AI Eats the World: Benedict Evans on the Next Platform Shift

Evans frames AI within the context of previous platform shifts (PCs, Internet, mobile), examining what patterns repeat and what's fundamentally different. Unlike past shifts where physical limits were known, we don't understand AI's theoretical ceiling or why it works so well, making forecasting uniquely challenging.

• •AI may be 'as big as the Internet or smartphones, but only as big' - not necessarily larger
• •Key difference: we don't know the physical limits of AI technology or have good theoretical understanding
• •The term 'AI' only applies to new things - databases and machine learning are no longer considered AI
• •AGI is either already here (just more software) or perpetually five years away
• •Historical platform shifts created winners/losers in tech but affected industries very differently

Discussion of bubble dynamics and the massive infrastructure investments by hyperscalers. Evans notes that transformative technologies deterministically lead to bubbles, and we're seeing rational actors betting that the downside of underinvesting exceeds the downside of overinvesting.

• •Very new, very big, world-changing things tend to lead to bubbles - if not now, then soon
• •Hyperscalers saying 'downside of not investing is bigger than downside of overinvesting'
• •Forecasting compute requirements is like predicting bandwidth in the late 90s - impossible to model accurately
• •Zuckerberg's claim about reselling excess capacity ignores that everyone would have excess simultaneously
• •Inference costs dropping 20-40x per year while usage increases, creating unpredictable dynamics

Evans highlights a critical disconnect: ChatGPT has 800-900M weekly active users, but only 10-15% use it daily. He explores why 5x more people have accounts but can't think of anything to do with it, comparing this to how spreadsheets transformed accounting but didn't help lawyers.

• •ChatGPT: 800-900M weekly users, only 5% paying, 10-15% using daily in developed world
• •Clear use cases exist for software development, marketing, and flexible knowledge work
• •Most people 'get it, know what it is, have an account' but can't find weekly use cases
• •Excel transformed accounting (month of work in 10 minutes) but didn't help lawyers with daily tasks
• •Products wrap technology in workflows - 400-500 SaaS apps in typical companies are 'unbundled Oracle'

Analysis of when AI works versus when it fails, focusing on validation challenges. Evans uses OpenAI's Deep Research as an example where numbers were wrong, and discusses how error rates determine whether AI augments or replaces human work.

• •Marketing use case works: generate 200 images, pick 10 good ones - more efficient than making 10
• •Data entry fails: if you must verify 200 numbers, might as well do it yourself
• •OpenAI's Deep Research demo had wrong numbers - either transcription errors or wrong sources
• •Raw chatbot asks you 'literally everything' - no institutional knowledge or workflow guidance
• •GUIs reflect careful thought about what questions users should be asked at each workflow step

Evans applies disruption theory to AI, noting that new platforms are typically bad at what mattered to old platforms but enable entirely new things. He questions whether we're looking for the wrong use cases by focusing on replacing existing tasks.

• •Apple II couldn't run banks, Netscape couldn't do professional video editing - wrong questions
• •New platforms do different things: 5B smartphones vs 700-800M consumer PCs
• •AI may not be good at many old tasks but enables things 'you would never have done before'
• •Discovery happens two ways: users with chatbots OR entrepreneurs building specific products
• •Software companies exist to unbundle general capabilities into specific workflows and solutions

Detailed analysis of how major tech companies are positioning themselves. Evans examines sustaining vs. disruptive dynamics, noting that mobile was transformative for Meta but 'just search' for Google, and explores what AI means for each player's core business.

• •Google: frontier models are table stakes, search optimization continues, can afford $100-250B/year CapEx
• •Meta: bigger questions around content, social, recommendation - imperative to own models
• •Amazon: commodity infra business plus potential for breakthrough recommendation/discovery
• •Apple: most intellectually interesting - 'we don't have YouTube or Uber either' defense
• •Apple's Siri demo from 2 years ago still can't be delivered by Google or OpenAI consistently

Evans dissects OpenAI's strategic challenges: 800M users but no network effects, no infrastructure ownership, and commodity-level models. They must simultaneously build product differentiation and secure compute infrastructure while competitors have both.

• •Models are commoditized for casual users - Claude, ChatGPT, Gemini benchmark similarly
• •ChatGPT dominates consumer usage through brand and distribution, not technical superiority
• •No network effects yet, no feature lock-in, no cost advantage - 'very fragile' position
• •Must scramble in two directions: build products (browser, social, apps) AND secure infrastructure
• •Getting monthly bills from Microsoft while trying to build defensible business value

Evans explores how different industries will experience AI transformation, from step one (make it a feature) to step three (complete industry redefinition). He uses examples from retail, content, and health insurance to illustrate varying levels of disruption.

• •Step 1: Make it a feature and absorb it. Step 2: Do new stuff. Step 3: Redefine the industry
• •Walmart manager: find metric → build dashboard → 'what should I worry about on Black Friday?'
• •Amazon: 'bought light bulbs, here's tape' → should show home insurance ads for people moving
• •Content business: recipe sites vs. Stanley Tucci talking about Italy - which survives LLMs?
• •Industries protected by 'boring, time-consuming tasks' face unbundling - didn't realize that was their defensibility

In closing, Evans addresses what would need to be true for AI to exceed the Internet's impact. He emphasizes that current AI isn't a replacement for actual people outside narrow guardrails, and we'd need to see fundamental capability shifts, though defining that threshold remains philosophically challenging.

• •We forget how enormous the iPhone and Internet were - some still claim smartphones aren't a big deal
• •Current AI is not a replacement for an actual person outside very narrow, tightly constrained guardrails
• •Would need to see something that shifts perception from 'does people-like things sometimes' to 'is actually a person'
• •Deliberately not giving falsifiable answer - unclear what that would even be
• •Larry Tesler's law: 'AI is whatever doesn't work yet' - becomes philosophy as much as technology