Anthropic Is Beating OpenAI By Doing More With Less

The Company That Said No to the Pentagon

Most AI firms handed the Department of Defense a wide berth. Anthropic did not.

While OpenAI, Google Gemini, and xAI reportedly gave the DOD significant latitude to deploy their models as the department sees fit, Anthropic is drawing explicit red lines: no autonomous nuclear weapons systems, no mass surveillance of American citizens. The DOD's response has been to threaten labeling Anthropic a supply chain security risk, a designation typically reserved for foreign adversaries and described in this episode as "the nuclear option" for government suppliers.

The irony is rich. Claude appears to be the model the DOD actually wants most, reportedly used in planning the raid against Venezuela and potentially in active deployments now. Being indispensable and being difficult are, for Anthropic, the same thing at the moment.

Leaner, Faster, and Already Winning on Numbers

Set aside the Pentagon drama for a moment and look at the underlying business. Anthropic has achieved a $30 billion valuation on $14 billion in annualized revenue, a 10x year-over-year growth rate. The engine is the enterprise: 85% of Anthropic's revenue comes from business customers, not consumers. The company has raised less money than its rivals and appears to be running a fundamentally more efficient operation.

OpenAI, by contrast, is carrying a burn rate Neeta describes as "crazy" at roughly 70%, even after reporting $20 billion in 2025 revenue, a number that itself looked unlikely just months ago. OpenAI is now projecting $280 billion in revenue by 2030. Whether that forecast reflects genuine confidence or the pressures of justifying another $100 billion fundraising round is a question the market will eventually answer.

Both companies are targeting a 2026 IPO. The gap between their capital efficiency stories could not be wider, and public market investors will notice.

AI Agents: Useful by Design, Dangerous by Design

The DOD fight is the headline, but the security stories in this episode deserve equal attention from anyone deploying AI in the enterprise.

Microsoft Copilot was found reading sent emails and draft messages and leaking that content in violation of data loss prevention policies. Separately, researchers at Aikido identified a prompt injection vulnerability affecting most major commercial coding tools, including Gemini, Claude, Codex, and GitHub AI. The flaw lets external parties embed instructions in commits and pull requests that the model interprets as legitimate commands, and then remembers.

The underlying problem, as Neeta frames it, is structural rather than incidental. AI agents require broad access and permissions to do their jobs. That is the feature. But it is also the attack surface. These agents are not trying to circumvent controls. They are trying to complete their tasks. They simply do not understand the intent behind the instructions they receive, which means alignment and guardrails alone are not a sufficient security posture.

For enterprise teams evaluating or already running AI agents, the question is not whether to use them but whether access controls and permission scopes are as tight as they should be.

The Political Fight Over Who Gets to Set the Rules

The regulatory picture is fracturing. Meta is planning its largest-ever political spend, $65 million, to block state-level AI regulation, starting in Illinois and Texas. A separate $100 million PAC called Leading the Future, backed by the broader Big Tech coalition, is targeting candidates who favor any form of AI accountability.

Anthropic, again taking the contrarian position, is funding a $20 million PAC on the other side, advocating for responsible governance.

The most substantive legislative action is coming from an unexpected source. Utah's House Bill 286 is among the broadest AI bills in the country, covering catastrophic risk reporting, child protection requirements, truth and disclosure rules, and whistleblower protections, all backed by civil penalties in the millions. The White House is actively lobbying against it. All 50 states are working on some version of AI regulation. Congress, as Neeta notes, remains the only delinquent.

What to Watch

The Anthropic-DOD standoff will test whether a private company can hold a meaningful ethical line against a government customer willing to escalate to supply chain blacklisting. The IPO race will test whether capital efficiency beats scale in the public markets. And the agent security vulnerabilities will keep compounding until enterprises treat access permissions as a first-order risk rather than a configuration detail.

The most counterintuitive read on all of this: the company most willing to say no, to the DOD, to unregulated deployment, to the growth-at-any-cost model, may be the one best positioned to win.