Dario Amodei Made One Mistake. Sam Altman Got $110 Billion. Here's the Full Story.

A U.S.-Israeli strike on Iran highlighted a hard reality: large language models are already embedded in real-time military workflows, and once that integration happens, governments can’t simply “phase out” a vendor—even after a presidential order. Hours after the president directed federal agencies to stop using Anthropic technology, U.S. Central Command reportedly used Claude for intelligence assessments, target identification, and combat simulations during the Iran operation. The episode ties back to a Pentagon dispute that escalated when Anthropic refused to grant unrestricted use for lawful military purposes; the company was then labeled a supply chain risk and given a six-month phase-out. The practical takeaway is blunt: operational dependence makes political and contractual remedies slow, and it shifts leverage toward whichever model provider can stay inside classified pipelines.

That dependence isn’t theoretical. Claude had been used in earlier operations, including a January mission tied to capturing Nicolas Maduro in Venezuela, and it has been deployed through Anthropic’s partnership with Palantir on Amazon’s top-secret cloud. The transcript argues that this is part of a broader shift in warfare: AI models have moved from “supplemental” tools to load-bearing components that compress the observe–orient–decide–act loop, turning tasks that once took analysts days into reactions feasible in real time. A public example cited is an Israeli product manager, Yonathan Bach, who built a strike-radar dashboard in about six hours using Claude to estimate strike likelihood based on flight data, aircraft movements, and news feeds—an illustration of how quickly capabilities can be operationalized.

The conflict also hinges on how autonomy is framed. Dario Amodei’s February 26 statement is portrayed as less an anti-war moral stance than a technical one: Anthropic supports most Pentagon use cases and argues partially autonomous weapons are vital, while also saying even fully autonomous weapons “may prove critical” once model reliability improves. The transcript claims this makes Anthropic’s position contingent on capability, not ethics, and argues that the Pentagon’s human-in-the-loop oversight already exists under Department of Defense directive 3000.09—meaning Anthropic’s request for clearer contractual safeguards may have been rejected to preserve procurement flexibility.

While Anthropic went public and then lost leverage, OpenAI’s path is described as quieter and more strategically aligned with defense procurement culture. After Anthropic’s designation, OpenAI announced a classified deployment deal and a $110 billion funding round—timed within hours of the supply-chain action. The deal includes red lines on mass domestic surveillance, autonomous weapons, and high-stakes automated decisions, while OpenAI’s architecture is described as cloud-only with embedded engineers, limiting integration into weapons hardware. The transcript argues that deference to the defense establishment—negotiating privately rather than publicly—helped OpenAI secure both classified access and a structural advantage.

The money and infrastructure reinforce that advantage. The funding round values OpenAI at $730 billion pre-money and $840 billion post-money, with major commitments from Amazon, Nvidia, SoftBank, and others, plus expanded cloud distribution through AWS for OpenAI’s enterprise agent platform. The transcript links this to Stargate, a SoftBank-led AI infrastructure buildout targeting 10 GW by 2029, and to massive GPU and cloud commitments across multiple chip architectures. Even with Anthropic’s strong revenue growth and broader cloud availability, the transcript argues the Pentagon’s designation threatens the most valuable asset: sticky enterprise and government contracts. In the end, the central claim is that the week’s drama is a proxy for a deeper contest over who gets to own the “grid” of AI compute—especially in classified, durable procurement pipelines—and whether enterprises will accept a single-model winner or push for a multi-model future.