Note to Readers
This week’s issue comes to us from upstate New York, and is written by my esteemed friend and great colleague Ricardo Tavares, who advises governments and companies on technology policy and has been doing so for decades. With Washington trying to figure out what to do about Chinese AI models, and Xi Jinping arriving in the United States in September, Ricardo takes us through what is actually at stake. He has been on both sides of negotiations like this one… Over to you, Ricardo.
AI models coming from China have become a hot item for discussion among Washington technology policymakers. The reason is simple: Chinese open-weight models are gaining significant market penetration among American startups and even some larger companies, thanks to lower costs and “good-enough” performance that keeps shrinking the distance with U.S. frontier models.
Despite the United States government seriously considering banning these models, they have proved to be quite useful lately. While everyone is talking about AI agents going wild, Hugging Face said it actually used a Chinese open-weight model – Z.ai’s GLM 5.2 – to fend off the much discussed cyberattack carried out by rogue OpenAI agents that had exploited a previously unknown vulnerability to break into Hugging Face’s systems. On the one hand, Anthropic’s own model, Fable 5, reportedly declined to help with the defense, its safety guardrails treating reverse-engineering the exploit the same as launching an attack. On the other hand, GLM 5.2 proved to be much more flexible — and saved the day.
Models of Change
The reason behind the calls for a ban is that Chinese open-weight models have raised two suspicions:
Tech Cold War hawks are suspicious of these models because they’re “Chinese” — they see them as Trojan horses used to penetrate American enterprises and households. Additionally, open-source enthusiasts make clear that “open weight” doesn’t mean “open source.” You can play with the model’s “weights” to adapt it to your needs, but you don’t have visibility into the underlying code, which reinforces suspicions. (Hawks, however, make clear that they’re not against open-weight models per se, but against the Chinese ones.)
Market analysts see Chinese AI labs and their latest models, such as DeepSeek v4, Moonshot’s Kimi K3, Alibaba’s Qwen 3.6-27B, Tencent’s Hunyuan 3.0, as well as GLM 5.2, following the same recipe Chinese companies have used in other market segments: launch relatively competitive products, push prices down, drive the competition out of business, and then raise prices. In fact, the timeline associated with this AI cycle is much more compressed and high-paced than usual, as the industry sees a new model launching every six days on average, a majority coming from China, while the latest announcements of Kimi K3 and DeepSeek v4 featured a significant API price increase – a departure from the loss-leader pricing that built their initial market share, showing a new level of confidence in their market take up.
Reports indicate the White House is internally divided. On the one hand, Commerce Department officials are reluctant to impose a ban on Chinese models; they believe it would be difficult to enforce it, affecting its credibility. On the other hand, cyber-defense officials favor measures closer to the import bans already used to block Chinese hardware from U.S. markets; for example, the Federal Communications Commission (FCC) recently introduced several bans on Chinese hardware, including routers. The two camps appear to be converging on “distillation” as the main point of concern. In their common view, Chinese AI labs are quickly catching up with American frontier models, and doing so on the cheap, thanks to a large extent to distillation and circumvention of export controls on advanced silicon designed by U.S. companies.
A Matter of Essence
Distillation is a legitimate machine learning (ML) technique where you train a smaller “student” model to mimic a bigger “teacher” model, using the teacher’s outputs as training data. It’s cheaper than training from scratch and is widely used, including by the original model makers themselves. However, the version at the center of the current controversy is when someone does this to a competitor’s proprietary model without permission: querying it thousands or millions of times through the API, harvesting the answers, and training a rival model on them — no code theft, no hacking, just systematic extraction through normal access channels, which is exactly what makes it so hard to police.
Treasury Secretary Scott Bessent is the top interlocutor of the Trump Administration with senior Chinese officials, negotiating trade, AI, and other matters with Beijing, an expanded mandate for his institutional role. Bessent has equated distillation with “IP theft,” warning that sanctions and Entity List designations will be on the table when Chinese firms conduct “covert, industrial-scale” distillation that crosses the line.
The United States has been complaining about Chinese market distortions for decades; distillation is just one more instance of that. Bessent will have a chance to raise this issue with his Chinese counterparts when they meet to discuss the AI bilateral relationship in a few weeks. It’s not likely that the White House will introduce any immediate measures against Chinese models before the planned visit of China’s leader Xi Jinping, announced by President Trump for September 24, 2026. But Bessent’s AI meeting and the Chinese leader’s presence in Washington represent an opportunity to address the matter.
More Competition
Nearly 200 Silicon Valley startups, united in an organization called the Little Tech Association, made public their position against banning Chinese open-weight models. Their position is easy to understand: if you can access distilled, frontier-adjacent models that are good enough, at low cost, why would you opt to license original frontier models from U.S. big-tech companies at a much higher cost? That doesn’t mean, however, that policymakers in Washington should accept this economic rationality, which cuts against the American model of leading AI by leveraging fast-building of compute capacity and R&D investment to secure leadership at the frontier.
Yet the startups have a point about AI models in the U.S. being controlled by a few players. A coalition of large technology companies, led by Nvidia, Microsoft and Meta and including Dell, Palantir, Mistral and Hugging Face, issued a public statement in defense of open-weight models, arguing they strengthen competition, security and U.S. technological sovereignty. Notably, Anthropic, the American lab most invested in closed, proprietary models, did not sign it, and neither did Amazon, its largest investor. OpenAI and Google were absent from the original list of 25 signatories and joined within a day.
The distinction between open-weight models generally and Chinese open-weight models specifically is critical here. It could well be that market forces, in particular the high compute costs to develop large language models, settle the AI market structure into an oligopoly. At any rate, fostering competition with open-weight models is good public policy, nonetheless. Having more American open-weight models matters in this picture.
Meta’s Move and Zuckerberg’s Wishes
Meta just did that by launching Muse Glimmer, an open-weight model designed to run agentic tasks on a personal computer. While other similar models are targeting business users, Meta is focused on consumers and small businesses. In a lengthy essay released this week, Mark Zuckerberg, Meta’s founder and CEO, made the case for open AI models, rejected the idea of a ban on Chinese models, stated that the U.S. must outcompete China, and may have complicated things for Washington. How?
He appeared to defend the ability of American companies to distill proprietary advanced models such as OpenAI’s and Anthropic’s to develop open-weight models, which today would violate terms of service and could result in lawsuits. Zuckerberg seems to advocate for developers of open-weight models in the U.S. to be able to do the same thing Chinese labs are doing with distillation. Below, I quote a paragraph from Zuckerberg’s essay, “The Future is for Everyone.” (The bold highlights are mine.)
“For the U.S. to lead in open source, we will need to rethink our policies in several areas, including distillation and data use in training. The ability for models to learn from other models is an important principle of how the open source ecosystem works. All AI models are derived from human knowledge. Some have tried to frame distillation as harmful, but I think it is important to protect the principle that you can learn from anything you can observe. This is how the world works, and the U.S. will not be able to lead if we restrict ourselves on this front.”
Difficult Conversations
Which brings us back to the U.S.-China open-weight models discussion, where we see a significant but not surprising lack of trust between the two AI superpowers. The “strategic competitors” or “geopolitical rivals” are the protagonists presenting a unique and unprecedented combination of superior economic, military, and AI capabilities. Banning open-weight models simply because they are “Chinese” accelerates the bifurcation of global digital markets that will hurt everyone — even if it seems inevitable.
Bessent and his Chinese counterparts will have an opportunity, in a few weeks, to sit down and talk about AI. We know that collaboration to reduce AI-existential risks will be part of the discussion; perhaps, ways to build trust to avoid a complete market bifurcation could also be discussed, for the sake of securing competitive markets.
Bringing Chinese open-weight models into compliance with American companies’ terms of service on distillation should be high on that agenda. China has enormous AI talent, no shortage of electricity, and is a leading AI research powerhouse; on the latter, it is dwarfing the United States by some accounts. It holds a place in AI development second only to the United States. But the claim that it’s purely “algorithm optimization,” paired with Huawei’s Ascend chips, that is keeping Chinese models near the U.S. frontier despite far weaker compute infrastructure and capital investment doesn’t hold up. Unauthorized distillation seems to be playing a key role in it, as well as training in clouds inside and outside China using the most advanced American-designed, Taiwan-manufactured AI-accelerator chips.
What is inevitable for sure is that the two elephants in the room need to have the difficult conversations necessary to find ways to compete on fair terms, a feat which only the most skilled economic diplomats in the world can possibly achieve.
We would love to hear your thoughts!
Ricardo Tavares
Guest Contributor, The Compute
Ricardo Tavares advises on public policy and regulation in AI, semiconductors and communications technologies. Since 2004 his work has concentrated on EMEA and South Asia. He has negotiated with governments on behalf of Google, Ericsson and ByteDance, designed the privatisation of state-owned telecoms operators for the World Bank, and advised Inmarsat, GSMA, Vodafone, Telenor Group and Confidential Core AI, a startup working on security for AI agents.
Wirelessly yours,
Ziad Matar
Co-founder, Tokenando
Editor-in-chief, The Compute
References
Fortune — “Hugging Face says it resorted to a Chinese AI model to battle a fully autonomous cyberattack because U.S. model guardrails hampered its defense,” July 20, 2026. https://fortune.com/2026/07/20/hugging-face-turns-to-chinese-open-source-ai-to-fend-off-autonomous-ai-cyber-attack-after-american-ai-guardrails-stymie-defense/
NPR — “How OpenAI’s and Anthropic’s AI models hacked other companies,” August 1, 2026. https://www.npr.org/2026/08/01/nx-s1-5914852/anthropic-openai-models-hack-cybersecurity
Axios — “White House draws new AI line on China,” July 24, 2026. https://www.axios.com/2026/07/24/white-house-ai-line-china
The Standard (Hong Kong, via Reuters) — “US, China to hold AI talks in September, sources say,” July 21, 2026. https://www.thestandard.com.hk/innovation/article/337836/US-China-to-hold-AI-talks-in-September-sources-say
Fox News — “Trump says Xi Jinping will visit the US on Sept 24 to discuss AI,” July 23, 2026. https://www.foxnews.com/politics/trump-says-chinese-president-xi-jinping-visit-us-sept-24
Little Tech Association — open letter (PDF), July 2026. https://framerusercontent.com/assets/fgypzPpKVEO0leVL64uweDUVphQ.pdf
Nvidia — “Open Weights and American AI Leadership” (PDF), July 2026. https://images.nvidia.com/pdf/Open-Weights-and-American-AI-Leadership.pdf
Forbes — “Nvidia Open Weights Letter Doubled To 50 Without Amazon And Anthropic,” July 25, 2026. https://www.forbes.com/sites/sandycarter/2026/07/25/huangs-open-weights-letter-doubled-to-50-without-amazon-and-anthropic/
Dataconomy — “DeepSeek Warns Developers Of Significant API Price Increases” (Aug 6, 2026): https://dataconomy.com/2026/08/06/deepseek-significant-api-price-increase-2026/
Wired - “China’s Open AI Models Are Challenging Silicon Valley’s Playbook,” July 22, 2026. https://www.wired.com/story/chinas-open-ai-models-are-challenging-silicon-valleys-playbook/
Wired - “The White House Is Trying to Figure Out What To Do About Chinese AI,” July 22, 2026. https://www.wired.com/story/the-white-house-is-trying-to-figure-out-what-to-do-about-chinese-ai/
Council on Foreign Relations (CFR) on LinkedIn - “On AI With Chinese Characteristics,” July 31, 2026. https://www.linkedin.com/pulse/ai-chinese-characteristics-council-on-foreign-relations-z3o3c/?trackingId=4cC8yU2%2FQ%2BqeHpkn0YCE%2FA%3D%3D
BCG - “The Great Divide: How the U.S. and China Are Splitting the AI World,” June 30, 2026. https://www.bcg.com/publications/2026/us-and-china-ai-strategy-causing-global-ai-divide
Meta - “The Future is for Everyone - The Path to a Positive AI Future,” Mark Zuckerberg, August 10, 2026. https://www.meta.com/thefutureisforeveryone/
Reuters - “Meta launches new AI model as Zuckerberg champions open-weight push,” August 10, 2026. https://www.reuters.com/world/china/meta-launches-new-ai-model-zuckerberg-champions-open-weight-push-2026-08-10/