Weinberg: AI is being consumed like media, not used like infrastructure

What happened

Gabriel Weinberg — the founder of DuckDuckGo, which has shipped its own AI chat product for two years — posted an essay arguing that the prevailing narrative about AI adoption is wrong in a specific, measurable way. The piece hit 456 on Hacker News in a few hours, which is unusual for a sober anti-hype take.

Weinberg's thesis is narrow and worth quoting directly: people are consuming AI the way they consume other media — episodically, for specific tasks, with long gaps in between — not the way they use infrastructure like email, search, or the OS file manager. The distinction matters because nearly every revenue projection from OpenAI, Anthropic, and the hyperscalers backing them assumes the second pattern, not the first.

He stacks the receipts. Pew's August 2025 survey put regular ChatGPT use among U.S. adults at roughly a third, and 'daily use' in single digits. Reuters Institute's 2025 Digital News Report found 7% of people use generative AI weekly for news — lower than the share that uses TikTok for news. Anthropic's own Economic Index, published earlier this year, showed that even among paid Claude users, the median session count per week is in the low single digits and the distribution is brutally long-tailed: a thin sliver of power users accounts for most of the volume.

Why it matters

The gap between vendor framing and observed behavior is not new — every consumer technology overshoots its early projections — but the scale of the AI overshoot is unusual. Goldman Sachs estimated $1 trillion in AI-related capex through 2027. Microsoft alone has guided to ~$80B in FY25. Those numbers price in an adoption curve closer to the smartphone (universal, multi-hour daily use) than to streaming video (universal account ownership, episodic engagement). Weinberg's claim is that the data fits the second curve, not the first.

The HN thread is unusually substantive — and unusually split. The pro-Weinberg camp points out that 'usage' in vendor metrics is gameable: a single Copilot autocomplete counts as a 'use,' a single Gmail 'help me write' click counts as a 'use,' and Apple Intelligence's notification summaries count as a 'use' the user did not request. Strip those out and the numbers compress hard. The skeptics — including several builders running internal AI tools — counter that B2B usage is the wrong place to measure: developers using Cursor or Claude Code for 6+ hours a day are real, just invisible to consumer surveys.

Both can be true. The honest reading of the data is that AI has a developer-and-power-user core that genuinely lives in it, surrounded by a much larger audience that opens it once or twice a week the way they open Wikipedia. That is a real business — Wikipedia is one of the ten most-visited sites on earth — but it is not a $4 trillion business.

The second-order problem is what this implies for the capex cycle. If the median user opens Claude four times a week for three minutes, the marginal GPU is not earning its depreciation. The hyperscaler bet has always assumed a behavior shift in which AI is *invoked implicitly* — agents acting on your behalf in the background, IDE autocompletes firing on every keystroke, summarization happening on every email. Weinberg's data suggests the implicit-invocation pattern is real for developers and largely absent for everyone else. Anthropic's Dario Amodei has been explicit that the path to revenue runs through agents, not chatbots; the chatbot numbers are what we have because the agents are mostly not shipped yet.

What this means for your stack

If you are building on top of foundation models, the practical implication is to stop modeling your DAU/MAU curve against the consumer-app comparables (Instagram, TikTok) and start modeling it against the media-consumption comparables (Spotify, YouTube — high reach, lower per-user intensity outside a power-user cohort). Your retention math should assume a barbell: a small group of daily heavy users who justify your inference spend, and a much larger group whose engagement looks like a magazine subscription. Pricing accordingly — usage-based for the heavy tail, flat-rate-with-floor for the dabblers — is the rational response.

If you are a buyer rather than a builder, Weinberg's data is ammunition for the procurement conversation. The default vendor pitch is 'every employee will save N hours per week.' The Anthropic Economic Index suggests the real distribution is 'a quarter of your engineers will save 8 hours, half will save 1, and the rest will not log in twice.' License the heavy users at the seat tier they need and put the rest on a metered plan. This is also why Microsoft's per-seat M365 Copilot uptake has been slower than guidance — the dabbler tail does not justify $30/seat/month even when the heavy tail does.

And if you are on the infrastructure side — running your own inference, picking between vLLM and TGI, sizing a Triton deployment — the same shape applies. Provisioning for peak concurrent users assumed under the 'AI is infrastructure' model is wildly different from provisioning for the 'AI is media' model. Burst-friendly autoscaling, aggressive cold-start optimization, and smaller always-on baselines look much better under Weinberg's curve than under the hyperscaler curve.

Looking ahead

Weinberg is careful to say he is not predicting the bubble pops — he is saying the timeline is wrong. The agent pivot, if it works, would convert AI from a thing you visit into a thing that runs in the background, and that *would* validate the capex. But the agent products that ship today are demos, and the gap between a Devin demo and a deployed Devin is the same gap that exists between a Pew survey of weekly ChatGPT use and the marketing deck's projected $200/month-per-knowledge-worker TAM. Until the median user's behavior changes — not the median developer's — the honest read is that we have built infrastructure for a workload that does not yet exist. The interesting question for 2026 is whether the workload arrives before the depreciation schedule catches up.