OpenAI Vs. Anthropic: How the Pentagon Picked Its Partner
Summary
- Anthropic chose two non-negotiable safeguards—no mass domestic surveillance and no fully autonomous weapons—even at the cost of its federal business. On February 26, Dario Amodei said, “These threats do not change our position. We cannot in good conscience accede to their request.” Trump then ordered a six-month federal phaseout of Claude, while Defense Secretary Pete Hegseth directed that Anthropic be designated a supply-chain risk.
- The supply-chain action could be existentially punitive, but its formal status remained uncertain at recording. Hegseth said no military contractor, supplier, or partner could conduct commercial activity with Anthropic, yet the hosts had seen only his social-media post—not a proceeding delivered to the company. Anthropic said it would fight the designation in court.
- OpenAI announced an agreement to deploy models in the Pentagon’s classified networks while claiming essentially the same red lines Anthropic was punished for defending. The contracts were unavailable, leaving two live explanations: either officials pursued a political vendetta against Anthropic, or OpenAI accepted substantive concessions behind legal language. Casey Newton’s assessment was that, except perhaps the secretary of defense, nobody in the situation knew which explanation was true.
- The dispute turns on whether “all lawful use” protects Americans when AI and privacy law leave major surveillance practices legal. Federal agencies may lawfully buy data about millions of Americans from brokers and potentially feed it into AI systems for analysis at scale—activity Casey called “functionally equivalent” to domestic surveillance. His framing: “This is the whole ballgame here.”
- OpenAI’s proposed “safety stack” may offer employees reassurance without solving the underlying provenance problem. A model cannot necessarily determine whether personal data was gathered legally or purchased from a broker; sources characterized the safeguards as “security theater.” That puts unusual weight on employee scrutiny, internal resistance, and eventual whistleblowing.
- For investors, the episode elevates political alignment from reputational noise to material counterparty and platform risk. A former Trump official called the campaign against Anthropic “attempted corporate murder,” while Kevin described it as possibly the most punitive federal action against a major American company this century. Early consumer reaction may cut the other way: some users were switching from ChatGPT to Claude, including Katy Perry purchasing Claude Pro.
Deep dive
1. Anthropic’s red lines triggered an extraordinary federal threat
Kevin dates the escalation to Thursday, February 26, with one day remaining before the Pentagon’s deadline. Amodei refused to compromise on mass domestic surveillance or fully autonomous weapons: “We cannot in good conscience accede to their request.”
Trump’s late-Friday statement ordered every federal agency to stop using Anthropic technology, allowing a six-month phaseout. Because it omitted both a supply-chain designation and the Defense Production Act, Casey initially thought the government might “walk away from this whole debacle and try to save some face.” Invoking the Defense Production Act, Casey said, would have effectively nationalized or partly nationalized Anthropic and forced it to make a version of Claude that did the government’s bidding.
Hegseth escalated a little over an hour later: “Effective immediately, no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic.” At recording, Casey’s caveat was crucial: no formal proceeding had surfaced, and Anthropic said it would litigate.
2. OpenAI secured a deal whose protections remain impossible to verify
After a leaked Thursday message expressing solidarity with Anthropic, Sam Altman announced Friday that OpenAI would deploy models inside the Pentagon’s classified network. OpenAI maintained that its agreement reflected the same constraints against mass domestic surveillance and autonomous weapon systems.
Casey locates the conflict in the Pentagon’s demand for “all lawful use.” Without meaningful AI regulation or a national privacy law, Anthropic argued that legality leaves dangerous practices available rather than establishing a sufficient safeguard.
Kevin’s close reading of Altman’s statement found room for interpretation: the Pentagon “agrees with these principles, reflects them in law and policy, and we put them into our agreement.” Amodei said Anthropic had been offered similar concessions, but that accompanying “legalese” would have made them ineffective.
Casey offered two unresolved explanations: officials may have pursued an ideological or personal vendetta against Anthropic—including reported bad blood between Emil Michael and Amodei—or OpenAI may have conceded what Anthropic refused. The unseen contracts may contain the answer.
3. Legal data purchases expose the weakness in technical guardrails
Casey’s concrete example was another federal-agency social-media dragnet applied to prospective immigrants. Officials might deny that it constitutes surveillance, but AI capable of reviewing every post in real time would make the practical distinction increasingly hollow.
Data brokers can legally assemble information on millions of Americans, and federal agencies can legally purchase it. That arrangement does not meet the legal standard for domestic surveillance, Casey argued, but is “functionally equivalent” in practice. Feeding the data into a model also does not let the model reliably determine how it was obtained.
OpenAI told employees it believed it had won stronger protections through a model-level “safety stack.” Casey recalled promised Sora safeguards involving Bryan Cranston that did not arrive on time; Kevin’s sources called the new proposal “security theater.”
The hosts framed the stakes as more than a contractual dispute: who controls powerful technology—the people who build it, or the militaries and governments where it is built?
4. Political loyalty is becoming a condition of doing business
Former Trump official Dean Ball called the pressure campaign “attempted corporate murder.” Casey compared the governing logic to China’s treatment of technology companies: align with the party or risk being crushed.
Kevin judged the threatened action “by a fairly wide margin” the most punitive taken against a major American company this century, possibly ever. He connected that chilling effect to Silicon Valley’s rightward shift, citing Tim Cook’s golden trophies for Trump and Greg Brockman’s $25 million PAC donation.
Employees at OpenAI, Google DeepMind, and other labs signed a letter supporting Anthropic’s position. Casey considered that activism meaningful and hoped workers would obtain and scrutinize the contracts—and blow the whistle if real deployments resembled domestic surveillance.
Casey applied the regulatory-capture accusation to OpenAI: it used “vibes, charm” and political relationships to enter a dispute involving its largest rival, ultimately accepting that government business requires operating on government terms. “That is regulatory capture.”
5. Altman’s credibility and Anthropic’s moral posture now face real tests
Casey highlighted the “limb” Altman has crawled onto: critics among his former co-workers said he told different people what they wanted to hear, a criticism tied to his 2023 firing. Casey hopes Altman truly secured both red lines but fears naivety or deception could conceal surveillance now and “autonomous murder bots” in perhaps three or five years.
Kevin identified three open questions: whether the supply-chain designation becomes formal and whom it binds; what OpenAI accepted that Anthropic would not; and whether consumers respond. One early signal was Katy Perry posting her newly purchased Claude Pro plan with a red heart.
Amodei had long given employees The Making of the Atomic Bomb, believing advanced AI could become as consequential to national security and global order as nuclear weapons. Kevin’s conclusion: this is precisely the moral conflict Amodei expected once government wanted powerful technology “on their terms.”