Mar 19, 2026 · 1h 1m · mad
Benedict Evans: OpenAI’s Moat Problem & the Future of Software
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of The MAD Podcast, technology analyst Benedict Evans joins Matt Turck to examine why foundation AI models face severe commoditization and how AI will dramatically increase total software volume rather than replace it. They discuss OpenAI's strategic challenges, macro CapEx investment cycles, and practical enterprise adoption.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 9.6% of the talking time here. How this is scored →
Matt as informed peer 5.4 Guest teaching 6.8 Guest disagreement 2.9 Matt pushing back 2.4
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 3:10 Direct challenge on cloud comparison
Benedict explicitly interrupts and rejects Matt's comparison between foundation models and cloud oligopolies, stating 'First of all, I pushed back slightly on that comparison.'
Hardest push from Matt ▶ 3:10 Cloud oligopoly counter-argumentMatt challenges Benedict's assertion that foundation models lack moats by bringing up AWS, Azure, and GCP as an example of highly profitable, undifferentiated tech oligopolies.
Biggest teaching moment ▶ 26:02 Dismantling the vibe-coding narrativeBenedict dismantles the host's framing around AI software generation by calling vibe-coded enterprise software a straw man and introducing a formal taxonomy of software systems.
Matt holds his own ▶ 45:22 Exposing the human-labor TAM fallacyMatt demonstrates sharp analytical insight by pinpointing the core flaw in market-sizing models that price AI software at human worker rates, which Benedict enthusiastically validates.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| OpenAI's Strategic Pivot and the 'Side Quest' Problem | 5 | 6 | 2 | 1 | Matt opens by referencing recent WSJ reporting regarding OpenAI's refocus on coding and business users to stop side quests. Benedict responds by reframing the core problem as a lack of network effects or winner-takes-all dynamics in foundation models. | |
| Cloud Oligopoly Analogy vs. Foundation Model Value Capture | 7 | 7 | 5 | 7 | Matt pushes back directly on Benedict's commodity claim by raising the cloud oligopoly analogy of AWS, Azure, and GCP. Benedict explicitly pushes back on the comparison, explaining structural differences between cloud infrastructure and foundation model ecosystems. | |
| The Chatbot Product Problem and OpenAI's Search for a Platform | 3 | 7 | 2 | 1 | Benedict delivers a sustained analysis on the limitations of raw chatbots, citing usage metrics where 80 percent of users rarely prompt them. He draws an analogy to TSMC to explain how value capture works when building on frontier technology. | |
| Why 'Better Models' Don't Solve the Application & UI Problem | 6 | 5 | 2 | 3 | Matt observes that despite massive model advancements in reasoning, better models have not solved the core product and UI challenges. Benedict agrees and elaborates on jagged capabilities and why incremental accuracy gains do not remove human verification requirements. | |
| User Engagement Depth, Memory Moats, and the TSMC Analogy | 6 | 7 | 3 | 2 | Matt asks whether user memory could create defensible moats for chatbot products. Benedict counters with empirical prompt usage stats showing minimal engagement and highlights the inherent difficulty of differentiating a simple input-output text box. | |
| Research-Driven Product Strategy vs. Customer Experience | 6 | 7 | 2 | 2 | Matt brings up Benedict's insight regarding foundation model companies building product inside-out from research discoveries. Benedict contrasts this technology-driven product development with Steve Jobs's philosophy of starting from customer experience. | |
| Prompt Data Advantages vs. Platform Ecosystem Building | 5 | 8 | 4 | 3 | Matt suggests OpenAI's vast prompt log data gives them a unique platform ecosystem advantage. Benedict reframes this, explaining that prompt logs merely reveal self-selected desire paths rather than underlying product opportunities. | |
| Historical Software Abstractions, VisiCalc, and OpenClaw | 4 | 7 | 3 | 2 | Matt asks if Benedict is more bullish on Anthropic or open-source solutions like OpenClaw. Benedict compares current OpenClaw enthusiasm to 1970s Homebrew Computer Club and desktop Linux, noting polished consumer UX remains far off. | |
| Software Taxonomy: Systems of Record vs. Vertical SaaS | 4 | 8 | 5 | 2 | Matt asks if Claude Code will enable individuals to build custom software, and Benedict dismisses the idea of users vibe coding their own complex software like ERPs as a straw man. He then presents a structured taxonomy dividing software into systems of record, vertical SaaS, and improvised tools. | |
| Ephemeral Software and Improvised AI Workflows | 6 | 6 | 2 | 2 | Matt suggests the concept of ephemeral software created on the fly. Benedict agrees and expands on improvised workflows, comparing AI-generated software expansion to price elasticity and the historical impact of spreadsheets on finance headcount. | |
| Disruption Disparities: Software's Selective Impact across Industries | 5 | 7 | 4 | 2 | Benedict critiques simplistic AI exposure benchmarks like GDP VAL, arguing industry disruption depends on specific structural nuances rather than generic scoring. Matt contributes contextual reference points on exposure metrics. | |
| Tech Bubbles, Capital Overinvestment, and Financial Risk | 6 | 7 | 2 | 3 | Matt inquires if the tech market must go through a dot-com style phase of destruction. Benedict analyzes capital overinvestment, vendor financing, and balance sheet leverage while comparing NVIDIA's cash generation to historical bubbles. | |
| Physical Compute Limits, CapEx Gravity, and the Infrastructure Peace Dividend | 6 | 7 | 3 | 2 | Matt asks about the potential economic peace dividend resulting from massive AI compute infrastructure spending. Benedict outlines financial gravity limits, noting hyperscalers spending half their revenue on CapEx cannot maintain that trajectory indefinitely. | |
| Dissecting AI TAM, Labor ROI, and Platform Shift Frameworks | 7 | 7 | 3 | 4 | Matt points out the fundamental fallacy in AI TAM models that assume software can be priced identically to human labor. Benedict strongly agrees, providing historical context on labor automation, price elasticity, and enterprise back-office realities. | |
| Enterprise AI Adoption: Hype vs. Practical Reality | 5 | 6 | 2 | 2 | Matt asks Benedict what he observes during conversations with enterprise executives regarding AI adoption. Benedict reports that enterprises have deployed pilot projects and back-office optimizations, but remain uncertain about transformative new capabilities. | |
| Building in the AI Era: Advice for Tech Founders | 5 | 6 | 2 | 1 | Matt asks what advice Benedict has for startup founders building in the AI era. Benedict explains that core software building principles remain unchanged: identifying real problems and distribution matters vastly more than writing code. |
Statements from this episode (28)
Evans: AI foundation models lack winner-takes-all network effects
“And the problem is that as far as we can see, there is no winner takes all effect or network effect in a foundation model.”
Evans: Microsoft Bing will never catch up with Google in search
“Doesn't matter how much money and how hard Microsoft works, Bing will never catch up with Google.”
Evans: Three to six organizations can create frontier AI models
“So that means that we've got, like, pick a number between three and six or maybe more organizations that can make a frontier model, and they keep leapfrogging each other every couple of weeks or every couple of months.”
Evans: AWS, Azure, and Google Cloud are fundamentally different businesses
“The first is if you actually look at the market shares, Google Cloud, Azure, and AWS are actually in quite different businesses. AWS is mostly infrastructure. Microsoft is mostly services. Google is mostly scrambling to catch up in a very distant certain place…”
Evans: Meta and Google do not need standalone LLM monetization
“Because if you are Meta or Google, You've got this whole other highly profitable business, which now needs to have LLMs inside it, powering all sorts of capabilities and features, and you probably want them to be your LLMs rather than somebody else's. But you …”
Evans: 10% of population uses LLMs daily, 50% monthly
“If you look at the usage data, something like 10% of the population is using these things every day, but another 50% are using it every week or every month.”
Evans: OpenAI has 900M weekly active users, but only 5% pay
“You've got nine hundred million weekly active users, but most of them are not using it every day and can't think of anything to do with it. And only five percent of them are paying for it.”
Evans: Incremental AI accuracy improvements don't reduce necessary human review
“But if you've got a bunch of use cases where you need the right answer, as opposed to sort of the right answer, then saying that the model is better doesn't mean anything. I mean, literally, it is literally meaningless. What you're telling me is, I asked the m…”
Evans: Submitting 1,000 prompts placed ChatGPT users in top 20%
“Turns out that if you did a thousand posts, if you did a thousand prompts last year, you're in the top 20%.”
Evans: Differentiating a chatbot is like differentiating a web browser
“The chatbot itself is kind of like trying to differentiate a web browser in that you've got an input box and an output box, and how can you make them different if the whole point is that you can type in anything and get anything out.”
Evans: AI product teams are strategy takers, not strategy setters
“You start from the technology. You don't control the product strategy, which is of course how science works, but you don't know what's going to happen. You don't know what's going to get built. You know, obviously you've got like Sam and Dario and so on are li…”
Evans: 'Agent' in AI is as fuzzy a term as 'metaverse'
“It's a very kind of, it's a slightly kind of fuzzy term. It's a little bit like saying metaverse. You know, you don't really know what somebody meant when they might've meant VR or VR is real. They might've meant games. Games are real. But when they said metav…”
Evans: AI app design is currently at the 'PDF catalog' stage
“And we're still at the stage of, you know, taking a PDF of your catalog and putting it in your company website as we try and work out what we should do with AI.”
Evans: AI is a continuation of 50 years of software abstraction
“Tell that next time you hear a software developer saying, like, AI is a completely different thing, and nobody has ever abstracted software like this before, like, yeah, we've been doing this for 30 years, 50 years.”
Evans: Anthropic momentum is fleeting as AI model leadership shifts weekly
“Not really. I mean, this week they've got all the fire, they've got all the juice, whatever the word is. I don't know. This week, next week, it'll be something else.”
Evans: Typical big US company uses 400 to 500 vertical SaaS apps
“And so this is why the typical big US company today has, depending on your numbers, like four to 500 vertical SaaS apps.”
Evans: AI coding will lead to far more total software volume
“There will be way more software, and that will pick up many more of those use cases, either that weren't automated before, either because they were too small, or because you could actually couldn't automate that thing with software before, and now with AI you …”
Evans: No one will vibe code their own ERP software
“No, no one will vibe code their own ERP or their own frame.io, but they may ask Anthropic or Gemini or ChatGPT, can you do this thing for me?”
Evans: Spreadsheets increased finance employment rather than causing job losses
“Spreadsheets did not result in a collapse in the number of people working in finance. They're quite the opposite. You have way more people in finance, because now it's possible to do all this more, all this new stuff that you couldn't have done before.”
Evans: Precise numeric scores for job AI exposure are ludicrous
“You want to take all the numbers off because it could tell yourself that that one is 96.5 and that one is 78. It's just ludicrous.”
Evans: Almost everyone in a tech bubble acts rationally
“Well, you know, everyone in, in, generally in a bubble, everybody's a rational actor. Almost everyone's a rational actor given their situation.”
Evans: OpenAI has commodity tech and no differentiation but massive mindshare
“If you're Sam Altman, you've got a commodity technology. You've got, you're competing with people who have giant legacy cash flows. You don't have your own infrastructure. Don't really have any differentiation, but you've got massive mindshare.”
Evans: Nvidia hit over $70B in trailing 12-month free cash flow
“I mean, I haven't looked at, I haven't updated my number here, but like, I think Q three last year, I think NVIDIA had something over seventy billion dollars of trading 12 months free cash flow.”
Evans: AI differs from past tech shifts because its physical limits are unknown
“The one way this is unquestionably different is that with all the other platform shifts, we knew what the physical limits of the science were.”
Evans: Corporate audit costs stayed flat since early 2000s despite software progress
“Average audit data, all sorts of data for audit costs since the Barnes Oakley, which has basically been flat since about since the early 2000, despite everything that's happened in software”
Evans: Initial enterprise deployments of Microsoft Copilot were largely unsuccessful
“Everyone deployed co-pilot and went, oh, okay, that wasn't very successful.”
Evans: Most SaaS companies are fundamentally just database wrappers
“Most SaaS companies are database wrappers, where somebody realized that here is this problem, and here is the people who have it, and here is a way of turning it 90 degrees, and this is your insertion point, this is how you build it and take it to market.”
Evans: The hardest part of building software isn't writing the code
“The hard part of writing software is not writing code. It's all the other stuff around, like, what should the code be doing? And how would we tell people that they should be using it? And what should we charge? And how do we go to market? And which bit of the …”