Pentagon Signs Eight Tech Giants for AI-First Fighting Force; Anthropic Excluded
TL;DR
Eight tech companies are now supplying AI for classified US military networks under the Pentagon's "AI-first fighting force" initiative. Anthropic is conspicuously absent — the company rejected a usage clause and was flagged as a security risk. This formalizes the 04-23 Pentagon-Anthropic standoff and crystallizes the policy reality: AI sovereignty for classified work is a binary contract gate, not a capability ladder. The companies that signed get the work; Anthropic, despite holding the most-capable cyber-offensive model (Mythos), does not.
Key findings
- Eight signing tech firms supplying AI for classified military networks (specific list per source).
- Anthropic excluded after rejecting a usage clause; flagged as a "security risk" in Pentagon framing.
- Direct continuation of the 04-23 Pentagon-Anthropic dispute — that disagreement has now resolved as exclusion, not negotiation.
- Frames AI procurement as policy-gated, not capability-gated — capability rankings (CAIS Dashboard, UK AISI cyber evals) do not predict classified-network access.
Relation to prior wiki knowledge
Resolves the 04-17 Anthropic Mythos policy thread. Mythos was the highest-scoring cyber-offensive open frontier model at the time. The 04-17 wiki note flagged the procurement question as open. That question is now closed: holding the strongest cyber model does not earn classified procurement, it earns exclusion when the usage clause is rejected.
Strengthens the 05-01 SemiAnalysis "supply is the binding constraint" thesis. Pentagon procurement is not awarding the most-capable model; it is awarding the most-compliant supplier. With supply already constrained, the labs that accept the contract terms capture the demand. This is the political-economy mirror of the SemiAnalysis margin argument.
Pairs with the 05-01 White House Mythos block. Two same-week stories: White House blocks Mythos for ~70 additional companies (citing compute), Pentagon excludes Anthropic from the AI-first contract. The Mythos access surface is shrinking on both sides — government use blocked at the top, classified contracts blocked at the bottom.
Pairs with Chinese AI startups onshoring (05-01). Both stories same day: US tightens classified procurement around domestic compliance signals; China tightens domestic registration around the same kind of signal. AI sovereignty as policy is now bilateral.
Why it matters
This is the first concrete US case of an AI capability ranking being overridden by procurement compliance. Capability evaluations (CAIS Dashboard, UK AISI cyber tests) measure what a model can do; procurement measures whether the lab will sign the clause. The two have decoupled. For a wiki tracking research → industry pipeline, this means frontier-capability narratives stop predicting government deployment, which has not been true at any prior point in the LLM era.
Open threads to track
- Which clause did Anthropic reject? The exact contract language matters — usage restriction, audit rights, deployment monitoring all have different implications.
- Who are the eight? Microsoft (OpenAI), Google, Amazon (Anthropic resale?), Palantir, Meta, OpenAI direct, Mistral US, xAI are plausible candidates; the source headline doesn't enumerate.
- Anthropic's market response. SemiAnalysis (05-01) framed Anthropic's growth as supply-bound. Pentagon exclusion removes one buyer; whether private demand absorbs the gap is a 90-day question.