On June 2026, Anthropic received a letter from the United States Department of Commerce, which ordered the organization to deny two of its newest artificial intelligence models, Claude Fable 5 and Mythos 5 to foreign nationals and organizations, including Anthropic’s non-US-citizen employees. The organization could not enforce the rule selectively as American export control law treats handing technology to a non-citizen as an "export." Within hours, it switched both models off for everyone. It was the first time Washington had used export controls against an AI model.

The next day, a Chinese organization called Zhipu AI released a free-of-charge model of its own, GLM-5.2, and put it online for anyone to download. The American model was just locked away and the Chinese one was being handed out.

These two events captured something that has been building for more than a year, that Large Language Models (LLMs) are no longer just products but have become instruments of state power where governments now license and restrict technology in a way they once managed weapons and strategic exports.

But why should a piece of software cause trouble between governments at all? The answer lies in the dual-use nature of these models. The same ability that allows a model to find and repair a flaw in a piece of code is also the one that lets it find and exploit it. They carry the imprint of whoever builds them, including their values, ideology and biases. They also have the ability to censor sensitive information, something that can become a national security threat. Moreover, they move data across borders so that information typed into a foreign model may end up on foreign servers, under foreign law. Since most capable models are held by a very few organizations, and in a handful of countries, whoever controls them can grant or withhold access at will, turning a product into a lever of foreign policy.

Under the above-mentioned backdrop, this article explores the geopolitical objectives of these LLM models with a focus on the anthropic ban and a predictive outcome of the future on the AI race.

The Initial Escalation between American and Chinese Models - Launch of DeepSeek

The competition between the US and China over LLMs goes back to 2025, with the launch of DeepSeek. DeepSeek launched the reasoning model called R1 and claimed that it had cost less than USD 6 million to train, a fraction of what American labs were spending. The model was open which meant that anyone could download it, inspect it, and run it themselves.

This had an adverse consequence in the US market as the chipmaker organization Nvidia lost roughly USD 600 billion in market value in a single day, the largest one-day loss in Wall Street history, as investors suddenly doubted whether the American approach of spending more on computing power was the only road to advanced AI. Venture capitalist Marc Andreessen called R1 "AI's Sputnik moment."

The USD 6 million figure was probably misleading as analysts pointed out it likely left out earlier research and hardware, but the symbolism landed regardless. For two years, the assumption behind American policy had been that strict controls on advanced chips would keep China a step behind. DeepSeek’s launch suggested that a Chinese lab could produce a competitive model and then give it away to the world. However, within weeks, governments from Italy, Australia and South Korea were banning the DeepSeek app from official devices, worried about user data flowing to Chinese servers and about the model's tendency to avoid politically sensitive subjects.

The Debate of Open vs Closed Models

The easier way to describe the AI race between the US and China is the question of open vs closed models. The leading American frontier AI organizations, including OpenAI, Anthropic, and Google continue to keep their most capable models under proprietary control. Users can access them through subscription services or APIs, but the underlying model weights remain unavailable.

China, moreover, has pursued a different strategy. Organizations such as DeepSeek, Alibaba (through Qwen), and Z.ai (through GLM) have released powerful open-weight models that developers can download, modify, and run on their own hardware.

However, not every Chinese frontier model is open, and not every American model is closed. Meta and Mistral have championed open-weight releases in the West, while firms such as Baidu continue to rely on proprietary systems. This has led to the emergence of an important divergence, Chinese firms have embraced open-weight releases more aggressively as a means of accelerating global adoption and ecosystem growth. The strategic logic extends beyond the model itself. By making advanced AI models freely available or inexpensive, Chinese firms reduce barriers to adoption and accelerate their integration into global digital ecosystems. As developers, businesses, and public institutions build applications around these models, complementary assets including software tools, documentation, technical expertise, and developer communities are getting built around them. This creates dependencies in a way that switching to alternative ecosystems would cost more. In this sense, the strategic advantage lies not only in developing frontier models but also in embedding them within the technological infrastructure upon which future AI innovation depends.

This strategy is working even inside the United States. A US congressional commission concluded in early 2026 that, “Chinese labs have narrowed performance gaps with top Western LLMs.” Further, Martin Casado, a partner at the venture-capital firm Andreessen Horowitz, told The Economist last year that among startups building on open-source AI, there was an "80% chance" they were quietly building on a Chinese model, simply because they were good and cheap. Chinese models have at times accounted for most of the top downloads on Hugging Face, the main platform where developers share AI systems.

In August 2025, OpenAI released its first open-weight models since 2019, in what analysts read as a direct response to China's growing dominance of the open ecosystem. Its chief executive, Sam Altman, had earlier admitted that the organization may have been "on the wrong side of history" on the question of openness. The closed-versus-open divide had become important enough that even the most famously closed lab in America felt it had to move.

The Anthropic Ban

This is the backdrop against which the Fable 5 shutdown should be understood. The US government argued that it discovered a method for circumventing the safeguards that restrict Fable 5’s cyber security capabilities, commonly termed a “jailbreak.” However, the method of the technique appears vague. The researcher engaged by Anthropic to review the underlying study, Katie Moussouris, reported that the model was presented with source code containing both known and deliberately introduced vulnerabilities; when instructed to review the code for security flaws, the safeguard engaged and the model declined, but when instructed instead to “fix this code,” it produced functional remediations.

Anthropic argued that this problem was narrow as it was a known weakness in the system, and the same results could come from the current open models. More than 150 cyber security professionals signed an open letter urging the government to reconsider their decision.

The Challenge of Controlling Frontier AI

The Fable and Mythos 5 ban illustrates a broader limitation of applying traditional export-control to artificial intelligence. Unlike physical technologies such as advanced semiconductors or manufacturing equipment, frontier AI models exist within a rapidly evolving and globally distributed ecosystem. Restricting access to one model does not necessarily prevent access to the underlying capability.

Several factors contribute to this outcome:

• First, the capabilities that governments seek to restrict are often inseparable from the capabilities that make frontier models valuable. The same ability that enables a model to identify software vulnerabilities is also essential for detecting and repairing them. Limiting one of them would inevitably affect the other.
• Second, the growing availability of capable open-weight models provides alternative sources of comparable functionality. As these models become more sophisticated, users who lose access to proprietary systems can migrate to alternative ecosystems.
• Third, advances in open and foreign models have significantly shortened the time required for competitors to narrow performance gaps. What might once have represented a durable technological advantage can now erode within months as new models are released and improved.
• Fourth, unilateral restrictions face inherent jurisdictional limits. While a government may control access to models developed within its own borders, it cannot exercise comparable authority over models developed, hosted, and distributed elsewhere. Once competitive alternatives exist beyond its regulatory reach, restrictions imposed by one state are less likely to deny access globally than to redistribute demand across competing AI ecosystems.

These factors suggest that restricting a single frontier model is unlikely to prevent the international diffusion of advanced AI capabilities. Instead, such measures may unintentionally accelerate the adoption of alternative models and strengthen the ecosystems built around them, thus shifting competitive advantage from technological leadership to broader patterns of global adoption.

The Export-Control Paradox

The Fable 5 ban also raises a broader strategic question - if open-weight models are already available outside the US, what does restricting access to a single American model ultimately achieve?

The timing of events made this question particularly striking. Shortly after Anthropic restricted access to Fable 5 and Mythos 5 in response to the Department of Commerce's order, Z.ai released GLM-5.2 as an openly available model. Although the two systems are not identical, the sequence illustrated how users denied access to one frontier model could turn to alternative ecosystems rather than lose access to advanced AI capabilities altogether.

This dynamic has broader implications than the fate of any individual model. As open-weight systems continue to improve, export restrictions imposed on proprietary models may influence not only technological access but also patterns of global adoption. Developers, startups, and public institutions seeking stable and unrestricted access to frontier AI may favor open ecosystems whose availability is less dependent on the regulatory decisions of a single government.

Several analysts have pointed to this possibility. Paul Triolo of DGA-Albright Stonebridge Group argued that governments and organizations affected by American restrictions would likely explore alternative AI providers, including Chinese open-weight models. From this perspective, restrictions intended to preserve technological advantage may also create incentives for competing ecosystems to expand their global user base.

At the same time, proponents of the restrictions offer an important counterargument. They contend that the very decision to control models such as Fable 5 demonstrates that American frontier systems continue to possess capabilities worth protecting. Even if comparable alternatives exist, limiting access to the most advanced models may slow competitors, preserve a technological lead, and reduce the diffusion of cutting-edge capabilities. Jordan Schneider, author of the ChinaTalk newsletter, has argued that the episode should, therefore, be interpreted as evidence that export controls remain strategically relevant rather than obsolete.

Both perspectives capture different aspects of the same reality. The US may retain a meaningful technological lead, making protection of frontier models a legitimate national security objective. However, the emergence of capable open-weight alternatives complicate the effectiveness of unilateral model-specific restrictions. Rather than denying access to advanced AI altogether, such measures may encourage users to migrate toward competing ecosystems that remain freely available. In this sense, the strategic challenge is no longer simply preventing access to frontier AI but ensuring that efforts to protect technological leadership do not inadvertently accelerate the global adoption of rival platforms.

Lessons for Middle Powers

The implications of these developments extend well beyond the rivalry between the United States and China. For countries that rely on foreign AI models and cloud infrastructure, access to frontier AI can become subject to geopolitical considerations beyond their control. Questions of technological dependence are likely to become an important aspect of foreign and industrial policy as artificial intelligence becomes embedded in economic activity, public services, and national security.

For many middle powers, the challenge is, therefore, not simply choosing between American and Chinese AI ecosystems but reducing excessive dependence on either. Sandoval, 2026 identifies four broad strategic approaches that states may adopt:

• The first is specialization, whereby countries focus on niche strengths such as language models, domain-specific AI, or safety research rather than attempting to compete across the entire AI value chain. India's investments in sovereign language models and public-sector AI initiatives reflect elements of this approach.
• A second strategy is alignment, in which states deepen technological cooperation with a preferred partner in exchange for more reliable access to advanced capabilities, accepting a degree of strategic dependence in return.
• Third is sharing, through regional or minilateral cooperation on computing infrastructure, datasets, and model development in order to reduce costs and pool technological resources.
• Fourth, states may choose to hedge by maintaining relationships with multiple AI providers and avoiding dependence on any single technological ecosystem.

These approaches are not mutually exclusive, and many countries are likely to pursue elements of each simultaneously. Their common objective is to increase resilience in an environment where access to frontier AI may reflect geopolitical dynamics as much as commercial considerations.

The Future of the AI Race

The long-term trajectory of AI competition remains uncertain. Open-weight models may continue to narrow the gap with proprietary frontier systems. Export controls could evolve into more coordinated arrangements among allies, or governments may adopt divergent regulatory approaches that fragment the global AI landscape. At the same time, more countries are likely to invest in sovereign AI capabilities, data governance frameworks, and domestic computing infrastructure as artificial intelligence becomes recognized as a strategic technology.

What is already evident, however, is that the geopolitical competition surrounding LLMs is no longer defined solely by technological performance. Access to frontier AI has become an instrument of statecraft, while the diffusion of open-weight models has introduced a parallel dynamic centered on ecosystem expansion and global adoption. The competition is, therefore, shaped by two complementary forms of power: the ability to control access to advanced capabilities and the ability to establish technological ecosystems that others choose to build upon.

This shift carries important implications for the international system. The strategic value of AI may depend not only on which country develops the most capable models, but also on which ecosystem becomes embedded within the digital infrastructure of governments, businesses, universities, and developers around the world. In that sense, the future balance of power in artificial intelligence is likely to be determined as much by patterns of adoption and technological dependence as by frontier model performance alone.

Disclaimer: The views expressed in this article are solely those of the author and do not reflect DSCI's views or opinions on the subject.

Ishaan Banwait interned at the Data Security Council of India as a part of the Global Trade Development and Government Affairs vertical. His primary research focus was on the geopolitics of technology including AI, cyber security and quantum.