Since October 2022, the United States has devoted significant resources to restricting China’s access to artificial intelligence (AI) and advanced semiconductor technologies.[2] The U.S. national security enterprise has undertaken a massive, multifaceted effort to choke off China’s access to cutting-edge AI and related technologies.[2] Export controls on the high-performance semiconductors used to train and inference state-of-the-art AI models have been at the forefront of this effort so far.[2] Although these controls marked the reversal of nearly 30 years of trade policy, they did not achieve all of their intended goals.[2]

"Furthermore, advanced chips are essential for the development and deployment of AI, including frontier AI models." The report argues that "while it’s unrealistic to expect U.S. export restrictions to keep China dependent on U.S. chips forever, the United States can slow China’s progress" through carefully designed export controls on semiconductors.

The October 2022 U.S. export controls ban the export to China of any AI chips equal to or more capable than Nvidia’s A100, aiming to "maximize the United States’ AI advantage over China by hindering China’s ability to develop or run AI models at scale." The rationale is that training and deploying frontier AI models requires aggregating enormous numbers of advanced chips to create very large amounts of computing power. The article estimates that if Nvidia were allowed to export three million H200 chips to China in 2026, this would give China more AI computing power than it could domestically produce until 2028 or 2029, underscoring how controls constrain China’s AI compute growth trajectory.

AI chips, such as GPUs, provide the compute required to train AI models and are a key input in AI scaling.[3] Growth in GPU clusters has been the main driver of compute growth in the past few years, and higher performance, lower latency, higher memory bandwidth GPUs make it feasible to do ever larger training runs.[3] AI scaling could therefore be constrained by the number of state-of-the-art GPUs that chipmakers can produce.[3]

Scaling laws describe how the performance of AI systems improves as the size of the training data, model parameters or computational resources increases.[2] Per the pretraining scaling law, outlined in this research paper, when larger models are fed with more data, the overall performance of the models improves.[2] To make this feasible, developers must scale up their compute — creating the need for powerful accelerated computing resources to run those larger training workloads.[2]

The article finds that "chip export controls have clearly had a significant negative impact on China’s ability to produce advanced chips" and that restrictions on chipmaking tool sales "have significantly slowed the growth of China’s chipmaking capability" such that China "remains a marginal producer of AI chips." However, it concludes that these controls "haven’t significantly limited China’s ability to train cutting-edge models"—Chinese firms like Alibaba and DeepSeek have produced large language models that score highly on benchmarks despite hardware limits. The author argues instead that export controls "may have limited China’s ability to deploy AI" at scale, noting DeepSeek had to restrict access to its R1 model API, likely due to insufficient inference compute, and citing Tencent Cloud’s Wang Qi that deficits of advanced chips could slow AI adoption by limiting inference or driving up costs.

Scaling laws have been fundamental to the remarkable success of foundation models, demonstrating a predictable relationship between performance and the expansion of model parameters and training data.[4] However, the continued application of these scaling laws requires ever-increasing amounts of data and computational resources, pushing the boundaries of what is currently feasible.[4] The computational demands have grown exponentially – from BERT-Large’s training requiring 64 TPU v3 chips to GPT-3’s training on 10,000 V100 GPUs, while training GPT-4 reportedly required over 25,000 NVIDIA A100 GPUs.[4]

Effective October 7, 2022, the United States of America implemented new export controls targeting the People's Republic of China's (PRC) ability to access and develop advanced computing and semiconductor manufacturing items.[4] The export controls directly restrict the PRC's ability to obtain, develop, and manufacture advanced semiconductor technology.[4] Intended restrictions of the export controls include limiting AI chip access, limiting Chinese design capability, stifling advanced chip manufacturing, and limiting access to chip manufacturing technology.[4] This initiative is key to the U.S.'s ambitions in preventing China's access to advanced computing and semiconductors along with limiting its ability to develop and manufacture its own to maintain a global edge in artificial intelligence capabilities.[4]

Export control policies have long targeted U.S. foreign adversaries’ access to advanced AI chips with the goal of limiting their ability to develop advanced AI models.[3] U.S. export controls are evolving from restricting advanced AI chips to targeting a broader set of technologies that includes SME, cloud infrastructure, and even frontier AI models.[3] Persistent challenges – most notably chip smuggling, inconsistent implementation among allies, economic costs to U.S. firms, and adversaries’ ability to develop alternative pathways – suggest that controls are likely to function less as tools to stop adversaries and more as instruments for slowing adversarial progress while preserving U.S. technological advantage over time.[3]

"Comprehensive empirical analysis finds US chip export controls provide 1–3 year delays on Chinese AI development but face severe enforcement gaps (140,000 GPUs smuggled in 2024, only 1 BIS officer for Southeast Asia) and unintended consequences (DeepSeek achieved GPT-4 parity at 1/10th compute, forcing efficiency innovations)." The analysis notes that export controls have "demonstrably disrupted Chinese AI development in the near term, creating chip shortages and forcing workarounds," but also that "expert assessments of the time delay imposed by export controls range widely, from 6 months to 5 years… For training models at the scale of GPT-4, most analysis suggests a 1–3 year delay based on reduced hardware access." Controls appear "most effective against large-scale model training requiring thousands of high-end chips" and "function less as tools to stop adversaries and more as instruments for slowing adversarial progress while preserving U.S. technological advantage over time."

The Biden administration’s export controls "choked off" China’s access to advanced Nvidia GPUs, widening what White House AI adviser David Sacks described as a 3–6 month lag of China’s AI sector behind the United States. The article reports that China’s AI industry is "deeply dependent" on American GPUs, with an estimated 75 percent of chips powering AI training in Chinese data centers running on Nvidia’s CUDA platform, and argues that Chinese companies "can ‘keep the lights on’ with domestic chips" but "can only sustain ‘frontier-level’ progress" with continued access to Nvidia hardware. After export-compliant "green‑zone" GPUs became available again, the author predicts China’s AI development "will likely go into overdrive" as firms resume full-scale model training and inference, implying that stricter restrictions had previously slowed their frontier progress.

The scaling hypothesis is a law (technically three laws) that predicts AI model performance based on three key factors… 3. The amount of computing resources available for training – the biggest models are trained on hundreds of thousands of advanced computing chips, or semiconductors.[6] The scaling law says that the reason models keep getting better is that we’re throwing more computing resources at them.[6]

US export controls on chips and hardware alone will not prevent China from further developing advanced AI.[7] This policy aims to slow down the progress of Chinese AI and give the US more time to advance domestic AI capabilities.[7] Evidence does suggest that export controls limit computational resources for Chinese companies.[7] But they did not stop DeepSeek from releasing its high-performing model for far cheaper than US competitors; these innovations were driven by optimizations in memory management and the use of synthetic data rather than access to the most advanced chips.[7] The assumption that chips are a permanent chokepoint has already been undermined by algorithmic adaptation, enforcement gaps in export controls and a grey market that is growing faster than the regulatory apparatus designed to contain it.[7]

The article states that "DeepSeek’s models are a stark illustration of why U.S. export controls on advanced computing chips, instead of impeding China’s AI progress, may actually be accelerating it." It explains that "forced to operate under a far more constrained computing environment than their U.S. counterparts, AI engineers in China are innovating in ways that their computing-rich American counterparts are not." As "a direct result of U.S. controls on advanced chips, companies in China are creating new AI training approaches that use computing power very efficiently," suggesting that progress in AI capabilities can continue despite restricted access to frontier chips.

In this policy analysis, John Villasenor argues that "blocking China from accessing relatively high-performing AI chips" is unlikely to help the U.S. maintain AI leadership; instead, "starving China’s supply of U.S.-designed AI chips will have the opposite effect" by pushing China to accelerate development of its own AI chip capacity and ecosystem. He notes that the more challenges China faces in obtaining advanced U.S. chips, "the more it will invest in growing its own capacity," citing Huawei’s strides and the likelihood that its AI investments will increase in response to shortages. The piece suggests aggressive export bans may have complex, potentially counterproductive effects on long-term AI capability rather than straightforwardly slowing progress.

Compute is a vital part of scaling; scaling is only possible through increasing compute.[8] Training models with more parameters or larger datasets requires more computing power.[8] Accordingly, compute greatly influences both model size and dataset size.[8]

The United States currently stands at a critical strategic crossroads regarding its policy on advanced semiconductor exports to China, particularly those used for artificial-intelligence (AI) systems.[6] Fundamentally, it restricts the export to China of cutting-edge graphics processing units (GPUs) used for AI applications.[6] These restrictions have created a complex set of responses and adaptations. While they have slowed China’s advancement at the cutting edge, they haven’t stopped it entirely.[6] The case for maintaining or strengthening chip export controls, championed by figures like Anthropic CEO Dario Amodei, rests on the prediction that transformative AI capabilities will emerge relatively soon; under this scenario, denying China access to cutting-edge chips could meaningfully diminish their ability to deploy advanced AI systems at scale.[6] Conversely, skeptics like Ben Thompson argue that in a longer timeline the primary effect may be to deny revenue to U.S. firms while accelerating China’s push for indigenous chipmaking capabilities.[6]

This legal advisory describes expanded U.S. export controls on advanced GPUs and certain AI models under the January 2025 "AI Diffusion rule," which placed new restrictions on GPUs under ECCNs 3A090 and 4A090 and on closed‑weight AI models trained with at least 10^26 computational operations. China is listed among "restricted, arms-embargoed countries" subject to stringent licensing requirements with a presumption of denial, and the memo warns that strict controls on transfers to China—including for Chinese nationals or Chinese-controlled firms abroad—affect any company that exports, reexports, or utilizes advanced GPUs. While it focuses on compliance rather than technical impact, it documents the regulatory intent to constrain advanced compute and AI model diffusion to China.

This blog breaks it down into three "laws" of scaling compute: 1️⃣ Pretraining Scaling: - Bigger models + more data + more compute = better baseline performance.[7] 3️⃣ Test-Time Scaling ("long thinking"): … Great for complex tasks (e.g., detailed coding, strategic planning), but it can demand 100× compute at inference.[7] Ever-Growing Compute Needs: From training to inference, compute requirements are skyrocketing - expect a surge in specialized, accelerated hardware.[7]

US views progress in AI as a significant axis of international competition. The restrictions laid out in Oct 2022 were progressively amended.[5] The first round of the export controls were launched on October 7th 2022, targeting China’s access to high-end semiconductors and related equipment used for training frontier AI systems.[5] Because of this, the goal was to identify chokepoints in the global supply chain and restrict access to equipment needed for large AI training runs, but without affecting the sale of semiconductor equipment used for consumer purposes.[5] Later regulations set worldwide licensing rules on the number of US-designed chips that can be sold to different countries, dividing the world into tiers with different limits on access to high-end compute.[5]

The article explains that current U.S. policy "aims to slow down the progress of Chinese AI and give the US more time to advance domestic AI capabilities" by restricting access to advanced chips and chipmaking equipment. It finds that "controls *have* severely limited China’s share of the global AI infrastructure market, because, lacking competitive hardware, Chinese cloud computing firms have been unable to establish much, if any, AI infrastructure outside of China." It also states that "chip export controls have clearly had a significant negative impact on China’s ability to produce advanced chips" and "restrictions on chipmaking tool sales have significantly slowed the growth of China’s chipmaking capability." However, "Chinese models have advanced despite export controls" and "we must conclude that chip export controls have not seriously slowed improvements in Chinese model quality," as firms like Alibaba and DeepSeek have produced "impressive large language models" that are competitive on benchmarks.

In the past few years, AI labs have adopted a “more is more” approach to scaling LLMs. By introducing more parameters, data and compute, the result was a smooth consistent improvement in model performance in the form of a power law.[1] Performance it turned out depends much more on scale than on the algorithm.[1]

Scaling laws offer a way to forecast model behavior by relating a large model's loss to the performance of smaller, less-costly models from the same training distribution.[10] By empirically fitting a power-law relationship between loss and resources such as data, model size, and compute, researchers can predict how much additional compute is required to reach a target performance.[10]

A DW News segment explains that the United States has spent years "trying to starve China of advanced chip technology" with the explicit goal "to slow China's progress in artificial intelligence." The report notes that powerful processors such as Nvidia’s Blackwell chips are critical for training cutting-edge AI systems and that both Blackwell and similar high-end GPUs are banned from export to China. The correspondent states that "without access to Nvidia's technology, China falls further behind in the artificial intelligence race," reflecting a widely held view among U.S. officials and industry analysts about the impact of chip access on AI capabilities.

For decades, Moore's law (transistor counts doubling every couple of years) and Dennard scaling… underpinned the idea that simply throwing more computing power at AI problems would keep delivering big gains in performance.[9] But as Moore's law slows and the cost and energy use of computation continue to rise, there are growing concerns that 'bigger is better' scaling laws will eventually hit practical limits.[9]

To prevent opposing access to transformative AI technologies, the Action Plan promotes strengthening the nation’s export control strategy.[1] It proposes the implementation of “creative” approaches to export control enforcement around AI, including location verification features on advanced AI-capable US chips and expanding end-use monitoring in high-risk regions.[1] It also calls for the enhancement of export controls on semiconductors to address perceived gaps in existing US export controls, such as new controls on semiconductor manufacturing sub-systems.[1]

This regulatory action imposes global restrictions and onerous licensing requirements on US exports of advanced integrated circuits.[9] SIA supports targeted export controls that protect national security interests but believes that broad, unilateral controls risk harming the U.S. semiconductor industry without effectively advancing national security.[9]

A short news post reports that the United States moved to "close a loophole" that may have allowed Chinese companies to access advanced Nvidia AI chips via overseas subsidiaries. It reiterates that these powerful processors are "critical for training cutting-edge artificial intelligence systems" and that the broader U.S. strategy has been to "starve China of advanced chip technology" in order to slow its AI progress. The post underscores that without such chips, Chinese entities face constraints in building advanced AI systems at parity with the United States.

In a LinkedIn commentary, Aaron Ginn argues that "GPU export controls didn’t slow China’s AI advancements, and they never would," responding to a Wall Street Journal letter supporting strict restrictions. He claims that despite increasing U.S. controls and Nvidia not shipping certain GPUs like the H20 to China after the ban, "China’s AI breakthroughs continue" and "undermine American AI". This reflects a skeptical view of the effectiveness of hardware export restrictions in materially slowing China’s AI progress, contrasting with analysts who emphasize compute constraints.

Bloomberg and Reuters reports from March 5 said the U.S. was drafting global AI chip export controls, which sent semiconductor stocks lower.[8] Commenters discuss that restricting access to advanced AI chips is intended to limit rivals’ ability to train frontier AI models, though investors debate how effective such controls will be and how much they will impact companies’ revenues.[8]

LLMs use a fixed amount of compute per token they output… What test time compute scaling ala o1 does is add another scaling axis.[5] If we pick the most efficient point across all axes for a given amount of total compute (training+lifetime inference) we get a large jump in performance vs. not using the test time scaling axis.[5]

The 2021 final report from the U.S. National Security Commission on Artificial Intelligence argued that if China leapfrogs the United States and its allies in chip technology, it will gain the upper hand militarily "in every domain of warfare"; restricting access to advanced semiconductors was framed as central to maintaining a global edge in AI capabilities.[4] This reflects a policy assumption that limiting rivals’ access to advanced chips can slow their progress in AI and associated military applications.[4]