OpenAI's DOD Deal, Anthropic's Blacklist, and What It Means for Government Buyers
OpenAI CEO Sam Altman said this week that "the government is supposed to be more powerful than private companies," hours after the Department of Defense cut ties with Anthropic and the White House directed agencies to halt use of Anthropic's tools. OpenAI then announced a new agreement with the DOD. Altman admitted the timing "looked opportunistic and sloppy," and said the intent was to cool things down as tensions rose.
Anthropic CEO Dario Amodei reportedly criticized Altman's relationship with the Trump administration in an internal memo, while Defense Secretary Pete Hegseth labeled Anthropic a "Supply-Chain Risk to National Security." The clash centered on how the DOD could use Anthropic's AI models. In the span of a week, one major vendor was blacklisted and another stepped into the gap.
Why this matters for government
This is a live stress test of AI procurement, risk, and continuity. It shows how fast vendor status can change and why agencies need contracts, architectures, and playbooks that can flex under political, legal, and mission pressure.
It also validates a basic principle: public accountability outranks private preference. Whether you agree with the personalities involved or not, agency decisions must map to law, policy, and mission readiness-full stop.
Key facts from the week
- Altman: Companies should not abandon the democratic process "because some people don't like the person or people currently in charge."
- DOD declared Anthropic a supply-chain risk; the White House told agencies to cease use of Anthropic technology.
- OpenAI announced a DOD agreement the same day and acknowledged the optics.
- OpenAI launched GPT-5.4 across ChatGPT, its API, and Codex, calling it its most "capable and efficient" model for professional work.
- Reported traction: ChatGPT at 900M weekly active users; OpenAI ARR at $25B; Anthropic ARR at $19B.
Immediate steps for agencies
- Coordinate with counsel, the CIO, and mission owners on any cease-use directives and vendor communications. Keep a written record of decisions and dependencies.
- Inventory AI usage by vendor, model, and data type. Flag systems that touch sensitive or mission-critical workflows.
- Activate your contingency plan: identify drop-in alternatives, required ATO updates, and any data export/migration steps.
- Issue an internal advisory covering user access changes, model substitutions, and help desk escalation paths.
Procurement and governance moves to make now
- Multi-vendor by design: Avoid single-model exposure. Use an API layer or gateway that lets you swap models without rewriting apps.
- Exit and off-ramp clauses: Include termination for convenience, data portability, model substitution, and service credits tied to material policy shifts.
- Safety and policy alignment: Require documented content policies, allow-lists/deny-lists, and configurable guardrails aligned to your mission rules.
- Data boundaries: Contract for data residency, dataset segregation, and strict limits on training with agency data.
- Assurance: Ask for third-party assessments that map to the NIST AI Risk Management Framework and relevant cyber controls.
- Ethics baseline: Reference the DOD's AI Ethical Principles in solicitations and performance measures.
- Operational resilience: Require RTO/RPO targets for AI services, incident response SLAs, and clear fallbacks if a model is paused or pulled.
How to pressure-test your AI stack
- Model swap drill: Can your team replace a vendor model in days, not months? Prove it in a tabletop or sandbox exercise.
- Data egress test: Can you export prompts, outputs, and fine-tunes in a usable format without vendor help?
- Policy toggle: Can you tighten safety filters or usage scopes across apps from a single control point?
- Human oversight: Do you have named approvers and reviewers for high-impact decisions, with logs and audit trails?
Market context agencies should note
OpenAI's scale continues to grow, with a reported $110B round at a $730B pre-money valuation and GPT-5.4 rolling into its products. Anthropic remains a strong competitor on capability and adoption, but its government posture is under scrutiny after last week's actions.
The lesson isn't to pick sides. It's to build systems and contracts that keep your mission on track even if a vendor's status changes overnight.
What leaders can do this quarter
- Set a multi-model policy and require it for all new builds.
- Add AI vendor risk to the enterprise risk register with clear owners and thresholds.
- Update acquisition templates with the clauses above and require a model replacement plan in every proposal.
- Stand up a cross-functional review board (legal, procurement, security, mission) to handle AI escalations within 48 hours.
For deeper guidance
If you need practical playbooks for public-sector adoption, procurement, and governance, see AI for Government. For structured capability building across policy, risk, and oversight, explore the AI Learning Path for Policy Makers.
The events of this week underline one idea: public institutions must keep control of mission outcomes, not hand the keys to any one vendor-no matter how impressive the model or how loud the headlines.
Your membership also unlocks:
