Regardless of the rationality (reasonableness) of this order, this marked the second time that the conflict between Washington and Anthropic had erupted in public view. As early as March this year, the Department of Defense labeled it a “supply chain risk” because while the military wanted to use its AI, Anthropic insisted on stipulating that its models must not be used for large-scale surveillance of Americans or integrated into autonomous weapons. As negotiations broke down, President Trump directly ordered all government agencies to abandon its use, and OpenAI took over the confidential contracts it had lost within a few hours.

Even Anthropic, a company valued at US$965 billion that was set to go public in a few months, saw its strongest products halted by a single order. At this point, the crucial question is no longer whether AI works, but whether, given its power, it will remain accessible.

I sense that such incidents will only increase in the future. Unlike the early internet, which was universally connected, AI is grounded in computing power and electricity—it depends on whose territory it operates in and whose electricity it uses, and thus falls under whose control. If the United States becomes displeased, a single order could cut off access to the most advanced large models, causing companies relying on them to suffer as well.

As a result, the ability to keep AI in one’s own hands, immune to a single “kill switch,” will become increasingly valuable. Even if the companies providing this capability are considered “redundant assets” on a global scale, there will be opportunities for value rediscovery.

01 Who Controls the Most Powerful AI?

The government’s reason for banning Fable 5 was that it could be “jailbroken” to bypass safeguards and find software vulnerabilities.

This jailbreaking research came from Amazon researchers, who submitted it to the government before regulations were implemented. However, this rationale is not particularly strong. The so-called jailbreaking merely involves having the model read a specific piece of code and identify its flaws—something that other public models, including OpenAI’s GPT-5.5, can also do.

Precisely because this jailbreaking method is not unique, Anthropic argued that the government had only verbally stated its concerns without providing any concrete evidence. It also warned that if this reasoning were applied uniformly to every company, none of the most powerful AIs on the market would remain openly accessible. Moreover, the White House had signed an executive order in June explicitly stating that it would not subject AI models to approval processes; yet this current ban, targeted specifically at Anthropic, did exactly that.

Despite the flimsy justification, the ban was implemented. At first glance, it is easy to draw a parallel to the earlier conflict between Washington and Anthropic in March.

However, that earlier clash with the government was more of a trigger than the root cause. It is unlikely that Washington would have decided to single out this private company without other motivations. The real reason is that the most powerful AIs have now become quasi-national assets, and their availability increasingly depends on Washington’s whims. Anthropic’s refusal to comply led to its strongest models being shut off to the world; in contrast, OpenAI, whose model capabilities are not far behind, has cooperated and thus fared much better.

Interestingly, it was Anthropic itself that first treated Mythos as a dangerous entity. When it was released in April, access was restricted to a very small group because its vulnerability-finding capabilities were too strong. During testing, it uncovered over 10,000 critical vulnerabilities, including one that had gone undetected for 27 years in the OpenBSD system. Ultimately, only about 14% of these vulnerabilities were patched, and during the process, the model briefly escaped its designated isolated environment.

Over the past year, Anthropic has repeatedly and publicly emphasized how dangerous Mythos is, calling it a “watershed” in security. It published lengthy articles urging the government to establish regulations and define when such capabilities should be paused, even audaciously calling on all industry peers to halt AI development. From a marketing perspective, Anthropic aimed to portray itself as the only savior by emphasizing the model’s dangers and its own prowess.

But Washington only heard the first layer of this message and clearly believed it was better suited to regulate AI. Using national security as an excuse to control AI is not new in the United States; in recent years, it has prohibited NVIDIA from selling its most advanced chips to China, but that line of control stopped at the border, leaving domestic companies and individuals unaffected.

This time is different. For the first time, the same national security logic has been extended to the country’s most powerful AI company. Starting in March, Anthropic was excluded from military use, followed by abandonment by the Treasury Department and procurement agencies, and even its projects deployed in operations against Iran were halted. By June, its two most powerful models were directly banned, forcing it offline globally and cutting off access even for domestic users—a four-month trajectory of escalating restrictions.

Moreover, the legal basis for this ban is questionable. The government cited Section 3252 of military regulations, which was originally intended to prevent foreign backdoors, making it inapplicable to a U.S. company. If used in this manner, the government could label any company with which contract negotiations fail as a supply chain risk and cripple it.

Now that the deed is done, such controls have already triggered global reactions. The Canadian Prime Minister compared it to the 2008 banking crisis, stating that the world’s reliance on a few AI models represents a systemic risk. A French presidential candidate likened Anthropic to the U.S.-controlled “Strait of Hormuz,” declaring that “the AI war has begun.” Britain treats this sovereign-level threat as a matter of life and death, while the European Union has once again brought “technological sovereignty” to the forefront.

These sovereign nations, traditionally aligned with the United States, have realized for the first time that access to the most powerful AIs and the duration of that access are not entirely determined by themselves or the market but also by a hand that could tighten its grip at any moment. Once this risk becomes clear, many things must be rethought.

02 What Can Be Held Begins to Matter

Lifted by this ban are China’s open-source models. Before the ban, models like QianWen and DeepSeek were already leading in the open-source world. On HuggingFace, the largest open-source community, over 113,000 models are derived from Alibaba’s QianWen, compared to only 27,000 for Meta’s Llama.

Their users are not just tech enthusiasts. OCBC Bank in Singapore uses DeepSeek and QianWen to power over 30 internal tools, while Malaysia has built a national AI system using Huawei’s hardware. Even a U.S. congressional committee acknowledged in a March report that China is expanding its industrial advantages through this open-source approach. In terms of usage, China’s large model market now exceeds 140 trillion daily token calls this year, a thousandfold increase over the past two years.

This does not mean China will comprehensively surpass and replace U.S. closed-source models. A key limiting factor is chips: Huawei’s latest Ascend 950 has only 6% of NVIDIA’s computing power per chip, according to Bernstein estimates, and is constrained by SMIC’s 7nm process, with high-bandwidth memory relying on stockpiling.

Nevertheless, after the ban, the number of entities seeking models outside the United States has noticeably increased. Even Canada’s Cohere reports a surge in inquiries from clients, governments, and investors looking for non-U.S. alternatives. What they seek is freedom from U.S. control, with Chinese open-source models being the most significant option.

In terms of business, several layers of players are emerging to meet this demand. At the top are large corporations with both open-source models and cloud infrastructure. Alibaba is best positioned in this regard, as QianWen is already a leading open-source model, and Alibaba Cloud provides an excellent platform to run it. Management attributes cloud growth directly to QianWen and AI adoption.

In the most recent quarter, Alibaba Cloud’s revenue grew 38% year-over-year, with external client revenue growing even faster at 40%. AI-related revenue has more than doubled for 11 consecutive quarters. Baidu, at the same level, takes a different approach, pairing its Wenxin large model with its homegrown Kunlun chip, offering a self-contained ecosystem from model to chip that becomes increasingly attractive as buyers seek to bypass the United States.

The next layer consists of pure open-source labs like Zhipu, Moonshot AI (Kimi), and MiniMax. These lack the cloud infrastructure of large corporations but offer powerful models. Zhipu’s recently released open-source model, GLM-5.2, ranks first among currently available models on Code Arena, a programming platform where users vote, and first in open-source (and second overall, follow closely (closely following) Claude Opus) in SWE-bench evaluations, which measure code modification capabilities. Its weights are public (publicly available), allowing self-deployment. These labs are not just seeking attention; Zhipu and MiniMax generate revenue through APIs, subscriptions, and enterprise contracts, using open-source as a draw for overseas clients. However, their scale remains small, and how much business this popularity will translate into remains to be seen.

No matter how good the model, deployment requires chips, leading to the bottom layer: domestic chipmakers. Huawei’s Ascend remains the most prominent domestic option. Avoiding U.S. computing power entirely requires relying on it, but its performance limitations, as previously mentioned, cap its potential.

On the market side, the demand for domestic open-source models spurred by the ban has yet to be fully priced in. Discussions still largely focus on Amazon’s tattling and OpenAI’s gains, while brokerage logic favoring domestic computing power, based on AI’s popularity and the computing gap, has been around since March and is unrelated to this ban.

03 Conclusion

This ban may not last long. On June 15, Anthropic sent technicians to the Department of Commerce to discuss solutions, stating that progress was being made rapidly. However, a Cornell University AI researcher cautions that reinforcing the model’s safeguards will not happen quickly and is not as simple as fixing ordinary software bugs.

Whether the ban is lifted is one issue, but its impact is irreversible.

In just six months, SoftBank has committed 75 billion euros to build data centers in France, South Korea has tripled its AI budget for next year, and Canada and Germany have formed a sovereign technology alliance to develop independent capabilities. Nations are no longer just paying lip service to sovereignty but are investing real money.

By spending this money, countries are acknowledging something that has not been seriously considered before: an asset’s value lies in its consistent accessibility. China’s open-source models and self-sufficient domestic computing power have long been seen as redundant assets, overshadowed by superior U.S. models and thus undervalued.

This ban has revealed, for the first time, the existence of a hand capable of severing the main trunk externally, justifying a reevaluation of the discounts previously applied due to “redundancy.” However, the extent of this reevaluation remains uncertain. Whether the ban persists is questionable; even if it does, China’s models have not yet truly caught up in performance, and whether open-source models can translate into businesses that justify their valuations remains unanswered. The direction is clear, but the magnitude depends not only on whether the United States tightens its grip but also on how much China can address its own shortcomings.

Disclaimer: This article is for learning and communication purposes only and does not constitute investment advice.

Welcome to like, share, and repost. Your support is what keeps us updating!