Why AI Feels Like the Internet in 1997 | Benedict Evans on a16z
We're at an inflection point where agentic coding has moved from 'kind of useful' to 'really changing everything.' The most actionable insight: Don't wait for certainty about AI's future—start experimenting with tools that automate repetitive tasks in your domain today. The companies and professiona
1h 0mKey Takeaway
We're at an inflection point where agentic coding has moved from 'kind of useful' to 'really changing everything.' The most actionable insight: Don't wait for certainty about AI's future—start experimenting with tools that automate repetitive tasks in your domain today. The companies and professionals who will win aren't those predicting the future perfectly, but those actively testing, learning, and adapting as capabilities evolve. Like mobile data before it, we're in extreme disequilibrium between supply, demand, and pricing that will settle—but the window to learn and position yourself is now.
Episode Overview
Benedict Evans, former a16z partner and tech industry analyst, reflects on the evolution of AI since his original 'AI Eats the World' presentation 18 months ago. The conversation explores how agentic coding has emerged as the first killer use case, why foundation models may be commodities rather than products, and the fundamental uncertainties about where value will accrue in the AI stack—drawing parallels to previous platform shifts like mobile, the internet, and PCs.
Key Insights
Agentic Coding: The First Killer Use Case
Software development has emerged as the first domain where AI has achieved true product-market fit, with customers 'pulling it out of your hands.' This wasn't entirely surprising—software developers were the ones experimenting with the technology, so naturally they applied it to software development first. The shift happened at the beginning of 2024, when agentic coding went from being 'kind of useful' to fundamentally changing how development teams work.
Foundation Models May Be Infrastructure, Not Products
Evans argues that foundation models lack the characteristics needed to capture value: no clear network effects, no sustainable competitive differentiation, and difficulty building defensible moats. Like hyperscalers (AWS, Azure, GCP), models may become commoditized infrastructure where customers don't care which one powers their applications. The real value will likely accrue further up the stack, in specialized applications and workflows built on top of these models.
The Chatbot Is a V1 Interface, Not the End State
The current chatbot interface is a limited first iteration that works well for some tasks but requires significant tooling, configuration, and domain expertise for most real-world applications. Just as people who excel at jobs aren't necessarily the best at designing tools for those jobs, we need product builders to create the right interfaces and workflows. Templates and 'skills' are like Excel templates—useful starting points that people eventually outgrow.
We're in Extreme Supply-Demand Disequilibrium
The current AI market mirrors mobile data in 2009-2010: extreme pricing imbalances where some pay $20/month for $10,000 worth of tokens while others accidentally rack up massive bills. This is temporary. With $1-2 trillion in capex coming and models getting 100-200x more efficient yearly, we'll reach a new equilibrium—but it's unclear if model providers will have pricing power when supply catches up.
Industry-Specific Questions Matter More Than Technology Questions
As AI matures, the critical questions shift from technology ('how good are the models?') to domain-specific implementation ('what does this mean for law firms, consultancies, or finance?'). Like Netflix, where all the important questions became Los Angeles questions rather than Silicon Valley questions, AI's impact will be determined by people who deeply understand specific industries, not just the technology itself.
Notable Quotes
"Agentic coding went from being kind of useful to really changing everything. It was going to be magic. And in 20 years time, we'll just say, 'Well, of course that's how it is. Computer's always done that.'"
"I don't think foundation models are a product. I don't think a chatbot is a product. I think the value will be further up."
"We are in this extreme scarcity. Like, we can't spend $10 trillion a year on AI infrastructure cuz there isn't $10 trillion a year there to spend on it."
"The moment that you understand something and you know how it works and what's going to happen is the moment you should move on to something else. You should always be looking for the places where we don't know what the answers are."
"Just because demand for tokens is infinite that doesn't mean that you can't get to a different price equilibrium because of course that's what happened with mobile data. Demand for bits is infinite. It's grown 1500-2000x in the last 15 years. But you still got your supply and demand price equilibrium."
Action Items
-
1
Start Experimenting with AI in Your Domain Today
Don't wait for certainty about AI's future direction. Identify repetitive or time-consuming tasks in your work and test whether current AI tools can automate or accelerate them. The learning curve matters more than picking the 'winning' platform—capabilities are evolving rapidly, and hands-on experience will help you adapt as the technology matures.
-
2
Focus on Problem-Specific Solutions, Not Generic Tools
Rather than trying to use chatbots for everything, identify specific workflows or processes where AI can be integrated with proper tooling, data, and configuration. Think about building or using specialized applications that combine AI with domain expertise, much like how Excel is powerful but eventually people need dedicated software for complex tasks.
-
3
Prepare for Pricing and Supply Dynamics to Shift Dramatically
Current AI pricing is in extreme disequilibrium and will change significantly as supply catches up with demand. If you're building on AI infrastructure, plan for scenarios where compute costs drop substantially and consider how that might change your business model or competitive positioning. Don't over-index on today's scarcity-driven pricing.
-
4
Ask Industry-Specific Questions, Not Just Technology Questions
If you're in professional services, finance, law, or other knowledge work fields, deeply examine what junior staff actually do, what clients pay for, and how AI might reconfigure those relationships. The answers require domain expertise, not just technical knowledge. Engage people who understand both the industry structure and the emerging capabilities.