Seizing on a historic boom of investment in artificial intelligence (AI), President Donald Trump unveiled his administration’s AI Action Plan at the end of July, with a goal of accelerating the development and diffusion of AI in the US and around the world while discarding the Biden administration’s cautious approach to regulating AI.
Trump’s plan covers wide ground from workforce and infrastructure to refining export controls viewed as barriers to AI development in the US and export of US technology abroad, although provisions addressing concerns over safety and national security risk management are also included. The plan targets environmental regulations that have delayed power and data center buildouts as well as a growing patchwork of state AI laws that could fragment the US AI market. Ignoring earlier criticism from Trump, it also recommends continued support for the Biden-backed Creating Helpful Incentives to Produce Semiconductors (CHIPS) and Science Act of 2022, which has achieved historic success in attracting investment to produce advanced semiconductors in the US. Missing, however, is any easing of the administration’s tough policy discouraging immigration, a problem because the US needs to keep relying heavily on foreign talent to compete with China’s strong pipeline of AI researchers.
The AI Action Plan demonstrates the administration’s bias in favor of more diffusion of US technologies, even when that will come to some extent at the expense of those technologies getting into the hands of China. By contrast, the Biden administration leaned more heavily into worldwide restrictions and closed models to avoid enhancing China’s AI capabilities. The administration also supports more open AI model development in the US, a contrast to the focus of most US AI labs today on closed models with clearer economic returns and less risk of misuse, including by China.
A surprise decision this week by OpenAI, creator of ChatGPT, to release two new openly shared AI models, is the kind of shift that Trump’s plan is designed to encourage. OpenAI’s move toward open models gives it ammunition to compete with Chinese AI companies that today dominate open models, and it is a sign that such competitiveness concerns are beginning to be seen as outweighing the risks of more powerful open models.
Regarding export controls, the Trump plan, accompanying an executive order on AI exports and other actions, signals a generally more open approach to exports of US AI technology. It seeks to more effectively implement existing sprawling semiconductor controls and synchronize the US approach with allies. But the plan does little to clarify how the administration will balance reducing red tape to boost global US AI technology adoption with the risk of diversion of US AI strengths to Beijing or convince other countries to join US controls despite the risk of Beijing’s wrath.
Push for Open AI Models
AI models from US firms like OpenAI, Google,[1] Anthropic, and xAI are consistently at the top of leaderboards for capability in domains like math, coding, and writing. But the highly publicized Chinese tool DeepSeek is only the tip of the iceberg in capable Chinese AI models, which include Alibaba's Qwen and Moonshot’s highly praised Kimi.[2] These tools are also different in a key way: Chinese models tend to be “open weight,” meaning anyone can freely download them, look under the hood to examine how they work, customize or “fine tune” them for specific applications, and run them on their own devices (or cloud) without sharing any data with the developer. By contrast, US firms keep their best models as a secret sauce that users can run only through them.[3]
US companies have good reasons to focus on closed models. First, with frontier model development costs rising at rate of 3.5 times per year into the hundreds of millions of dollars per model, releasing the results for free would make it challenging for developers to recoup the costs. Second, US companies worry that freely downloadable models can be used by their rivals or—more worryingly—stripped of their safety protections by criminals, terrorists, or military adversaries. The risk of misuse, from fraud to generation of novel biological weapons, rises as models become more capable.
All these tradeoffs are complex, but Chinese dominance of open weight models changed the calculus for the US. If Chinese firms continue to push out the best open models that are not far behind closed US models, advanced capabilities will proliferate no matter what the US does. Some users with sensitive data and limited budgets—such as those in government, health, and research—may prefer open source models on their own devices or supplied by cloud providers,[4] despite their lower capabilities today. Many startups building applications on top of AI are building not on US models but on Alibaba’s open Qwen. Both Qwen and DeepSeek outperform OpenAI’s just released open weight models in intelligence as well, illustrating that competing with China in this space will be an ongoing challenge. As the Trump AI Action Plan says, these open models could “become global standards in some areas of business and in academic research worldwide.” OpenAI clearly has China in mind when saying in its latest release that “broad access to these capable open-weights models created in the US helps expand democratic AI rails.”
Accordingly, the plan strongly favors the US developing more open weight and even fully open source models. The administration leaves it up to model developers to weigh the risk of developers in and outside the US choosing to build on more open Chinese models that encode censorship or (even worse) against the risk of misuse of releasing powerful models without durable safety protections.
Export Controls
One of the most consequential debates in Washington today is over how to balance the benefits of exporting US AI chips and models against the risks that exports to third countries become diverted to China or are otherwise leveraged against US interests. The Biden administration’s AI diffusion rule would have separated the world into a small group of allies with unlimited access to US AI hardware and models while attempting to constrain computing power and model deployment in most countries. The Trump administration rescinded that rule as hampering “global adoption of U.S. technologies.” Until a new rule is put in place, firms in most countries can purchase unlimited US-designed AI chips without a license and even rent access to them to Chinese AI firms without breaking the law.
The Trump plan does not say what if anything will replace the Biden era AI diffusion rule to handle these concerns that US technology could fuel Chinese AI capabilities. It does not even generally call for more US export controls, so the administration may use other methods like pressuring chip importing countries to ban re-export or rentals of US AI chips to China. Instead it aims for better enforcement of existing controls, including the backing of a controversial congressional proposal to embed trackers in US AI chips, together with upping pressure on US allies to adopt controls in line with those from the US. It suggests plugging loopholes, but only in a narrow category of real concern: expanding controls on equipment like German lasers and mirrors that can help China develop homegrown alternatives to US and allied chipmaking equipment. If China can make this equipment good enough, constraints on advanced chipmaking in China will be useless, and the US will lose its ability to constrain China through export controls of chips.
Bringing Allies on Board: Worthy Goal, Uphill Battle
The Trump administration’s call for multilateral adoption of US controls is positive because the US cannot impose effective controls alone. But the Biden administration struggled to convince even a small group of allies to impose controls. Today it is even more challenging for two reasons. To begin with, the administration’s export control policy has left allies unclear whether it intends to impose a harder line or pursue a grand tech bargain with Beijing that could broadly relax US export controls on China. The administration sent mixed signals when it imposed and then quickly removed export controls on chip electronic design automation (EDA) tools to China earlier this year. Its reversal was apparently designed to gain short-term leverage over Beijing. The US also blocked and then unblocked Nvidia’s H20 AI chip, but the administration has left it ambiguous whether the reversal is related to the trade negotiations or a broader shift in US export control policy.
Such confusion takes a toll. Much of the Biden administration’s diplomatic efforts with allies like Japan and the Netherlands involved discussions over where to draw the line on exports of chipmaking equipment to China. Future negotiations with allies will be more challenging because of their doubts that the US has a firm line it will stick with.
Those reversals and the reticence of allies to impose strict export controls on China stem in part from a common concern—Chinese retaliation. Allies have for years warned the US about the costs they faced when China retaliated for their controls on chips. The Biden administration did not take these concerns as seriously, in part because Beijing refrained from retaliating against the US controls until the end of President Biden’s time in office. But China now knows it holds powerful cards, and it is willing to play even against the US.
Recent dealings between the Trump administration and China have been anything but consistent. Once China targeted the administration with retaliation, the US lifted certain export controls in exchange for Beijing granting licenses for things US industry needs, like critical minerals. Even if administration officials insist that the core of export controls on advanced chips are not up for easing, the US will have a hard time arguing that allies should take the hit of Beijing’s retaliation if the US caves when targeted.
Notes
1. Google’s flagship Gemini models are closed, but Google also has a class of open weight models known as Gemma.
2. When an AI model is trained, a set of “weights” in a gigantic equation shape what the model outputs for a given input. The weights can be likened to recipes, where closed models have a secret sauce rather than giving the recipe away for free to be replicated and tweaked.
3. Meta, long the main champion of open models in the US, has struggled to keep pace with Chinese advances and is reportedly shifting to closed models.
4. Companies today can run DeepSeek and other Chinese models on major US cloud computing providers.
Data Disclosure
This publication does not include a replication package.
