Where the Chips Fall: U.S. Export Controls Under the Biden Administration from 2022 to 2024

Commentary by Barath Harithas and Andreas Schumacher

Published December 12, 2024

Introduction

This piece offers a practical guide to navigating the thicket of U.S. export controls on China from 2022 to 2024. The shifting rules and continuous churn of updates have left observers grasping for stable ground. Moreover, where they exist, explanations are often buried in technical minutiae.

By cutting through the opacity, this paper provides a clear account of the motivations driving U.S. actions, gaps in enforcement, as well as the action-reaction dynamic between the U.S. and China.

In September 2022, National Security Advisor Jake Sullivan declared that securing "as large a lead as possible" in force multiplier technologies like AI was a national security imperative. This charge was led by the implementation of export controls. Far from a blunt instrument to hamper China's immediate progress, they were engineered to degrade its AI capabilities over time.

September 2022 Export Controls: Strangling with Intent

The 2022 export controls were a calculated exercise in brinkmanship designed to throttle China's AI ambitions. Recognizing the escalating computational demands of frontier AI, reliant on thousands of the most advanced microprocessors and memory chips, the United States moved to weaponize its dominance over critical chokepoints in the global semiconductor supply chain, which China was dependent on. It was a bold act of bureaucratic foresight-the United States had anticipated the importance of AI before ChatGPT 3.5 made it undeniable just a month later.

The fact that the complex supply chains needed to produce advanced semiconductors were concentrated in the United States and a small number of allied countries further provided a singular opportunity for export control revision.

The controls targeted the sale of the following:

The United States operated with transitive precision. First, it blocked direct access to the advanced semiconductors needed for frontier AI development. Then, it denied Beijing the tools-the software and machinery-needed to design and fabricate the chips themselves. Finally, it preempted any efforts to indigenize by obstructing access to the critical components required to build the machinery domestically. The cascading sequence of restrictions aimed to trap China in a technological cul-de-sac and sustain the United States' dominance.

The October 2023 Export Controls Update: Closing Loopholes from 2022

However, a familiar pattern of gaps began to take shape. These deficiencies broadly fell into three categories:

1. Technical gaps due to technocratic blind spots/overreliance on static criteria, facilitating simple workarounds;
2. Legal circumvention whereby companies skirted restrictions, sliding just under thresholds, to continue operations within the letter of the law; and

• Illicit evasion of export controls (i.e., illegal acts such as smuggling), undermining their intended effect and challenging enforcement mechanisms.

Technical Gaps from 2022

The 2022 export controls were fundamentally designed to limit the total processing power (TPP) available alongside an interconnect speed threshold.

The assumption was that AI capabilities were driven by raw parallel processing power and the ability to network large numbers of them at the data-center level and in supercomputing facilities-a straightforward equation of performance = quality × quantity. It also was the debut of the Sullivan doctrine of "Small Yard, High Fence"-it allowed the United States to restrict just the most advanced semiconductors to China while allowing U.S. businesses to continue selling less advanced chips and manufacturing equipment to China.

However, there were two critical oversights. First, the rule underestimated the ability to scale AI models by brute-forcing large quantities of less performant chips. Radically inefficient, perhaps, but an effective workaround. Second, the U.S. Bureau of Industry and Security (BIS) overemphasized interconnect speed as a key chokepoint, failing to anticipate how easily it could be circumvented. Manufacturers could simply understate interconnect specifications or market chips with "configurable" speeds, allowing for post-sale adjustments.

These reflected a broader pattern of underestimating China's ability to work around technical thresholds. The most dramatic demonstration came in September 2023 with Huawei's Mate 60 Pro smartphone, which contained a 7-nanometer (nm) processor manufactured by Semiconductor Manufacturing International Corporation (SMIC) -well below the 14/16 nanometer threshold set by export controls. Without access to leading-edge extreme ultraviolet lithography equipment, SMIC deployed a resolution enhancement technique-called multi-patterning-using deep ultraviolet (DUV) lithography equipment. This approach-previously used by Taiwan Semiconductor Manufacturing Company Limited (TSMC) for their early 7 nm production-compensates for DUV's limited resolution by breaking complex chip patterns into multiple simpler exposures, like creating a detailed image by overlaying several less detailed stencils.

2023 Update: BIS recognized that its 2022 approach had over-indexed on individual chip specifications. This recognition led them to first abandon the interconnect speed parameter, a criterion that had proven too easy to game, and second, to reduce the TPP threshold from 4,800 to 1,600 Giga operations per second. This was an acknowledgment that even less performant chips, when used en masse, can achieve significant computational capabilities.

Two additional criteria were introduced to close potential loopholes. A performance Density requirement (TPP/die area) was added to target chips that, while physically smaller and thus having lower absolute computing power, achieved computing efficiency on par with the most performant chips. Additionally, a datacenter criterion was implemented to address chips designed for networking in large-scale systems, recognizing that advanced AI training relies not on individual chip performance but on the efficient clustering of many chips within datacenters.

In addition, following U.S. diplomatic efforts in March 2023, the Netherlands agreed to restrict exports of the most advanced DUV machines (specifically the NXT:2000i) capable of manufacturing chips at 14 nm and below, as well as tools that could be used for multi-patterning techniques to potentially achieve 7 nm or below.

Legal Circumvention

NVIDIA introduced modified versions of its flagship H100 and A100 chips for the Chinese market. The H800 (300 GB/s) and A800 (400 GB/s) were specifically engineered to remain below the 600 GB/s interconnect speed limit while still offering considerable capabilities.

2023 Update: Modified chips such as the Nvidia H800 and A800, amongst others, were added to the controlled list. In fact, U.S. secretary of commerce Gina Raimondo directly criticized Nvidia for skirting trade rules at the Reagan National Defense Forum, stating, "If you redesign a chip around a particular cut line that enables them to do AI, I'm going to control it the very next day."

Illegal Evasion

The 2022 controls primarily targeted entities within China, leaving open the possibility of Chinese firms obtaining restricted semiconductor technologies through subsidiaries or partnerships abroad. There was also evidence of significant stockpiling in China during the period.

2023 Update: BIS extended restrictions to an additional 43 countries beyond China, targeting nations with strong ties to China. This move aimed to prevent the diversion of semiconductor technologies through third countries.

2024 Export Controls Update

Despite best efforts, the broad trio of deficits from 2023 did not disappear. They simply shapeshifted:

Technical Gaps from 2023

While past controls focused heavily on GPUs and raw computing power, they underestimated the crucial role of memory bandwidth in AI system performance. Modern AI training faces what experts call the "memory wall," which refers to the mismatch between the speed of GPU processors and the slower rate at which data can be supplied to them due to limited memory bandwidth.

Think of each GPU as a super-fast assembly line worker in a factory. No matter how quickly the worker can process computational operations, their productivity is severely limited if they cannot receive data quickly enough. They need constant access to their instruction manual (model weights) and raw materials (input data and intermediate calculations). Then enters high-bandwidth memory (HBM).

Previously, traditional dynamic random access memory (DRAM) was like having a single-story warehouse located across town from the factory (i.e., logic processor)-requiring data to travel relatively long distances. HBM revolutionized this by essentially building a vertical automated warehouse directly above the factory floor. Imagine 8-12 floors of storage (DRAM layers) stacked on top of each other, connected by thousands of high-speed elevators (through-silicon vias) and direct access points (micro-bumps). This architectural transformation increased data access speeds more than 130-fold-from 25 GB per second with traditional DRAM to 3.35 TB per second with modern HBM.

The memory wall becomes more acute when considering inter-GPU communication in large GPU clusters required for frontier AI training. These clusters are like thousands of factory floors that need to constantly coordinate their work, sharing gradient updates and synchronizing model parameters. Without HBM, adding more GPUs becomes counterproductive-it will be like trying to feed a superhighway through a country road.

In addition, HBM production demands highly sophisticated manufacturing capabilities and complex integration with GPUs. Only a handful of companies worldwide can produce these components at scale. SK Hynix led the HBM market in 2023 with a market share of 53 percent, followed by Samsung and Micron at 38 percent and 9 percent. Controlling HBM technology could be more effective than controlling GPUs alone. It addresses both a crucial technical bottleneck and leverages a naturally concentrated supply chain.

2024 Update: Given the memory wall in AI development, the 2024 controls ban HBM sales to China from HBM2E (enhanced version of second generation HBM) and above, targeting a key vulnerability. While leading-edge GPUs now use HBM3 memory, Chinese manufacturers have not yet mastered HBM2E. With Huawei's flagship AI processors still dependent on Samsung-produced HBM2E and SK Hynix and TSMC planning HBM4 production by early 2025, these controls aim to keep China at least two generations behind the curve.

Legal Circumvention: Offshore Manufacturing Loopholes

While a piece of semiconductor manufacturing equipment cannot be exported to China from U.S. factories, the same equipment produced in an overseas facility can legally be sold to an advanced logic fab in China, provided no U.S. persons are involved in its manufacturing, sale, installation, or servicing, even if the customer is on the U.S. entity list. This effectively created a backdoor in the control regime.

2024 Update: To address this, the United States introduced a Foreign Direct Product Rule (FDPR) for semiconductor manufacturing equipment (SME). The FDPR subjects products made abroad to U.S. export restrictions if produced using U.S.-origin technology, software, or equipment. This move aims to prevent companies from circumventing restrictions by relocating production or relying on foreign facilities to supply restricted technologies to entity-listed Chinese firms.

Extraterritorial measures are not new to U.S. semiconductor export controls. The 0 percent de minimis rule for advanced lithography, introduced in October 2023, had already set a precedent for such restrictions. The 2024 rule, however, introduced two potent new mechanisms. First, the expansive "single chip de minimis" provision extends U.S. control to any equipment produced in facilities containing even a single U.S. chip-a dramatic expansion of extraterritorial reach. Second, BIS created a new classification (ECCN 3B993) covering "node-agnostic" tools used in both legacy and advanced production, particularly targeting the multi-patterning techniques China has used to push older DUV technology beyond its intended limits. The controls also encompass equipment needed for advanced memory production and developing technologies like nanoimprint lithography that could offer potential workarounds to existing restrictions.

Imperfect and Incomplete Blacklists of Entities

Another challenge lies in identifying, ex ante, which companies are part of the Huawei complex. Corporate structures in China are often opaque, with numerous subsidiaries, affiliates, and indirect partners falling outside the scope of existing controls.

For instance, for chip fabrication, previously, while known Huawei affiliates like PXW (PengChip) and Fujian Jinhua were flagged in the entity list, others such as SwaySure, PXT (Pensun), Qingdao Si'En and Pengjin Hi-Tech operated with limited or no restrictions under the guise of independence.

2024 Update: BIS designated 16 entities under Footnote 5 (FN5) of the Foreign Direct Product Rule (FDPR). FN5 sharpens the FDPR mechanism by targeting high-value Huawei-affiliated firms critical to Beijing's indigenization campaign, aiming to replicate the "Huawei effect," rendering these companies commercial pariahs and excising them from global supply chains. BIS also expanded the SME product scope under FN5 to cut off foreign tool supplies to these priority targets.

In addition, BIS added 140 other companies to its Entity List, including SwaySure, PXT, and Qingdao Si'En, subjecting these firms to U.S. licensing restrictions and the presumption of denial. The list revisions also included investment firms such as JAC Capital and Wise Road Capital, who have promoted China's advanced chip development via strategic investments and acquisitions.

Limitations in Enforcing End-Use Restrictions

Export controls under the Biden administration were designed to narrowly target leading-edge semiconductor production in China. One particularly egregious example, however, illustrates the limits of such an approach. An SMIC Advanced Logic Fab and SMIC Legacy Fab were connected by a wafer bridge with an automated overhead track, allowing wafer transfer between them. Functionally, this created a single, continuous cleanroom and effectively one fab.

2024 Update: Red Flag 27 addresses this by imposing a new end-user verification requirement, ensuring no covered transactions occur with facilities physically connected to sites producing advanced-node integrated circuits.

Illegal Evasion

Investigative reports have exposed large-scale chip smuggling operations, with some bulk orders exceeding $100 million. Importantly, these transactions included not only thousands of advanced semiconductors but also entire servers, which are required at the datacenter level for large-scale AI development.

2024 Update: BIS added eight new Red Flags to enhance compliance efforts. Additionally, BIS has imposed a new "knowledge standard," requiring companies to demonstrate due diligence to ensure exported items are not diverted to prohibited end-uses or unauthorized destinations.

Conclusion

The success of these controls will hinge on their multilateralization. Japan and the Netherlands have already aligned with equivalent restrictions, and the United States has offered carveouts to countries like Germany, South Korea, and Singapore to incentivize their participation in harmonized controls. Yet, this process faces two significant challenges.

First, domestic political instability in key allied nations complicates cooperation. South Korea's recent declaration of martial law has created an unpredictable policy environment, while fragile political coalitions in Japan and Germany risk undermining their ability to make decisive commitments. Second, export control agreements do not occur in isolation, they are influenced by the broader context of U.S. trade policies and diplomatic actions. Aggressive U.S. tariffs or heavy-handed negotiation tactics could alienate allies, making harmonization more difficult. Moreover, export controls are no longer a one-sided game. Beijing's recent moves, such as imposing restrictions on critical minerals and launching a probe into a four-year-old NVIDIA acquisition, signal a shift from strategic restraint to active countermeasures.

In this context, the United States must look beyond refining its control mechanisms and focus on building durable coalitions that not only safeguard national security interests but, more crucially, drive the innovation and competitive edge needed to stay ahead of China.

Barath Harithas is a senior fellow with the Economics Program and the Scholl Chair in International Business at the Center for Strategic and International Studies (CSIS) in Washington, D.C. Andreas Schumacher is a former visiting technology fellow with the Economic Program and the Scholl Chair in International Business at CSIS.

Programs & Projects