On August 4th, the White House met with executives from the leading frontier AI companies. The meeting was initially aimed to clarify the June 2nd Executive Order (EO). According to the EO, within 60 days of its release, a new voluntary framework will be introduced for U.S. organizations to submit frontier models for governmental review prior to release. The temporary ban of Anthropic’s strongest models (Fable 5 and Mythos 5) to foreign nationals in June, as well as OpenAI’s limited release at the request of the government, expedited the need for greater clarification of the EO.
This is not the only regulatory area pertaining to AI that is undergoing major shifts and uncertainty. First, while there is great flux in the governance of exporting U.S. frontier models, there also are increasingly vocal arguments and movement toward banning the import of Chinese AI models and hardware. Second, as we described in the previous blog, Chinese AI models are dominating global market share. Many contend this is due to their open weight approach, which is cheaper and requires less compute, compared to the closed weight approach of most U.S. companies. Silicon Valley is divided on this, with the duopoly of Anthropic and OpenAI largely pushing for closed models while others argue for open models.
As we noted in the first blog in this series, AI is at a significant inflection point as autonomous agents circumvent guardrails and gain unauthorized access to systems. In fact, in the short time since that post, the UK’s AI Security Institute revealed new capabilities that combine autonomy with deception, representing a “shift in the risk landscape”. Meta also announced that their AI models broke loose and gained unauthorized access. How governments respond to this risk landscape, and whose governance structure dominates, has significant implications for AI security, safety, and geopolitics.
Open Versus Closed Weight Systems
Global diffusion of Chinese AI models is strongly influenced by their open-weight model approach to AI. As a form of soft power, China’s strategy targeting global AI adoption is working. At the World AI Conference in mid-July, Chinese President Xi Jinping warned of the dangers of a single country dominating AI technology, while formally introducing the World Artificial Intelligence Cooperation Organization (WAICO). It includes 29 founding members and is the latest Chinese push for the open system approach to AI.
As President Xi pushes for global AI adoption of Chinese models, U.S. frontier companies generally pursue a closed weight approach, which refrains from disclosing training data, model weights, and other proprietary information. The most powerful models are only available through purchase. This approach is largely driven by IP considerations as well as security and financial ones, thus impacting global adoption and access.
Many U.S. tech companies seek to shift the current paradigm, with some taking the lead to reshape the governance landscape. Nvidia’s Jensen Huang is among those who advocate for open-weight AI models. Following the Hugging Face breach, Huang organized prominent tech companies to sign a letter advocating for open-weight models, attempting to shape U.S. policy and American AI leadership across the globe. Microsoft cosigned the letter and released its own defense of open weight models. The letter has over 270 signatories.
The charm offensive appears to have paid off. In the August 4th meeting, White House advisors told the tech executives that open-weight models would be exempt from government review. The following week, Meta released their manifesto pushing for open-weight models alongside the release of Muse Glimmer and upcoming release of the parameters of the more powerful Muse Spark.
This is an important shift as the U.S. government has been pursuing a larger role in the release and accessibility of U.S. frontier models. For example, Anthropic introduced Claude Fable for general use, which included guardrails in high-risk areas. Nevertheless, within a week of release, the U.S. government introduced an export control, prohibiting the use of Anthropic’s most capable models by foreign citizens. The ban was short-lived, and at the end of June, Anthropic announced the U.S. Department of Commerce had lifted the ban. Although news that open-weight models would be exempt from government review added some clarity, and the pathway to open weight releases such as Meta’s this week, there is still significant uncertainty. The broader AI framework currently is not planned for public release, leaving much speculation about what it contains, such as whether closed-weight systems still require voluntary submission to the government prior to release.
Interestingly, the momentum may be shifting in the opposite direction in China. As Chinese models catch up – and potentially surpass – those made in the U.S., security and commercial motivations are gaining traction. For instance, China is now exploring limiting foreign access to their most powerful frontier AI models, and has held discussions on the matter with Alibaba, ByteDance, and Z.ai. Many Chinese frontier AI companies have recently introduced closed source frontier models due to the simply financial reality of needing to increase revenue.
AI Supply Chain Warfare
In addition to the open versus closed weight debate, the other large governance transformation can be succinctly summarized as AI supply chain warfare. Security risks, coupled with concern of AI hegemony and dependencies, are fueling the next range of governance shifts focused on export controls.
Over the last year, U.S. AI policy has oscillated between the easing of regulations last year, to June’s EO asking companies to provide government access to frontier AI models prior to release. In fact, the bipartisan AI “Kill Switch” bill, proposed in the U.S. House of Representatives following the Hugging Face breach, would require frontier AI companies to report incidents and ensure the ability shut down AI systems.
Given the stakes and geopolitical competition, the U.S. government is increasingly looking toward export controls and policy updates to minimize Chinese access to frontier models. The U.S. Department of Commerce’s Bureau of Industry and Security has formally launched an investigation into whether Chinese companies accessed banned, U.S. advanced chips, as well as into the accusations surrounding distillation and IP theft. Chinese officials have rejected the accusations, and remarked, “China will take all necessary measures to firmly safeguard its legitimate rights and interests against any actions that substantially harm China’s interests.”
Furthermore, recent reports indicate that the Trump administration is looking to ban frontier Chinese AI models, especially in light of the Kimi K3 release. Some frontier AI executives also warn of the risks of Chinese AI models due to the security concerns. For instance, the U.S. government sanctions Zhipu AI, but many of their fine-tuned models are integrated into dominant AI coding platforms used across Silicon Valley.
At the same time, some U.S. tech executives are pushing for greater use of Chinese models by U.S. companies as they are cheaper and require less compute. In fact, on July 22, almost 200 companies signed a letter urging the Trump Administration to refrain from banning the use of Chinese AI models.
While a decision on banning Chinese AI models remains in flux, the Trump administration continues to expand export controls on the Chinese hardware that powers or is adjacent to AI. A new FCC ban targets Chinese humanoid robots and power inverters. China responded with their own sanctions. The tit-for-tat restrictions may continue, with reports that the Trump administration is considering bans on Chinese AI companies and optical transceivers, which are used in data centers. As they play out, we will likely also see export ban circumvention, similar to reports of Nvidia chip smuggling and related arrests.
Looking Ahead: The September Summit
At the Black Hat cybersecurity conference, U.S. National Cyber Director, Sean Cairncross, affirmed that open source will play a vital role in spreading U.S. preferences and policies globally. As he noted, “We are extremely interested in looking at ways to build U.S. open source, make it competitive, make it the preferential adoption by planet Earth.” This highlights the duality of the AI race. It still very much remains the U.S. versus China, but now the open versus closed systems debate adds another layer of complexity. The interplay between these layers will determine global AI governance and market share.
In September, President Xi is expected to visit the United States. This summer’s eventful drumbeat of AI breakthroughs and inflection points will likely be on the agenda. The Summit may well be a bellwether into the future of AI governance and whose norms and policies will shape the future.