A rational conversation on where AI is actually going | Benedict Evans

What It Covers

Benedict Evans, independent tech analyst and former a16z partner, argues AI ranks alongside the internet and mobile as a platform shift — not larger. Drawing on his "AI Is Eating the World" presentation, he examines where value accrues in the AI stack, why job displacement fears are overstated, and how enterprise adoption timelines constrain the pace of change.

Key Questions Answered

• •AI Adoption Timeline: Treat current AI development as equivalent to 1997 internet — most applications haven't been built yet, adoption is uneven, and roughly 60% of 13-18 year olds still report zero usage. Daily active users remain a minority even in tech-forward demographics. Expecting rapid universal transformation misreads how platform shifts actually propagate through economies and organizations historically.
• •Foundation Model Pricing Power: Foundation model companies likely lack durable pricing power because no winner-takes-all network effects have emerged between competing models. With three to six large labs selling functionally similar outputs, commodity pricing dynamics should eventually apply — similar to how global mobile telecoms generate $1 trillion revenue but flat stock returns over 25 years despite exponential data consumption growth.
• •Task vs. Job Distinction: The critical analytical question for any profession is whether AI automates the task or the actual job. McKinsey clients pay for organizational diagnosis and political navigation, not PowerPoint slides. Amazon retrieves known SKUs but cannot determine which SKU you need. Identifying where this split falls in a given profession determines actual displacement risk versus productivity augmentation.
• •Enterprise Adoption Lag: Enterprise software sales cycles run 18+ months, meaning even companies ready to deploy AI face multi-year implementation timelines. Replacing core systems like SAP takes 3-10 years sector by sector. The assumption that companies will rapidly fire staff after purchasing AI tools fundamentally misunderstands how large organizations evaluate, procure, and integrate new technology infrastructure.
• •Distribution as the Dominant Moat: When AI model outputs become commodities, distribution determines market outcomes. Google deploys Gemini across existing search surfaces; Meta embedded its models across all social platforms before most users noticed. OpenAI's late-2024 strategy of launching products across every surface reflects recognition that default placement and user inertia — not model quality differences — will determine consumer market share.
• •Professional Services Demand Increases: Contrary to expectations, leading AI labs including OpenAI and Anthropic are expanding headcount and investing in consulting-style forward-deployed engineering teams. Reimagining internal workflows requires dedicated project teams of 5-10 people working 1-2 months, followed by separate implementation projects. Organizations lack idle internal capacity for this work, driving demand for external professional services rather than eliminating it.