Evans / Lenny’s Podcast — A rational conversation on where AI is actually going (2026-05-31)
Benedict Evans is an independent analyst and former partner at Andreessen Horowitz, where he spent years as their in-house “thinker” tracking the most important technology trends. For the past six years, he’s been publishing deeply researched presentations on where tech is heading, most recently focused on AI’s transformation of the economy. His work is read by founders, investors, and operators trying to make sense of a noisy field. His most controversial opinion: AI is as big a deal as the internet or mobile—and only as big.
(Channel description, Lenny’s Podcast, 31 May 2026.)
A ~1:19:50 long-form interview published by Lenny’s Podcast (host Lenny Rachitsky) on 31 May 2026. Guest Benedict Evans — independent analyst, ex-a16z partner, former equity researcher — discussing his most recent biannual AI Eats the World deck (~80 slides, published the day before the interview). The wiki’s first Benedict-Evans-authored source and the fifth source under the Lenny's Podcast channel, joining Caldwell (April 2024), Schoening (May 2026), Ries (May 2026), and Spiegel (April 2026).
This is the wiki’s first independent-analyst-altitude long-form anchor on the what-people-aren’t-pricing-in question, sitting between the practitioner-altitude cluster ( YC, HelloPrint), the consulting-firm strategic-altitude cluster ( McKinsey, McKinsey), and the academic-economist altitude (Brynjolfsson, Dell’Acqua). Evans’s distinctive rhetorical move is deflation: AI is as big as the internet or mobile — and only as big.
TL;DR
Twelve substantive contributions.
1. The 1997 for AI framing. Evans’s most controversial opinion: ‘AI is as big a deal as the internet or mobile and only as big a deal as the internet or mobile.’ Pushing against industrial-revolution analogies on one side and just-another-tech-cycle dismissal on the other. The framing’s load-bearing implication is timing: ‘if you’re going to make the internet comparison it’s like we’re in 1997. Most stuff kind of doesn’t work yet. Most of the stuff that people are going to do hasn’t been built yet and it’s not really clear how any of it’s going to work when it does work.’ Evans’s 80-slide deck, in his own facetious paraphrase, ‘is saying we don’t know.’
2. The task vs job distinction as the operational lever. ‘What is the hard part of the job? Is the hard part of the job writing the code line by line? Is the hard part of the job making the PowerPoint? Or is the hard part of the job something else? Is it the task or the job?’ The McKinsey-deck worked example: a Claude-Code-generated McKinsey deck is ‘a bunch of dog crap. That’s not what you’d get from McKinsey. But even if it was, that’s not what you paid them for. What you actually pay Bain to do is to go and walk all over your enterprise, your company, and work out, yes, but why is it that you didn’t do that? And how do the politics of this work? And what do you actually need to do?’ Same argument applied to Amazon (gets you the SKU, but knowing what SKU you want is a separate job) and software (Claude writes the code, but what code do you want? what features? who’s the customer?).
3. The expert-systems analogy as the critique of O*NET-style task-decomposition. ‘You can’t look at a senior partner at a law firm and say, well 17% of their work could be automated. This is horseshit … ironically this is the logical systems problem. The expert systems problem … you start building up logical steps. So you make an edge detector and then you make a third detector and you make an eye detector and you make an ear detector and 15 years later you’ve got 700 steps and it doesn’t work. This is what happens when you try and look at a profession and sort of break it down by which bits can be automated and which can’t.’ This is the wiki’s first independent-analyst-altitude direct attack on the methodology underlying Brynjolfsson’s task-level AI-exposure analysis and the Dell’Acqua-style task-level frontier studies — Evans accepts the jagged-frontier concept while rejecting the percentage-of-job arithmetic.
4. The Excel made bankers work longer, not shorter / Jevons-paradox argument. ‘Young people won’t believe this but before Excel, junior investment bankers worked really long hours and now thanks to Excel, Goldman’s associates all work at lunchtime on Fridays. Like — well, why is that not what happened? You could make the same point in software development. Before IDEs and libraries and operating systems, developers had to write all the code. Now if you write an iPhone app, 90% of the code is written for you by Apple. So we’ve got like a tenth as many engineers now? Well, no.’ The empirical anchor: the US accountant-headcount chart rising continuously through 20th-century adding-machines → punch-cards → mainframes → ERP → cloud → spreadsheets, and continuing to rise into the 21st century. Evans’s own corollary on AI labs themselves: ‘Anthropic, OpenAI … everyone’s just increasing headcount. Like the companies you would think would be least likely to add humans are adding many, many humans.’
5. Forward-deployed-engineer ≈ Accenture-outsourced-developer-in-SF — and why AI labs invest in consulting partnerships. ‘A forward deployed engineer is like an Accenture outsourced software developer who lives in San Francisco.’ The mechanism: companies don’t have spare consultants on staff to do the workflow-redesign work the AI rollout requires. ‘You’re supposed to reimagine all of the internal workflows of your company and work out which of them could be automated really quickly with AI. That’s a project. That’s a project that needs like five or 10 people to sit down and spend a month or two working it out. And then actually doing it is another project. Okay. So we need to plug these three vertical systems into these two horizontal systems and build in a bunch of new workflows and train people to do that. Well, guess what? Who’s going to do that? Because you don’t have a bunch of people sitting around not doing anything.’ The funny part Evans flags: ‘the most cutting-edge AI labs are the ones most investing in these folks.’ Convergent with McKinsey on agentic-era software delivery and Sternfels on consulting reinvention — Evans frames the same workflow-integration layer from outside the client-services firm.
6. The foundation models = commodity utilities, not platforms thesis. ‘The models don’t seem to have network effects. So there doesn’t seem to be a winner-takes-all effect where one of these will run away ahead of the others. So you should have competition indefinitely. You have competition indefinitely. You don’t have really radical differentiation in what the product is. Then why would you have pricing power?’ The analogy structure: foundation models look more like AWS than Windows — every developer doesn’t have to standardise on one because the customers don’t care which one runs underneath. Sam Altman’s ‘AI on a meter like water or electricity’ line gets the response ‘my dear sweet child, you need me to explain the margin structure of the utility industry to you … when you watch television the TV company isn’t paying a percentage of your monthly bill to the electricity company.’ The telecom worked example: global mobile-data consumption ~1,500–2,000× higher than 2010; telco stocks have ‘gone nowhere in 25 years because it’s an X-growth low-margin commodity utility.’ If foundation models look like that, the value accrues up-stack to apps.
7. Distribution becomes the moat once product is commodity. Same line of argument extended: when models are commodity, distribution wins. ‘You could see that in OpenAI’s strategy late last year — people called it everything yesterday — they were just kind of trying everything to work out how they would get distribution before Google and Meta and Amazon spray it everywhere and get everybody using that one.’ Survey-data anchor: pre-llama-news, meta-AI was tracking ‘up there between ChatGPT and Gemini’ in user surveys ‘which, if you’re in tech, people have completely written off — but it was like they’d sprayed it on every surface and it wasn’t that bad. It was fine.’ The Apple Intelligence WWDC-2024 vision gets explicit credit: ‘the most compelling vision of a personal AI assistant I’ve still seen. They then couldn’t ship it, but then neither was anybody else.’ Same thesis as [[2026-04-26-how-to-win-when-software-is-not-a-moat-evan-spiegel-snapchat-ceo|Spiegel’s software-is-not-a-moat]] at a different altitude.
8. AGI is a sliding goalpost — the Larry Tesler quote as the schema. ‘A quote I used in my presentation late last year was an AI scientist called Larry Tesler who said AI is whatever machines can’t do yet. Because once machines can do it people say, well, that’s just software.’ Evans applies the schema to current AGI definitions: ‘now clearly you can see people redefining AGI to mean the stuff that works now … it can do a certain percentage of economically valuable work. Well, that’s a very different thing to it has a soul and it’s alive — because a database can do that. An IBM mainframe in 1975 could do a meaningful percentage of economically valuable work that was previously done by people.’ Same argument for super-intelligence: ‘last year I thought super-intelligence was like really good but not as good as AGI and now it’s like — oh no no we’ve already got AGI but super-intelligence that’s really hard.’
9. The anti-AI backlash is a big fuzzy mess analysis. Evans separates electricity-bill concerns (locally real but contained to specific siting decisions) from the water-consumption claim (‘completely fake’). The specific number Evans cites: a Livermore-Lab end-2024 study estimated US data-centre water consumption at ‘about 0.017% of US water consumption.’ (‘Now obviously if you live in a small town and you’ve got one well and they capped the well and gave all the water to the data center then you’re really pissed off — but that’s a planning problem.‘) Energy comparison: ‘data centers are what like 5% of US energy and might grow at 1% a year for the next five years.’ On the social-media-backlash analogy: ‘some of the backlash around social was true and some of it was sort of true and some of it wasn’t.’ The Jonathan Swift quote Evans returns to: ‘you can’t reason somebody out of an idea they won’t reasonably into.’
10. The 5–10-year sector-by-sector replacement speed model. ‘Typical big-company enterprise software sales cycle is 18 months if you’re lucky. Enterprise sales cycle is shorter than the venture-backed startup software funding cycle — longer — like it takes you longer to get an enterprise deal than it takes you to go between rounds. So no, people aren’t just going to tear out SAP and replace it with XYZ. Maybe in three, five, ten years yes that whole estate will look radically different and all those jobs will have changed, but it will take three, four, five, ten years and it will take time sector by sector and it will take time for people to work out — oh, you could do that thing with this.’ The Frame.io reference: ‘there’s nothing there that you couldn’t have done at least five years earlier and maybe ten years earlier … the delay was somebody realizing, oh we could — that problem exists inside that industry and oh this is the way that we would solve it. It didn’t all happen the day after Google Docs.’
11. The mobile shift did not change half of the internet industry analogy applied to AI value capture. ‘The funny thing about mobile is that some companies missed it completely. And for some of them, it really didn’t change anything. Like for Google, it didn’t change anything. For Meta, this was great. Amazon, like what does this change? Yahoo Mail fails to make the jump. eBay, you could argue about individual names. The point is that we went through that shift and it didn’t change anything for half the internet industry.’ The implication Evans is gesturing at: equivalently many AI-era ‘incumbents stay roughly where they are’ outcomes are plausible against the ‘completely transformative’ default reading. Quoting Steven Sinofsky (who used to run Windows at Microsoft): ‘incumbents always try and make the new thing a feature, and sometimes they’re right, sometimes it’s a feature.’
12. The just dive in, don’t shout from Bluesky career-advice posture. Evans’s prescription for people worried about their jobs: ‘Don’t stick your head in the sand and say I hate all of this stuff because that gives you a great feeling of moral superiority and you can go on Bluesky and shout at everybody about how evil AI is. Like — great, I’m happy for you. But that’s not going to help. What helps is you diving into this completely submerging yourself in it and coming out understanding what you can do with it, how this changes things, how you can be a great hire.’ The associate-going-into-law-firm-interview scenario: ‘going to the interview and saying, well I think AI is — and I’m never going to use it — is probably not the right mood.’ Convergent in posture with Koomen’s just-build-and-ship practitioner stance.
Evans’s own self-described AI usage
A useful negative data point. ‘I’m sort of the lawyer looking at ChatGPT. The stuff that I would do that I would automate are sort of precise information retrieval tasks, which is precisely the thing that this is kind of worst at … I use it for proof-reading. I use it for images. I used it redecorating my apartment — that worked fantastically well at that. Here’s a picture of this room. Repaint it. Add this light and this table and this rug.’ The aphorism he attributes to ‘somebody a couple of years ago’: ‘AI is good at stuff that computers are bad at and bad at stuff that computers are good at.’ The relevance: an independent-analyst-altitude observer whose job is ‘synthesise a whole bunch of other stuff into a whole bunch of new ideas’ finds the synthesis-across-many-sources use-case still unreliable in May 2026 — sits in tension with practitioner-altitude reports of AI-assisted research as transformative.
The under-asked question Evans poses
In response to ‘what’s a question about AI that you think not enough people are asking?’ Evans names two. First: ‘I’m not sure how many people are asking whether model labs have pricing power. I think a lot of people are just presuming that the situation today will continue.’ Second, the larger one: ‘What’s the task and what’s the job? What is just the thing that becomes a button versus what are people actually hiring you for?’ — closing with the global-recorded-music revenue chart (U-shaped: dropped by half 2000–2015, recovered to ~75% of peak by recent) as illustration that ‘you only even know what the question is after it’s been asked and you built a billion-dollar thing that lots of people use.‘
Linked entities and concepts
Already-promoted entities referenced in the conversation: Lenny’s Podcast (channel), OpenAI, Anthropic, Google, Amazon, McKinsey & Company, Bain & Company (as comparison), Y Combinator (Evans mentions a16z; Y Combinator appears via Lenny’s reference to Dan Shipper). Andrew Ng is invoked indirectly via the harness terminology.
Concept pages this source informs (Process step 6 targets): jagged-frontier (Evans uses the term explicitly with nuance — accepts the concept, rejects the percentage decomposition), ai-employment-effects (lump-of-labor argument + 5–10-year sales-cycle speed model + AI-labs-still-hiring observation), automation-vs-augmentation (task-vs-job framing), foundation-models (commodity-utility thesis), agent-harness (namechecked), enterprise-ai-adoption (5–10-year sector-by-sector replacement), dynamic-capabilities (the entire deck is a strategic-foresight exercise; multiple W&W cells touched).
Dangling (single-source mention, deferred per the second-source promotion rule): Benedict Evans (this is his first appearance in the wiki — the entire content of this source is his thinking, so the strict reading of the rule would defer entity-page creation to a second source; the spirit of the rule is preserved by the substantial body coverage on this page); Lenny Rachitsky (host; named in 10+ Lenny-channel sources but never as author: since the channel convention is author = channel); Larry Tesler (the AI is whatever machines can’t do yet quote attribution); Mark Andreessen (referenced repeatedly; not yet an entity page); Dario Amodei (referenced; appears as author of Amodei et al. May 2026 — likely candidate for promotion in a focused entity-creation pass); Steven Sinofsky (one quote); Eric Schmidt (commencement-speech aside); Dan Shipper (referenced as a recent Lenny guest); Frame.io (Evans’s worked example for the delay was somebody realising you could do that thing with this); House of a16z / Andreessen Horowitz (Evans’s former employer); Llama / Meta-AI (briefly).
W&W cells touched (per Warner & Wäger process model). The deck-and-interview as a whole is a strategic-sensing + strategic-foresight exercise; specific cells the conversation surfaces:
digital-sensing/digital-scouting— the biannual AI Eats the World deck is itself a digital-scouting practice at independent-analyst altitude, and Evans’s ‘I read this whole thing called O*NET’ anecdote is digital-scouting under critique.digital-sensing/digital-scenario-planning— the 1997 for AI framing is scenario-style situating; the task-vs-job lens is a scenario-decomposition tool.digital-seizing/balancing-digital-portfolios— Evans’s central where will value accrue: models vs apps question is portfolio-balancing at the industry-investment altitude.digital-seizing/strategic-agility— the 5–10-year sector-by-sector speed model is itself a pacing-of-strategic-response argument.digital-transforming/redesigning-internal-structures— the workflow-redesign needs five-to-ten consultants for one-to-two months observation is structural-redesign at the firm level.strategic-renewal/business-model— the foundation-model commodity-utility claim is a business-model-renewal claim about the layer at which value capture happens.contextual/external-triggers— the anti-AI backlash analysis is a contextual-trigger analysis (and Evans’s ‘big fuzzy mess’ characterisation is the analytical move).