The NSPM-11 Vendor Landscape: Two Positions, One Policy Decision, What Each Means for Government AI

The procurement story is already covered. The regulation pillar published its NSPM-11 brief on June 5, the compliance implications, the 90-day DoD directive timeline, the contract termination mechanism. If you need the regulatory picture, that’s the right starting point.

This brief asks a different question: what does NSPM-11 mean for the design of AI vendor usage policies, and which position does each major vendor now occupy in government AI procurement?

What NSPM-11 Actually Requires

The memo, issued by the White House and covered in detail in the regulation pillar, directs federal agencies to terminate contracts with AI vendors whose usage policies include restrictions that conflict with military operational requirements. The contract termination mechanism is reported by secondary sources as either termination for default or termination for convenience; the specific legal language requires verification against the memo’s primary text.

What’s not disputed: Anthropic’s standard commercial Claude models include usage policy restrictions on certain military and autonomous weapons applications. Under NSPM-11’s framework, those restrictions create termination exposure for federal agency contracts using standard Claude.

The Two Confirmed Vendor Positions

Position 1: Compliant. OpenAI, Google, and xAI do not, per publicly available policy documentation and Breaking Defense’s reporting on the Maven Smart System replacement evaluation, include usage restrictions that conflict with NSPM-11’s requirements. They’re positioned for government procurement. All three are named as candidates for replacing Claude in the Pentagon’s Maven Smart System.

Position 2: Effectively Excluded. Anthropic’s standard commercial Claude models are in this position. The usage restrictions that created Anthropic’s compliance advantages in civilian enterprise markets, the same restrictions that enterprise risk teams have cited as a reason to trust Claude with sensitive deployments, are the restrictions that create NSPM-11 exposure in government markets.

There is a third scenario worth noting with explicit qualification: secondary sources report a limited NSA waiver allowing continued use of Anthropic’s Mythos model under specific classified conditions. This has not been confirmed against NSPM-11’s primary text and should not be treated as established policy. If it’s accurate, it would complicate the two-position picture, suggesting the memo includes a carve-out mechanism for classified-application-specific model approvals. Until primary source confirmation exists, the operative model is two positions, not three.

What Maven Smart System Represents

Per Breaking Defense’s coverage, the Pentagon’s Maven Smart System used Claude as its primary LLM. Maven is one of the Department of Defense’s most visible AI programs, it gained public attention in 2018 when it became the center of Google’s employee protest over Project Maven’s autonomous targeting applications, and it has evolved significantly since. Its current scope covers AI-enabled intelligence analysis, targeting support, and operational decision-making assistance.

The Maven replacement evaluation isn’t just a procurement decision. It’s the first major test of NSPM-11’s enforcement mechanism in practice. The vendor selected will demonstrate which compliance posture, and which usage policy architecture, the DoD is prepared to accept as the new standard for government AI contracts. That sets a precedent that will affect vendor policy design conversations for years.

The Policy Architecture Fork

Here’s the strategic problem NSPM-11 creates for AI vendors not already positioned in one camp or the other.

Usage restrictions in AI vendor contracts serve multiple functions simultaneously: they manage liability exposure, they address genuine safety concerns about specific applications, they respond to customer demand from enterprise risk teams, and they comply with voluntary commitments made to policymakers and oversight bodies. Those functions don’t disappear because a government memo says to remove the restrictions.

Vendors who want to serve both the enterprise civilian market (which values restrictions) and the government defense market (which NSPM-11 says must not have them) face a genuine architectural challenge. The options are:

Separate product lines, A government-specific model with different usage terms than the commercial product. This is the most defensible structure from a policy standpoint but it requires maintaining separate model governance, separate safety evaluations, and potentially separate model weights for government vs. commercial deployment. OpenAI’s separate model approach for high-stakes verticals is the closest existing example of this architecture, though it wasn’t designed specifically as an NSPM-11 response.

Tiered access architecture, One model, two access tiers, with the government tier unlocking capabilities that the commercial tier restricts. This is administratively simpler but creates questions about whether the commercial restrictions are meaningful if the underlying model can operate without them.

Full policy alignment, Remove restrictions entirely and accept the loss of enterprise customers for whom those restrictions were a feature. This is the path that maximizes government market access but may foreclose civilian enterprise segments that have made procurement decisions based on vendor safety commitments.

What Compliance Teams at AI Vendors Should Audit Now

If your organization develops or deploys AI under a commercial contract with any federal agency, three audits are now material:

First, review your current usage policy language against NSPM-11’s contract termination criteria. Identify any clause that could be read as restricting military deployment, autonomous weapons development, or related applications.

Second, review your government contract terms specifically. Many AI vendors have separate government contract terms that may already accommodate NSPM-11’s requirements without touching commercial policy language. If that separation exists and is documented, NSPM-11’s exposure may be limited to contracts using standard commercial terms.

Third, assess what changing your usage policy language would mean for your voluntary safety commitments, your enterprise customer base, and your position in ongoing policymaking conversations. This is a legal and strategic decision, not a compliance checkbox. It requires counsel with both government contracts expertise and AI policy background.

What to Watch

Maven vendor selection: When the DoD announces a Maven replacement, the selected vendor’s usage policy architecture becomes the de facto NSPM-11 compliance template. Watch for that announcement and read the contract terms if they’re public.

The waiver question: If the reported NSA Mythos waiver gains primary source confirmation, it changes the vendor landscape meaningfully. A workable waiver mechanism could create a path for restricted-policy vendors to serve classified applications, which would alter the strategic calculus significantly for Anthropic and any vendor that follows its policy model.

Congressional response: NSPM-11 is an executive action. It can be contested, modified, or legislatively constrained. Several members of Congress have expressed concern about the Pentagon’s AI procurement speed and oversight. Watch whether NSPM-11’s contract termination mechanism attracts legislative scrutiny in the coming months.

The structural reality NSPM-11 established is this: in the current policy environment, AI safety commitments that restrict government applications are not neutral design choices. They have market consequences. Vendors and enterprise teams deploying AI in or adjacent to government contexts need to understand that their vendor’s policy position is now a procurement variable.