’s proposal has implications beyond AI safety. He proposes outside evaluators, common capability thresholds, government-enabled coordination among frontier labs (including a narrow antitrust waiver), and eventually an AI agreement with China. Whatever their safety merits, these measures also fit the altered economic interests of closed frontier labs: coordinated restraint reduces the cost of slowing by ensuring domestic rivals slow together, while common compliance requirements raise barriers to challengers. The question for the US government is whether a framework so well suited to the labs’ altered commercial interests also advances the national interest.

Recent statements from the Trump administration make that emerging divergence explicit. Only days before Amodei published his essay, Treasury Secretary Scott Bessent dramatically declared that “there is no day after tomorrow if China wins” the AI race. “If they were to pull away from us on AI”, he added, “then nothing else would matter”. Trump subsequently rejected calls to pace the frontier, arguing that the US must preserve its lead over China because “whoever wins AI wins”. He has since hardened that position (it has become his latest slogan), dismissing Amodei’s safety concerns as a “hoax” and arguing that opposition to AI and data center development benefits China.

Trump’s former AI czar David Sacks sharpened the institutional distinction: Anthropic and OpenAI are free to pace themselves, but voluntary corporate restraint does not entitle them to an antitrust waiver, a regulatory approval process, or government-enforced restraints on their competitors. Taken together, Trump and Sacks seem to suggest that Washington can accept caution by individual companies—but will not make slower development a part of US AI strategy.

The reason is straightforward. If China cannot be persuaded to slow its own frontier, Washington has two principal unilateral levers for preserving the US lead: slow China or accelerate America. That means technological denial on the one hand, and rejecting domestic pacing and continuing to push the American frontier forward on the other. The two are not mutually exclusive, and for an administration that regards AI leadership as existential, the strongest competitive strategy would be to pursue both.

How these strategies affect the alignment between Washington and the frontier labs becomes clearer when AI policy is disaggregated into development, deployment, and diffusion. Development is where the divergence is sharpest. Amodei argues that recursive self-improvement makes frontier development itself dangerous, but slowing the frontier would by definition restrain future US AI capability. Pacing development would therefore turn a domestic safety measure into a geoeconomic gamble over the future balance of technological power with China.

Deployment produces a more conditional alignment. The government and the frontier labs share an interest in preventing unsafe uses, and Washington may accept (or demand) restrictions on dangerous public or commercial deployment so long as frontier models remain available for national defense and intelligence purposes it regards as legitimate. But they may disagree over which uses meet that standard. For example, Anthropic’s February 2026 dispute with the Department of War showed that a frontier lab can be a national asset without being a national instrument. If Washington increasingly protects and depends upon frontier labs whose restrictions can limit strategic deployment, it may eventually demand corresponding access—or outright control. Although such direct state intervention is not yet visible on the horizon, the move toward pacing could hasten its arrival, possibly opening the door to extreme measures like nationalization of the frontier labs.

Diffusion is where the frontier labs and the government retain the greatest common interest, corresponding to the strategy of slowing China through technological denial. At the hardware layer, US export controls already restrict China’s access to the advanced chips and semiconductor manufacturing equipment needed to develop frontier systems. Further tightening could target smuggling, third-country diversion, and Chinese access to advanced compute through overseas data centers.

At the model layer, additional security measures could target unauthorized AI “distillation”, through which access to a commercially available American proprietary model can become a channel for strategic capability transfer. Frontier capabilities can thereby cross borders without chips, source code, or model weights themselves being transferred, allowing Chinese firms to capture some of the value generated by enormous US investments without replicating the underlying expenditure. In response to such concerns, a September joint advisory from the FBI, NSA, and CISA accused Chinese labs—including DeepSeek and Moonshot AI—of extracting capabilities from US models at industrial scale.

The distillation allegations therefore strengthen the case for tightening access controls on deployed models and controlling diffusion, but they do not necessarily support slowing domestic frontier development. In fact, if Chinese labs can use deployed American systems to accelerate their own progress, the US government may have still greater reason to pursue the complementary strategy of accelerating America: pushing the frontier forward while closing the channels through which its advantages escape.

Amodei draws a different conclusion from this remaining alignment over diffusion. It supplies him with an answer to his strategic problem: how can US labs slow development without allowing China to close the gap? Rather than combine technological denial with faster American development, he pairs coordinated restraint at home with intensified denial abroad. If these measures slowed China sufficiently, he believes, US labs would gain the needed room to pace their development. Pacing could then be made strategically asymmetric: American labs would slow, but Chinese labs would slow more. The problem is that sustaining such asymmetry would be extremely difficult. Domestic compliance rules may prove easier to enforce than technological denial against Chinese development, leaving American labs “pacing” while their Chinese competitors continue “racing”.

To make that asymmetry sustainable, Amodei turns to a third, cooperative mechanism, namely a US-China AI agreement. He treats technological denial as a means to this end, following a two-stage logic: denial would first widen the American lead, creating both the breathing room required for domestic pacing and leverage for negotiating reciprocal restraints on frontier development. A sufficiently comprehensive and enforceable agreement could then convert temporary asymmetry into mutual restraint, reducing the absolute pace of development without sacrificing America’s relative position. Rather than eliminating the strategic difficulty, however, this shifts it from maintaining an asymmetrical lead through technological denial to securing verifiable Chinese restraint.

Yet China’s own security debate points instead toward acceleration under Party direction. In a recent article, Minister of State Security Chen Yixin identified AI risks ranging from cognitive warfare and data leakage to cyberattacks. But his prescribed answer is tighter Communist Party control over the internet and data, converting political authority into effective AI governance. China is therefore responding to AI risk through stronger political supervision rather than slowing the frontier—it is pursuing indigenous development, Party-directed deployment, and selective international diffusion when open models strengthen Chinese influence.

Why, then, would Beijing accept comprehensive, coordinated pacing while US semiconductor export controls already restrict its access to the frontier? An agreement that preserved China’s disadvantages in AI chips while limiting its opportunity to catch up would offer Beijing little incentive to accept it. China could instead demand relief from those controls or simply refuse. The first obstacle is therefore whether any bargaining space exists between an American proposal intended to preserve the US lead and a Chinese government determined to overcome it.

The response from Chinese state media illustrates that difficulty. An editorial in the Global Times called Amodei’s proposal a “Cold War script” and portrayed its technological barriers and regulatory restraints as an attempt to curb China’s AI development, preserve US technological “monopolistic hegemony”, and exclude China from global AI governance. It also described the idea of combining global AI safety with containing China as a “fundamentally unrealizable fantasy”, arguing that pacing the frontier would lose its meaning if China and its companies were excluded from the global AI ecosystem. Part of the editorial’s animus may also reflect its timing: only days earlier, Anthropic had alleged that Chinese firms routed customer queries—including sensitive personal, corporate, and possibly military information—through Claude as part of distillation efforts, thereby exposing a potential Chinese data-security vulnerability.

More fundamentally, the underlying message is clear: Beijing is unlikely to regard restraints imposed alongside US export controls as neutral safety governance, but rather as an attempt to lock in American technological dominance. In short, Amodei’s sequence contains an internal contradiction, since the very controls he expects to generate leverage may make the necessary US-China agreement harder to obtain.

Even if sufficient bargaining space could be found, enforcement would present a second problem. Secret capability development would be exceptionally difficult to monitor, and each side would fear that the other was advancing behind a public promise of restraint. That leaves little room for a major US-China AI agreement along Amodei’s preferred lines during Xi Jinping’s planned visit to Washington on 24 September. Narrow cooperation on testing, incident reporting, crisis communication, or particular misuse risks may be conceivable, but enforceable reciprocal limits on overall capability development are another matter entirely.

Absent such restraint, “pacing the frontier” could initially proceed only as a unilateral US policy. The choice then turns on how the US should use the technological advantage that hyperscaling created. Amodei treats the US lead as leverage for securing eventual bilateral pacing with China; Trump treats it as an opportunity to accelerate further. In other words, a durable alignment between pacing and US national security would ultimately rest on an enforceable agreement with China.

The frontier labs may have sound reasons to pace themselves, but they have not shown why their preferred pace should become American strategy. Doing so would require Washington to reconstruct through regulation an alignment that market incentives no longer sustain, even as the national security imperative continues to favor acceleration. Indeed, while the commercial logic of hyperscaling may be weakening, the geoeconomic logic of “racing China” is not.

(Photo Credit: iStock)

Disclaimer: The views expressed in this IOG Economic Intelligence Report do not necessarily reflect
those of the API, the Institute of Geoeconomics (IOG) or any other organizations to which the author belongs.

IOG English Newsletter

Subscribe to the English IOG Newsletter to stay up to date on IOG’s activities across various briefings, op-eds, YouTube videos, podcasts, as well as new publications, events, research activities, and more.

Click here to subscribe