Introduction
Picture this: a company developing a powerful digital technology capable of being used to kill people decides to regulate its operations and make it safer. How would the State respond? Realistically, it might commend the company’s responsible approach to technology. But what if the State instead blacklists the company because it considers the company a national security threat? No, that is not a hypothetical dystopian scenario, but the current reality of Anthropic vs. the Pentagon.
Refusal to Perform Military Functions
Anthropic was founded by former OpenAI employees to build a safer AI frontier. The company is committed to its safety standards, which is why it refused to remove two major restrictions on the use of its AI models when demanded by the U.S. military.
First, it refused to permit mass domestic surveillance of Americans, arguing that such surveillance would threaten fundamental democratic rights and individual privacy. Second, it refused to allow the use of its models for fully autonomous lethal warfare, in which AI could select and engage targets without meaningful human involvement. Anthropic argued that its current models are not reliable enough to make life-and-death decisions independently.
The Pentagon saw these restrictions as a potential liability to its military functions and operations. It wanted access to Anthropic’s models for all lawful purposes, without the company imposing its own restrictions on particular military applications. After negotiations broke down, the government designated Anthropic a supply-chain risk under federal procurement law and directed the removal of its technology from relevant military systems. The Pentagon also instructed defence contractors not to use Anthropic’s products in work covered by the designation.
Anthropic vs. Pentagon
In March 2026, Anthropic challenged the government’s actions through two separate legal proceedings in two different federal courts.
The first case was filed in the U.S. District Court for the Northern District of California under 10 U.S.C. § 3252, a provision governing certain Department of Defense procurement actions. Anthropic argued that the government’s decision was unlawful retaliation for refusing to remove Claude’s safeguards. It also argued that the government had exceeded its legal authority and violated Anthropic’s First Amendment and due-process rights.
On August 27, 2026, the district court ruled largely in Anthropic’s favour. It found that the government’s actions under this provision were unlawful, including because the designation violated the governing statutory framework and the company’s constitutional rights. The court vacated the designation issued under that provision and barred enforcement of the associated directives.
The second case concerned the Federal Acquisition Supply Chain Security Act of 2018 (FASCSA). Anthropic challenged the separate supply-chain-risk designation in the U.S. Court of Appeals for the D.C. Circuit.
On September 25, 2026, the D.C. Circuit ruled 2–1 in favour of the Pentagon, upholding the designation under this separate legal framework. The majority reasoned that the military could reasonably consider Anthropic’s ability to restrict its models a supply-chain risk if those restrictions prevented the models from performing tasks the government considered lawful and necessary for national security. The court did not require evidence that Anthropic intended to sabotage the government’s systems.
The dissenting judge disagreed, arguing that the government had interpreted the supply-chain law too broadly. In her view, a company enforcing its own safety restrictions does not automatically constitute the kind of supply-chain threat contemplated by the statute.
The two rulings therefore addressed separate legal provisions and reached different conclusions about the government’s actions. Neither ruling should be understood as resolving every aspect of the dispute.
The legal battle is also continuing. On October 13, 2026, Anthropic asked the D.C. Circuit to reconsider its September decision through a panel rehearing or rehearing en banc. As of October 29, the court has not issued a decision on that request.
The Akamai Deal
On September 24, 2026, Anthropic announced an $11.6 billion, seven-year computing agreement with Akamai, a day before the D.C. Circuit upheld the Pentagon’s supply-chain-risk designation.
One unusual detail about the agreement was that Akamai granted Anthropic a warrant to purchase up to approximately 7.7 million shares, representing roughly 5% of Akamai’s outstanding common stock, at an exercise price of $111.33 per share. Approximately 2% is associated with the initial agreement, while the remaining potential equity interest depends on Anthropic expanding its investment commitment. The warrant does not mean Anthropic already owns those shares.
This arrangement highlights that the relationship between Akamai and Anthropic is more than a conventional supplier-customer relationship. Anthropic is buying computing capacity while also receiving the opportunity to become a shareholder in its infrastructure provider.
The Akamai agreement is part of a much larger infrastructure strategy. Anthropic has also committed substantial sums to Amazon and Microsoft for computing capacity over the coming years. These are long-term commitments that can create significant financial obligations even if demand, usage patterns or business conditions change.
The scale of these agreements raises an important question: how much freedom does an AI company retain when its ability to develop and operate its products depends on long-term commitments to a relatively small group of powerful infrastructure providers?
Loss of Corporate Autonomy and Government Intervention
Given its agreements with some of the world’s largest technology companies and the government’s intervention in its affairs, Anthropic faces growing constraints on its strategic autonomy.
This does not mean that Anthropic has legally lost its independence. Rather, it highlights the difference between formal corporate independence and the practical freedom to make decisions.
Anthropic depends on a limited number of powerful companies for the computing infrastructure required to train and operate increasingly capable AI models. Its long-term infrastructure commitments can also limit its flexibility. Even if Anthropic suddenly decided to slow its AI development, it could still face substantial contractual obligations, depending on the terms of its agreements.
Adding the governmental aspect makes the situation more complicated. The Pentagon wants access to AI systems for its military operations without restrictions that it considers incompatible with national security. Anthropic, meanwhile, maintains that certain safeguards are essential to prevent potentially dangerous applications of its technology.
This creates pressure from three directions: the company’s own safety standards, the financial obligations associated with its infrastructure commitments, and governmental demands for access to its models.
The result is a tension between legal independence and practical autonomy. Anthropic can retain the formal right to make its own decisions while facing powerful financial and political incentives that shape which decisions remain viable.
Conclusion
Anthropic’s case leaves behind a potentially disturbing question: when a company develops a powerful technology with potentially lethal applications and attempts to restrict its use in the interest of public safety, should the State encourage those safeguards or seek to override them in the name of national security?
The answer is not as simple as choosing between corporate self-regulation and government control. Companies developing frontier AI have incentives to protect their products, reputations and commercial interests, but those incentives alone cannot guarantee responsible use. Governments, meanwhile, have legitimate national security responsibilities, yet their demands for unrestricted access can create their own risks.
The debate over AI governance therefore requires more than voluntary commitments from technology companies. It also requires clear laws, independent oversight, enforceable safeguards and meaningful accountability for both private companies and the State.
Ultimately, the dispute between Anthropic and the Pentagon raises a fundamental question about the future of artificial intelligence: can companies keep increasingly powerful AI systems safe when the governments and infrastructure providers on which they depend have competing priorities?
The answer may determine not only who controls AI, but also whether the institutions responsible for developing and deploying it can remain accountable for its consequences.
