Episode At a Glance
- Podcast: a16z show
- Episode: "The Economics of AI Usage and What's Next For SaaS | Benedict Evans on a16z"
- Guest: Benedict Evans
- Host: Erik Torenberg
- Published: June 8, 2026
- Duration: 1 hr 33 sec
Episode Overview
- Summary: Benedict Evans argues that AI has moved from broad excitement to a sharper question: coding works, but usage economics, SaaS disruption, infrastructure spending, and model value capture are still unsettled.
- Central question: If AI usage keeps rising, who captures durable profit: model labs, infrastructure owners, or application and workflow companies?
- Core argument: AI will probably create far more software, but pricing, adoption, and value capture are still in disequilibrium.
- Why it matters: SaaS buyers, founders, and investors must separate real workflow change from temporary scarcity, hype, and unsolved ROI math.
👉 Core Insights
1. Coding is the first AI market with obvious pull
Evans says the past year narrowed the AI conversation. Instead of asking whether every broad use case will work at once, the industry found a concrete one: agentic coding. Developers were already experimenting with the tools, and software development became the place where customers pulled the product forward.
That does not answer what happens to engineering teams, junior roles, or company structure. Evans is careful on that point: the market is too young, the tooling changed too recently, and pricing is still unstable. The first mass-market AI proof point is not every job at once; it is coding with visible demand.
2. AI pricing looks like early mobile data
Evans compares today's token economics to mobile data around 2008-2010. Users wanted flat-rate plans, networks had real marginal costs, and sudden usage spikes forced carriers to rethink caps, bundles, throttling, and fair use.
The same mismatch appears in AI: one person can pay a flat monthly fee and consume enormous token value, while another can experiment for a few days and receive a shocking bill. AI pricing is still trying to align marginal cost, perceived value, and scarce capacity.
3. Foundation models may be infrastructure, not the final product
Evans does not claim models are definitely commodities. His position is more cautious: the argument for commoditization is strong enough that model labs need to explain why it will not happen.
He points to missing network effects, limited sustainable differentiation, and the fact that enterprise buyers often do not care which cloud or model powers a product. If the model is abstracted away behind a SaaS workflow, it may be essential without being strategically controlling. Models can be essential without controlling the final application layer.
4. SaaS is more likely to be recomposed than erased
AI makes software cheaper to build and enables analysis that older systems could not perform. That will hurt some SaaS companies. But Evans expects the result to be more software, not less software.
His enterprise map has big horizontal systems, vertical SaaS, internal tools, Excel, email, shared files, and now LLM-assisted tools. AI becomes another option for where a workflow lives. The SaaS shock is about where workflows move, not whether software disappears.
5. The hardest enterprise work is discovering the real workflow
Evans emphasizes that many business processes are not documented, not in training data, and not easily explained by the people doing them. Official process maps often miss incentives, politics, exceptions, and tacit knowledge.
That is why consultants can still matter: they are allowed to interview across silos and find how the company actually works. The hardest enterprise AI work is often discovering how a company actually runs before automating it.
6. AI capex has a ceiling
Large technology companies may feel that underinvesting in AI is existentially risky. Evans accepts that logic, but he also points to financial gravity.
He compares AI infrastructure to telecom and oil-and-gas-scale capital spending. Hundreds of billions can be rational in global infrastructure, but trillion-dollar annual escalation cannot continue forever. The question is not whether AI infrastructure is valuable; it is how much spending can be sustained.
👉 Stories from the Conversation
1. Anthropic's coding focus
Evans contrasts OpenAI's broad product push with Anthropic's narrower coding focus. Whether Anthropic chose that path deliberately or stumbled into it, the outcome was clear: coding worked while many other consumer and enterprise uses remained fuzzier.
Speaker: Benedict Evans
Why it matters: Focused workflow pull may matter more than broad platform ambition in the first durable AI markets.
2. Mobile data as the token pricing analogy
Evans returns several times to mobile data. Flat-rate pricing made sense to users, but the network still had capacity costs. Then smartphones, 3G, and YouTube created demand patterns that forced carriers to rebuild pricing around caps, bundles, throttling, and fair use.
Speaker: Benedict Evans
Why it matters: Explosive infrastructure usage does not guarantee that infrastructure providers capture the best profit pools.
3. Graduate recruiting can live in many software layers
Evans uses graduate recruiting to show how enterprise workflows move across tools. A large firm hiring thousands of graduates may need dedicated software. A small company hiring five people may use email and a shared Google Sheet. The middle can shift between Workday, Excel, vertical software, or now an LLM-built tool.
Speaker: Benedict Evans
Why it matters: AI enters a fragmented software landscape rather than replacing one clean category.
4. Consultants expose the process that is not written down
Evans says consultants create value by interviewing across teams, finding why official strategies are not followed, and discovering incentives that managers may not see from the org chart.
Speaker: Benedict Evans
Why it matters: Enterprise AI adoption depends on messy organizational discovery, not only model capability.
👉 Memorable Quotes
| Quote | Speaker |
|---|---|
| "Agentic coding went from being kind of useful to really changing everything." | Benedict Evans |
| "The pricing has got to get back into alignment with the cost." | Benedict Evans |
| "The answer is more software, like way more software." | Benedict Evans |
| "All the decisions are really exception handling." | Benedict Evans |
| "In 20 years time, we'll just say, well, of course that's how it is." | Benedict Evans |
👉 Data Highlights
| Value | Label | Explanation |
|---|---|---|
| 6 months | Coding shift window | Evans says agentic coding did not work in the same way six months earlier, making the market structure too young to predict. |
| $20/month vs $10,000 | Token pricing mismatch | He contrasts flat-fee access that can consume high token value with API experiments that can create unexpectedly large bills. |
| 1,500-2,000x | Mobile data traffic growth | Evans uses the rise in mobile data traffic since the smartphone pricing shock as a comparison for AI demand growth. |
| $1 trillion | Mobile network revenue | He says mobile networks collectively have about a trillion dollars in revenue while much of the profit moved up the stack. |
| $200 billion/year | Mobile network capex | The mobile network capex figure anchors his comparison between valuable infrastructure and value capture. |
| 300-400 SaaS apps | Large-company SaaS footprint | Evans says a typical large U.S. company may have 300 to 400 SaaS apps plus many internal applications. |
👉 Points of Debate
1. Will AI value sit in models or applications?
Host view: Erik asks whether AI looks more like cloud, where infrastructure captures value, or like the internet, where applications and higher layers capture margins.
Guest view: Evans argues that foundation models lack obvious network effects and may become infrastructure unless they find leverage up the stack.
Where they agree: The current market is too early and too supply-constrained to prove the final value chain.
2. Does AI kill SaaS or create more of it?
Host view: Erik pushes on whether software investors should worry about a SaaS apocalypse as AI makes software easier to build.
Guest view: Evans says some SaaS companies will be damaged, but AI also creates new categories, more tools, and more workflow choices.
Where they agree: The disruption is real, but blanket derating is too blunt because nobody knows which workflows will move.
3. Should companies overspend on AI infrastructure?
Host view: Erik raises the claim that underinvesting may be riskier than overinvesting for large tech platforms.
Guest view: Evans accepts the existential pressure but points to financial gravity: there is a ceiling on sustainable capex, even for trillion-dollar companies.
Where they agree: Strategic fear can justify huge spending for a while, but ROI and physical limits still matter.
4. Can AI automate the job or only the tasks?
Host view: Erik asks about new AI-native interfaces and systems designed for agents rather than humans.
Guest view: Evans says the central boundary is exception handling: AI handles average, describable tasks better than undocumented judgment and novel decisions.
Where they agree: Enterprise adoption depends on discovering where automation belongs inside real workflows.
If you want more podcast briefings like this, search for PodFaro and use it to turn long conversations into structured notes, quotes, and decision-ready summaries.
Top comments (0)