Are Consumer GPUs a Problem for US Export Controls?
Executive Summary
Advanced AI models are trained and run with methods that require vast amounts of computing power (“compute”). Typically, these models require thousands of expensive, specialized AI chips. The most advanced AI chips are subject to US export controls and, therefore, cannot be legally exported to countries of concern, notably China. Most of these AI chips are graphical processing units (GPUs) designed specifically for use in data centers (“data center GPUs”). However, some US-designed consumer GPUs (also called “commodity GPUs” or “gaming GPUs”) are also suitable for some compute-intensive workloads and are largely exempt from export controls. As a result, they are potential substitutes for data center GPUs in China.
While the possible use of consumer GPUs to circumvent US export controls is a known issue, the significance of this issue has, to our knowledge, not been comprehensively analyzed. This report addresses this gap, including by analyzing three scenarios where Chinese actors aim to acquire as many high-end consumer GPUs as possible, given their constraints.
The report makes four key points. One, there is evidence that Chinese actors use US-designed consumer GPUs for AI workloads today. Two, under the current policy and enforcement landscape, Chinese actors could likely acquire millions of consumer GPUs this decade. Three, those consumer GPUs would likely be sufficient for achieving significant AI capabilities. Four, the Bureau of Industry and Security (BIS), other policymakers, and GPU makers can implement measures to reduce flows of consumer GPUs to China and other countries of concern; we outline a number of recommendations for these actors below.
Key Findings
Chinese actors are likely using US-designed consumer GPUs for small or moderate-size AI workloads today. News reports indicate that Chinese workshops are reconditioning consumer GPUs to make them more suitable for AI workloads and/or data centers by upgrading memory chips or form factors. Reports also indicate that such GPUs are being stockpiled by Chinese resellers.
A well-resourced effort to acquire consumer GPUs could enable a supercomputer of hundreds of thousands or even millions of GPUs over the next ten years. However, we estimate that the number of GPUs acquired could be reduced by half an order of magnitude with improved policy and enforcement. For comparison, cutting-edge AI models are typically trained with tens of thousands of data center GPUs. Consumer GPUs have a price advantage over data center GPUs, with high-end consumer GPUs costing less than a tenth of data center GPUs and being about three to six times more cost-effective in terms of theoretical performance. However, the stockpiler’s access to the consumer GPU market may be partially limited, and—if US export controls are tightened—increasingly so in the future.
Such quantities would provide enough compute to develop significant AI capabilities, though falling short of what is needed to keep parity with the frontier of AI progress. Though consumer GPUs are cheaper, we estimate that data center GPUs would be about three times faster than consumer GPUs when training large models in 2025, growing to about five times in 2035. For comparison, the stockpile would likely amount to about a fifth or less of the forecasted total compute available from AI chips produced indigenously in China. Nevertheless, such an effort would be sufficient to quickly replicate OpenAI's GPT-4 and, within five years, to train a model with the amount of compute we anticipate OpenAI's next major model, "GPT-5", to utilize. These forecasts are based on quantitative simulations detailed in Appendix B.
Whether or not Chinese actors will make major efforts to acquire consumer GPUs for use in data centers hinges largely on the availability of domestic alternatives. If China can significantly increase the production of its indigenous AI chips and build a thriving software ecosystem around them, there is little incentive to use consumer GPUs except in the short term. This effect is possible because it seems likely that indigenously produced Chinese AI chips will be about as performant as high-end consumer GPUs. Other factors relevant to whether consumer GPUs are used to circumvent export controls are the viability of AI chip smuggling and access to compute via Western and “third country” cloud services.
On a technical level, consumer GPUs could potentially be used as alternatives to data center GPUs. There are currently a number of practical issues with using consumer GPUs for large AI workloads, notably related to driver software, form factor, and limited access to support and warranty due to driver licenses prohibiting the use of many of these GPUs in data centers. However, those issues are likely manageable, given a substantial engineering effort. Given such an effort, consumer GPUs would likely only suffer from somewhat inferior throughput, mainly due to their lower theoretical performance and memory bandwidth.
Recent trends and business incentives may suggest a growing gap between data center and consumer GPUs in terms of their suitability for large AI workloads. Data center GPU performance is improving at a faster pace than consumer GPU performance. GPU makers are generally incentivized to make consumer GPUs less useful for large AI workloads in order to encourage enterprises to purchase more expensive products, like data center GPUs. This phenomenon has, to some extent, already occurred, with Nvidia artificially limiting the AI performance of some consumer GPUs in recent years. However, these trends are uncertain and could reverse, particularly if consumer GPUs provide the sole means for GPU makers to capture a portion of the considerable Chinese AI chip market.
Recommendations
While high-end consumer GPU exports to China are subject to licensing requirements, they qualify for a special License Exception at the discretion of BIS because they are not “designed or marketed for use in data centers”. As a result, today, Chinese entities likely have partial, but not unfettered, access to consumer GPUs through legal channels and some access through illicit trade. To prevent Chinese actors intending to develop advanced AI capabilities from acquiring consumer GPUs, we outline recommendations for BIS, other policymakers, and industry below.
The Bureau of Industry and Security: As the agency responsible for administering and enforcing the AI chip controls, BIS is exceptionally well positioned to address these concerns. Moving forward, BIS should:
Retain the data center/consumer GPU distinction, for now: It is highly difficult to design precise technical classifications that are robust to “gaming” by GPU designers, and to our knowledge, none have been proposed that are sufficient. Therefore, BIS should retain the distinction between chips designed or marketed for data centers and other chips for the time being, while seeking to eventually replace the distinction with a technical classification.
Monitor and assess consumer GPU flows: This issue is multifaceted and fast-moving. As such, BIS should keep continuously monitoring and assessing the flow of consumer GPUs to countries of concern and whether they are used for large AI workloads.
Deny License Exceptions to listed entities: Listed entities pose a higher risk to US national security, and it is, therefore, especially important to restrict their access to compute. BIS should deny License Exceptions for exports to listed entities or those closely associated with such entities via ownership or partnerships.
Deny License Exceptions for consumer GPUs suitable for large AI workloads: Future generations of consumer GPUs may be better suited for developing advanced AI systems. BIS should assess those GPUs and deny License Exceptions for any future GPU that is highly suitable for large AI workloads.
Require robust due diligence: A handful of US companies dominate the consumer GPU market. With stronger due diligence processes, these companies can help identify more instances of smuggling and other red flags. BIS should require these exporters to commit to appropriate due diligence measures as a condition for granting License Exceptions.
Other policymakers: To support BIS in its mission, other policymakers (e.g., members of Congress) should:
Provide guidance for BIS: Policymakers should make it clear that BIS can and must restrict high-end consumer GPU exports if it assesses their export to pose a risk to US national security.
Work to increase BIS’s budget: BIS is currently understaffed and reliant on outdated tools and methods. To ensure that BIS is properly equipped to fulfill its mission, policymakers should work for Congress to approve BIS’s requested additional funding: $223M for the fiscal year 2025, up from $191M.
GPU makers: The GPU industry has specialized knowledge and insights that can help detect smuggling and prevent the misuse of consumer GPUs in ways that conflict with US interests. To address these concerns, GPU makers should:
Void the warranty for consumer GPUs used in data centers: Voiding the warranty for consumer GPUs that are suspected to have been used in data centers in countries of concern discourages the use of consumer GPUs in data centers. Nvidia’s consumer GPU driver license already prohibits use in data centers.
Rescind support for consumer GPUs used in data centers: Publicly committing to offering no support (e.g., customer support or software updates) for entities suspected of using consumer GPUs in data centers would also serve to discourage such use.
Exercise heightened scrutiny when exporting: GPU makers and other exporters should avoid exporting high-end consumer GPUs to distributors if there is a reasonable risk that the GPUs will be used in a data center.
Design consumer GPUs to be more clearly distinct from data center GPUs: To help BIS more effectively distinguish data center GPUs from consumer GPUs and thereby make more targeted export controls, GPU makers should explore technical solutions to make consumer GPUs more distinct from their data center counterparts.
A Note on Our Methodology
The takeaways and recommendations in this report are grounded in quantitative simulations of consumer GPU stockpiling in various scenarios, combined with qualitative considerations. We assess the technical viability of using consumer GPUs for large AI workloads—and associated performance penalties—by reviewing the literature and interviewing distributed computing experts.
The report studies three scenarios through 2035:
In the baseline scenario, Chinese actors (“the stockpiler”) legally import consumer GPUs while avoiding provoking stricter export controls from the US and its allies and partners. We consider this scenario to be likely if controls on Chinese chip making are largely successful and there are no major changes in export policy and enforcement related to consumer GPUs.
In the massive investment scenario, the stockpiler invests substantially into legally importing consumer GPUs while retaining broad access to the consumer GPU market. We consider this scenario to be implausible since imports of that scale would likely trigger countermeasures by BIS, and it should, therefore, be seen as an upper bound for what could be gained by stockpiling consumer GPUs.
In the improved enforcement scenario, the stockpiler legally imports consumer GPUs, albeit with reduced access to the consumer GPU market, and is also forced to acquire consumer GPUs through illicit trade.