Congress Enters the Chip Wars

Few corners of American artificial intelligence (AI) policy have grown stranger over the past year than exports of advanced semiconductors. For years, a bipartisan consensus held that restricting China’s access to cutting-edge AI chips was essential to maintaining America’s technological edge in the geopolitical competition that may define this century. The Trump administration has upended that consensus, embracing a new theory that American interests are better served by selling chips to China—capturing billions in revenue for American chipmakers and a 25 percent tariff for the U.S. Treasury—than by ceding the market to domestic Chinese alternatives. Congress has now responded with a bipartisan effort to impose guardrails on the administration’s new policy.

The AI OVERWATCH Act, a bill that passed the House Foreign Affairs Committee on Jan. 21 by a vote of 42 to 2, would give Congress authority to block exports of advanced AI chips to adversary nations, treating these semiconductors with the same oversight as arms sales. It is a bipartisan effort led by a Republican committee chairman, criticized by the White House AI czar, attacked by conservative influencers, and praised by the national security establishment—all while China has simultaneously approved and throttled imports of the very chips at issue.

The struggle over the bill reflects deeper tensions between Congress and the president over who controls export policy, between national security hawks and commercial interests, and between two superpowers racing to dominate AI.

Primer on AI Chips and U.S. AI Chip Policy

To understand the AI OVERWATCH Act, one must first understand the technology at stake and the evolution of U.S. policy.

Advanced AI chips, primarily graphics processing units (GPUs) designed by American companies such as Nvidia and AMD, are the computational engines that power artificial intelligence. These chips perform two essential functions for AI systems: training, the computationally intensive process of developing AI models by processing vast datasets; and inference, the ongoing task of running those models to generate outputs for users. Both require enormous parallel processing capability, and demand is growing rapidly as AI systems become more sophisticated and widely deployed.

The United States and its allies dominate the advanced semiconductor supply chain. U.S. companies control chip design; U.S. partners (particularly TSMC in Taiwan) control manufacturing; and U.S., Japanese, and Dutch companies control production of semiconductor manufacturing equipment (SME). China’s challenge is not merely that its domestic chipmakers lag in designing cutting-edge processors, but that they cannot produce the quantity of advanced chips that China’s AI sector demands. According to Matt Pottinger, former deputy national security adviser, “China’s lack of access to advanced semiconductors and its inability to produce them at scale constitutes the central bottleneck preventing China from overtaking the United States in AI.” Restricting China’s access to American-designed AI chips—and working with allies to restrict the export of the equipment needed to fabricate them—was therefore a central objective of U.S. technology policy from the first Trump administration through the Biden administration.

When President Trump assumed office for a second time, he inherited a debate over whether to prohibit the export to China of H20 chips, which Nvidia had designed specifically to fall below existing export restrictions thresholds. In April 2025, the Trump administration announced it would ban the sale of H20s—only to reverse course a few months later, announcing it would allow such sales. In December 2025, it went a step further, announcing it would permit sales of the far more powerful Nvidia H200 and AMD MI325X chips. The reversals marked a notable departure from the long-standing U.S. position that national security export controls are not up for negotiation, instead treating chip restrictions as a flexible instrument of trade policy.

White House AI and Crypto Czar David Sacks articulated a theory behind the shift: It is better to have Chinese AI systems running on American chips—and maintain U.S. leverage—than to incentivize China to build indigenous capabilities. This position holds that overly restrictive export controls have been counterproductive, allowing domestic Chinese chipmakers, such as Huawei and Cambricon, to gain market share while depriving American companies of revenue. Under the new approach, American chipmakers recapture sales and the U.S. Treasury collects a 25 percent tariff on covered chips ultimately destined for China (though critics have questioned the tariff’s legality).

Consistent with the new policy, in January, the Commerce Department’s Bureau of Industry and Security published a final rule revising the license review posture for H200 and MI325X chip exports to China from “presumption of denial” to “case-by-case review.” The rule established technical thresholds and requirements, including a certification framework mandating exporters to attest that, among other things, their Chinese customers conduct robust know your customer (KYC) practices to prevent the Chinese military from accessing the chips. The rule caps total shipments at 50 percent of each chip’s U.S. sales—but even this threshold would allow China to acquire nearly 900,000 H200-equivalent chips, roughly twice what Chinese fabs are expected to produce domestically in 2026 and sufficient compute to train models matching or exceeding current American frontier systems.

Commerce Secretary Howard Lutnick has defended the rule’s conditions, telling Congress in February that Nvidia “must live with” detailed licensing terms. But critics argue the safeguards are unenforceable. Janet Egan and James Sanders at the Center for a New American Security concluded that the certifications face “significant verification challenges”—chips can be redirected once inside China, and the most likely purchasers have documented relationships with China’s military. Tencent, for example, is identified by the Defense Department as a “Chinese Military Company.”

Chris McGuire at the Council on Foreign Relations criticized the rule, stating that it “inadvertently demonstrates that there is no version of an AI chip export policy to China that is simultaneously permissive, implementable, enforceable, and protective of US national security.” According to McGuire, the rule’s safeguards will not prevent China’s AI industry from closing the gap with the United States and will not prevent China’s military from using these chips—arguing instead that the only sensible policy is prohibiting the export of all AI chips to China.

The Structure and Substance of the AI OVERWATCH Act

The AI OVERWATCH Act is Congress’s most recent answer to the administration’s new export policy. It follows several earlier bills seeking to impose restrictions on chip exports, most prominently the GAIN AI Act, which would have required chipmakers to prioritize domestic customers before exporting to China and other countries of concern. That bill passed the Senate but was dropped from the final defense bill after lobbying by the White House and Nvidia. The AI OVERWATCH Act takes a different approach. The most recent version of the bill (dated Jan. 21) establishes a two-tier framework for congressional oversight of advanced AI chip exports.

The first tier covers chips such as the H200 and MI325X—the same chips the administration’s January rule moved into a more permissive licensing posture for China—which qualify as “covered integrated circuits” based on their processing performance and memory bandwidth. The bill’s processing-power thresholds are high enough that the H20 likely falls outside those triggers, though the bill’s separate memory-bandwidth triggers may still capture it. Before the Commerce Department can approve any license to export covered integrated circuits to a “country of concern”—defined as China (including Hong Kong and Macau), Russia, North Korea, Iran, Cuba, or any other D:5 country designated by the secretary of state—it must notify the House Foreign Affairs Committee and Senate Banking Committee at least 30 days in advance (60 days during summer recess). During that window, Congress may pass a joint resolution blocking the sale. This mechanism creates a high hurdle for Congress, requiring veto-proof majorities to ultimately stop a determined president, but it forces the administration to expend political capital on every controversial sale.

This structure borrows directly from the Arms Export Control Act (AECA), which has governed congressional oversight of major conventional weapons sales since 1976. Under the AECA, the executive branch must formally notify Congress before approving significant arms transfers, and Congress has a defined window to enact a joint resolution of disapproval. The notification requirement serves a function beyond interbranch mechanics: It brings consequential national security decisions into public view, providing a focal point for debate and allowing journalists, advocacy organizations, and citizens to engage with elected officials on decisions that might otherwise be finalized behind closed doors. The AI OVERWATCH Act adapts this model to semiconductors, treating advanced AI chips as strategic assets with military and intelligence implications that warrant the same oversight as arms sales.

The second tier is more restrictive. The bill designates the most advanced chips as “restricted integrated circuits” and imposes a mandatory denial requirement—no exceptions, no waivers, no congressional review. This category captures two groups. The first includes any chip exceeding a total processing-performance or performance-density thresholds that would cover chips more powerful than the H200, such as Nvidia’s top-tier Blackwell chips. The second group addresses the risk that chipmakers will engineer products to fall just below control thresholds, as Nvidia did with the H20. Any chip “first marketed for sale after January 1, 2026” that meets the processing-performance or performance-density thresholds for covered integrated circuits faces mandatory denial, even if it falls short of the higher restricted threshold. Existing products such as the H200 remain subject to notification-and-review, but future chips cannot be designed around the rules.

The bill also seeks to accelerate deployment of American AI infrastructure in allied and partner markets. Under current export rules, even transfers to close allies can require complex licensing, creating friction that could push partners toward Chinese alternatives. The “Trusted United States Person” framework, as outlined in the bill, addresses this by exempting qualifying U.S. cloud providers (for example, Amazon Web Services, Microsoft Azure, Google Cloud, and Oracle) from certain licensing requirements when deploying chips to countries outside the restricted list, provided the chips remain under the ownership and control of the U.S. provider and the provider meets security and ownership standards. The bill also directs the Commerce Department to consider expanding the program to allied-country entities. The apparent goal is to anchor partners to the American AI stack rather than Chinese alternatives.

And finally, the bill includes transitional provisions designed to reset the status quo. It would terminate all existing licenses for covered chip exports to countries of concern, thereby canceling any such approvals issued before enactment (including under the administration’s January rule), and impose a temporary freeze on new licenses. The Commerce Department would be required to deny applications until 14 days after the administration submits to Congress an “American Artificial Intelligence Victory Strategy,” a whole-of-government plan for maintaining the U.S. lead over China. The bill also locks in its technical thresholds for both covered and restricted chips for at least 24 months after that strategy is submitted; any subsequent modifications require interagency approval, congressional consultation, and a determination that the change would pose no adverse impact on national security. Taken together, these provisions ensure that the current framework remains in place long enough for Congress to evaluate the results—and would run out the clock on most of the second Trump administration.

Unusual Commercial and Political Dynamics

For Nvidia and AMD, the commercial stakes are significant. The bill’s notification requirements add uncertainty and delay, while the mandatory denial provision forecloses a market worth tens of billions of dollars as new chip generations launch. But not everyone in the technology sector opposes the legislation. Some frontier AI laboratories have endorsed restrictions on national security grounds, with Anthropic CEO Dario Amodei comparing chip exports to China to “selling nuclear weapons to North Korea.” The two camps’ commercial interests also diverge: Chipmakers profit from China sales, while AI developers compete against Chinese firms that would benefit from American hardware.

The partisan dynamics are no less scrambled. The bill is a Republican production: It was introduced by Rep. Brian Mast (R-Fla.), who chairs the House Foreign Affairs Committee, most cosponsors are Republicans, and it is endorsed by conservative organizations including the Foundation for Defense of Democracies and American Compass—the latter praising it as “how America First policy looks.”

Yet the bill has drawn criticism from within the Trump administration. Sacks quote-tweeted a social media post characterizing the bill as an assault on presidential authority, responding simply: “Correct.”

A few days before committee markup of the bill, a coordinated social media campaign emerged against the legislation. Within a two-day window, around a dozen conservative influencers—including Laura Loomer, Brad Parscale, and Ryan Fournier—posted strikingly similar criticisms. The posts shared specific phrases, framings, and even an identical typo, rendering “AI” as “Al” (using a lowercase L) in two posts—a mistake suggesting cut-and-paste from a shared source. Several characterized the Republican-led bill as a Democratic effort that would enable House Minority Leader Rep. Hakeem Jeffries (D-N.Y.) to “greenlight” chip sales to China.

Model Republic documented the pattern of the bill’s criticism on social media and noted that several accounts had ties to Influenceable, a public relations firm that has previously coordinated undisclosed paid political influence campaigns. The publication stopped short of identifying who funded the effort but noted that Nvidia CEO Jensen Huang publicly opposed the bill and that Nvidia stands to lose billions—though Nvidia has not confirmed any connection to the campaign.

Mast drew his own conclusions about the campaign: “Jensen Huang, since you’re not on X, I’ll tag your company @nvidia. You and your paid minions are fighting to sell millions of advanced AI chips to Chinese military companies like Alibaba and Tencent. I’m trying to stop that from happening. Jensen, if you ever feel like debating the actual facts, I’ll meet you anytime, anywhere.”

The House Foreign Affairs Committee passed the bill 42 to 2, with one voting present, on Jan. 21. Sens. Elizabeth Warren (D-Mass.) and Jim Banks (R-Ind.) have also announced plans to introduce a Senate companion. The bipartisan support suggests that members of both parties concluded that AI chip exports to China are too consequential to be negotiated away behind closed doors.

The China Variable

The political contortions in Washington are matched by equally difficult calculations in Beijing. While American policymakers debate the risks of selling chips, Chinese leadership is confronting the invidious trade-offs of buying them. That calculation has driven a rapid evolution in Beijing’s posture, shifting from urgent solicitations to strategic stalling.

During the Biden administration, China elevated chip access as a priority in bilateral relations, lobbying hard for more powerful hardware. But the Trump administration’s oscillation—banning the H20 in April only to reinstate it in July—handed Beijing a tactical opening. China responded by directing domestic companies not to purchase H20s, citing security concerns, but more likely seeking to protect domestic industry from competition and pressing for access to more advanced chips. When the administration approved the far more powerful H200 in December, Beijing again hesitated, convening emergency meetings and instructing tech giants to pause their orders.

Some analysts view Beijing’s hesitation as leverage, not rejection. AJ Kourabi of SemiAnalysis noted that “China is no doubt trying to encourage and establish silicon self-sovereignty, while also getting the best possible chip in the meantime.” McGuire agreed: China’s refusal to allow imports “is almost certainly a negotiating tactic,” given that any U.S. chips would add critical compute capacity to Chinese firms whose domestic supply is severely constrained.

Chinese AI developers face genuine compute shortfalls, and Huawei’s Ascend chips remain significantly behind Nvidia’s products in both performance and manufacturing scale. Matt Pottinger, testifying before the House Foreign Affairs Committee, quoted Chinese AI scientists acknowledging the gap. At a summit hosted by Tsinghua University in January, Alibaba’s top AI scientist forecast only a 20 percent probability at most that the world’s leading AI company will be Chinese within five years, citing the compute deficit as the primary constraint. Quoting another Chinese scientist, Pottinger noted that “the gap [between the U.S. and China] may not be narrowing—it might even be widening.”

Yet Beijing is also reluctant to purchase its way out of the problem on Washington’s terms—paying a 25 percent tariff to its primary adversary, which serves as a stark reminder of China’s technological inferiority in this sector, and deepening dependence on a supply chain that Washington has demonstrated it is willing to weaponize. “Even if we introduce Nvidia chips, our determination to pursue independent innovation will remain unwavering,” Wei Shaojun, vice chairman of the China Semiconductor Industry Association, told Chinese state media in January. “The purpose of importing is to catch up better, and catching up will eventually lead to running alongside or even taking the lead.”

The result—on both sides of the Pacific—has been a halting approval process. Beijing has granted conditional approvals to major tech companies while reportedly requiring domestic chip purchases alongside imports; Washington has imposed stringent licensing terms that have delayed shipments. Chinese firms have expressed demand for millions of chips worth tens of billions of dollars, but both governments are treating approvals as instruments of leverage, and conditions remain subject to change.

Separation of Powers and the Constitutional Framework

Beyond the policy debates, the bill has also drawn constitutional objections, including the claim that it encroaches on the president’s Article II foreign affairs powers. But Congress’s track record in this space cuts the other way.

Article I of the Constitution grants Congress the power to “regulate Commerce with foreign Nations.” As the Congressional Research Service explained, “Congress has delegated to the executive branch some of its express constitutional authority to regulate foreign commerce by controlling exports.” This delegation reflects the practical judgment that the executive branch is better suited to make the granular, time-sensitive determinations that export controls require—with Congress providing the policy framework. The authority the president exercises over dual-use and defense exports flows from statutes enacted by Congress—primarily the Export Control Reform Act (ECRA) and the AECA. Congress has enacted four comprehensive dual-use export control statutes since 1949, each iteration refining the objectives, procedures, and constraints that channel executive discretion.

The AI OVERWATCH Act’s notification-and-review mechanism is borrowed directly from the AECA, which has required congressional notification for major arms sales since the mid-1970s. Under Section 36 of the act, the executive must notify Congress 30 days before approving significant defense sales (reduced to 15 days for NATO members and other allies), and Congress may block any notified transaction by enacting a joint resolution of disapproval. This framework has governed billions of dollars in arms transfers for 50 years. Congress has introduced approximately 300 resolutions of disapproval under this process. Though none has ultimately succeeded in blocking a sale—overcoming a presidential veto makes that outcome rare—the framework, structured as a joint-resolution mechanism requiring bicameral passage and presentment (rather than a legislative veto), has never been struck down by courts on constitutional grounds.

Nor is the AI OVERWATCH Act’s mandatory denial provision for frontier chips constitutionally novel. Congress has previously enacted statutory prohibitions on certain exports. Section 40 of the AECA prohibits the export of munitions to state sponsors of terrorism, permitting presidential waiver only upon notice to Congress and determination that a transaction is “essential to the national security interests of the United States.” Congress exercised similar authority in the National Defense Authorization Act for fiscal year 1999 (P.L. 105-261), when it responded to concerns regarding technology transfer to China by legislatively reclassifying commercial satellites as “defense articles” and transferring jurisdiction from the Department of Commerce to the Department of State. Legislative roll-backs of executive efforts to liberalize high-tech export controls are not unprecedented. The AI OVERWATCH Act’s prohibition on exporting the most advanced AI chips to adversary nations operates within this established pattern.

The constitutional questions surrounding the bill are therefore less dramatic than they first appear. The real debate is not whether Congress can impose oversight requirements on chip exports but whether the specific framework the bill establishes is appropriate.

The AECA Model and Its Adaptation

The formal 30-day notification period is not where the AECA’s most meaningful oversight occurs. The window is too short and the procedural hurdles too high for Congress to evaluate complex transactions under narrow time constraints. The formal process is a backstop. The more important work happens earlier.

Since 1976, the State Department’s practice has been to provide informal notification to Congress in a confidential process before formal notification begins. The current “tiered review” system gives Congress advance visibility—typically 20 to 40 calendar days depending on the system and destination. During this window, committee staff request classified briefings, examine end users and strategic rationale, and raise concerns. If a chair or ranking member of a committee of jurisdiction places a “hold,” the State Department “generally will not formally notify an arms transfer,” according to a 2020 inspector general report. Concerns are resolved through negotiation before the statutory clock begins.

This system persists, despite frequent executive frustration with it, because both branches benefit. The executive avoids public confrontations with Congress over arms sales that could harm bilateral relations with the country involved and gains a channel to address concerns before they harden into opposition. It also avoids provoking Congress into statutory reforms that could further constrain executive discretion—a risk demonstrated by bipartisan threats to eliminate emergency waiver authority after the Trump administration’s 2019 bypass of congressional review for Saudi arms sales. Congress, for its part, obtains meaningful oversight without requiring floor votes on every contested transaction. The informal process transforms what could be an adversarial dynamic into a collaborative one—at least when both sides want it to work. When they do not—as in February 2025, when the Trump administration bypassed a committee hold to formally notify $8 billion in arms sales to Israel—Congress has no statutory remedy to enforce a hold beyond pursuing a joint resolution of disapproval.

Like the AECA, the AI OVERWATCH Act does not codify an informal notification-and-hold process—and codification would likely be counterproductive (and could raise constitutional concerns if it purported to give binding legal effect to committee holds). The value of the AECA’s informal system lies precisely in its voluntary, confidential character; it works because both branches, however reluctantly, have an interest in making it work. That foundation may not exist for the AI OVERWATCH Act. The bill is designed to constrain an administration that opposes its underlying policy.

Without an informal process, Congress would face an uncomfortable choice of allowing transactions it has not had time to meaningfully evaluate to proceed, or blocking exports based on incomplete information. Neither serves the national interest. Legislation can create oversight mechanisms, but it cannot compel good faith. Whether this framework functions effectively will depend on whether both branches recognize that it requires cooperation—even amid profound disagreement about the policy itself.

The Mandatory Denial Outlier

The notification-and-review framework described above governs covered chips like the H200. The “restricted integrated circuit” tier is different—it represents the bill’s most significant departure from the AECA model. For frontier chips, the AI OVERWATCH Act prohibits the Commerce Department from granting licenses at all—and provides no emergency waiver permitting the president to bypass the prohibition. The AECA, by contrast, allows the president to waive congressional review upon certifying an emergency requires an immediate sale. But the absence of such authority in the bill is arguably defensible given that it applies exclusively to adversary nations, whereas the AECA covers a wide range of U.S. security relationships. No plausible scenario requires emergency exports of advanced AI chips to China, Russia, or Iran. An emergency waiver would serve only to circumvent oversight, not address a genuine crisis.

Still, the mandatory denial creates rigidity that may prove difficult to sustain as technology evolves. The bill’s drafters appear to have recognized this. Beginning 24 months after the administration submits the required American Artificial Intelligence Victory Strategy, the Commerce Department may modify the technical thresholds defining both covered and restricted chips—but only with interagency approval, 30 days’ advance consultation with Congress, an updated strategy submission, and a determination that the change would pose no adverse impact on national security. This mechanism allows adaptation to changing circumstances while requiring the executive to build consensus across agencies and with congressional stakeholders before relaxing controls.

There is a strong policy argument, however, for extending the AECA’s model to frontier chips rather than imposing a blanket prohibition. Some experts argue that advanced chip exports to adversaries—especially China—should be barred altogether; on that view, a notification-and-review framework is too permissive because blocking a sale ultimately requires Congress to enact a joint resolution into law, often over a presidential veto. But mandatory prohibitions are calibrated to a specific threat assessment and a specific political moment. If the technology landscape shifts, a statutory ban persists through legislative inertia even after the policy logic supporting it has eroded, while a notification-and-review framework accommodates changing circumstances. The AECA has endured for 50 years across vastly different geopolitical contexts in part because it encodes a process rather than a policy outcome.

A notification-and-review framework for AI chips, including informal holds but with no emergency waiver, would still give Congress a meaningful oversight lever and a path to block specific transactions. But it would also allow the executive to withdraw a proposed export quietly after learning that Congress would object—preserving diplomatic flexibility and avoiding a statutory prohibition designed to be revisited just as the current administration is leaving office, which telegraphs distrust and invites resistance. The informal process works in part because it lets both branches reach the right outcome without forcing a confrontation. But that alternative depends on an executive branch willing to treat congressional oversight as legitimate rather than an obstacle to be circumvented—a condition the bill’s drafters likely concluded they could not assume.

What Happens Next?

The AI OVERWATCH Act faces a long path to becoming law. It must pass the full House and Senate and either receive the president’s signature or survive a presidential veto. Still, the 42-2-1 House committee vote and the announcement of a bipartisan Senate companion bill suggest that Congress is prepared to assert itself in the chip wars.

The bill’s significance extends beyond its prospects for enactment—and beyond the strange political dynamics surrounding it. For 50 years, congressional oversight of arms sales has functioned not primarily through joint resolutions of disapproval but through informal consultation, negotiation, and mutual restraint. The AI OVERWATCH Act borrows the AECA’s formal architecture, but the bill’s effectiveness will depend on norms it does not codify. Whether those norms can develop when the executive opposes the framework itself remains uncertain. Making oversight meaningful will require something harder to legislate.