The Pentagon opens its classified networks to AI giants

The Pentagon brings generative AI into the era of classified networks

The U.S. Department of Defense has just taken a symbolic and industrial step in adopting generative artificial intelligence. According to The Verge, the Pentagon has concluded a new series of agreements with several major names in the sector to deploy AI capabilities in classified environments. The list of selected partners is revealing of the market’s real state: OpenAI, Google, Microsoft, Amazon, Nvidia, xAI and Reflection are part of this wave of contracts. Anthropic, however, does not appear on it.

Beyond the announcement, the signal is strong. Until now, discussions about generative AI applied to defense had fluctuated between experimentation, institutional communication and ethical debates. With these agreements, Washington is showing that the matter is entering a far more concrete phase: that of deployments on sovereign, secure infrastructure potentially connected to the most sensitive decision-making chains.

The Pentagon’s choice is not only about models. It also reshapes the hierarchy among cloud, chip, infrastructure and software providers in one of the world’s most strategic markets. For European, and particularly French, players, this development provides a valuable indication of the direction major administrations are taking: AI will not only be a matter of performance, but of trust, authorization and integration into classified systems.

A list of partners that says much about the new balance of power

According to information reported by The Verge, the agreements concern the use of AI in environments involving defense secrecy. The lineup is broad, but it is far from insignificant. First are the major cloud and software-capability providers already well established in the U.S. administration: Microsoft, Amazon and Google. They are joined by central players in the current generative wave, such as OpenAI and xAI, as well as Nvidia, whose presence underscores that at this level of criticality, the hardware layer is as strategic as the model itself.

The presence of Reflection also draws attention. Less visible in the media than the other companies mentioned, the company is entering a landscape in which defense contracts often serve as an indicator of the suppliers considered credible for highly constrained uses.

Conversely, the absence of Anthropic stands as a political and industrial message. The company, often seen as one of OpenAI’s main competitors in the field of foundation models and security, is not part of this first wave. The context matters: in recent months, discussions around military uses of AI have highlighted fault lines between technology providers and defense institutions. In this context, not being selected for classified deployments carries significant weight, because this type of contract is not limited to immediate revenue; it also determines access to data, use cases and institutional legitimacy that is difficult to regain.

From prototype to critical system: the real turning point is operational

Since late 2022, generative AI has primarily advanced through public demonstrations, office-suite integrations and sector-specific pilots. Administrations, for their part, have moved more slowly, held back by security, compliance and traceability requirements. The agreements revealed by The Verge indicate that the Pentagon now believes that certain building blocks are sufficiently mature to be deployed on classified networks.

This shift is significant. Working on a classified network does not mean installing a chatbot in a secure facility. It entails a complete chain: isolated hosting, access control, auditability, log management, data segmentation, guarantees regarding updates, human oversight and compatibility with the Department of Defense’s cybersecurity policies. In this context, providers are no longer selling just a high-performing model, but an industrial capability to operate under extreme constraints.

The logic is also geopolitical. Armed forces and intelligence agencies are seeking tools capable of accelerating document analysis, intelligence synthesis, software assistance, searches across massive corpora, planning and maintenance. But they want to do so without exposing their data to open infrastructures or dependencies deemed excessive. This is where generative AI changes status: it becomes a possible component of the sovereign information system.

The central issue is no longer merely “which model is the best?”, but “which provider can operate reliably, in a controlled manner and with authorization in a classified environment?”

Why Anthropic’s absence is a stronger signal than it appears

On paper, Anthropic remains one of the market’s most credible players. Its Claude models are regularly cited among the most high-performing for professional uses, and the company has invested heavily in its image as a safety-focused provider. Yet its absence from this wave of contracts is difficult to read as anything other than a strategic misalignment with the Pentagon’s immediate expectations.

The issue is sensitive, because the boundary between “responsible” AI and AI that can be “deployed” in defense is becoming increasingly concrete. Companies wishing to work with military institutions must balance stated principles, pressure from their employees, public reputation and commercial opportunities. OpenAI itself has changed its position on uses related to national security. Google, after the controversies surrounding the Maven program, has returned to the defense field in a more structured way. Microsoft and Amazon, for their part, have long been familiar with these markets.

In this landscape, Anthropic’s exclusion can be interpreted as a sign that competition is no longer based solely on the theoretical safety of models, but on the capacity to accept the defense sector’s political, contractual and operational conditions. For investors as well as public-sector customers, this distinction is becoming decisive.

Cloud, chips, models: defense is reshaping the entire value chain

Another lesson from these agreements lies in the diversity of the selected companies. The Pentagon is not choosing a single generative AI champion; it is selecting an ecosystem. This approach reflects the market’s actual structure. Foundation models are worth nothing without compute, secure cloud, software orchestration and integration into business workflows.

The presence of Nvidia is particularly significant. The group currently dominates the AI accelerator market, with a share that far exceeds the majority of the data-center GPU segment according to industry estimates. In a classified context, access to hardware, its supply chains and its support becomes an issue of sovereignty as much as performance. The Pentagon is therefore sending a clear message: the battle for defense AI is being fought as much in secure datacenters as in conversational interfaces.

For Microsoft Azure, Google Cloud and AWS, these contracts reinforce an already visible dynamic: major public AI contracts will be awarded to those able to combine model, infrastructure, compliance and local presence. OpenAI benefits here indirectly from its ties with Microsoft, while xAI gains institutional credibility by entering this select circle.

Viewed from Europe, this development echoes debates on digital sovereignty, trusted cloud and strategic autonomy. In France, initiatives around the trusted cloud, SecNumCloud infrastructures or AI-augmented defense projects show that the same questions are emerging: where do the models run, who controls the data, and what level of auditing is actually possible?

What this changes for France and Europe

For French and European players, the U.S. announcement serves as a real-world test. It shows that the next AI value cycle will not be limited to productivity assistants or public APIs. The most strategic markets will go to solutions capable of operating in closed, regulated and critical environments.

This trend may accelerate several developments in Europe:

• demand for models deployable on-premise or in certified clouds;
• stronger traceability requirements and data governance;
• the rise of specialized providers in defense and security integration;
• pressure on European champions to demonstrate not only their performance, but also their ability to obtain authorization.

In the context of the European AI Act, the issue becomes even more delicate. Europe is seeking to regulate high-risk uses while avoiding industrial decline. Yet the Pentagon’s contracts show that the most advanced states will drive innovation toward areas where regulation, national security and industrialization converge. Companies capable of addressing these three dimensions will have a lasting advantage.

The next step could be the most decisive: once the first classified deployments are stabilized, administrations will require complete chains of trust, from the chip to the model, including cloud, observability and governance. At that point, the hierarchy of the generative AI market will no longer be dictated by benchmark rankings or announcement effects, but by the ability to become a certifiable building block of state infrastructure. This is precisely where the sector’s real shift is now taking place.