Last December, the Chinese AI firm DeepSeek reported training a GPT-4-level model for just $5.6 million, challenging assumptions about the resources needed for frontier AI development. This perceived cost reduction, and DeepSeek’s cut-rate pricing for its advanced reasoning model R1, have left tech stocks plunging and sparked a debate on the effectiveness of U.S. export controls on AI chips.

Some argue that DeepSeek’s efficiency breakthroughs mean the controls have backfired and must be lifted. But this view overlooks the bigger picture: DeepSeek’s success in fact underscores the need for smarter export controls. DeepSeek exploited gaps in current controls, such as exports of chips to China that matched U.S. performance despite the initial October 2022 rules, chip smuggling, inadequate oversight on chip manufacturers like TSMC, and slow regulatory updates that enabled stockpiling. 

Moreover, export controls must keep pace with AI developments. As deployment becomes increasingly important for capabilities, restricting exports of chips optimized for deployment workloads, like the Nvidia H20, is vital. Maintaining U.S. leadership in computing power is one of the best tools for countering Chinese AI ambitions, though it must be part of a broader strategy.

Putting DeepSeek’s efficiency advances into context

AI has a long history of algorithmic innovations. Over the last 12 years, algorithmic advances have halved the computing power needed to achieve the same performance roughly every eight months — a staggering 262,000-fold reduction in compute requirements. On top of algorithms, hardware improvements double the performance for the same cost every two years. Combining these two trends, AI capability gets significantly cheaper over time. For example, in 2017, training an image classifier to 93 percent accuracy cost more than $1,000. By 2021, that cost had dropped to only $5, a reduction of over 99 percent. 

DeepSeek deploys its reasoning model R1 at only 4 percent of the cost of OpenAI’s o1. However, it’s unclear if DeepSeek truly has an edge, as companies like Google haven’t yet released pricing for their reasoning models,​​ and some analysts suggest DeepSeek’s profit margins might be thin. And while DeepSeek’s recent advances are impressive, ongoing efficiency gains in AI development are following predictable industry trends, making capabilities increasingly accessible. This raises the stakes for maintaining a strategic advantage in compute, which determines the extent to which efficiency gains can be scaled.

AI chips matter more than ever

DeepSeek may have achieved V3 with a smaller compute budget than others, but the amount of compute still matters. Efficiency gains generally have two effects: they enable re-creating capabilities with less compute (as seen with R1) and achieving more advanced capabilities with the same resources. Companies like OpenAI, Google, or Anthropic — each with access to hundreds of thousands of cutting-edge AI chips — can leverage these same discoveries to train even more powerful models (assuming they haven’t already made similar breakthroughs that remained private for competitive and security reasons). 

o1 and R1 reveal a new scaling law where performance improves with more compute not only during training, but also during deployment. From 2012 to 2023, scaling focused on training ever-larger models on larger datasets, requiring more compute. In 2024, an additional paradigm emerged: reinforcement learning (RL) that trains models to generate chains of thought, enabling longer “thinking” and better responses. Companies are rapidly adopting RL-based reasoning, which remains early in its scaling curve and might continue advancing parallel to pre-training improvements. Many industry leaders increasingly expect significant gains ahead, with growing confidence in reaching “Artificial General Intelligence.”

While Chinese chips will inevitably improve, export controls can influence the pace of advancement and help maintain or increase the technological gap as the U.S. and partners also improve.

Finally, AI development isn’t just about a single training run, it involves experimentation and rapid iteration. Training an AI model is like a chemical experiment, where access to compute allows companies to use and try many different mixtures before landing on a final formula, with failed experiments and their resources discarded. Deployment plays a key role in this process, enabling capability feedback loops by generating synthetic data and refining reasoning through repeated interactions, similar to AlphaGo’s self-play. The more compute a company has, the faster it can achieve large-scale deployment that meets user demands for response quality and speed.

Notably, DeepSeek’s efficiency gains may stem from its extensive pre-export-control compute access, as it operates Asia’s first cluster of 10,000 Nvidia A100 chips. DeepSeek’s founder, Liang Wenfeng, openly acknowledges that “the embargo on high-end chips” remains the company’s primary constraint. With greater access to advanced AI chips, American companies can run far more experiments than their competitors, giving them a significant edge.

The U.S. must strengthen export controls

DeepSeek claims to have trained V3 on Nvidia H800s, chips designed to comply with October 2022 U.S. export controls but which match the performance of the restricted H100. While DeepSeek’s success reveals gaps in early export rules, it does not show that they can’t work. The Department of Commerce realized its mistake a month after the October 2022 controls but only revised rules to ban H800 exports in October 2023. Had Commerce been faster and established working controls earlier, DeepSeek would have faced greater difficulty training the model, needing to use H20s with a 6.7 times worse computational performance than the H100. 

In October 2024, reports revealed that TSMC had produced, at minimum, hundreds of thousands of export-controlled AI chip dies — the integrated circuit used in AI chips — for Huawei, a violation later addressed by the January 2025 foundry due diligence rule. Despite Huawei’s attempts to develop AI chips, it is currently uncompetitive and is falling further behind due to semiconductor equipment controls. Its 2022 chip, the Ascend 910B, offers only a 1.2 times improvement over the first-generation Ascend 910, whereas industry leaders are tripling performance between generations. 

The weaker quantity, quality, and software ecosystem of domestic Chinese chips explains DeepSeek’s reliance on Nvidia rather than Huawei to train V3. However, maintaining this lead requires vigilant enforcement of export controls. While Chinese chips will inevitably improve, export controls can influence the pace of advancement and help maintain or increase the technological gap as the U.S. and partners also improve.

Critics argue that export controls backfire by forcing Chinese companies like DeepSeek to innovate more efficiently, but this view is flawed. First, efficiency gains are a natural aspect of AI development, with leading U.S. AI companies constantly optimizing performance within fixed compute budgets (leading to smaller models like Claude 3.5 Haiku and OpenAI’s GPT-4 mini). Export controls impose cumulative constraints rather than creating absolute barriers. Second, while compute scarcity may incentivize efficiency innovations, it also constrains experimentation and scaling, limiting the discovery and impact of advances. If less compute truly drove better innovation, we would expect startups with limited compute to lead in AI, not companies that have invested over $500 million in AI chips. Third, export controls need time to take effect. DeepSeek acquired its 10,000 A100 cluster before restrictions and trained V3 on H800s, an initial mistake now corrected.

Steps toward smarter controls

As AI technology evolves rapidly, export controls must become more targeted and responsive to new AI developments. One key example is the growing importance of scaling AI deployment compute, as seen with reasoning models like o1 and r1. AI chips with high memory bandwidth are essential for AI deployment, which led to the December 2024 controls on high-bandwidth memory (HBM) units. These units can be packaged into chips to enable high memory bandwidth. Notably, reports of the U.S. government’s plan to ban HBM exports to China emerged in July 2024, but the restrictions weren’t implemented until December. This extended delay allowed China to stockpile HBM units — likely accumulating enough to enable Chinese domestic chip production for a while. The U.S. should have moved faster. Now, the U.S. faces a new challenge that must be quickly addressed: chips with high memory bandwidth, such as the Nvidia H20, can be exported. Failing to close this gap would allow China to bypass HBM export controls, undermining U.S. efforts to limit China’s AI deployment capabilities.

Second, the U.S. must increase the capabilities of the Commerce Department’s Bureau of Industry and Security (BIS), responsible for export controls. In January 2025, the U.S. government issued the AI diffusion framework to address critical gaps such as chip smuggling and Chinese entities building data centers in other countries, further elevating BIS’ role. Howard Lutnick, President Trump’s nominee for Commerce Secretary, has emphasized that he’s “thrilled to empower BIS,” which is promising as BIS remains chronically underfunded and understaffed. An empowered BIS would hire technical staff with chip hardware expertise and build internal capabilities to detect and prevent export control violations. 

While U.S. algorithmic advantage has weakened for now, China remains constrained by access to advanced AI chips, which increasingly matter for AI development and deployment. The U.S. must act decisively to maintain this edge by swiftly and strategically updating export controls in response to new AI developments and empowering BIS. The alternative — allowing unrestricted flow of advanced AI chips to China — would squander America’s compute advantage at a critical moment in AI development.

---