What is a Browser? What is a Bubble?
ChatGPT has an Answer
What is a Browser? What is a Bubble? ChatGPT Has an Answer
The web front door has been morphing into AI for a while now, this week it accelerated. With Atlas, OpenAI didn't bolt a chatbot onto Chrome; it turned the browser into an answer and action layer. And that has everyone - from Wikipedia editors to Google's ad machine - reaching for the smelling salts.
Here's the context:
1) The browser is becoming an agent, not a link map.
Atlas launched as a native ChatGPT browser: it is able to summarize in‑page via a right side AI panel, draft in‑place, and in preview, click and complete multi‑step tasks for you. Per OpenAI's own docs, Agent Mode can open tabs, navigate, fill forms - an outcome‑first model of the web.
The backlash is immediate. Anil Dash called Atlas "anti‑web" for rendering AI summaries that look like pages without obvious pathways back. Wikipedia's Marshall Miller found that an 8% "rise" in traffic was actually bots evading detection - and after reclassification, human visits are down roughly 8% YoY. That's what happens when conversations and answers replace links: fewer impressions, fewer clicks, less feedback data. For the user a huge gain in value and productivity. For web sites a minus, unless they migrate their business model from search to AI.
Cloudflare's Matthew Prince is pressing the U.K. CMA to force Google to unbundle its search and AI crawlers so publishers can block AI use without disappearing from search. That's the crux: if assistants own the session, who pays the sources? Matthews remedy may penalise Google but in a world where AI is becoming the user interface a wholesale migration of the paid link ecosystem to AI will be required if traffic is to hold up and grow.
2) Bubble talk vs. buildout - and why "Minsky" isn't destiny.
Paul Kedrosky warns of a Minsky moment - credit migrating from "good" to "bad" projects via vendor financing and SPVs until the music stops. It's a valuable alarm: he has watched the 'circular deals' and thinned coverage ratios.
But this week's data argues that the investments into AI are funding real infrastructure based on real customers and growing demand and thus revenues. Dwarkesh Patel shows NVIDIA's 2025 earnings could cover multiple years of TSMC's capex. Google's decade‑old TPUs are finally gaining outside traction. Anthropic locked a multi‑year Google Cloud chips pact precisely because compute is the scarce input to a booming service.
In other words, cash is coming from two external sources, not just accounting loops: investors (a16z lines up $10B across growth/AI/defense) and customers (GPU clusters, TPUs, cloud contracts with hard dollars attached). A Minsky moment is a theory of unstable credit regimes - not a synonym for "lots of spending." The test is simple: are customers paying for capacity at rising scale? So far, yes.
3) The web's economics must reprice - fast.
Fast Company is right: one generative answer compresses an entire results page of ad inventory. Google will adapt (it's already jamming ads into AI Overviews), but the pie slices differently based on AI market share of primary consumer use.
The fix isn't to ban AI answers; it's to instrument them and include relevant links. The web needs AI to include it. But it also needs receipts: durable citations, usage‑based licensing, and verifiable payouts to knowledge origins. Regulators are already in "harms first" mode - the FTC's staff post centers on fraud, surveillance, and discrimination. The more the assistant mediates reality, the more provenance, consent, and settlement become needed product features.
The infrastructure build out will continue so long as the demand for ever smarter AI doesn't dissipate.
"We want to create a factory that can produce a gigawatt of new AI infrastructure every week." - Sam Altman
There are things to look out for:
Atlas adoption: do users stay in the AI pane, and does Agent Mode work without brittle misfires?
Pay the sources: does the CMA force crawler unbundling - and do Reddit/Wikipedia‑style usage deals become standard?
Capex vs. revenue: do chip rental prices and utilization stay tight, validating the buildout - or does secondary GPU pricing sag?
Google's ad pivot: can "ads as answers" replace the link‑page cash cow without starving the open web? Or can OpenAI, Anthropic and others build a link based revenue model?
Bottom line: The internet's UI is shifting from navigation to delegation. Will the money - and the credit - shift with it?
Essay
💸 The imminence of the AI bust [correct]
Exponential view • Azeem Azhar • October 18, 2025
Essay•AI•MinskyMoment•Hyperscalers•Capex