Skip to content
a16za16z

Foundation Models are a Commodity | Benedict Evans on a16z

Erik Torenberg speaks with tech analyst Benedict Evans about the current state of AI, what has changed over the past year, and which questions remain unanswered. The conversation covers coding agents, foundation models, AI infrastructure spending, software economics, and the tension between today's AI excitement and the long-term realities of technology adoption. Evans discusses why coding has emerged as AI's first breakout use case, how previous platform shifts can help frame the current moment, and why many of the most important questions about AI remain unresolved. Along the way, they explore the future of software, enterprise adoption, consumer behavior, and whether AI models ultimately capture value themselves or become infrastructure for the next generation of applications. Timestamps: 00:00 - Intro 01:05 - AI Adoption Accelerates 06:00 - OpenAI Strategy And Usage Gap 09:27 - Platform Shifts And Value Capture 30:43 - Automation And Jevons 33:27 - Ads And Shopping Agents 39:41 - Enterprise Stack Rewired 49:57 - Capex Commodities And Magic Resources: Follow Benedict Evans on X: https://x.com/benedictevans Follow Erik Torenberg on X: https://x.com/eriktorenberg Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends! Find a16z on X: https://twitter.com/a16z Find a16z on LinkedIn: https://www.linkedin.com/company/a16z Listen to the a16z Show on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX Listen to the a16z Show on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711 Follow our host: https://x.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see http://a16z.com/disclosures.

Benedict EvansguestErik Torenberghost
Jun 4, 20261h 2mWatch on YouTube ↗

CHAPTERS

  1. What’s changed since “AI Eats the World”: product strategies diverge and coding hits PMF

    Evans reflects on the last 12–18 months of generative AI and argues the big update is strategic divergence: not just “bigger models,” but clearer product focus. Agentic coding is the standout use case with undeniable product-market fit, while many other consumer and enterprise uses remain tentative or experimental.

  2. Why coding was the first breakout use case—and why job implications are still unknowable

    Coding’s success is partly path-dependent: developers were the earliest heavy users and naturally tried to apply LLMs to software creation. Evans cautions that conclusions about engineering org structure and career ladders are premature because this shift only recently started working well and pricing/capacity constraints distort outcomes.

  3. OpenAI’s “everything all at once” phase vs. refocus—and the consumer usage gap

    Evans describes OpenAI’s attempt to build value above the model via many adjacent product bets (ads, commerce, payments, browser, social). The core challenge persists: outside of coding and a few niches, most users aren’t daily—there’s a big gap between Valley power users and mainstream weekly dabblers.

  4. AI adoption through the lens of past platform shifts: acceleration, friction, and pricing shocks

    Comparing AI to PCs, the internet, and mobile, Evans argues adoption always accelerates because each wave stands on prior infrastructure. Early phases are messy and unreliable, and AI is also experiencing a familiar “pricing shock” moment similar to mobile data’s transition to sustainable pricing models.

  5. Where value gets captured: infrastructure layers vs. platforms vs. apps

    Evans frames a central uncertainty: do model providers capture lasting value like Windows/iOS, or do they become commoditized infrastructure like telcos/ISPs/chip layers? He notes many historic infrastructure builders transformed the world but failed to capture the bulk of profits, which migrated up the stack.

  6. Why prediction is hard right now: multiple paths before the S-curve narrows

    Evans resists deterministic forecasts, arguing history helps generate questions more than answers. At this stage, many outcomes are plausible, and clarity only arrives once adoption steepens and the market “narrows” toward a dominant structure.

  7. “Foundation models are a commodity” argument: weak differentiation, chatbot limits, and abstraction

    Evans lays out a chain of reasoning for commoditization: model differentiation seems hard to sustain, the chatbot UI is a limited V1, and the real productization requires tooling, data, workflows, and interfaces that model labs can’t build for every domain. He likens model choice to cloud choice in SaaS—often abstracted away from customers.

  8. Next questions: on-device/cheaper models, and what AI means for specific industries

    Evans highlights new focal points: when “good enough” models shift workloads off the most expensive frontier systems, and how AI changes pyramid-shaped professional services (law, consulting, banking). He argues the hardest, most important implications become industry-specific, not purely technical.

  9. Automation, Jevons paradox, and the search for what becomes newly possible

    Framing AI as automation, Evans explores elasticity: cheaper work can mean doing the same for less, more for the same, or entirely new categories of activity. The important frontier isn’t “old work, faster,” but new capabilities that were previously impossible or not even imagined.

  10. Ads, shopping, and discovery agents: redefining how products are understood and sold

    Evans argues commerce and advertising are promising because LLMs can interpret what products “are” and why people buy them, not just correlate SKUs. This enables richer shopping agents (visual search, comparisons, personalized style guidance) and helps explain improving ad conversion and recommendation performance.

  11. Enterprise stack rewired: more software, new abstraction layers, and where probabilistic systems fit

    Evans predicts “way more software,” not less, as AI becomes embedded both as a feature within systems of record and as a cross-system synthesis layer. The core design tension is placing probabilistic LLM behavior alongside deterministic databases and workflows—especially around exception handling and judgment calls.

  12. Capex, scarcity, and the ‘magic’ problem: how much investment is sustainable, and what’s the ROI?

    Evans notes current AI capex is approaching “global infrastructure scale,” with big tech spending levels far above traditional capital-intensive industries by revenue percentage. The present moment is extreme disequilibrium—pricing, usage, and ROI measurement are still unstable, and some benefits look like consumer surplus that gets competed away.

  13. Closing synthesis: AI becomes invisible infrastructure—then ‘of course computers can do that’

    Evans ends with a historical reminder: every major computing wave felt magical and uncertain, then became mundane. His baseline expectation is that AI will embed everywhere, create winners/losers, and eventually fade into the background as an assumed capability of modern computing.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.