Library/The Economics of AI Usage and What's Next For SaaS | Benedict Evans on a16z

a16z

The Economics of AI Usage and What's Next For SaaS | Benedict Evans on a16z

1h 00mJun 8, 2026
Key Themes
agentic codingAI pricingplatform economicsmodel commoditizationenterprise softwareworkflow automationcapex constraintsconsumer surplus
Summary

Benedict Evans argues AI’s most durable near-term value is concentrating in coding, application layers, and enterprise workflows, while foundation models themselves risk commoditization.

This conversation traces how AI has shifted from broad excitement to a narrower set of high-signal use cases, especially agentic coding. From there, Evans develops a platform-economics view of the industry: pricing is still in a temporary disequilibrium, model providers face strong commoditization pressure, and much of the durable value is likely to accrue higher up the stack in software, tools, workflows, and services. The discussion expands into enterprise software, capex constraints, consumer surplus, and the possibility that AI produces more software, more competition, and new categories of work rather than simply replacing old ones.

1
Agentic coding is the current anchor use case.

The episode repeatedly returns to coding as the area with the clearest product-market fit. That matters because it suggests the AI market is not evenly distributed across all use cases; instead, real traction is concentrated where users can directly verify value and workflows are already digital.

2
Early AI pricing looks like a temporary mismatch.

Evans compares token economics to the early mobile-data era, when demand rose faster than pricing and network policies could adapt. The implication is that current bill shock and pricing confusion may ease as the market settles into a more stable equilibrium.

3
Value is likely to move up the stack.

A central argument in the episode is that foundation models may be necessary infrastructure, but not the primary capture point for profits or durable differentiation. If that proves true, apps, tooling, workflows, and vertical software could become the main places where economic value accumulates.

4
The biggest AI opportunity may be new work, not just automation.

Rather than treating AI only as a way to make existing processes cheaper, Evans emphasizes that the more important gains may come from tasks and products that were previously impossible or too expensive. That broader lens is useful because it shifts attention from simple labor replacement to new categories of software and services.

5
Enterprise software may become more competitive, not less.

The episode argues that AI makes software easier and faster to build, which can increase the number of tools and intensify competition. At the same time, pricing remains difficult and organizational workflows are messy, so adoption will likely be uneven and shaped by specific domains rather than a single universal AI layer.

6
AI infrastructure spending has real limits.

Evans argues that major cloud and AI companies are investing heavily, but those outlays cannot grow forever because of financial and physical constraints. This suggests the current capex wave may eventually normalize, even if competitive pressure keeps spending elevated in the near term.

7
A lot of near-term AI value will be hard to measure.

The conversation repeatedly notes that productivity gains, better analytics, and improved support may be real even when they do not show up cleanly in revenue or cost-line metrics. That makes AI harder to evaluate using traditional short-term business accounting alone.

8
Model commoditization is the base case unless disproven.

In the final chapter, Evans frames the commodity thesis as a logical argument rather than a certainty. The point is that frontier models may require large, sustained investment while much of the value gets captured elsewhere, which would leave model providers competing like infrastructure businesses rather than owning the entire stack.

Select any chapter text to Deep Dive with AI
01Intro and the shift to coding-first AI strategy

The conversation opens with Benedict Evans’ updated view that agentic coding is now the clearest AI product-market fit. He argues the industry is narrowing from broad generative-AI enthusiasm toward a few high-signal use cases, while the long-term winners in models and applications are still unclear.

Agentic coding has moved from promising to transformative.
AI attention is narrowing toward software development and a few other high-signal use cases.
Models may become increasingly commoditized while value shifts higher up the stack.
OpenAI is portrayed as broader and more exploratory, while Anthropic is more coding-focused.
The episode frames AI as an early platform transition with major questions still unresolved.
02Pricing Crunch and Where AI Value Accrues

Evans compares current AI pricing to the early mobile-data era: demand surged faster than pricing and capacity models could adapt. He argues foundation models may end up as commoditized infrastructure, with more durable value moving into applications, tools, and vertical software.

Token pricing resembles early mobile-data bill shock.
AI demand has outpaced pricing and capacity adjustments.
Foundation models may function like infrastructure rather than primary value-capture layers.
The chatbot interface is likely only a first-generation product layer.
The market is still in temporary disequilibrium rather than final equilibrium.
03What Comes After Coding

The discussion widens from coding to the broader ways AI could reshape industries such as professional services, media, advertising, e-commerce, and enterprise software. Evans emphasizes that the most important changes may come from enabling new things, not just automating old ones.

Current AI economics appear temporary and highly fluid.
AI’s impact will vary widely by industry and workflow.
Professional services may be reorganized as lower-level tasks are automated.
Advertising and e-commerce may improve as systems understand products and intent better.
The biggest opportunities may come from previously impossible products or workflows.
04AI & enterprise software economics: pricing, workflows, and capex limits

The episode examines how AI changes enterprise software economics, making software cheaper and faster to build while increasing competition and blurring the line between deterministic systems and probabilistic models. Evans also argues that AI infrastructure spending faces financial and physical limits, and that near-term gains may show up as productivity and consumer surplus more than obvious new revenue.

AI reduces the cost and time required to build software.
LLMs can be embedded into systems of record or used as top-layer synthesizers.
Pricing remains difficult because outcomes are hard to map to enterprise work.
AI may create more software, not less.
AI infrastructure spending cannot rise indefinitely.
Near-term AI gains may be hard to measure directly in revenue terms.
05Will Models Become Commodities?

Evans closes by arguing that the evidence points toward model commoditization unless someone can explain why that pattern breaks. He compares frontier AI economics to mobile infrastructure and says the practical challenge is turning AI into useful products for real businesses through service layers and intermediaries.

The commodity argument is presented as a chain of reasoning, not certainty.
Frontier model economics resemble mobile infrastructure economics.
Value may accrue elsewhere in the stack rather than at the model layer.
Consultancies and IT services may help businesses apply AI in practice.
AI will likely become normal and familiar in hindsight.