The Pentagon just leaked your AI vendor playbook

AWS, Google, Microsoft, NVIDIA, OpenAI, SpaceX, Reflection AI. Oracle was bolted on hours later, bringing the count to eight. The DoD''s Chief Digital and AI Office published the rationale in plain English: ''build an architecture that prevents AI vendor lock and ensures long-term flexibility for the Joint Force.'' That single sentence is the entire 2026 enterprise AI procurement memo, free of charge — copy it into your next vendor MSA verbatim. Notably absent: Anthropic, whose Claude has been running on classified networks for months via Palantir''s Maven toolkit. The administration has been trying to ban Anthropic from federal work; lawsuits are active. A single political relationship just disappeared the most-deployed frontier model on the most sensitive workloads in the U.S. government.

Reflection AI has 60 employees, no released frontier model, and its code agent Asimov is still on a waitlist. It''s also reportedly raising at a $20B valuation right now, after NVIDIA, Sequoia, Lightspeed, and Eric Schmidt put $2B in at $8B in October. The Pentagon just slotted that 60-person team alongside AWS on the same contract vehicle. The signal isn''t about Reflection''s product — it''s about American sovereign open-weight models as a strategic category. Sovereign-AI buyers in the Gulf, India, and EU follow the DoD''s lead within 60 days. Every hyperscaler procurement team in Frankfurt and Riyadh just got a new name on their RFP.

The published scope: ''streamline data synthesis, elevate situational understanding, and augment warfighter decision-making in complex operational environments.'' Translate that to commercial: agent-driven analysis on your most sensitive data — M&A diligence rooms, financial close, regulated R&D, audit workpapers, claims adjudication on PHI. The same eight vendors will be fighting for the same workloads in BFSI, pharma, and energy by Q4 2026. Whoever the DoD picked for IL7 today is whoever your CISO will pre-approve for your tier-1 data tomorrow. The procurement battle is not theoretical — it''s already settled at the top of the federal stack.

If your AI strategy is single-vendor on Anthropic — or single-vendor on anyone — you just watched the worst-case scenario play out in real time. Claude was deployed, working, classified-network certified. One political shift later, it''s on the outside of the contract list and in court. The CIOs who survived this week are the ones who already have parallel inference fabric on at least two of {OpenAI, Anthropic, Gemini, Llama-class open weights}, with prompts and fine-tuning artifacts portable across all of them. The ones who didn''t are now writing emergency 90-day migration plans. This is exactly the model-portability clause we''ve flagged for two consecutive premium briefs — the difference is now you have the most authoritative case study possible.

The DoD refused to pick a winner. Three hyperscalers, three foundation labs, one open-weight challenger, one launch-vehicle company turned satellite-AI vendor. That''s the architecture. Mid-market and enterprise buyers about to sign three-year exclusive AI contracts in Q2 2026 just got a free second opinion from the most paranoid buyer alive: don''t. Force vendor diversity at the contract layer. Force model portability at the architecture layer. Treat single-vendor AI exclusivity the way you''d treat single-vendor cloud exclusivity in 2018 — a bet you''ll regret in 18 months.