Trump imposes multi-vendor AI at the Pentagon: the end of Anthropic's monopoly on US defense

Trump imposes multi-vendor AI at the Pentagon: the end of Anthropic's monopoly on US defense

🔎 Why Washington decided to break up AI monopolies

On June 5, 2026, Trump signs a memorandum that redraws the map of the Pentagon's AI vendors. The objective is simple: to prohibit national security agencies from relying on a single vendor. This memo does not come out of nowhere. It is the direct political response to the power struggle between Anthropic and the Department of Defense, a conflict that lasted nearly four months.

It all starts in February 2026, when Secretary of Defense Pete Hegseth designates Anthropic as a "supply chain risk." The reason: Anthropic refuses to give the military unrestricted access to Claude, notably requesting safeguards prohibiting use in combat without human supervision.

The June 5 memorandum transforms this bilateral conflict into structural policy. From now on, no AI vendor will be able to lock down defense contracts alone. OpenAI, Google, xAI, and other players enter the dance.

The key points

• On June 5, 2026, Trump signs a memorandum requiring national security agencies to work with multiple AI vendors, putting an end to the possibility of exclusive contracts.
• This memo follows the designation of Anthropic as a "supply chain risk" by the Pentagon on February 27, 2026, after the startup's refusal to lift usage restrictions on Claude for the military.
• An executive order on June 2 creates a framework for the voluntary review of advanced AI models, led by the NSA, with a 30-day deadline.
• OpenAI distances itself from this review framework, creating an unprecedented split with the White House on the issue of safety.
• 8 of the 10 largest US companies use Claude: the issue of multi-vendor goes far beyond the military framework.

Recommended tools

The June 5 memorandum: what it actually changes

The text requires every national security agency to maintain contractual relationships with at least two distinct AI providers. It's written in black and white in the official White House communique.

The memorandum is explicitly addressed to the Secretaries of Defense and Homeland Security, the Attorney General, and the Director of National Intelligence, according to the Associated Press analysis picked up by Yahoo.

In practice, this means that any contract awarded to a single provider will have to be complemented by a backup agreement with a competitor. The Pentagon will no longer be able to say "we only use Claude" or "we only use GPT". It will have to demonstrate a diversity of providers.

This is a paradigm shift. Until now, DoD AI contracts followed the classic logic of defense procurement: one prime contractor, subcontractors. The memo imposes a multi-vendor logic at the level of the models themselves.

The White House describes this text as a "historic directive on AI in the national security enterprise," with massive investments in American AI infrastructure. The message is clear: national security relies on a diversity of providers, not dependence on a single player.

Anthropic vs Pentagon: the origins of the conflict

To understand the June 5 memo, we need to go back to Anthropic's announcement of Claude Mythos in April 2026. This advanced model triggered a legal dispute with the Trump administration over a contract with the Pentagon, as reported by the Associated Press.

Anthropic's position is consistent with its historical stance. The startup demands explicit restrictions: no use of Claude in combat without human supervision, no mass surveillance via AI. Safeguards that Dario Amodei considers non-negotiable.

The Pentagon, for its part, wants unrestricted access. For the DoD, a model that refuses certain military uses is a model that fails to fulfill its contract. Hegseth's position is radical: if Anthropic does not fully cooperate, it is a risk to the supply chain.

On February 27, 2026, Trump directs all federal agencies to stop using Anthropic's AI technology. Mayer Brown's legal analysis describes this designation as a major precedent: never before had an AI provider been excluded from the federal market in this way.

Just Security details the legal implications: the "supply chain risk" designation is a powerful regulatory tool that goes beyond the simple contract. It practically prohibits any federal agency from working with Anthropic, unless granted an explicit exemption.

On June 5, Bloomberg confirms that the memorandum is designed to directly address this conflict. The message to Anthropic is clear: if you don't play by the Pentagon's rules, the market will pass you by — and other providers will take your place.

The June 2 executive order: the voluntary review framework

Three days before the memorandum on multi-vendor, Trump signs another crucial text. On June 2, 2026, an executive order creates a voluntary review framework for advanced AI models, with a focus on cybersecurity risks.

This framework is led by the NSA. The principle: before deploying an AI model in a sensitive infrastructure, the provider can submit its model to a 30-day review. It is voluntary, but the political pressure to participate is obvious.

The timing is not insignificant. According to the NYT, Sam Altman was scheduled to visit the White House on June 3, the day after the signing. Altman's visit fits into a clear dynamic: OpenAI positions itself as the cooperative partner, in contrast to Anthropic.

This review framework raises a fundamental question: what does "voluntary" mean when the President of the United States has just banned your competitor from the federal market? The distinction between incentive and constraint becomes blurred.

The framework applies to advanced models, those that could pose significant risks in the event of misuse. In practice, this targets the models at the top of the agentic leaderboard: GPT-5.5 (score 98.2), Gemini 3 Pro Deep Think (95.4), Claude Opus 4.7 Adaptive (94.3).

OpenAI diverges from the White House on safety

The most surprising aspect of this sequence is OpenAI's position. On June 3, 2026, POLITICO reveals that OpenAI is distancing itself from the review framework of the executive order.

This is a significant split. OpenAI, once portrayed as the White House's "teacher's pet," refuses to align with a framework it deems insufficient or inadequate. Altman's position is subtle: OpenAI cooperates on the ground (visiting the White House, publicly supporting Dario Amodei when Anthropic is attacked) but rejects the proposed regulatory framework.

This divergence illustrates a tension found in the corporate world. AI providers want to cooperate with the government, but not at the expense of their own internal safety framework.

For OpenAI, the risk is twofold. On the one hand, aligning too closely with the administration on a framework deemed weak could harm its safety reputation. On the other hand, publicly distancing itself could put it in the political crosshairs — exactly what happened to Anthropic.

The parallel with Anthropic is striking. Sam Altman offered his public support to Dario Amodei when Trump ordered agencies to stop using Anthropic tools, according to the BBC. But this support has its limits: OpenAI will not follow Anthropic into an open confrontation with the government.

The winners and losers of the new landscape

With multi-provider imposed, the AI defense deck is being reshuffled.

OpenAI emerges strengthened from the conflict. GPT-5.5 tops the agentic leaderboard with 98.2 points. The company has shown it can navigate its relationship with Washington without breaking ties. Its models — GPT-5.5, GPT-5.4 Pro (91.8), GPT-5.4 (87.6) — offer a complete range for various military uses.

Google is positioned as the natural alternative. Gemini 3 Pro Deep Think reaches 95.4 in agentic and 92 in general. Google has the advantage of cloud infrastructure (Google Cloud) already massively deployed within the government. The diversity required by the memo works in its favor.

Anthropic is the big political loser. Banned from the federal market since February, the June 5 memorandum institutionalizes its exclusion. Claude Opus 4.7 Adaptive remains a leading model (94.3 agentic, 90 general), but it is unusable by the Pentagon. The question is whether this exclusion is temporary or permanent.

Smaller players indirectly benefit from the multi-provider approach. Moonshot AI's Kimi K2.6 (88.1 agentic, 84 general) and Z.AI's GLM-5 (82 agentic) could find openings in secondary contracts. The memo doesn't say "work with OpenAI and Google." It says "work with multiple providers." This leaves room for diversity.

xAI with Grok 4.1 (79 agentic, 90 general) and DeepSeek with DeepSeek V4 Pro Max (88 general) remain outsiders for US defense contracts, for obvious sovereignty reasons. But the multi-provider principle creates a precedent that could serve them in other markets.

Why multi-vendor AI has become a business necessity

The Anthropic-Pentagon conflict is not just a Washington affair. According to 365i's analysis, 8 of the 10 largest US companies use Claude in their workflows.

This is the crucial point that many commentators missed. Anthropic's designation as a supply chain risk is not just about the Pentagon. It creates a precedent that worries the private sector.

A company that has built its entire AI infrastructure around Claude wakes up one day to find its provider designated as "at risk" by the federal government. What does the IT department do? Does it continue using Claude? Does it replace it? And with what?

A multi-vendor approach is not just a matter of getting a better price. It's an insurance policy against political risk. No political event should be able to shut down your workflows. This is exactly the logic behind multi-agent architectures collaborating with multiple AIs.

This logic also applies to deployment channels. A company that deploys its AI agents on a single messaging platform takes the same kind of risk. Multi-platform gateway architectures that allow you to connect Discord, Slack, WhatsApp, and more become tools of resilience, not just convenience.

This is also why orchestration platforms like ruflo are gaining traction. When you orchestrate agents across multiple models, the loss of one provider becomes a minor incident, not a major crisis.

The table below illustrates the vulnerability of a single-vendor strategy:

What the memo means for the global AI industry

The June 5, 2026 memorandum goes beyond US defense. It establishes a principle that will spread: relying on a single AI provider is a systemic risk.

For European companies, the signal is twofold. On the one hand, US regulation can exclude a provider overnight. On the other hand, multi-vendor becomes the norm, which favors open ecosystems and interoperability standards.

For China, this is one more argument to accelerate the development of sovereign models. If the United States itself does not trust a single provider, the logic of technological sovereignty is reinforced.

For AI startups, the memo creates an opportunity. Large defense contracts will no longer be "winner-takes-all". There will be room for specialists, niche models, custom solutions — as long as they fit into a multi-vendor architecture.

The agentic benchmark model provides a glimpse of the current hierarchy, but the memo could accelerate the emergence of specialized models for specific military tasks (satellite imagery analysis, cryptanalysis, scenario simulation) where a general score is not the relevant criterion.

❌ Common mistakes

Mistake 1: Thinking the memo only targets Anthropic

The memorandum is written in a neutral manner and applies to all providers. Anthropic is the trigger, not the sole target. Tomorrow, it could be OpenAI or Google that finds itself in conflict with the government. The text is designed to prevent any future dependency, not to punish a specific player.

Mistake 2: Confusing the June 5 memo with the June 2 EO

These are two distinct texts. The June 2 executive order creates a framework for voluntary model review (led by the NSA, 30 days). The June 5 memorandum imposes multi-provider requirements on national security agencies. The two are politically linked but have different legal scopes.

Mistake 3: Believing Anthropic is dead in the US market

The federal ban is serious, but the private market remains immense. 8 of the 10 largest US companies use Claude. Nothing in the memo prohibits the private sector from working with Anthropic. The real question is how many companies will follow the political signal and diversify, even without being obligated to do so.

Mistake 4: Ignoring OpenAI's position on safety

OpenAI is not just a simple executor of the White House. Its public divergence over the review framework shows that even the most politically aligned provider has its limits. Reducing the landscape to "Anthropic = safety, OpenAI = compliance" is a superficial reading.

❓ Frequently Asked Questions

Does the memorandum literally require using at least two providers?

The text requires national security agencies to maintain relationships with multiple AI providers. It does not set an exact minimum number, but the intent is clear: zero monopoly, zero single dependency.

Can Anthropic return to the federal market?

Theoretically yes, but this would require either a change in Anthropic's stance on military usage restrictions, an explicit waiver from the DoD, or a change in administration. None of these options are imminent.

What happens to private companies using Claude?

Nothing legally binding. But the political precedent is strong: if the Pentagon considers a provider "at risk," corporate risk departments at major companies will start asking questions. The indirect pressure is real.

Is the June 2 executive order really voluntary?

The review framework is formally voluntary. But in a context where a competitor (Anthropic) was banned for non-cooperation, the pressure to participate is considerable. It is "voluntary" with a gun on the table.

Which models are covered by the NSA review framework?

The text targets "advanced AI models" presenting significant cybersecurity risks. In practice, this aims at the models at the top of the leaderboard: GPT-5.5, Gemini 3 Pro Deep Think, Claude Opus 4.7 Adaptive, GPT-5.4 Pro.

✅ Conclusion

The June 5, 2026 memorandum marks the end of the era of exclusive AI contracts at the Pentagon — and potentially the end of the single-vendor strategy across the entire sector. The lesson for any organization deploying AI today is simple: vendor diversity is not a luxury, it is a resilience requirement. To build robust architectures, understanding multi-agent patterns has become indispensable.