Anthropic CEO Dario Amodei sent an internal memo of 1,600 words to all employees last Friday. The memo was exposed by the tech media The Information today, immediately igniting the entire Silicon Valley.

The core of the memo is just one sentence: Anthropic was blacklisted by the Trump administration, not because of unresolved security terms, but because they did not make donations.

The Distance of 25 Million Dollars

Dario named OpenAI in the memo.

He stated that the real reason the Trump administration disliked Anthropic was that the company did not donate to Trump and did not offer "dictatorial praise." Meanwhile, OpenAI had provided both money and gestures.

In September 2025, OpenAI President Greg Brockman and his wife donated 25 million dollars to Trump's MAGA Inc political action committee. According to Federal Election Commission filing documents, this donation was the largest single contribution in that cycle for MAGA Inc, accounting for nearly a quarter of their half-year fundraising total. Brockman later posted on social media, stating that the donation was to "support policies that drive American innovation" and praised the Trump administration for being "willing to engage directly with the AI community."

CEO Sam Altman took a different route. He did not directly donate large sums to MAGA Inc but donated 1 million dollars to Trump's inauguration committee in December 2024. More importantly was the gesture: the day after Trump took office, Altman stood behind the presidential seal in the White House Roosevelt Room and announced a 500 billion dollar AI infrastructure project, Stargate, stating to Trump on camera, "This would not have been possible without you." At the White House technology dinner in September 2025, he thanked Trump directly, saying, "Thank you for being such a business-friendly and innovation-friendly president."

Interestingly, this Sam Altman publicly wrote in 2016, "For anyone familiar with the history of Germany in the 1930s, watching Trump's actions is chilling." He compared Trump to Hitler and discussed the "big lie." Before the 2024 election, he donated 200,000 dollars to help Biden's re-election.

In contrast, Anthropic did nothing. No donations, no attendance at the dinners, and no standing behind the presidential seal to express gratitude.

This is the first time a CEO of a top company in the AI industry has publicly stated that your treatment in Washington depends on how much money you give to the White House.

Dario also directly criticized the contract OpenAI signed with the Pentagon. He said the reason OpenAI accepted the contract was that "they care about placating employees, while we care about truly preventing misuse," and referred to the public relations rhetoric around the contract as "naked lies."

The White House's response indirectly confirmed Dario's claims. A government official told the news site Axios, "You can't believe Claude won't secretly execute Dario's personal agenda in a classified environment." Treasury Secretary Scott Bessent directly tweeted a rebuttal, stating that no private company should influence U.S. national security terms.

The response did not mention a single word regarding the technical discrepancies in security terms. It was all personal attacks.

After the memo leaked, former Google CEO Eric Schmidt publicly supported Dario, saying, "Dario is right; this is one of the most important decisions our society faces." Meanwhile, OpenAI CEO Altman acknowledged in an internal memo that signing the Pentagon contract just hours before Anthropic was blacklisted "looks both opportunistic and hasty."

Those Betting on Both Sides, Palantir's Awkward Position

The other half of the memo dismantles the position of Palantir, the largest defense data analytics company in the United States, valued at approximately 350 billion dollars.

Palantir does not build models or chips; it acts like a government-exclusive OpenClaw, serving as an intermediary layer that integrates large models like Claude and GPT into your chat tools and workflows. Palantir integrates others' AI models into the military's classified data systems, allowing these models to read intelligence, run analyses, and perform target identification. 54% of its revenue comes from government contracts, with the U.S. market accounting for 74% of total revenue. Its relationship with the Pentagon is not one of cooperation, but of symbiosis.

By the end of 2024, Anthropic entered the Pentagon's classified network through Palantir, with Claude becoming the first frontier AI model deployed within U.S. military classified systems. In July 2025, the Pentagon awarded Anthropic a contract worth 200 million dollars. The contract ceiling for the Maven intelligent system, Palantir's flagship AI project for intelligence fusion and target identification, mentioned a ceiling of 1.3 billion dollars. The U.S. military's six major combatant commands and NATO rely on Claude.

Then things began to go wrong. In January 2026, during the U.S. military's operation to capture Venezuelan President Maduro, Claude participated in intelligence analysis through the Palantir platform. Anthropic later questioned Palantir whether Claude was used for the firing phase. Palantir forwarded the question to the Pentagon. The military deemed this to be an AI vendor reviewing military operations after the fact, causing the relationship to irreversibly fracture.

At a critical moment when negotiations stalled between Anthropic and the Pentagon, Palantir did something to pour fuel on the fire: it pitched the Pentagon a self-developed "classifier" security solution, claiming it could automatically determine whether each use of Claude crossed the line through machine learning. The implication was that even if Anthropic was unwilling to sign an unlimited contract, they could still control its model. This solution effectively gave the Pentagon an out—since Palantir claimed it could manage things, Anthropic's security terms became redundant.

Dario dismantled this solution in the memo. He said, "About 20% is true, 80% is showmanship." The reasons being: models cannot determine if they are in a loop of autonomous weapon systems; they do not know whether the data they analyze comes from foreign sources or U.S. citizens; they do not know if the data was obtained with user consent or through gray market channels; jailbreak attacks are frequent and easily executed. Palantir's classifier cannot answer any of the four questions.

Dario stated that Palantir's real understanding of its position on Anthropic is: "You have some disgruntled employees, and you need to provide them with something to placate them."

On March 3, Palantir CEO Alex Karp fired off unmentioned critiques at a Washington Defense Tech Summit hosted by top Silicon Valley venture capital firm Andreessen Horowitz: "If Silicon Valley thinks it can take away everyone's white-collar jobs and then screw over the military, you are out of your mind (retarded)." Everyone knew who he was talking about. However, what he did not mention was that the company he referred to as "uncooperative" was precisely his own platform's most core AI provider.

Palantir sold the Pentagon a set of AI security layers, with the AI running in those layers being Anthropic's Claude, while Anthropic's CEO stated that this security layer was merely showmanship. Palantir found a reason for the Pentagon to kick out Anthropic, but after kicking out Anthropic, it was Palantir that suffered the most.

Dismantling the Engine Hurts More Than Imagined

Reading this, one might naturally think: Claude is just the model base, wouldn’t it work to switch to OpenAI’s GPT or xAI’s Grok? Just like switching the default model in OpenClaw?

It's not that simple. Reuters cited two informed sources today stating that there are a large number of prompts and workflows within the Maven intelligent system that are built around Claude. This is not just a matter of changing an API address. The prompt chains, workflows, output formats, and security audit processes are all fine-tuned to Claude’s behavioral patterns. Changing models means reconstructing and testing a complete process used for military intelligence analysis and target identification. Informed sources indicated Palantir would need to "rebuild part of the software."

The Maven contract ceiling of 1.3 billion dollars is effective until 2029. The deployment scope spans the U.S. military's six major commands and NATO. The timeline for the National Geospatial-Intelligence Agency, the primary geographical intelligence agency in the U.S., is for Maven to start delivering "100% machine-generated" intelligence to combatant commanders by June 2026. Now that the engine is to be changed, this timeline is likely to slip. Analysts at Wall Street investment bank Piper Sandler said: Anthropic is deeply embedded in the military and intelligence systems; "connecting and negotiating alternative technologies requires time and resources that could have been used for growth opportunities."

Michael Burry, the prototype for the film The Big Short and Palantir's most well-known short seller, added a sting: The six-month transition period precisely highlights that the stickiness lies in Claude's technology, not on Palantir's platform. If Claude could be switched as easily as swapping models in OpenClaw, why bother with a six-month transition period at all?

Wall Street does not care about these issues. After Anthropic was blacklisted, the tech stock boutique investment firm Rosenblatt raised Palantir's target price from 150 dollars to 200 dollars, and UBS also upgraded its rating. On March 4, Palantir's stock price rose by 3.28%. During the same period, CEO Karp and co-founder Peter Thiel sold over 400 million dollars' worth of Palantir stock between February 20 and March 3. Analysts urged buying while the founders were selling.

On the same day the memo was exposed, a twist in the story occurred.

Dario stated on March 4 at the Morgan Stanley Technology Conference that Anthropic is "trying to cool things down and reach a mutually viable agreement" with the Pentagon. He asserted that Anthropic and the Pentagon "have far more common ground than differences." According to insiders, during the five days of being blacklisted, Anthropic executives had privately expressed regrets about the previous communication methods to the Pentagon.

However, the leak of the memo might have stirred the pot again. Axios reported that White House officials believe Dario's attacks on the Trump administration in the memo "could ruin reconciliation chances." A government official’s exact words were: "You can't believe Claude won't secretly execute Dario's personal agenda in a classified environment."

Interestingly, OpenAI was also lending a hand. Altman, when signing the Pentagon contract, proactively requested that the government "offer the same terms to Anthropic" and publicly opposed designating Anthropic as a "supply chain risk." He called it a "very bad decision."

The Pentagon gave Anthropic 48 hours to make a decision. In the time since they dismantled the engine for Palantir, it has already been over a week. Yet now, both sides have begun talks again.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。