Six months ago, the AI bubble narrative was winning. Hundreds of billions pouring into data centers with no clear path to profitability. OpenAI's own CEO admitting investors were overexcited. Journalists comparing the build-out to dot-com excess. The bear case was clean, intuitive, and popular.

Then Claude Code happened. And the numbers stopped cooperating with the skeptics.

The Atlantic's Roge Karma published a significant piece this week that amounts to a public recantation of the bubble thesis. The headline: 'Maybe AI Isn't a Bubble After All.' Coming from a journalist who wrote one of 2025's most cited AI skepticism pieces, this is not a casual observation. It is a data-driven course correction.

The Numbers That Changed Everything

Here is what moved the needle. Anthropic's annualized revenue went from $14 billion to $30 billion in two months. That is not a growth rate. That is a vertical line. As Axios's Jim VandeHei pointed out, Anthropic grew four times as much during Q1 2026 than Google did over three years during its peak expansion in the early 2000s. Faster than Zoom during the pandemic. Faster than Standard Oil during the Gilded Age.

If this sounds like it cannot be real, the rest of the sector suggests it is. OpenAI's annualized revenue jumped nearly 20 percent from December to February. Google, Microsoft, and Amazon reported cloud revenue growth of 48, 39, and 24 percent respectively, driven largely by AI. CoreWeave's annual revenue grew 168 percent. Micron's nearly tripled.

The cause is not mystery. It is Claude Code and the agentic coding wave that followed. When Anthropic shipped its November update, AI crossed from interesting gadget to life-changing tool for software developers. Other companies, OpenAI's Codex, Anysphere's Cursor, followed with competitive products. But Claude Code was the threshold moment.

The Research Flipped Too

This is the detail that should worry remaining skeptics. Last year, the think tank METR ran an experiment showing developers completed tasks 20 percent slower with AI tools. That study was central to the bubble argument. Recently, METR re-ran the exact same experiment with the latest tools. Result: developers completed tasks almost 20 percent faster with AI. A 40-point swing in the same study, same researchers, same methodology. Some power users refused to participate without AI tools.

Goldman Sachs confirmed the enterprise side. In mid-April interviews with 40 software companies, Goldman found many were overrunning their initial AI budgets by orders of magnitude. Some companies are already spending 10 percent of total engineering labor costs on AI tools. As Goldman analyst Gabriela Borges told The Atlantic, the speed at which enterprises are adopting is surprising.

The Demand Problem Reversed

Six months ago, the worry was too many data centers and not enough demand. Today, demand is rising so fast that AI companies cannot build infrastructure quickly enough. Anthropic has been forced to limit Claude Code usage during peak hours. OpenAI scrapped its video app to free up compute. Even Nvidia's fourth-best chip, released in 2022, costs more today than it did three years ago.

This is the structural reversal that makes the bubble comparison fragile. In a real bubble, supply outpaces demand until the whole thing collapses. In the current AI market, demand is outpacing supply. That does not mean the economics are sustainable forever. Anthropic and OpenAI are still not profitable, spending everything on next-generation models. Anthropic expects to turn a profit in 2028, OpenAI in 2030. The revenue growth needs to continue for years.

The Bear Case Is Not Dead, but It Is Injured

The remaining skeptics have a legitimate argument: coding is unusually well-suited for AI automation, and the productivity gains may not transfer to legal, marketing, finance, and other knowledge work. That is a fair point. But it is a much weaker argument than 'AI is a bubble,' which is what they were saying six months ago.

Here is my read. The AI sector has crossed the revenue threshold that separates speculative mania from real industry. The growth rate is still outrageous and will eventually slow. But the question is no longer whether AI companies will make money. It is how much money they will make and who will capture the most. That is a very different question than 'is this all going to crash.'

The Atlantic piece matters because it is a bellwether. When the publication that helped build the skeptic consensus starts recanting with data, the narrative equilibrium has shifted. The bubble camp needs a new argument.

What to watch: Anthropic's path to profitability (2028 target), whether coding productivity gains translate to non-technical knowledge work, and whether the compute squeeze forces pricing increases that slow adoption.

Analysis based on reporting by The Atlantic (Roge Karma, May 1), Axios, Goldman Sachs research, and METR studies.