Benedict Evans: AI Eats the World

Updated 26 May 2026

https://www.ben-evans.com/s/2026-Spring-AI.pdf

China + open source will like make AI model commodity, but you need infra to run it. OpenAI / Anthropic / Cloud providers getting infra commitments, but will compete on pricing. Value capture is at the app layer? If AI is electricitym electric company doesn’t capture value, people who use electricity do? Using AI to solve real problems is valuable. Need to be close to problems.

Evans’ answer to “where does value capture happen?” is essentially: not at the model/infrastructure layer — probably up the stack, in software and applications. Here’s how the argument builds:

Models are likely commodities (no value capture there)

Frontier LLMs by aggregate benchmark score show that for most general and consumer use, models are very similar, and crucially there are no network effects. Evans draws an explicit parallel: mobile networks are a trillion-dollar industry, but all the use-cases and value-capture are built by other people — commodity infrastructure rarely captures value up the stack. Telcos built the pipes; Google and Meta captured the value.

Infrastructure/capex spending is a race with unclear ROI

LLMs are capital-intensive, have no network effects, and may be commodities with low margin — a stark contrast to software, which is capital-light, benefits from network effects, and commands monopoly-like high margins. The open question he poses: will LLMs have pricing power, or will they just be infrastructure?

The provisional thesis: value migrates up to applications

His provisional thesis is that chat is a terrible UX, general use needs “apps,” labs can’t build all the apps, models are commodities with no network effects, and therefore models will just be infra — with innovation moving up the stack.

Where specifically up the stack?

Evans points to a few candidates:

• Coding, analytics & productivity, marketing, customer support, and back-office processes are working first — and enterprise AI spending is overwhelmingly in coding relative to other categories.
• Whoever owns the use case layer — proprietary data, workflows, tooling, GUI, GTM — captures the value. The hard part of building AI software is use cases, tooling, GUI, process, GTM, and proprietary data — the question being whether the model goes at the top of the stack or the bottom.
• Systems integrators and consultants (Accenture et al.) are currently capturing a lot of the deployment value, because automation takes a lot of manual labor to deploy.

The structural analogy

The key question for any industry is: what can you and your competitors do with this, and can this unlock or break something crucial in your business model — and to presume radical uncertainty.

Bottom line: Evans thinks the capex war at the model layer is largely a commodity trap. Value capture happens wherever someone builds the software layer — workflows, vertical applications, proprietary data — on top of the infrastructure. The analogy he keeps returning to is telcos vs. the internet app layer: the pipe builders didn’t win; the app builders did.