a16z Interview With Benedict Evans On AI Models, SaaS, And Value Moving Up The Stack
They built this amazing piece of global infrastructure... and they didn't make any money from it because all the value moved up stack.
Watch the recap video here
Source Moment
In the selected clip, Benedict Evans uses mobile networks as the analogy for large language models. Mobile data became essential infrastructure, but the strongest profit pools formed in apps and platforms above the network; Evans asks whether foundation models face the same pressure as companies build AI into products, workflows, and industry software.
Context
a16z, a venture firm and media publisher, published this June 8, 2026 YouTube interview with technology analyst Benedict Evans about coding agents, AI model pricing, infrastructure spending, and SaaS; the important question is whether AI profits stay with model providers or move up into apps, distribution, and company workflows.
Big Ideas
- Model value may move above the model layer: Evans compares AI models with mobile networks. The network became huge and necessary, but the best profits moved to apps and platforms above it. His caveat is that the analogy is a frame, not a forecast; the open question is whether foundation models gain pricing power or become infrastructure sold closer to marginal cost.
- AI can mean more enterprise software: Evans's SaaS answer is expansion and recombination. Cheaper software creation can create more tools, more competition, and more ways to move work between big systems, vertical apps, spreadsheets, email, and internal tools. The hard part is understanding exceptions and tacit company process, not generating code alone.
- After coding, AI becomes an industry workflow question: Coding has the clearest current product pull. Evans says the next questions move into law, finance, consulting, advertising, retail, media, and other domains where outsiders may not know what work gets done, what customers pay for, or how margins change.
Full Recap
Benedict Evans is looking at AI after coding agents became the first clear daily-use market. Coding shows that LLMs can move from demos into real work, but it does not settle the broader economics. The bigger question is where value lands as model companies, chip suppliers, software companies, and enterprise workflows adjust.
Evans's strongest analogy is mobile data. Mobile networks became essential global infrastructure, yet much of the attractive profit moved to companies building apps, platforms, and services above the network. He applies that question to foundation models: do model companies become the powerful product layer, or do they become expensive infrastructure that enables value elsewhere?
For SaaS, Evans expects more software and more workflow recombination. Companies already run on a messy stack of big systems, vertical apps, spreadsheets, internal tools, email, and unwritten process knowledge. AI can sit inside existing apps, above several systems, or inside new internal tools. The value depends on who understands the workflow and who controls distribution.
- 10:31-22:48:Pricing Crunch, Platform History, And Model Commoditization: Evans connects AI token pricing with earlier infrastructure markets. Current prices are shaped by scarce capacity, high demand, and unclear value, but the longer-run question is whether models have durable pricing power. The mobile-network analogy supports his central claim: essential infrastructure can be huge while profit moves up to products and workflows built on top.
- 38:18-48:43:Enterprise Software Gets Rebuilt Around Workflows: Evans says AI should make software cheaper and faster to build, which points toward more software rather than less. Enterprise work already moves between SAP-like systems, vertical SaaS, internal tools, spreadsheets, email, and shared files. AI can change those boundaries, but workflow knowledge and exception handling decide where the useful product sits.
- 22:48-38:18:What Comes After Coding: Evans moves from developer tools to industries. In law, finance, consulting, advertising, retail, and media, the AI question depends on what work is being automated, what customers pay for, how teams are structured, and which tasks become cheaper or newly possible. People outside those industries may know the technology but miss the workflow.
- 00:44-05:53:Coding Becomes The First Clear AI Use Case: Evans says agentic coding has become the clearest AI product-market fit because software developers are pulling it into real work. That makes coding the first strong proof point for AI usage, while other daily-use markets remain less settled.
- 48:43-55:07:The AI Capex Ceiling: Evans treats AI infrastructure spending as enormous but finite. Chips, data centers, power, model efficiency, demand, open-source models, and edge compute all shape the ceiling. Large capex can support a major infrastructure market without proving infinite spending or permanent model-company pricing power.
- 55:07-60:32:Models May Become Commodities, But AI Still Becomes Magic: Evans does not claim certainty that models become commodities. He says the burden is on model companies to explain why similar models running on similar chips keep pricing power. His closing frame is that AI may become normal computing infrastructure that feels obvious later.
- 05:53-10:31:OpenAI, Anthropic, And Product Strategy: Evans contrasts OpenAI's broad product experimentation with Anthropic's tighter coding focus. The section shows model labs searching for durable product surfaces above the raw model, with coding as the clearest current example.
Technical Need To Knows
- Agentic coding: AI-assisted programming where the model can plan, edit files, run tools, and iterate toward a software task. Evans treats this as the first AI use case with clear product pull.
- Foundation model: A large general-purpose AI model that can support many downstream products. The feature's value-capture question is whether these models become durable products or commodity infrastructure.
- LLM: A large language model, the AI system behind products such as ChatGPT and Claude. LLMs are the base technology being pushed into coding, enterprise software, search, customer support, and other workflows.
- Token pricing: AI services often charge based on tokens, which are chunks of text or data processed by a model. Evans says current prices are still distorted by scarce capacity, high demand, and uncertain value.
- Capex: Capital expenditure on long-lived infrastructure such as chips, data centers, and power. AI capex can be enormous while still facing financial and physical limits.
- SaaS: Software as a service, or software delivered online through subscription or cloud access. Evans frames AI's SaaS effect as more workflow recombination, more tools, and changed pricing pressure.
- Marginal cost: The extra cost of serving one more unit of usage. Evans uses it to ask whether model competition eventually pushes prices toward infrastructure-like economics.
- Network effect: A product becomes more valuable as more people use it. Evans suggests foundation models may lack the network effects that helped some platforms capture value.