"So, About That AI Bubble" is a long-form feature by Rogé Karma, a staff writer at The Atlantic, published 2026-05-01. It updates Karma's own September 2025 Atlantic piece on the AI-bubble debate, arguing that the burden of proof has shifted: where the bull case once rested on speculation, it now rests on revenue and observable productivity gains, while the bear case rests on the claim that those gains will not generalize beyond software.
Summary of argument
Karma's central claim is that the structure of the AI-bubble debate inverted over roughly six months. In his September 2025 piece, the bull case rested on speculative claims about future AI capability; by May 2026, he argues, the bull case rests on revenue and observed productivity, and the bear case rests on speculative claims that those gains will not generalize. He states the inversion directly: "Six months ago, people arguing that AI was a bubble were pointing to real-world facts, whereas people arguing against the bubble hypothesis were making speculative promises about the future. Today, the roles have reversed." Karma identifies himself as a former bubble-thesis sympathizer who changed his view on the basis of new evidence, and he is explicit that the burden-of-proof framing is his own synthesis rather than a consensus position.
The piece marshals expert commentary (Mollick, Azhar, Borges, Mertens, Denain, Kedrosky) and recent empirical studies to support that thesis, while giving the bear case its strongest available framing through Paul Kedrosky.
Claude Code and the chatbot-to-agent shift
Karma identifies Claude Code's November 2025 update as the moment "AI seemed to cross some invisible threshold between interesting gadget and life-changing technology." Ethan Mollick (UPenn) is quoted: "For years now, we've been in an era of chatbots that mostly say things. Now we've officially crossed into the era of agents that can actually do things." Karma frames the change as a move from chatbot-as-conversation-partner to agent-as-autonomous-worker. He notes that OpenAI's Codex and Cursor shipped competitive coding agents in the months that followed.
Revenue and demand
Karma's evidence that revenue is catching up to capital expenditure centers on a set of growth figures:
- Anthropic: annualized run rate $14B → $30B in two months (early 2026), with Q1 growth he describes as 4× Google's during its peak expansion years. If the current rate continued, Anthropic would be the world's highest-revenue company by early 2027, per a projection by Patrick Pfau (Pfau projection).
- OpenAI: annualized revenue up 20% from December to February.
- Cloud providers (Q1 2026 vs. Q1 2025): Google +48%, Microsoft +39%, Amazon +24%, which Karma attributes to AI customers.
- CoreWeave: +168% annual revenue growth.
- Micron: roughly 3× revenue growth.
Azeem Azhar is quoted: "It's very important to emphasize that this pace of revenue growth is absolutely not normal."
Karma reports that demand now exceeds supply across compute, data centers, and chips. Anthropic limited Claude Code use during peak hours because of compute constraints; OpenAI scrapped its Sora video app to free GPUs for higher-margin uses; and even Nvidia's fourth-best AI chip, released in 2022, cost more in 2026 than three years earlier, per Dylan Patel of SemiAnalysis on Dwarkesh.
On enterprise adoption, Karma cites Ramp data showing the paid AI subscription rate among US businesses rising from about 25% in early 2025 to more than 50% by the time of writing. Goldman Sachs interviews with 40 software companies in mid-April 2026 found many "overrunning their initial budgets by orders of magnitude," with some spending up to 10% of total engineering labor costs on AI tools. Gabriela Borges of Goldman is quoted: "The speed at which we're seeing companies adapting these tools is actually quite surprising."
The METR study reversal
Karma highlights a reversal in METR's randomized study of developer productivity. The 2024 version found developers using AI completed coding tasks 20% slower, a result that figured prominently in his September 2025 bubble piece. The same researchers re-ran the study in early 2026 with current tools and found the same developers completed tasks 20% faster with AI. Karma notes the new estimate is likely conservative, because some of the heaviest AI users declined to participate without their tools.
The bear case and the generalization question
Karma gives the bear case its strongest framing through Paul Kedrosky (SK Ventures, MIT). Kedrosky's argument is that coding has features that make it unusually suited to AI — abundant training data, a narrow output space, and objective evaluation — while other knowledge work such as legal and marketing lacks objective evaluation and domain-specific training data. On this view, capital expenditure for compute and data centers is being deployed in anticipation of demand that may never materialize for non-coding tasks. Kedrosky draws an analogy to the 2006–2007 real-estate market: "Market hype leads to more demand. More demand makes you think you need more supply. Before you know it, you've built more homes than anyone can actually afford."
Against this, Karma uses a SemiAnalysis-reported framework under which all knowledge work decomposes into Read → Think → Write → Verify. On this account, coding has more data and better verification but is not categorically different from other knowledge work; it is easier first, with other work following on a delay. As empirical support, Karma cites a MIT study (Mertens et al.) that attempted to quantify AI ability across roughly 3,000 white-collar tasks taking humans three to four hours. The AI completion rate rose from 50% (mid-2024) to 65% a year later, with the authors projecting 80–95% by 2029. Mertens told Karma: "This pace of improvement isn't quite as fast as what we've seen with AI and coding. But it's still really, really fast." Ethan Mollick frames the range: "There's clearly a spectrum here, with coding on one end and things with really hard-to-judge outputs, like short-form fiction writing, on the other. But a lot of knowledge work — law, finance, consulting, marketing — falls somewhere in the middle."
Capability frontier
Karma cites Mythos as evidence that the underlying model frontier is still moving quickly, describing it as having "blown away just about every benchmark of AI progress" and as having "discovered cybersecurity vulnerabilities that had gone undetected by humans for decades." He notes OpenAI's GPT-5.5 followed. Jean-Stanislas Denain of Epoch AI is quoted: "On basically every indicator we have, we were already seeing a big acceleration in the pace of AI progress. And that was before Mythos."
Profitability
Karma is explicit that Anthropic and OpenAI are not profitable, spending revenue plus more on next-model training. He reports Anthropic targets profitability in 2028 and OpenAI in 2030, and frames the open question as whether the current growth rates are sustainable for two to four years.
Reception and provenance
This is a long-form Atlantic feature rather than a research paper; it argues a thesis by assembling expert quotes and recent empirical studies. The burden-of-proof framing is Karma's own synthesis, not a consensus position. The bull case's central empirical pillar — that knowledge work generalizes from coding — is not yet established for legal, marketing, or finance writing; the MIT 3,000-task study is supportive, but its projection to 2029 is the authors' extrapolation.
The Anthropic ARR figures are cited in the piece from Axios (Mike Allen/Jim VandeHei and April 13, 2026 Axios reporting) and are corroborated by SemiAnalysis ($44B+ run rate, inference margins around 70%, per May 1 ChinaTalk-cited reporting on Anthropic).
Several of the claims in the piece carry differing confidence. The Anthropic ARR figure ($14B → $30B in two months, Q1 2026) and the METR reversal (20% slower → 20% faster on coding tasks, same researchers and methodology) are corroborated and treated as high confidence; the Claude Code November 2025 inflection point is likewise high confidence. The burden-of-proof shift is Karma's normative framing and is treated as a position rather than a fact. Whether AI productivity gains generalize from coding to broader knowledge work remains unestablished for downstream tasks, though the empirical observation that task-completion rates are rising on the MIT benchmark is well supported.
Relationships
- supports: AI Bubble Debate — position update, with the shift-in-burden-of-proof framing
- supports: AI Coding Agents — Claude Code as inflection point, METR study reversal
- supports: Anthropic — $14B → $30B ARR data, Q1 growth comparison
- supports: AI Labor Disruption — Meta layoffs, Zuckerberg "single very talented person" quote
- related: MarketBench: Evaluating AI Agents as Market Participants — Fradkin & Krishnan (April 2026) — Karma's framing assumes models can coordinate and self-execute; MarketBench shows the gap
- related: Claude Mythos Preview — Mythos as bull-case evidence for the capability frontier