Three percent. That's the share of households that pay for any AI service, according to analyst Ed Zitron in a recent long-form piece at wheresyoured.at. Generative AI tools - ChatGPT, Gemini, Claude - have been freely available globally for years. The bottleneck isn't access. Something else is going on.
Zitron's argument is methodical where most AI criticism is vague. He pulls specific numbers from specific filings and earnings calls and shows how they contradict each other.
Anthropic's Revenue Math Doesn't Work
Here's one example. On March 9, 2026, Anthropic's CFO stated the company had generated "$5 billion in lifetime revenues." Thirty-five days earlier, on February 12, the company claimed "$14 billion in annualized revenue." These two numbers cannot both be true. The second figure comes from multiplying a four-week API usage window by thirteen - a window that happened to coincide with a major model launch and an unusual spike in token consumption (tokens are the units AI models process - roughly one token per word). That's not sustained revenue. That's a marketing construct.
The same pattern shows up at OpenAI. The Wall Street Journal reported the company presents two versions of its financials: one that includes training costs (losing money) and one that excludes them (nearly profitable). Break-even is projected for the 2030s. OpenAI is currently burning a projected $121 billion on compute over two years. These are not the economics of a business that has found product-market fit.
What "AI Agents" Actually Are
The other major target in Zitron's analysis is the "AI agent" category, where almost every major announcement turns out to be a chatbot connected to an API. Goldman Sachs partnered with Anthropic months ago and still cannot specify what tasks the agents will actually perform. OpenAI's Frontier platform and Meta's agent initiatives reduce to the same thing: text generation plus access to external systems.
That's not nothing - but it's also not what's being sold.
On the coding side, the numbers are similarly uncomfortable. One financial services firm increased code production from 25,000 to 250,000 lines per month using AI coding tools. The result was a one-million-line code review backlog, more security vulnerabilities, and more demand for senior engineers to review the output - not fewer. Academic literature documenting genuine productivity gains from AI-assisted coding remains thin.
Meanwhile, Meta built an internal leaderboard called "Claudeonomics" tracking employee token consumption. One employee burned $80,000 in compute. The Information estimated Meta's total token consumption at 60 trillion monthly - potentially $900 million in annual spend. This is corporate experimentation with no measured return, disguised as demand signal.
The Normalization Problem
What Zitron is really writing about is how business journalism covers AI. Announcements get treated as achievements. Projections get treated as results. The "early days" framing - meant to explain away current limitations - doesn't hold up when the tools have been globally available for years and adoption has plateaued at 3% of paying households.
Hyperscalers have committed $600 billion in planned data center spending. Microsoft spent $37.5 billion in capital expenditure in a single quarter on AI infrastructure. Only 5 gigawatts of global data center capacity is actually under construction. The gap between announced investment and physical reality is enormous.
None of this means LLMs are useless. But there's a meaningful difference between "useful for some tasks" and "the next industrial revolution." The financial infrastructure being built assumes the latter. The usage numbers suggest the former.