The Pentagon's AI Vendor Shuffle: Nvidia, Microsoft, AWS In — Anthropic Out五角大楼 AI 供应商洗牌：英伟达、微软、AWS 进，Anthropic 出

The Setup

The U.S. Department of Defense just signed deals with Nvidia, Microsoft, Amazon Web Services, and Reflection AI to deploy their AI hardware and models on classified military networks. The agreements cover Impact Level 6 and IL7 environments — the highest security tiers for data deemed critical to national security.

This follows earlier deals with Google, SpaceX, and OpenAI. The Pentagon is building an “AI-first fighting force” with a deliberately diverse vendor stack, and it’s accelerating fast: 1.3 million DOD personnel are already using GenAI.mil, the secure enterprise platform for generative AI.

The Anthropic Backstory

You can’t understand these deals without the breakup. The Pentagon wanted unrestricted use of Anthropic’s AI tools. Anthropic said no — it insisted on guardrails against domestic mass surveillance and autonomous weapons. The DOD tried to brand Anthropic a “supply-chain risk.” Anthropic sued and won an injunction in March.

Now the Pentagon’s statement reads like a pointed reply: “The Department will continue to build an architecture that prevents AI vendor lock-in.” Translation: if you won’t play by our rules, we’ll find vendors who will.

Key Takeaways

• The military AI stack is being built right now. Nvidia provides the hardware, Microsoft and AWS provide the cloud infrastructure, OpenAI/Google/Reflection AI provide the models. This is a full-stack procurement.
• Anthropic’s principled stand has a cost. By refusing unrestricted military use, Anthropic lost access to the largest single AI procurement in history. Whether that’s noble or naive depends on your view, but the revenue is real and it’s going to competitors.
• “Vendor lock-in prevention” is the new procurement doctrine. The Pentagon is deliberately spreading contracts across competing firms — good for the ecosystem, terrible for any single company hoping to dominate defense AI.

Why It Matters

For AI investors, this is a signal: defense spending on AI is not theoretical anymore. It’s billions in committed contracts flowing to public companies (Nvidia, Microsoft, Amazon) and late-stage startups (OpenAI, Reflection AI). The companies that say “yes” to defense work are building moats that companies saying “no” cannot access.

For the broader AI safety debate, the Pentagon’s moves validate Anthropic’s fears. If the DOD is deploying AI on classified networks for “decision superiority across all domains of warfare,” the line between defensive and offensive use is thin. Anthropic’s guardrails weren’t paranoid — they were prescient.

What to watch:

• Whether Anthropic’s court battle expands beyond the injunction into a broader defense procurement ruling
• Which AI company is next to sign (Meta? Mistral? DeepSeek is obviously not in the running)
• Whether Congress attaches AI safety conditions to future defense appropriations