On February 26, 2026, the American company Anthropic rejected the Pentagon’s proposal for a new $200 million contract regarding the use of the Claude model—an artificial intelligence system that the U.S. military has employed in combat operations, including strikes on Iran and the capture of Venezuelan President Maduro.
The company’s leadership stated that the contract concerns the potential use of the company’s artificial intelligence for mass surveillance or in fully autonomous weapons.
The next day, Donald Trump and Secretary of Defense Pete Hegseth announced that government agencies have 6 months to cease using Anthropic products.
The company was designated a “supply chain risk”—a status previously applied exclusively to foreign sanctioned companies like Huawei and assigned to an American corporation for the first time.
The conflict between the Pentagon and Anthropic became the first public signal of this structural vulnerability—the state, which seeks to maintain technological superiority over China, does not control the artificial intelligence infrastructure on which this superiority is built.
The American technological model has a structural vulnerability. Artificial intelligence is created and controlled by private companies that shape the architecture of models, determine the rules for their use, and regulate access to application programming interfaces. Tools on which military planning depends remain subordinate to corporate sector decisions.
The architecture of Claude is built on the concept of Constitutional AI—a system of embedded ethical constraints that check every request against a set of rules and can block the execution of tasks classified by the model as potentially harmful.
For the Pentagon, this means that an algorithm involved in combat planning is capable of refusing in a critical situation, and the developer’s update policy can change the model’s behavior without warning the customer.
According to a leaked internal letter from Anthropic CEO Dario Amodei, at the end of negotiations, the Pentagon agreed to all of the company’s conditions except for one specific phrasing.
This was the prohibition on “analysis of bulk acquired data,” the very contract line that protected against using the model to work with arrays of commercial data from American citizens.
As noted by the Pentagon’s Chief Digital Officer Emil Michael, the internal constitution of an AI model cannot dictate to generals and soldiers the order of actions in the military management system.
In response, the U.S. DoD began promoting a new model of cooperation with technology companies. This model involves purchasing artificial intelligence systems along with access to their architecture and training procedures.
It also allows for modification for military tasks and autonomous deployment without relying on the developer’s commercial updates and services.
Anthropic’s position on autonomous weapons reproduces the principle that the Pentagon has formally enshrined in its own directives—maintaining human control over decisions on the use of force.
The problem is that this principle was formulated in the past decade—in conditions where no strategic competitor to the USA integrated AI into military planning at a speed comparable to the PRC.
The policy of refraining from delegating lethal powers to machines is being reviewed at a time when competing states demonstrate readiness to delegate more powers to machines than America does.
The conflict around Anthropic is the first practical manifestation of this collision—the Pentagon faces a situation where corporate restrictions of the American AI model hinder military planning, while the PRC deploys its own systems without internal corporate resistance and without public ethical constraints.
None of the leading military powers—including the USA, Russia, the United Kingdom, and Israel—has joined the international campaign to ban lethal autonomous systems. Anthropic’s position essentially attempts to implement at the corporate level restrictions that states have rejected internationally.
This approach also contradicts the economic logic of technology companies. Developing large language models requires tens of billions of dollars in investment, and creating a separate version of a system for a single government customer means losing access to a large-scale commercial market.
As a result, structural tension forms between the state and the technology sector. Washington seeks to integrate artificial intelligence into military planning and cyber operations, while companies strive to maintain control over their products and avoid responsibility for their application in combat.
The Pentagon views artificial intelligence primarily as a tool for preparing for conflict with China. In July 2025, the DoD’s Chief Digital and Artificial Intelligence Office (CDAO) signed contracts worth up to $200 million with four companies: Anthropic, OpenAI, Google, and xAI.
These contracts are intended to prototype and scale AI applications in combat planning, intelligence, and cyber operations.
Military applications are outpacing civilian use. According to studies, most economic sectors still experience a significant gap between the theoretical capabilities of AI models and their actual implementation.
The defense sector, however, seeks to apply these capabilities at the limits of what is available, making it the first to face control problems that the broader market will encounter later.
A key direction is automated analysis of vulnerabilities in Chinese critical infrastructure—energy grids, telecommunications, and industrial control systems.
Such systems can reduce the time between vulnerability detection and its integration into military planning from weeks to hours, replacing manual work of cyber operators with algorithmic analysis in real time.
The Pentagon’s dependence on commercial products is exacerbated by another factor—the active use of American models by Chinese laboratories to improve their own systems.
Chinese developers apply distillation techniques, where the capabilities of a large model are reproduced in a more compact and cheaper version. In such a scheme, even limited access to commercial models can gradually turn into a source of technological learning for autocracies.
In the case of American models, this process has begun to go beyond internal use. Anthropic stated that Chinese laboratories DeepSeek, MiniMax, and Moonshot AI were conducting large-scale extraction of Claude model capabilities through massive requests to its interfaces.
According to the company, over 24 thousand fictitious accounts were created for this purpose, and more than 16 million interactions with the system were conducted, allowing gradual reproduction of the model’s functional capabilities in their own algorithms.
The control problem exposed by the conflict between the Pentagon and Anthropic has a deeper technological dimension. Claude entered the Pentagon’s classified networks not directly, but through integrators—Peter Thiel’s company Palantir and the Amazon Web Services Top Secret Cloud infrastructure.
The cloud architecture of this system assumes that the computational infrastructure, agent memory, system prompts, and model updates remain on the developer’s servers.
A customer integrating such a system into their processes receives functionality but does not control the means of production—disabling access stops operations, and switching to an alternative requires months.
CENTCOM used Claude for intelligence data assessment, target identification, and combat scenario modeling during strikes on Iran on February 28—less than a day after Trump’s order to cease using Anthropic products.
The Pentagon began the largest American military operation since 2003, realizing that over the next months it would lose access to the system on which this operation was planned. The system was so deeply embedded in classified networks that immediate replacement proved impossible.
Palantir, which provides the integration wrapper for any model on classified networks, remains the only irreplaceable node in this architecture regardless of which model the Pentagon chooses.
Simultaneously with the internal conflict around Anthropic, China is expanding the availability of its own open-architecture artificial intelligence models.
According to MIT data, Chinese open-source models have already surpassed American ones in total downloads; Alibaba’s Qwen family has overtaken Meta Llama and become the most widespread open model family in the world.
The Qwen 3.5 family, released in February 2026 under the Apache 2.0 license, provides performance comparable to the previous generation of Claude, with the possibility of autonomous deployment on local hardware—the cost of use is approximately 13 times lower.
For Washington, this creates a threat of a different type than cloud dependence on Anthropic. Open architecture eliminates the possibility of sanctions or export pressure from the USA—the user country receives a model that cannot be remotely disabled.
At the same time, the “openness” of Chinese models does not mean their neutrality. According to a study funded by the Swedish Agency for Psychological Defense, none of the ten tested companies using Chinese base models, including Qwen, proved completely free from embedded Beijing information directives.
Internal directives of Qwen models, discovered by researchers, require presenting information about the PRC in a positive light and avoiding critical assessments—these constraints persist in derivative models built on Qwen, in eight languages, from English to Hindi.
The PRC’s generative AI regulations from 2023 require all models operating in China to undergo approval by the Cyberspace Administration and comply with party censorship rules. Chinese officials also openly discuss using AI exports to expand Beijing’s “discursive power” on the international stage.
The architectural asymmetry of the two models creates a paradox. The American cloud model assumes that all requests, data, and updates pass through the developer’s servers—the customer receives functionality but loses privacy and autonomy.
The Chinese open model in local deployment does not transmit any data to the developer—the weights are auditable, requests are not logged, the system does not “phone home.” For countries choosing between two ecosystems, the Chinese model offers full control over data with zero dependence on an external supplier.
The price of this autonomy is embedded Beijing information influence, which persists in the model’s weights even after local deployment and is not subject to complete removal through retraining.
Washington finds itself in a situation where its own cloud model is architecturally more vulnerable to developer control than the Chinese alternative.
The internal dynamics of the Chinese AI sector confirm this logic of subordination. On March 3, 2026, during the same week as the resolution of the conflict around Anthropic, the technical lead of the Qwen project, Lin Junyang, unexpectedly resigned from Alibaba along with the post-training head Yu Bowen.
A colleague from the team publicly noted that the departure was not voluntary. Alibaba restructured the Qwen team, replacing the vertically integrated development model, in which an autonomous group controlled the entire cycle from training to release, with a horizontal structure under direct corporate management.
Processes on both sides of the Pacific are moving in the same direction: in the USA, the state subordinates a private company through the mechanism of defense contracts.
In China, the corporation under state pressure subordinates an autonomous development team. In both cases, developer autonomy is reduced in favor of state or quasi-state control over the technology.
When the capabilities of American models are gradually reproduced in Chinese systems, the advantage on which American military planning relies is reduced.
This creates favorable conditions for autocracies that integrate artificial intelligence into military planning and political control systems, combining technological borrowings with centralized state management.
The resolution of the conflict with Anthropic was prepared by the Pentagon’s Chief Digital Officer Emil Michael—a former Uber top manager appointed by Hegseth to reform the DoD’s technology policy.
Michael conducted negotiations with Anthropic and OpenAI in parallel, providing the Pentagon with a backup option before the deadline expired.
The conflict also had a political dimension. OpenAI President Greg Brockman and his wife donated $25 million to the super PAC MAGA Inc., Altman—$1 million to Trump’s inaugural fund.
Anthropic, on the contrary, directed $20 million to Public First Action—an organization that advocates for AI regulation and against the administration’s deregulatory course.
In the same letter, Amodei directly stated that the real reason for the conflict is the absence of donations in support of Trump and refusal of “praises,” unlike OpenAI’s leadership.
The industry perceived this tactic as a targeted operation: over 900 employees of OpenAI and Google signed an open letter warning that the Pentagon is trying to pit AI companies against each other, using the conflict with one of them as leverage on the rest.
The chronology of events on February 27 confirms this calculation. The deadline for concluding the agreement with Anthropic was set at 17:01. Before that, Trump published a post on Truth Social, calling Anthropic “left-wing radicals trying to pressure the Pentagon.”
After the deadline expired, Hegseth officially declared the company a threat to the defense supply chain. On the same day, OpenAI signed a contract with the Department of Defense on terms that Anthropic refused.
On February 27, 2026, the day the contract was signed, OpenAI completed a funding round of $110 billion at a company valuation of $730 billion. Amazon invested $50 billion, SoftBank increased its stake to 13% through a $30 billion investment, Nvidia contributed $30 billion.
Market concentration around one supplier occurred instantly—the company that agreed to the Pentagon’s terms received an unprecedented amount of capital from the largest technology corporations on the same day.
Altman himself publicly acknowledged that blocking Anthropic creates a dangerous precedent, noting that as the more powerful side, the government bears greater responsibility for the conflict’s consequences.
A paradoxical consequence of the conflict was the strengthening of Anthropic’s market position. After the blocking announcement, the company’s projected annual revenue increased from $9 billion at the end of 2025 to $20 billion.
The Claude app rose to first place among free Apple apps in the USA, overtaking ChatGPT for the first time, and the number of paid subscribers doubled. The company that the administration intended to punish received a massive consumer base, reducing its dependence on government contracts.
For the Pentagon, this means that the mechanism of pressure on the technology sector through disconnection from defense orders has limited effectiveness—a company with a sufficient commercial base can ignore state demands without financial consequences.
An attempt to institutionally resolve the conflict failed. California Congressman Sam Liccardo proposed an amendment to the Defense Production Act (DPA) that would prohibit the government from blocking technology companies for establishing ethical restrictions on the use of their products.
The amendment did not gain votes in the committee—the voting went along party lines. As Liccardo noted during the meeting, the only problem with the Pentagon’s approach is that a law regulating military AI use does not exist—legislation is years behind the technology.
The failure of the amendment confirmed that institutional mechanisms for restraining executive power in the AI sphere are currently not functioning. While the American political system is unable to regulate the conflict around AI control, the Pentagon has lost access to a key model on classified networks.
The scale of Chinese technological expansion unfolding in parallel already has tangible dimensions. According to Microsoft research, Chinese systems account for approximately 18% of the AI market in Ethiopia and 17% in Zimbabwe.
In politically isolated markets, the concentration is significantly higher, reaching about 56% in Belarus, 49% in Cuba, and 43% in Russia.
The geography of dissemination reflects Beijing’s strategy. American companies build business around paid access and corporate licenses, limiting their penetration into markets with lower solvency.
Chinese developers combine open models with state subsidies and infrastructure investments, offering solutions at a significantly lower price. While Washington resolves issues of control over its own models, Beijing is forming an alternative technological ecosystem in the Global South.
The decision to cease using Claude in government structures and defense contractors showed that transitioning to an alternative model for work in military operations takes up to three months.
This vulnerability gap accelerated the formation of the Pentagon’s new technology policy—selecting a limited circle of companies ready to integrate their technologies directly into military infrastructure and coordinate usage rules with the government.
The NATO contract with OpenAI, signed on March 3, 2026, extends this dependence from the national to the allied level—the key military AI architecture of the alliance is tied to one supplier.
The conflict around Anthropic demonstrated that temporary loss of access to a model paralyzes military planning for months, and this risk now concerns the entire NATO infrastructure.
