AI Eats the World? A Reality Check with Benedict Evans
Tech analyst Benedict Evans reviews his 'AI Eats the World' presentation roughly 18 months after writing it, reflecting on what has and hasn't changed. He argues that agentic coding has emerged as the only clear product-market-fit use case so far, while fundamental questions about value capture, model commoditization, and broader adoption remain unresolved. Drawing on analogies to mobile, the internet, and PCs, Evans suggests foundation models may end up as commodity infrastructure rather than capturing value up the stack.
Summary
In this A16Z podcast conversation, tech analyst Benedict Evans revisits his 'AI Eats the World' presentation to assess what has played out over the past 18 months and what remains uncertain. The most significant development he identifies is the emergence of agentic coding as a category with genuine product-market fit — evidenced by Cursor's revenue growing from a $9 billion to a $47 billion run rate in a short period. He notes that OpenAI initially pursued a scattershot strategy of building everything at once, while Anthropic focused on coding and found traction. However, Evans emphasizes that most fundamental questions from two or three years ago remain unanswered: whether models can differentiate, whether they can capture value up the stack, and how to drive daily consumer usage beyond the current ~10% daily active user ceiling.
Evans draws extensive parallels to previous platform shifts — PCs, the internet, and mobile — arguing that AI follows a pattern of accelerating adoption built on prior infrastructure, but with the important difference that we don't know the physical or capability limits of LLMs the way we knew bandwidth or device cost constraints in earlier eras. He uses the mobile data analogy extensively: telecoms built trillion-dollar infrastructure that transformed lives, yet all the value migrated up the stack to companies like Google, Meta, and Apple. He argues foundation models risk the same fate — becoming commodity infrastructure without pricing power, especially as efficiency improvements and open-source competition intensify.
On the question of whether models can become 'operating systems' with leverage over the application layer, Evans is skeptical. He argues there is no clear network effect for LLMs, the chatbot UI is a limited V1 interface, and model companies cannot build every vertical application themselves. He compares the situation to Windows and iOS, but notes that models lack the control mechanisms those platforms had. He suggests the more likely outcome is that models resemble hyperscalers — powerful infrastructure without decisive control over what gets built on top.
Evans discusses the broader question of AI's impact on industries like law, consulting, finance, and advertising. He argues these questions are increasingly 'half AI, half industry-specific' — meaning the right people to answer them are lawyers, consultants, and industry insiders, not technologists. He uses Netflix and Hollywood as examples of how technology enables transformation, but the important decisions become domain-specific rather than technical. He also highlights that much organizational knowledge is implicit and undocumented, making it hard to automate workflows that nobody has ever written down.
On capex and infrastructure spending, Evans notes that the big four tech companies are on track to spend 50% of revenue on capex — far exceeding even capital-intensive industries like telecoms (15-20%). He acknowledges this level of spending cannot grow indefinitely due to financial constraints, but stops short of predicting when a correction will occur. He frames the current moment as one of extreme disequilibrium between supply, demand, and pricing that must eventually normalize.
Evans closes by contextualizing AI within the longer arc of technological change, referencing a 1950s IBM ad showing engineers with slide rules being replaced by computing power. He argues every major technology wave — PCs, internet, mobile — seemed uniquely transformative at the time, and AI will follow a similar pattern: producing disruption, displacing some jobs, and eventually becoming so normalized that people forget the world was ever different.
Key Insights
- Evans argues that agentic coding is the only AI use case with genuine product-market-fit right now, evidenced by Cursor's run rate growing from $9B to $47B, while no other domain has achieved equivalent traction.
- Evans claims that roughly 10% of users are daily active AI users and 30-40% are weekly users, suggesting the technology has not yet achieved the kind of habitual daily integration seen in mature platform shifts.
- Evans contends that foundation models are likely to become commodity infrastructure — analogous to mobile network operators — where enormous capital is deployed, usage grows massively, but value migrates up the stack to application-layer companies.
- Evans argues there is no clear network effect for LLMs, no sustainable differentiation mechanism beyond willingness to spend on compute, and no equivalent of the iOS/Windows leverage that would give model companies control over the application ecosystem.
- Evans frames the current pricing and capacity crunch as directly analogous to mobile data in 2009-2010, where consumers received surprise $10,000 bills and networks scrambled to align supply, demand, and pricing — a transitory disequilibrium, not a stable state.
- Evans points out that the big four tech companies (Microsoft, Meta, Google, Amazon) are on pace to spend ~50% of revenue on capex this year — roughly double the capital intensity of telecoms and comparable in absolute dollars to the entire global oil and gas industry.
- Evans argues that many of the most important questions about AI's industry impact are no longer purely technology questions — they are 'half AI, half industry-specific' questions that require domain expertise (e.g., understanding how law firm pyramids actually work) to answer meaningfully.
- Evans claims that organizational workflows are largely implicit and undocumented, meaning a significant portion of what consultancies like McKinsey sell is the ability to discover how an organization actually operates versus how it's supposed to operate — a problem LLMs cannot easily solve from training data alone.
- Evans argues that LLMs will be strong where tasks can be clearly described and where 'the average answer' is what's wanted, but weak where the value lies in doing something differently from how most people would do it or solving problems nobody has yet recognized exist.
- Evans suggests the 'SaaS apocalypse' framing is premature because while some SaaS companies will be disrupted, AI is likely to produce an order-of-magnitude increase in total software, just as SaaS itself dramatically expanded the software market rather than simply replacing prior categories.
- Evans argues that OpenAI's late-2024 strategy of pursuing everything simultaneously — ads, e-commerce, browsers, social video — looked like asking ChatGPT for 15 business ideas and then building all of them, and that Anthropic's narrower focus on coding proved more effective.
- Evans notes that mobile data traffic has grown 1,500–2,000x over 15 years while telecom stocks have been flat for 20 years, making the case that building critical, heavily-used infrastructure does not guarantee capturing the economic value that infrastructure enables.
Topics
Full transcript available for MurmurCast members
Sign Up to Access