The AI Company That Said No

Published 13 April 2026

I've been using Claude daily for over a year now. It handles my emails, manages my tasks, drafts my documents. So when Anthropic, the company behind Claude, went to war with the US government earlier this year, I paid attention.

The short version: the Pentagon wanted to use Claude for autonomous weapons and mass surveillance. Anthropic said no. The government's response was to blacklist them, designate them a "supply chain risk," and order every federal agency to stop using their technology.

Anthropic sued. A federal judge sided with them, calling the government's actions "an attempt to cripple" the company.

And in the same quarter, Anthropic passed OpenAI in revenue.

What the numbers actually say

Anthropic hit $30 billion in annualised revenue in April. Up from $9 billion at the end of 2025. OpenAI sits at $25 billion.

But revenue alone doesn't tell you much. The efficiency gap is what caught my attention.

Anthropic is projected to spend around $30 billion on training costs through to 2030. OpenAI's projection? $125 billion. Four times more, for less revenue.

Anthropic expects to be profitable by 2027. OpenAI has pushed their breakeven target to 2030.

Over 1,000 enterprise customers now spend more than $1 million annually with Anthropic. That number doubled in under two months. 80% of their revenue comes from businesses, not consumers. That matters because enterprise revenue sticks around.

What they've actually built

I think the revenue story makes more sense when you look at what Anthropic has shipped in the last six months.

Claude Chat is the bit most people know. The conversation interface. But unlike OpenAI chasing consumer scale, Anthropic focused on enterprise distribution. Claude is the only frontier AI available on AWS, Google Cloud, and Azure. OpenAI is only on Azure. If you're building anything on cloud infrastructure, that flexibility matters.

Claude Cowork is where I spend most of my time now. It's not a chatbot. You describe what you want done, and Claude does it. My morning email digest runs automatically. I point it at a folder of receipts and come back to a formatted spreadsheet. It connects to Slack, Notion, whatever you use, or just operates your screen directly when there's no plugin.

They shipped Dispatch in March, which lets you send tasks from your phone while Claude works on your desktop. I was sceptical about whether that would actually be useful. Turns out it is. Being able to kick off a task from the pub and have it waiting when I get back has changed how I think about delegation.

Claude Code is for developers. Terminal-based, lives alongside your IDE, handles git workflows. The thing I appreciate is that it asks permission before modifying anything. You stay in control. Anthropic claims their best engineering customers have seen 10-20x productivity gains. I can't verify that, but the developers I know who use it don't shut up about it.

Why I think this matters

When you pick an AI provider, you're making a bet. On their values. On their roadmap. On whether they'll still be around in five years.

Anthropic just demonstrated something I wasn't sure was possible: you can have principles and win commercially. You can tell the US government no, get dragged through federal court, and come out ahead.

I'm not saying they're perfect. Nobody is. But the contrast with OpenAI's approach is hard to ignore. The AI landscape is consolidating fast. The choices you make now will shape what's possible for your business later.

If you're evaluating AI tools and haven't given Claude a proper look, now's probably the time. Not because of the headlines, but because of what's actually in the product.

Latest News

29% of Fortune 500 now live with AI

New data from a16z confirms what many suspected: enterprise AI adoption is real. 29% of Fortune 500 companies are now live, paying customers of AI startups. Coding tools dominate by an order of magnitude, with support and search as the other clear winners. Tech, legal, and healthcare are leading the charge.

AI is "good enough to pass, not good enough to impress"

MIT tested 41 LLMs across 11,000 real work tasks. Result: 65% hit "minimally sufficient" quality. But when graded against higher standards, success rates dropped below 50% regardless of how much time the model had. AI handles the routine stuff. The complex work still needs humans.

The deskilling problem nobody's talking about

When Claude went down earlier this month, some developers admitted they couldn't continue working. "Outsourced half my brain to it," one posted. Researchers are warning about "cognitive debt" — AI boosting output while quietly eroding skills. The risk is highest for early-career workers who never build the baseline.

Anthropic launches Managed Agents

Hot on the heels of OpenAI's Frontier platform, Anthropic released Managed Agents — out-of-the-box infrastructure for building enterprise AI agents. Sandboxed environments, permission controls, multi-hour autonomous runs. Notion demoed using it for client onboarding. The message: the hard part of agents isn't the model, it's the plumbing.