Not long ago, Anthropic was the AI company that defined itself by what it wouldn’t do. Founded by former OpenAI researchers who left over safety concerns, the San Francisco startup built its brand on caution, on the idea that artificial intelligence was too dangerous to rush into the hands of the powerful without guardrails. It published research on AI alignment. It talked about existential risk. It hired philosophers.
Now it’s working with the Pentagon.
The shift, first reported by The New York Times, marks one of the most significant ideological reversals in recent Silicon Valley history — and one that tells us far more about the trajectory of the AI industry than any product launch or funding round. Anthropic, valued at roughly $60 billion following its latest fundraising, has begun providing its Claude AI models to the U.S. Department of Defense through a partnership with Palantir Technologies, the defense contractor co-founded by Peter Thiel. The arrangement allows military and intelligence analysts to use Claude for tasks ranging from processing intelligence reports to logistics planning.
For a company that once positioned itself as the responsible alternative to OpenAI’s move-fast approach, the Pentagon deal represents a striking recalibration. And it raises a question that the entire AI industry will eventually have to answer: Can you be a safety-focused AI lab and a defense contractor at the same time?
Anthropic’s leadership says yes. Dario Amodei, the company’s CEO, has argued in recent months that engagement with the U.S. government — including the military — is not only compatible with safety principles but essential to them. The logic goes something like this: if American AI companies don’t provide their technology to democratic governments, authoritarian rivals like China will fill the gap with systems built without ethical constraints. Better to have safety-conscious AI in the Pentagon than to cede the field entirely.
It’s a compelling argument. It’s also convenient.
The defense and intelligence community represents an enormous market. The Department of Defense alone spent more than $17 billion on AI and related technologies in fiscal year 2025, according to budget documents. For Anthropic, which burns through cash at a prodigious rate — the company reportedly spent more than $2 billion on compute costs last year — military contracts offer a revenue stream that doesn’t depend on consumer subscriptions or enterprise software sales. The Palantir channel is particularly attractive because it provides an existing infrastructure for classified work, allowing Anthropic to avoid building out its own government-grade security apparatus from scratch.
The partnership works through Palantir’s AI Platform, or AIP, which integrates large language models into workflows across government agencies. Palantir has similar arrangements with other AI companies, but the Anthropic deal stands out because of what Anthropic used to represent. When Google employees revolted over Project Maven — the Pentagon’s drone-imagery AI program — back in 2018, it was exactly the kind of principled resistance that Anthropic’s founders later invoked when discussing their own approach to AI development. That era feels very distant now.
To understand how Anthropic got here, you have to understand the competitive pressures that have reshaped the AI industry over the past 18 months. OpenAI signed its own defense contracts in late 2024, dropping a previous policy that had prohibited military applications. Google deepened its ties with the Pentagon. Meta’s open-source Llama models found their way into defense applications with no restrictions at all. One by one, the informal taboo against military AI work eroded.
Anthropic held out longer than most. But holding out has costs.
The company’s investors — including Google, Salesforce, and a constellation of venture capital firms — poured money into Anthropic on the expectation of massive returns. Those returns require revenue. And in an industry where the biggest customers are governments and Fortune 500 companies, turning away the single largest buyer of technology on the planet starts to look less like principle and more like financial negligence. Anthropic’s board, sources familiar with the discussions told The New York Times, debated the Pentagon question for months before concluding that a carefully scoped engagement was both ethically defensible and commercially necessary.
The internal debate wasn’t smooth. Several employees objected, and at least a handful departed. Anthropic’s own Responsible Scaling Policy — a framework the company developed to govern how its models are deployed as they grow more capable — didn’t explicitly address military use when it was first published. The company has since updated its usage policies to draw lines around what it considers acceptable defense applications (intelligence analysis, logistics, cybersecurity) and what it won’t support (autonomous weapons systems, lethal targeting decisions). Whether those lines hold under pressure from a customer as powerful as the U.S. military remains to be seen.
The broader context matters here. Washington has made clear, through executive orders and bipartisan legislation, that it views AI superiority as a national security imperative on par with nuclear deterrence during the Cold War. The current administration has pushed aggressively to recruit Silicon Valley talent and technology for defense purposes, and companies that participate are rewarded with favorable regulatory treatment, security clearances, and access to classified datasets that can improve their models. Companies that don’t participate risk being shut out of the government’s AI infrastructure entirely — a disadvantage that compounds over time as competitors train on government data and build relationships with procurement officers.
This is the strategic trap that Anthropic’s leadership recognized. Opting out doesn’t just mean lost revenue. It means lost influence over how AI gets used in the most consequential settings imaginable.
Dario Amodei made this case explicitly in a lengthy essay published on Anthropic’s blog earlier this year, arguing that “safety and national security are not in tension — they are deeply complementary.” He wrote that Anthropic’s engagement with defense agencies would be governed by the same safety principles that guide its commercial work, and that the company would retain the right to refuse specific use cases that violated its policies. The essay was thoughtful and detailed. It also read, to some critics, like an elaborate justification for doing what the market demanded all along.
The criticism has come from predictable quarters — and some less predictable ones. The electronic advocacy group Electronic Frontier Foundation called the Anthropic-Palantir arrangement “a betrayal of the company’s founding ethos.” Several prominent AI safety researchers who had praised Anthropic’s cautious approach posted pointed critiques on social media. But perhaps more notable was the muted response from much of the AI research community, which suggests that the normalization of defense work is already well advanced. What was controversial five years ago is now simply the cost of doing business.
Palantir, for its part, has been open about its ambitions. CEO Alex Karp has argued for years that Silicon Valley’s squeamishness about defense work is naive and ultimately dangerous. In Palantir’s worldview, the marriage of AI and military power isn’t something to agonize over — it’s an inevitability that democracies need to get right before autocracies do. Anthropic’s participation validates that thesis in a way that no previous partnership has.
The financial mechanics of the deal are worth examining. Palantir acts as the prime contractor, handling security requirements and integration with government systems. Anthropic provides the underlying AI models and receives licensing fees. The structure insulates Anthropic from some of the operational complexities of government work — classified facility requirements, personnel clearances, compliance with defense acquisition regulations — while still giving it access to defense revenue. It’s a model that other AI companies are watching closely. If it works, expect similar arrangements to proliferate.
There’s a deeper tension here that goes beyond any single company. The AI safety movement, which Anthropic helped pioneer, was built on the premise that powerful AI systems pose risks that require careful, deliberate management. Transparency. External oversight. Slow, cautious deployment. Military applications, by their nature, push in the opposite direction. They demand speed. Secrecy. Capability maximization. The question isn’t whether Anthropic can maintain its safety commitments in a defense context — it’s whether the institutional pressures of military contracting will gradually reshape those commitments into something unrecognizable.
History offers mixed precedents. The relationship between the U.S. military and the technology industry has always been symbiotic and fraught. The internet itself emerged from DARPA research. Silicon Valley’s semiconductor industry was built on defense contracts. But the post-Vietnam, post-Snowden tech workforce developed a deep skepticism of military entanglement — a skepticism that shaped hiring, culture, and corporate policy for decades. That era is ending. The AI arms race has rewritten the social contract between technologists and the state, and Anthropic’s Pentagon pivot is simply the latest and most visible evidence of the change.
So where does this leave Anthropic’s safety mission? The company insists it’s intact. And in some narrow sense, that may be true. Anthropic continues to publish safety research. It still employs one of the largest alignment teams in the industry. Its models include more built-in safeguards than most competitors. But the frame has shifted. Safety is no longer the reason Anthropic exists. It’s a feature — one among many — that the company offers alongside performance, scale, and now, military utility.
That’s not necessarily a failure. It may be the only realistic outcome for a company that needs tens of billions of dollars to compete in the AI race. But it is a transformation, and it deserves to be named clearly. Anthropic is no longer the AI safety lab that happens to build commercial products. It is a major AI company — with defense contracts, a $60 billion valuation, and all the compromises those things entail — that also does safety research.
The distinction matters.
For the rest of the AI industry, Anthropic’s decision removes the last major holdout against defense work among leading foundation model companies. The debate about whether AI labs should work with the military is effectively over. The new debate — harder and more consequential — is about the terms. What guardrails actually hold? Who enforces them? What happens when a model trained for intelligence analysis gets repurposed for something its creators never intended?
These aren’t hypothetical questions. They are the defining policy challenges of the next decade. And they will be answered not by philosophy papers or corporate blog posts, but by the daily decisions of engineers, procurement officers, and military commanders operating under pressure, in classified environments, far from public view.
Anthropic chose to be in the room where those decisions get made. Whether that makes the world safer — or simply makes safety rhetoric more palatable to the people building weapons — is a question that won’t be resolved anytime soon. But the bet has been placed. And there’s no walking it back.
WebProNews is an iEntry Publication