Back to the Future: From Freeze-in-Place to Sliding Scale Chip Controls

The Trump administration’s decision to grant NVIDIA licenses to sell its H20 chips to China is an attempt to wind the clock back to before October 2022, reversing the Biden administration’s paradigm shift on export controls that turbocharged China’s technological self-reliance drive. But the geopolitical climate has evolved dramatically in the past three years, and the Trump administration’s vision of keeping its chief adversary hooked on a US-made AI technology stack does not comport with China’s mission to shed itself of US dependencies. There are multiple variables at play that will test whether the H20 reversal will ultimately be remembered as a shallow transactional move or a more enduring policy shift on US tech controls and the US-China bilateral relationship more broadly.

A reverse paradigm shift on chip controls

NVIDIA, the world’s dominant AI chip designer, has (for now) shifted from being the chief target to the chief influencer of US technology controls. CEO Jensen Huang met with US President Donald Trump at the White House on July 10 before taking off for his third visit to China this year. Four days later, NVIDIA released a press release saying it would be resuming licensing applications for H20 sales to China after receiving “assurances” from the White House that licenses will be granted. On the sidelines of the China International Supply Chain Expo in Beijing, Huang said he’s looking to start shipping H20s very soon while launching a new China-compliant RTX Pro GPU, described by NVIDIA as “ideal for digital twin AI for smart factories and logistics.”

The about-face on the H20 rips open major questions about the trajectory of US technology controls and the US-China bilateral relationship. On one level, the H20 reversal is a function of Trump’s transactionalism: His administration has framed the move as a bargaining chip with China and has evidently been receptive to NVIDIA’s persuasion on easing up chip controls. It is also an implicit acknowledgement of the efficacy of China’s coercive leverage on everything from critical raw material licenses to anti-monopoly probes.

However, at a more fundamental level, the move represents a clash of worldviews on how export controls should govern AI competition. Under the Biden administration, the US pursued a “freeze-in-place” strategy that aimed to create as wide a lead as possible with China in AI development by imposing increasingly tight restrictions on the level of AI chips they could access from the US. This approach downplayed the traditional “foreign availability” argument that discourages the implementation of export controls where equivalent products are otherwise available in the target market.

The Trump administration is now arguing that this policy will simply cede China’s large domestic market to Huawei as soon as it can provide an alternative product that meets or exceeds that fixed threshold. By approving H20 licenses, the administration aims to restrict Huawei’s blue ocean market opportunity. By this logic, the administration appears primed to shift to a “sliding scale” approach that could free up NVIDIA and other US chipmakers to sell even more advanced chips to China as Huawei and other Chinese chipmakers continue to upgrade.

Sliding scale v. Freeze-in-place

The dueling perspectives on “sliding-scale” v. “freeze-in-place” came into sharp relief last summer, when it became popular among Washington pundits to declare that US chip controls have failed. As Huawei, China’s premier AI chip designer, continued to fill headlines with breakthroughs in its Ascend AI chip series, doubts grew over whether US-led chip and AI controls were slowing China down enough to justify the costs of forcing US companies out of the China market. The chip controls have produced patchy results, but to declare them a failure outright at this juncture is likely premature (see Silent Saboteurs: Loaded Assumptions in US AI Policy). The picture is obfuscated by a number of factors.

First, the US has been slow to patch up loopholes, giving China room to stockpile critical inputs like high-bandwidth memory chips and leaving room for foreign equipment suppliers like Tokyo Electron and ASML to continue servicing machinery that SMIC is relying on to produce advanced chips. Perhaps the biggest distortion stems from the tardy US revelation that Huawei had long been using two obscure Chinese chip design firms—Sophgo and PowerAir—as proxies to illicitly procure logic dies for its Ascend GPU series from TSMC. The “Sophgo subterfuge” has significantly complicated efforts to estimate China’s indigenous capacity to produce advanced AI chips: internal US government sources assess Huawei Ascend production capacity for this year at 200,000 units while others project 700,000 or more when factoring in contributions from Huawei’s TSMC shell game (for comparison, NVIDIA has shipped more than 3.6 million Blackwell GPUs so far this year.)

Skepticism of US chip controls was further amplified by the DeepSeek freakout. For some observers, the globally competitive performance of DeepSeek’s R1 reasoning model discredited the central premise of Biden administration’s chip control policy: that constraining China’s access to advanced chips could ensure a durable and expanding US lead at the cutting edge of AI development. Others looked at DeepSeek with more nuance: DeepSeek trained its model with a collection of NVIDIA chips that were no longer (legally) accessible to Chinese companies, and more time was needed to test the impact of hard-hitting semiconductor manufacturing equipment (SME) controls on China’s indigenous chip production, especially given the Sophgo intervention. For this latter camp, DeepSeek was a sign that US chip controls needed to expand to cover less advanced inference chips like the H20 to limit China’s ability to deploy DeepSeek and other models at scale.

The mixed results forged a forking path for US chip controls. The US could declare failure and shift to a “sliding scale” approach to controls in the hope that US companies could deepen Chinese dependencies on available American technology. Or it could double down on “freeze-in-place” policies, with a focus on containing China to its home market. In effect, this would mean the US would increasingly tighten extraterritorial measures that strip China of foreign support for its chip and AI production while preemptively blocking Huawei out of global AI infrastructure markets. The ultimate objective would be to keep China’s premier AI company largely confined to a slowing home market while the US races ahead in trying to lock in AI infrastructure dependencies in the rest of the world.

Foundational vs. transactional chip controls

The Trump administration, lacking a well-coordinated tech policy across the White House, National Security Council, and Commerce Bureau of Industry and Security (BIS), has been pursuing both policies in parallel. BIS, for example, has drafted several measures that align with the containment theory:

• A pending move to limit foreign foundries like TSMC, Samsung, and SK Hynix from operating their “in China for China” facilities by revoking their Verified End User license exemptions for purchases of restricted advanced semiconductor equipment and spare parts
• A pending BIS rule that would expand the BIS Entity List by automatically capturing any entity that is 50% or more owned by a listed entity
• Applying the Foreign Direct Product rule to forcibly align Japan and the Netherlands with US restrictions on servicing semiconductor manufacturing equipment at advanced fabrication facilities in China.
• BIS guidance that deemed Huawei Ascend chips in violation of US export controls
• A pending rewrite of the US AI Diffusion Framework and/or bilateral AI deals that would condition access to US AI compute on the adoption of economic security requirements that would limit Huawei’s ability to build AI infrastructure in third markets

But Trump’s trade war threw cold water on the containment pathway. Trump triggered an early crisis with Beijing at the beginning of April when he responded to China’s retaliation to US “Liberation Day” tariffs with 125% tariffs on Chinese imports. After the London reset, BIS issued guidance that using Huawei Ascend chips would violate US export controls. This then prompted China to restrict licenses for rare earth materials and magnets, threatening production lines around the world. Amid the tit-for-tat escalation, the US went hunting for additional leverage to force China to come through with critical raw material licenses. Here is where US chip controls got blurred between foundational controls to contain China’s chip and AI development and transactional controls to bargain with Beijing.

Foundational controls encompass US controls on semiconductor manufacturing equipment aimed at throttling China’s indigenous chip production. In contrast, broader US controls on electronic design automation (EDA) software that were deployed ahead of the Geneva reset and subsequently pulled back were transactional. Emerging economic security requirements to block Huawei’s access in third markets would also arguably fall into this foundational category.

But which bucket did the H20 fall in? Last week, Trump administration officials claimed—for the first time—that the original H20 ban was intended as a bargaining chip all along. Commerce Secretary Howard Lutnick even suggested that the H20 rollback was a concession to Beijing in return for rare earth magnet access, contradicting a statement Bessent made to the US Senate immediately after the London talks that “there would be no quid pro quo on chips for rare earths.”

Notably, the Trump administration sent an “is informed” letter to NVIDIA restricting the H20 and chips of equivalent performance on April 9—the same day the US escalated to 125% tariffs on China and shortly after China added seven rare earths to its dual use export control list. But no final rule has been implemented to enforce the April 9 letters. This raises a question whether the Trump administration will simply do away with the licensing requirement. Alternatively, the White House could follow through with the final rule in the Federal Register with the intent of dialing up or down permissiveness of licenses as a point of leverage with Beijing.

H20 acrobatics

The NVIDIA H20 chip was introduced in late 2023 in direct response to tightening US chip controls. To make it compliant with US export controls for the China market, NVIDIA designed the chip to have significantly reduced raw compute power, making it ill-suited for the compute-intensive workload of training large AI models, but with enhanced memory characteristics that make it ideal for inference, i.e., the real-world deployment of AI models. In justifying the H20 rollback, Trump administration officials have echoed NVIDIA’s basic argument that chip controls for the China market should be eased because US chips will be inherently more competitive than Chinese-made ones, thereby creating deeper Chinese dependencies on the US tech stack and depriving Huawei of space to build strength at home and scale out into third markets to threaten US tech hegemony.

The assumptions behind this logic need unpacking.

Assumption 1: The Trump administration’s openness to applying a sliding scale implies a gradually opening market for advanced chip sales to China. This would entail giving primacy to the traditional “foreign availability” argument that any policy to restrict a technology should consider whether equivalent products are already widely available in the target market. By this logic, Huawei advances in GPU and next-generation AI data center and supercomputing platforms would translate into dynamic easing of US measures to allow companies like NVIDIA to compete head-to-head in China’s home market. The Trump administration’s willingness to move down this path still needs to be tested, especially if evidence grows that such access is building a bridge to Chinese self-sufficiency rather than deepening Chinese dependencies on US AI technology.

Assumption 2: The sliding scale applies to advanced chip sales to China, but not to foundational semiconductor manufacturing equipment (SME) controls constraining Chinese chip production. Given the administration’s stance that keeping Chinese AI developers hooked on US-made GPUs and software ecosystems is intended to slow Huawei’s advanced chip progress, controls on the SME that SMIC is using and that Huawei is relying on to expand its chip production capacity would logically remain in place. SME companies like Tokyo Electron and ASML may attempt to use the “sliding scale” momentum to reinforce their own long-standing arguments that SME controls are only accelerating Chinese toolmakers like Naura and AMEC. But a company like NVIDIA competing with Huawei would want such controls to remain in place. We suspect the Trump administration will favor keeping the SME controls intact.

Assumption 3: The technological superiority of NVIDIA’s AI stack will overwhelm its Chinese customers’ geopolitical incentive to shift to indigenous suppliers. According to Secretary Lutnick, the Trump administration’s decision to lift the H20 ban is explicitly designed to “get Chinese developers addicted to the American AI stack,” thereby undercutting Huawei’s opportunity to build a competitive alternative in their home market unopposed. In the near term, NVIDIA will have no trouble finding buyers for the H20: Its mature CUDA ecosystem remains far easier to deploy and maintain than Huawei’s nascent in-house CANN framework. Chinese developers—whose global competitiveness hinges on reliability and rapid iteration—are thus overwhelmingly incentivized to stick with NVIDIA for now, despite growing geopolitical pressure to switch.

Huawei is reportedly working to ease the transition to Ascend for CUDA-based developers in China by building an intermediary translation software that would enable CUDA‑based models to run on Ascend chips. If successful, this solution could serve as a bridge across the CUDA moat, enabling Huawei to remain competitive in China’s AI hardware market as it continues to work with domestic developers on parallel efforts to build CANN into a viable alternative. This approach mirrors Huawei’s HarmonyOS strategy, in which Android compatibility provided a transitional bridge for developers to host their apps on Huawei devices while they simultaneously worked on HarmonyOS-native versions in preparation for Huawei’s 2024 launch of its wholly Android-independent HarmonyOS Next.

While the technical challenges of applying the HarmonyOS strategy to build a bridge to CUDA independence are substantial, there is a shared incentive among not just Beijing and Huawei, but also China’s broader AI development community, to vigorously support this effort. While Chinese firms will exploit the H20 reversal to build precious compute capacity, this partial concession will not be sufficient to disincentivize investments in Huawei’s AI stack. Not only will the experience of the last three years continue to encourage hedging against a potential future US crackdown, but the sliding‑scale approach itself incentivizes continued parallel support for domestic chipmakers: If the US pursues a policy of gradually relaxing controls to ensure US competitiveness as more performant domestic chips emerge, promoting domestic chip progress gains the added benefit of forcing the US to roll back restrictions further. Unless US controls are fully rescinded—an unlikely outcome—the equilibrium will remain unstable, with Chinese developers balancing immediate performance needs against mounting political and economic pressures to indigenize.

Assumption 4. Beijing will remain non-interventionist despite a national security mandate to steer Chinese AI developers toward homegrown technology. Cognizant that a heavy state hand in AI development could impede its ability to keep pace with the US, China’s leadership has been trying to strike a balance between facilitating innovation by startups like DeepSeek and applying state tools to steer growing adoption of indigenous technology. Beijing’s toolkit includes Autonomous and Controllable criteria to identify key targets for self-reliance in critical supply chains and restrict foreign technology as domestic producers prove capable of replacing foreign players. The Cyberspace Administration of China can also apply cybersecurity and data security measures to restrict foreign technology in critical IT systems. Moreover, Beijing can apply environmental criteria to discriminate against foreign suppliers. A case in point is the National Development and Reform Commission’s 2024 implementation of energy efficiency rules that technically bar H20 installations. Even if left unenforced, such measures signal a long‑term intent to tilt the market toward Huawei.

Assumption 5: Inference compute capacity is not relevant to US controls on chip sales to China. In AI, quantity has a quality of its own. One of the most prominent arguments against the administration’s “foreign availability” justification for the H20 reversal is that it overestimates Huawei’s ability to meet the massive demand for large-scale compute in China. While restrictions on China’s access to the most cutting-edge chips may inhibit Chinese developers’ competitiveness in training massive foundation models, freeing up the sale of the H20 will relieve a second, arguably more immediately relevant, bottleneck on China’s AI ambitions—scaling up inference compute to enable the widespread deployment of AI across China’s domestic economy.

Although Chinese developers like DeepSeek continue to impress with their ability to do more with less, deploying reasoning models like R1, and AI more generally, across China’s economy at scale will inevitably require access to massive amounts of parallel computing power. Lifting controls on inference-optimized chips like the H20 therefore risks supporting Beijing’s strategy to build asymmetric strength in the deployment of AI, particularly as a force-multiplier for its existing strengths in smart manufacturing and logistics.

Over the past two years, China has doubled down on this strategy, unveiling “AI+” initiatives aimed at promoting widespread adoption within practically every industry vertical in China. Beijing has also dramatically expanded its support for the development and adoption of “embodied intelligence” to establish China as a leader in AI-enabled smart manufacturing, logistics and autonomous vehicle applications. Realizing these ambitions will require a combination of large-scale cloud-based compute to power digital simulations, reasoning, and other compute-intensive workloads along with powerful edge AI compute platforms for low-latency, on-site inference.

NVIDIA is positioning to compete aggressively in both of these markets in China. In the same July 14 press release that revealed the H20 reversal, NVIDIA announced the launch of a new China-compliant RTX Pro GPU for smart factories, logistics, edge workstations and AI PCs. While NVIDIA would have pushed the RTX Pro into China regardless of the H20 policy, the sliding‑scale carveout for inference chips reinforces its opportunity to expand its foothold in China’s AI market from the cloud all the way through to downstream deployment in industrial and commercial AI applications.

Assumption 6: There is still more to come from the US’s block and tackle moves against Huawei in overseas markets. The Trump administration has been consistent in conveying its desire to lock in US AI infrastructure dependencies around the world and to preempt Chinese competition in this domain. The question ahead is: What tactics will the White House employ to block out Huawei AI contracts overseas? Our watch list includes:

• Details emerging from the annex of the US-UAE bilateral agreement on building AI infrastructure, which are supposed to include cybersecurity and other economic security requirements aimed at shutting out Chinese suppliers for UAE AI buildouts. The Trump administration has yet to release the details of the UAE framework.
• Economic security requirements written into bilateral trade agreements: US trade negotiations with partners remain in limbo in the lead-up to the August 1 tariff deadline as Trump attempts to maximize demands. But the prolongation means these will likely be multi-phase negotiations in which China-specific economic security requirements get incorporated ad-hoc, likely modeled in part after the UAE annex.
• Compute governance requirements to limit Chinese access to controlled chips and large-scale compute: Concerns around Chinese entities remotely accessing large-scale compute has drawn scrutiny from Trump administration officials on countries like Malaysia and Singapore over lack of compliance and enforcement of US chip controls. This could translate into stricter licensing requirements for countries suspected of circumvention as the US administration rolls out ad hoc revisions to the Biden-era AI Diffusion Framework. Alternatively, Congress could opt to take on this issue via the Remote Access Security Act, which would amend the text of US chip controls to extend their coverage beyond physical chip exports to include remote access to covered chips via foreign AI cloud platforms. Reports of large-scale smuggling of restricted chips has also built bipartisan momentum within Congress for the Chip Security Act, which would require AI chipmakers to verify—either through embedded technical tracking mechanisms or through regular on-site audits—that each chip remains at its authorized destination. Proponents of chip-based mechanisms argue that integrating location-verification features directly into the hardware offers the most robust, tamper-resistant solution, ensuring continuous compliance without relying on external oversight. However, recent amendments to the bill reflect growing momentum for the physical auditing approach favored by Trump administration AI and Crypto Czar, David Sacks.
• BIS ICTS investigation into data centers with long-arm provisions. The ICTS Office under BIS is rumored to be preparing an investigation into data centers as its next target. Such restrictions would primarily focus on restricting Chinese suppliers in the US market but could serve as the groundwork for long-arm measures premised on trusted networks to reinforce primacy of US hyperscalers and limit Chinese participation in AI data center buildouts abroad.

The road to a Trump-Xi summit

Under the influence of America’s top AI chip producer and under pressure from China’s rare earth controls, the Trump administration is trying to turn the clock back on US chip controls. But the train on US tech controls left the station in October 2022. While the strategy of keeping China dependent on the US AI technology stack had merit and was arguably viable before the paradigm-shifting October 7, 2022 “freeze-in-place” chip controls, the world is in a very different place today. China’s technological self-reliance push is driven by a degree of geopolitical existentialism that simply cannot be willed away by a techno-idealist vision for global AI diffusion that contradicts Beijing’s core national security principle: do not put your country’s security and economic stability at the mercy of your chief geopolitical adversary (for a deeper discussion on this clash of AI worldviews, see our Foreign Affairs essay, The Real Stakes of AI Competition).

NVIDIA CEO Jensen Huang believes progress is inevitable and resistance is futile: As he described in China last week, “there’s no stopping the pace of domestic innovation. NVIDIA can make a great contribution. AI is a vertically complex stack—chips, systems, networking, software, algorithms, applications. Each layer must innovate. If one layer lags, smart engineers at other layers compensate, and the whole stack moves forward.” We imagine that US policymakers prioritizing cutting down the Chinese AI stack, not enabling it to move forward, will take issue with this vision. Meanwhile, US administration officials broadcasting a strategy to get China “addicted” to US AI technology are only adding fuel for Beijing’s national security hardliners to double down on a self-reliance mandate.

Nonetheless, China has a window to probe the limits of Trump’s transactionalism. Beijing has demonstrated the efficacy of its critical raw material restrictions to not only force the US to the negotiating table but also extract concessions in US tech controls—an area that Biden-era officials have long argued are off limits in US-China bilateral negotiations.

Less attention has been paid to how Beijing has been able to leverage its long-arm anti-monopoly power in US-China negotiations. The day before the H20 rollback, China’s anti-monopoly arm, the State Administration for Market Regulation (SAMR), conditionally approved an acquisition by US-based Synopsys, a leader in electronic design automation software for chipmaking, and US-based engineering design firm Ansys. Given China’s track record of scuttling semiconductor-related deals and the high sensitivity around this particular deal following the US’s weaponization of EDA export controls, it appears a quid pro quo may have been made by China to allow the deal to go through in return for US restraint on broad EDA restrictions. Notably, the conditions that SAMR imposed on the approval included an order to divest part of the company as well as supply guarantees (though US companies would likely prioritize compliance with US export controls over SAMR conditions if the software in question were restricted.) Looking ahead, we need to watch for whether SAMR also drops an ongoing probe into NVIDIA-Mellanox over a prior transaction (See our Barron’s article, China Takes a Shot at NVIDIA. It Could Backfire on Beijing) and test whether the SAMR conditions for divestment for approving the Synopsys-Ansys deal also entails Beijing positioning a Chinese buyer for the divested asset as it tries to build up its homegrown EDA capabilities.

Now that the transactional line on tech controls has been breached, Beijing may as well try to press for further concessions while pushing a narrative of “peaceful co-existence” with the US to shift the dynamic from confrontation to recognizing respective spheres of influence. In commenting on the H20 rollback in a July 16 interview with Bloomberg, Treasury Secretary Scott Bessent said “we’ve done a lot on tariffs and export controls, now it’s time to move onto other issues.” He also said he will be meeting with the Chinese Vice Premier within weeks with an eye toward a Trump-Xi summit in September-October. This implies that the US president believes the time is ripe to test the potential for a broader bargain with Beijing. But that ambition comes in spite of multiple inconvenient truths:

• US-China trade imbalance cannot be solved quickly without deep, structural reforms in China to stimulate consumer demand, pare back state-led investment, and curb nonproductive growth. Bessent has theorized that there can be a “great rebalancing” if China consumes more and the US manufactures more, but we are not yet at the point where external pressures are compelling Beijing to go down a politically painful reform path.
• Tech controls remain in play. Beijing cannot assume that inconsistency from the Trump administration translates into an enduring truce on tech controls. Beijing will try to use its leverage in critical supply chains to compel US restraint, especially if it can count on Trump’s desire for a bilateral summit to keep a lid on moves being prepared by BIS. But Beijing will remain hyperalert to moves to block Huawei’s access to AI infrastructure transactions, China-focused economic security requirements in trade deals, emerging ICTS measures, entity listings and other measures on the table as US-China technology competition continues to intensify. Major technological advancements by Chinese AI champions can still capture media attention and lead to a reflexive tightening of controls.
• Clawbacks are in motion. Beijing is alert to ongoing US efforts to claw back Chinese-owned assets (See The Clawback: Reclaiming Strategic Assets from Beijing). Top of mind is a pending takeover by BlackRock and Mediterranean Shipping Co of CK Hutchison’s global port portfolio (minus Chinese mainland ports). Beijing is angling to insert state-owned COSCO into the transaction once the exclusivity period for the current parties ends on July 27, 2025. We will need to watch whether the Trump administration sees these clawback deals as bargaining chips in his negotiation with Xi even though they are aimed at mitigating the risk of Chinese ownership in critical supply chains. There are already early signs that Trump may be flexible on a potential TikTok deal that would entail leasing the app’s algorithm from ByteDance and allowing ByteDance to retain a stake in a new US entity.
• Taiwan is simmering. Beijing is also eyeing a more assertive posture by Taiwan’s president and the US administration’s multi-faceted efforts to draw semiconductor and electronics assets away from Taiwan and to the United States. If Beijing perceives that its ability to contain independence-leaning forces in Taipei is compromised and that the US is actively hollowing out Taiwan’s manufacturing base, it may be compelled to step up its coercion tactics on Taiwan. Such an escalation would reduce the space for a trade bargain.

This is the moment for China to make the most out of Trump’s transactional instincts. Beijing’s strategy will be to express outrage over any perceived escalatory move by the US, including measures to block Huawei’s overseas expansion and ongoing efforts by the US to claw back Chinese-owned assets in critical supply chains. China’s Ministry of Commerce will meanwhile have its hand on the critical raw materials licensing dial, ready to punish or reward depending on where negotiations go. Even though Beijing likely faces an uphill battle with Trump in getting significant tariff relief, both sides may still explore whether a deal can be struck on market reciprocity. This could entail the American president showing flexibility on expanding access to Chinese FDI and a willingness to compromise on an ICTS agenda aimed at purging Chinese suppliers from connected markets. But if the Trump administration’s national security compromises continue to stack up, we will likely see stronger pushback from Congress and quiet attempts from individuals within the administration that are less prone to transactionalism to steer the US-China relationship back toward a national security course.

The odds may be stacked against a US-China grand bargain. Nonetheless, the politics surrounding the H20 reversal and Trump’s desire to get a summit with Xi on the books will no doubt unnerve US trading partners. So long as trade deals with Trump remain elusive and the prospect of a Xi-Trump bargain lingers, key US trading partners will be compelled to hedge in the coming weeks. In practice, this may mean staving off hard commitments with the US on economic security measures that would trigger Beijing while getting comfortable with the idea of trade negotiations stretching for longer. For Beijing, this is precisely the point. It needs to disrupt US deal-making with trading partners aimed at China’s economic isolation while keeping a lid on US tech controls. US-China relations are going to be especially bizarre in the run-up to a potential summit. But if this is simply an experimental phase of deal-probing in the broader US-China crisis-reset loop, then the limits of Trump transactionalism will be revealed in due time.