Summary

Generative AI may be a platform shift, but Evans sees no evidence yet that it exceeds the internet or smartphones. His “centrist” position is that AI is “as big a deal as the internet or smartphones, but only as big a deal as the internet or smartphones.” The upside remains unusually unknowable because neither intelligence nor why these models work so well has a usable theory.
AI demand is enormous yet sharply uneven, making workflow adoption the central commercial question. ChatGPT has “800 or 900 million” weekly active users, but only about 5% pay; Evans cites roughly 10–15% of developed-world users engaging daily and another 20–30% weekly. The investor question is why five times more people understand the product yet “can’t think of anything to do with it this week or next week.”
A bubble is likely, but neither timing nor ultimate infrastructure demand can be modeled with confidence. “If we’re not in a bubble now, we will be,” Evans argues, while distinguishing 1997, 1998 and 1999-style conditions is impossible in real time. Compute requirements may fall 20, 30 or 40 times a year even as usage explodes, reproducing the impossible bandwidth forecasts of the late 1990s.
The product opportunity above the models is likely to center on specialized products that encode workflows, validation and institutional knowledge. “People buy solutions, they don’t buy technologies”: law firms want legal-discovery software, not translation and sentiment-analysis API calls. The discussion points toward purpose-built applications around general models rather than raw model access alone.
OpenAI’s 800–900 million weekly users constitute distribution, not yet a durable moat. Evans sees brand and default status but no clear network effect, feature lock-in, proprietary infrastructure or cost advantage: “You get a bill every month from Satya.” OpenAI must race simultaneously toward a defensible product ecosystem and infrastructure involving NVIDIA, Broadcom, AMD, Oracle and new pools of capital.
The incumbent impact is asymmetric: Google can absorb AI, Meta and Amazon face deeper product questions, while Apple may remain insulated unless computing itself changes. Google can fund frontier models and make AI a feature of search and ads; Amazon could finally improve discovery beyond commodity retail. Apple is endangered if apps disappear, but even an LLM-first world may still demand “a nice big color screen,” camera and battery—in other words, something much like an iPhone.
The deepest disruption will expose businesses whose profits depend on routing, bundling or friction rather than their stated product. Evans’s progression is feature adoption, new capabilities, then potentially “pull the whole industry inside out”; newspapers discovered they were partly light-manufacturing, local-distribution and trucking companies, while an LLM might erase defenses built on tedious administration. To call AI bigger than the internet, however, he would need to see something “actually a person” outside narrow guardrails: “What we have right now isn’t that.”

Deep dive

1. AI may be a platform shift, but history does not reveal its winners

Evans frames the presentation around two questions: what happens inside technology when a platform changes, and which outside industries are transformed rather than merely assisted. The internet radically changed newspapers, while for cement it was “just kind of useful” and did not fundamentally alter the business.
His deliberately moderate claim is that generative AI may be “as big a deal as the internet or smartphones, but only as big a deal as the internet or smartphones.” Those shifts already created and destroyed industries, rearranged technology’s dominant companies and produced new billion- and trillion-dollar businesses.
Today’s “AI” label may itself be temporary. Like databases, the web and smartphones, mature capabilities disappear into ordinary products: Otis once marketed the infrared beam as “electronic politeness,” whereas today “it’s just a lift.” In general usage, Evans says, “AI seems to mean new stuff,” while AGI means “new, scary stuff.”
Historical taxonomies are useful but not predictive. Mobile moved computing from the web toward apps, put smartphones in the hands of five to six billion people versus fewer than one billion consumer-PC users, and enabled TikTok and modern online dating. Yet in the mid-1990s, “you can know it but not know it”: Amazon was a bookstore, Netscape had launched, and Google and Facebook’s founders had not built their companies.

2. AI’s unknowable ceiling breaks conventional technology forecasting

Previous shifts contained uncertainty but had visible physical limits: telecom operators could not deploy universal gigabit fiber in 1995, and an iPhone would not suddenly gain a year of battery life, unroll into a projector or fly. AI lacks an equivalent roadmap because “we don’t really have a good theoretical understanding of why it works so well”—or of human intelligence itself.
Evans finds a revealing contradiction in an OpenAI livestream that promised human-level, PhD-level AI researchers the following year, then promoted APIs that would enable hundreds or thousands of software developers “just like Windows.” Either AI becomes a “god in a box” that eliminates the need for software companies, or it is a new software substrate from which many more products get built; the industry often argues both positions at once.
The AGI debate resembles his theologian’s joke: either the Messiah came and little visibly changed, or arrival remains perpetually five years away. Sam Altman says PhD-level researchers are here, while Demis Hassabis rejects that characterization; Andrej Karpathy puts the horizon around a decade. With no model of the ceiling, forecasts reduce to “I feel like,” and Evans does not offer a falsifiable date.

3. Overinvestment is likely even when every spender is acting rationally

Evans’s deterministic call is that “very new, very big, very exciting, world-changing things tend to lead to bubbles.” Marc Andreessen’s distinction—1997 was not a bubble, 1998 was not, 1999 was—captures the timing problem: “If we’re not in a bubble now, we will be,” but nobody knows which year this resembles.
Compute forecasting resembles estimating global bandwidth in the late 1990s. A spreadsheet could combine users, page sizes, video bit rates and viewing time, then infer router sales, but plausible inputs would create a hundredfold outcome range. AI demand adds similarly unstable variables across capability, efficiency and usage intensity.
Hyperscalers can still invest rationally because AI is already increasing the value of search, advertising, cloud and consumer products. Their stated calculus is that failing to invest poses more downside than overbuilding—an argument that “always works well until it doesn’t,” particularly once leverage, cross-leverage and circular revenue meet a falling market.
Efficiency may already be improving by 20, 30 or 40 times a year, and another model evolution might deliver equivalent results with one-hundredth of today’s compute; usage, however, is rising simultaneously. Zuckerberg’s suggestion that Meta could resell excess capacity misses the correlation: if Meta no longer needs it, “everybody else [is] going to have loads of spare capacity as well.”

4. Adoption divides between obvious power users and people awaiting a product

Evans sees immediate deployment in software development, marketing, flexible knowledge work and narrow enterprise point solutions. Marketers can create 300 assets where they once made 30, while consultancies such as Accenture, Bain, McKinsey and Infosys can apply models to specific processes inside large companies.
The mass-market numbers tell a different story: ChatGPT has “800 or 900 million weekly active users,” around 5% pay, perhaps 10–15% of developed-world users engage daily and another 20–30% weekly. For someone using it for hours, the revealing question is why roughly five times as many informed users cannot identify a useful task this week.
Excel supplies Evans’s analogy. For an accountant, changing a 10-year DCF’s discount rate collapses days of recalculation into moments; for a lawyer, the spreadsheet is useful but not the daily job. Torenberg adds that many people may need these capabilities embedded in workflow, UX, tooling and a product that shows them what to do.
Validation determines whether probabilistic output saves labor. Generating 200 marketing images and selecting 10 is efficient; copying 200 figures from PDFs is not if a person must verify every number. Evans cites OpenAI Deep Research’s mobile-market output as an example: figures were wrong because of both bad transcription and bad source selection, so “I might as well just do it myself.”

5. The winning product will tell users what to ask

Rejecting generative AI because it cannot execute every existing task is like rejecting an Apple II in the late 1970s because it cannot run a bank. The question is not only whether the new system can perform an incumbent system’s defining workload, but what it enables that could not be done before.
The opportunity is therefore not only automating old tasks but discovering actions nobody previously attempted. Entrepreneurs can identify one such capability, understand an industry and put a button around it; users do not have to derive the workflow from a blank chatbot. Torenberg describes this as “unbundling ChatGPT,” while Evans’s examples show why the product layer matters.
His Everlaw example makes the stack argument concrete. Machine learning can supply translation and sentiment analysis, but law firms still buy cloud legal-discovery software rather than assembling AWS API calls. “People buy solutions, they don’t buy technologies,” because domain-specific process, interface and distribution remain above the model.
A graphical interface does more than expose hundreds of features: each screen’s seven relevant buttons embody years of institutional knowledge about what the user should decide next. A raw prompt throws that work back onto the user, like receiving “infinite interns” who know neither venture capital nor which source—quarterly reports, Bloomberg or PitchBook—the assignment requires. “It’s asking you absolutely everything.”

6. Model parity leaves OpenAI with scale but a fragile moat

Torenberg’s a16z reflection is that the firm’s regret was “not going bigger”: voice, image generation and other specialties produced more independent winners than expected. Huge markets can support multiple companies even within one category, while categories themselves will be bundled, split and recombined; in 1995, Evans had four or five browsers because even the web’s purpose remained unsettled.
General benchmarks now place frontier models relatively close together, yet consumer usage is radically unequal. Heavy users distinguish Claude’s tone or GPT-5.1 from “GPT-4.9 or whatever the hell it’s called”; weekly users often do not. Claude has almost no consumer usage in Evans’s framing, while ChatGPT leads Meta and Google despite broadly comparable benchmark performance.
For casual users, the underlying model may therefore be a commodity. OpenAI’s 800–900 million weekly users rest on “the power of the default and the brand,” without an established network effect, uncopiable memory, broad ecosystem, owned infrastructure or cost advantage. It must urgently build upward into browsers, apps, social video and platforms while building downward into compute supply: “We’re going to build all of them yesterday.”

7. Each hyperscaler faces a different strategic equation

For Google, frontier capability may simply become a required cost of remaining Google. Gemini can trade benchmark leadership with GPT-5.1 from month to month; maintaining that position might cost, in Evans’s deliberately loose range, $100 billion or $250 billion annually. Google can pay, improve search and advertising, invent the defining AI interface—or copy it as Android copied the smartphone pattern.
Meta has larger questions around content, social experience and recommendation, making model control strategically imperative. Amazon can sell commodity infrastructure while using LLMs to improve discovery: it is excellent at delivering a requested SKU but “terrible at telling you what SKU you want.” AI might infer intent and create demand rather than merely correlate previous purchases.
Publishers, brands and marketers may not yet know their questions. If an LLM answers a recipe request directly, what happens to the recipe site that depended on Google’s routing? If a shopper points a phone at a living room and asks what to buy, the discovery path—and who captures the commercial intent—could differ fundamentally from search or Amazon’s current catalog.
Apple’s issue depends on whether AI is a service or a change in computing’s nature. Craig Federighi’s challenge—Apple does not own YouTube or Uber either—is stronger than it sounds; Microsoft lost the development environment yet sold an order of magnitude more Windows PCs because accessing the web still required a PC. Even if apps disappear into an LLM, users may still pay for the best screen, camera and battery, potentially leaving the iPhone’s hardware position more intact.

8. AI will reveal what each business was really selling

Evans’s adoption ladder has three stages: make AI a feature, use it to create something new, then watch a newcomer potentially “pull the whole industry inside out.” For a Walmart manager in the Bay Area or D.C., that could progress from “find me that metric,” to “build me a dashboard,” to asking on Black Friday, “What should I be worried about?”
Amazon’s analogous leap is from recommending packing tape after someone buys light bulbs to inferring that the buyer is moving and showing a home-insurance offer—an intent its purchasing correlations might miss. Content faces the same distinction: does the user want a Bolognese recipe or Stanley Tucci discussing Italian cooking; a slide deck or a week of advice from Bain partners?
Platform changes expose hidden jobs and moats. Newspapers emphasized journalism but discovered that light manufacturing, local distribution and trucking were economically central. Evans offers US health insurance as a deliberately hedged thought experiment: if profitability depends partly on making processes “boring and difficult and time-consuming,” an LLM that removes mind-numbing work attacks an unacknowledged defense.
The “killer use case” for 3G turned out to be having the internet everywhere, not the narrow applications analysts tried to name. AI may produce the same retrospective clarity, but Evans’s threshold for calling it bigger than the internet is higher: capability that is “actually a person” beyond narrow guardrails. Today’s systems sometimes perform person-like tasks brilliantly, but “what we have right now isn’t that. Will it grow to that? We don’t know.”

Deciphering Secrets of Ancient Civilizations, Noah's Ark, and Flood Myths | Lex Fridman Podcast #487 AI to AE's: Grit, Glean, and Kleiner Perkins' next Enterprise AI hit — Joubin Mirzadegan, Roadrunner