Anthropic Sounds the Alarm on China’s AI Race
Anthropic, the company behind the Claude family of AI models, published a lengthy policy paper that calls for aggressive American action. The message is direct, the timeline is tight, and the stakes, according to Anthropic, could not be higher. The firm wants Washington and its allies to tighten controls on advanced chips and restrict access to American AI models. The stated goal is to curb china ai progress before 2028, a year Anthropic believes will mark the arrival of truly transformative AI systems. But a closer look at the arguments reveals a more complicated picture.
The Core of Anthropic’s Warning
Anthropic argues that AI will deliver “transformational economic and societal impacts” within just a few years. The company believes that where the most capable systems are built first will shape global norms for decades. If democratic countries take the lead, the argument goes, AI will develop under rules that protect individual freedoms. If authoritarian regimes set the pace, those same tools could enable large-scale repression.
This is not a niche policy debate. It is a question about the future of governance, privacy, and human autonomy. Anthropic frames the issue as a narrow window of opportunity. The company insists that democratic nations have only a limited amount of time to act before the window closes for good.
Two Paths for 2028
Anthropic sketches two distinct scenarios for the year 2028. In the first, the United States has successfully defended its advantage in computing power. Democracies set the rules and norms around AI deployment. In the second scenario, China overtakes the United States. Authoritarian regimes shape the trajectory of the technology, and the most advanced models become instruments of automated repression on a massive scale. This binary framing is deliberate. It is designed to create urgency. It also invites scrutiny, because the real world rarely offers such clean choices.
The Timeline Debate
Anthropic picked 2028 as the inflection point. Is that date based on reliable forecasting or strategic messaging? Many independent researchers believe that transformative AI could arrive earlier or later depending on breakthroughs yet to come. The specific year may matter less than the underlying claim that the next few years will be decisive. Even skeptics of Anthropic’s policy proposals tend to agree that the pace of progress has accelerated dramatically since late 2022.
Export Controls on AI Chips
The first major pillar of Anthropic’s plan involves tighter export controls on semiconductors used for AI training and inference. Nvidia’s high-end GPUs, such as the H100 and H200 series, are the primary targets. These chips power the largest training runs at companies like OpenAI, Google, and Anthropic itself. Restricting their flow to Chinese entities has been US policy for several years now. Anthropic claims these controls have been “incredibly successful.”
How Chip Restrictions Actually Work
Export controls operate through a combination of legal prohibitions and licensing requirements. US-based companies cannot sell certain advanced chips to Chinese buyers without government approval. The rules also apply to third-party countries where transshipment might occur. Enforcement relies on customs inspections, corporate compliance teams, and intelligence monitoring. In theory, the system creates a wall around the most capable hardware. In practice, the wall has gaps.
The Success Question
If export controls have been so successful, why do many experts estimate Chinese researchers are only several months behind their American counterparts? That narrow gap raises uncomfortable questions about the effectiveness of the entire approach. Anthropic addresses this by arguing that Chinese labs have only come this far by exploiting loopholes and using techniques like model distillation. The implication is that without those workarounds, China would be much further behind. But the fact remains that the gap is measured in months, not years. The data suggests that current controls have slowed progress but have not stopped it. Anyone serious about using policy to curb china ai advancement must grapple with this reality.
Cutting Off Access to American AI Models
The second pillar of Anthropic’s proposal is more controversial. The company wants to restrict Chinese access to American AI models, either through licensing requirements, API restrictions, or outright bans. This goes beyond hardware. It targets the software and knowledge embedded in the models themselves.
The Distillation Loophole
Model distillation is a technique where a smaller model learns from the outputs of a larger, more capable model. It is a standard practice in the AI industry. Companies use it to create efficient versions of their own systems. But it can also be used to extract knowledge from a competitor’s model without permission. In February of this year, Anthropic accused DeepSeek, a Chinese AI lab, of using distillation to siphon knowledge from Anthropic’s own Claude models. That accusation sits at the heart of the company’s push for tighter controls.
Distillation is difficult to detect and even harder to prevent. When a user queries a model and receives a response, that response contains patterns, reasoning structures, and factual knowledge. A determined actor can collect millions of such responses and use them to train a competing model. This is not hacking in the traditional sense. It is more like reverse engineering through repeated interaction. Closing this loophole would require monitoring how models are accessed and who is using them, which raises its own set of privacy and operational challenges.
Why Distillation Matters for Policy
If distillation allows Chinese labs to replicate American advances within months, then hardware controls alone will never be sufficient. The knowledge leaks through the API layer. Anthropic’s proposal to cut off model access is a logical response to this vulnerability. But it also highlights how difficult it is to truly curb china ai progress when the technology itself is designed to be queried and shared.
The Innovation Blind Spot
Anthropic’s policy paper contains an assumption that has drawn sharp criticism. The company implies that China can only advance by copying American innovations. Recent developments tell a different story.
DeepSeek R1 and Indigenous Progress
The release of DeepSeek’s R1 model in early 2025 sent shockwaves through the AI community. Independent evaluations placed its performance on par with the best American models. This was not a distillation product. It was a model built with original architecture choices and training methods. If anything, DeepSeek R1 demonstrated that Chinese AI labs are capable of genuine innovation when given adequate resources and talent. Dismissing this as mere copying ignores the evidence and weakens the credibility of Anthropic’s entire argument.
Domestic Silicon Development in China
Chinese organizations have also made significant strides in developing their own AI chips. Reports indicate that several domestic designs now offer competitive performance for specific workloads. Beijing has actively encouraged this shift, even discouraging Chinese tech companies from buying and using Nvidia chips. The goal is clear: reduce dependence on foreign hardware and build a self-sufficient AI ecosystem. If this trend continues, export controls on American chips will become less relevant over time. The policy window for using hardware restrictions to curb china ai advancement may already be narrowing.
You may also enjoy reading: 7 Windows Desktop Apps I Built to Fix My Workflows.
The Self-Awareness Problem
There is an irony in Anthropic’s position that has not gone unnoticed. The company accuses Chinese labs of using distillation to copy its models. Yet the entire AI industry, including Anthropic, has trained its models on vast amounts of content created by others without permission. The Register pointed out this lack of self-awareness in its coverage of the DeepSeek accusation. The tension is hard to ignore.
Anthropic built Claude using data scraped from the public internet. That data includes copyrighted books, news articles, code repositories, and personal blogs. The company did not pay creators for that content. It did not ask permission. Now it is demanding that Chinese actors be blocked from doing the same thing to its own models. Many observers see this as a double standard. It does not mean Anthropic is wrong about the risks of Chinese AI dominance. But it does mean the moral high ground is more crowded than the company suggests.
Europe’s Digital Sovereignty Push
Another complication for Anthropic’s plan is the position of European nations. Many European policymakers view both American and Chinese AI supremacy as a threat to democracy. They see little benefit in swapping one dominant power for another. This has fueled a concerted push for what is called digital sovereignty.
European governments and companies are investing in homegrown AI infrastructure, cloud services, and model development. The goal is to reduce reliance on US technology while also guarding against Chinese influence. If Europe chooses to chart its own course, it may resist American efforts to enforce a unified bloc against China. This could create cracks in the coalition that Anthropic’s plan depends on.
The Trump Administration’s Shifting Stance
Anthropic’s proposals face an uncertain reception in Washington. The Trump administration has shown a constantly shifting attitude toward China on technology policy. Export controls were reportedly not a high priority during the president’s recent trip to Beijing. More notably, the US has now cleared around ten Chinese firms to purchase Nvidia’s second-most powerful AI chip, the H200.
That decision undercuts the entire logic of the tightening approach. If Chinese entities can legally buy advanced hardware, the idea of using export controls to curb china ai progress becomes much harder to sustain. It suggests that geopolitical realities and commercial interests often override the clean policy frameworks that companies like Anthropic advocate for.
What Realistic Policy Looks Like
A more balanced approach to managing the US-China AI competition would acknowledge several uncomfortable truths. First, export controls can slow but not stop determined adversaries. Second, innovation is occurring on both sides of the Pacific, and dismissing Chinese capabilities only weakens American preparedness. Third, the moral authority of US companies to demand intellectual property protection is undermined by their own data practices. Fourth, allies like Europe may not follow Washington’s lead if they perceive American dominance as their own risk.
Policymakers who want to curb china ai advancement effectively should focus on areas where the US retains genuine advantages: fundamental research, talent attraction, and the ability to set global technical standards through collaboration rather than coercion. They should also invest heavily in domestic chip fabrication capacity, not just restrictions on foreign buyers. And they should acknowledge that AI governance is a global conversation, not a unilateral command.
Anthropic is right about one thing: the next few years will shape the trajectory of AI for a long time to come. But the best way to influence that outcome is not through a narrow focus on cutting off access. It is through building systems that are so good, so transparent, and so aligned with human values that the rest of the world chooses them freely.
