Pentagon Awards Seven AI Contracts, Excludes Anthropic Over Weapons Clause

The Pentagon announced seven AI vendor contracts on 1 May 2026, selecting Amazon Web Services, Google, Microsoft, Nvidia, OpenAI, SpaceX, and newcomer Reflection AI. One prominent name was absent: Anthropic. The company's refusal to accept standard Defense Department contract language — and the Trump administration's subsequent designation of Anthropic as a supply chain risk — has drawn a sharper line than any prior public statement between safety-oriented AI development and the expanding militarization of large language models.

The Clause That Broke the Deal

Pentagon procurement contracts for AI services include a standard clause permitting the government to deploy acquired models for "all lawful purposes." For the six companies that signed — and for Nvidia, OpenAI, and the others on the list — that language was apparently acceptable. For Anthropic, it was not.

The company's objection centered on what "all lawful purposes" can encompass. Anthropic argued that the formulation is broad enough to authorize domestic mass surveillance of US citizens and the deployment of its models in fully autonomous weapons systems — both of which conflict with the company's published usage policies. Rather than negotiate a carve-out, Anthropic declined to participate, a decision that the Trump administration answered in March 2026 by formally designating the company a "supply chain risk," which in practice barred it from the May procurement.

The designation framing is notably aggressive. Supply chain risk classifications are typically reserved for foreign entities suspected of posing security threats. Applying the label to a California-headquartered AI safety laboratory because it refused unrestricted government use of its technology represents a pointed use of procurement policy as pressure — and a preview of how the current US administration intends to manage AI laboratories that resist defense cooperation.

The EU Parallel

The decision resonates beyond US domestic politics precisely because the EU AI Act has now, through the 7 May omnibus deal, codified its own set of prohibitions in adjacent territory. The Act bans AI systems that autonomously select and engage targets in violation of international humanitarian law — a provision aimed squarely at fully autonomous lethal weapons. The "all lawful purposes" clause Anthropic refused would, under the company's reading, permit uses that the EU framework explicitly prohibits.

That creates a direct governance divergence. European AI providers operating under the AI Act cannot legally offer their systems for autonomous lethal targeting, regardless of what a customer's government might classify as lawful. US providers that signed the Pentagon contracts face no equivalent restriction under US federal law.

For European AI companies considering US government contracts, and for US companies operating in the EU market, the gap is not merely rhetorical. A company that contracts its AI to the Pentagon under "all lawful purposes" terms while also deploying the same system in Europe may find itself navigating incompatible legal obligations — a tension that the EU AI Act's extraterritorial reach, which applies to any AI system whose outputs affect EU persons, makes practically unavoidable.

Safety-First as Competitive Position

The procurement episode also illuminates a strategic question about how AI safety positioning functions in a bifurcated geopolitical market. Anthropic has consistently described its mission as the responsible development of AI for the long-term benefit of humanity — language that implies willingness to forgo contracts whose terms conflict with that mission.

From a European regulatory perspective, Anthropic's refusal looks like alignment with the spirit, if not the jurisdiction, of the EU AI Act's autonomous weapons prohibitions. From a US government perspective, it looks like an unacceptable constraint on defense prerogatives. The Trump administration's supply chain risk designation suggests Washington views the latter framing as authoritative.

The practical consequence for Anthropic is exclusion from a significant and expanding procurement market. The US Defense Department's AI spending has accelerated sharply since 2024, and the seven winning vendors are positioned to embed their models into defense infrastructure, intelligence workflows, and autonomous systems development over multi-year contract horizons. Anthropic's absence from that process is not merely a missed revenue opportunity — it is a structural divergence from the deployment trajectory that many of the lab's competitors are now locked into.

What to Watch

The question of whether Anthropic's position becomes a model for other AI developers — particularly European ones subject to the AI Act — or whether competitive pressure to access government contracts erodes stated safety commitments is one of the defining tensions in AI governance heading into the latter half of 2026. The EU's sandbox reviews and market surveillance authorities will have their own opportunities to observe how models deployed under US defense contracts behave when used in contexts that overlap with EU jurisdictions. The gap between what US procurement rules permit and what EU law prohibits has never been wider, or more consequential.