Feb 28, 2026 · 26m · another-podcast
The end of the network effect
0:00 / 0:00
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Benedict Evans and Toni Cowan-Brown analyze the competitive dynamics of generative AI, examining why foundation models lack traditional network effects and how incumbent tech giants hold an advantage through superior distribution and platform lock-in.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 99.9% of the talking time here. How this is scored →
The hosts as informed peer 8.2 Guest teaching 1.3 Guest disagreement 1.3 The hosts pushing back 2.2
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 7:54 Toni challenges Evans on vertical integration substitutability
Cowan-Brown presses Evans on whether vertical integration replaces absent network effects in AI, prompting Evans to firmly reject the premise.
Hardest push from the hosts ▶ 8:27 Benedict pushes back against capex flywheelsEvans firmly dismisses marketing claims of AI flywheels, asserting that spending more capital on infrastructure is not a self-reinforcing virtuous loop like Amazon's.
Biggest teaching moment ▶ 13:30 Evans educates on research lab product management workflowsEvans cites internal product handoffs from OpenAI and Anthropic to explain how AI product roadmap setting is inverted compared to traditional software development.
The host holds their own ▶ 24:30 Evans analyzes Apple and Google operational moatsEvans demonstrates tech industry depth by showing how Apple and Google's dominance relied on unique structural operating models rather than pure operational cadence.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Historical Network Effects Versus Generative AI Scale | 8 | 1 | 1 | 2 | Evans lays out deep historical context on computing network effects from Windows/Intel to Google Search and contrasts it with the lack of inherent network effects in LLM foundation models. Cowan-Brown acts as an agreeable co-host prompting Evans's essay framework. | |
| The Abstraction Stack: Infrastructure Commodity Versus Consumer Platform | 9 | 1 | 1 | 2 | Evans dissects the abstraction stack, comparing AI foundation models to TSMC and cloud infrastructure rather than consumer platforms like Windows or iOS. He points out that infrastructure providers lack consumer-facing leverage unless higher-level lock-ins emerge. | |
| Vertical Integration, Moats, and Non-Virtuous Flywheels | 8 | 2 | 2 | 3 | Evans directly refutes the idea that vertical integration equates to a network effect, criticizing AI companies' pseudo-flywheels that merely turn capital into compute. Cowan-Brown interjects with queries about user value creation. | |
| Product Management Dilemmas and Strategy-Taking Labs | 8 | 1 | 1 | 2 | Evans cites product leaders like Fidji Simo and contrast them with Steve Jobs's product philosophy to demonstrate that AI labs are strategy-takers waiting on unexpected research breakthroughs. He emphasizes how undifferentiated models create a commodity dynamic similar to early web browsers. | |
| Branding, Marketing Strategies, and Market Share Disparities | 8 | 2 | 2 | 3 | The conversation shifts to consumer branding and distribution, contrasting Claude's niche tech cachet and Super Bowl ads with ChatGPT and Meta AI's massive distribution advantages. Evans points out the instability of commoditized technology paired with skewed adoption. | |
| Searching for Sustainable Advantage and Episode Conclusion | 8 | 1 | 1 | 1 | Evans concludes that sheer execution or hiring talented people does not substitute for structural defensibility or proprietary lock-ins. Both hosts wrap up the episode discussing the ongoing search for sustainable competitive advantages in AI. |
Statements from this episode (12)
Evans: There are no network effects in building AI models yet
“There are no network effects in building the models yet. There may be in the future, but at the moment we don't know.”
Evans: Big four platforms spent $400B CapEx last year, targeting $650B+
“Like the big four platform companies spent Four hundred billion dollars on CapEx last year, and they've announced something in the region of six hundred and fifty billion dollars this year, maybe more.”
Evans: AI model infrastructure likely settles into 3 to 6 company oligopoly
“So it seems quite likely that all of that will settle out at some point with some kind of oligopoly, where like the laws of gravity and the laws of financial gravity kick in, and while this is the amount of money that the revenue that you can get, and this is …”
Evans: Controlling lower stack infrastructure does not grant control over applications
“The whole point of a stack is that it's abstracted. Like Cisco didn't get much say in what the websites were, even though it might all have been running on Cisco readers.”
Evans: OpenAI and Anthropic lack distribution and cannot out-innovate startup ecosystem
“You don't have the existing feature set that Google and Meta and Microsoft and Apple and Amazon and everybody else have where they can add it and Salesforce and, you know, Workday and, you know, Figma and everyone else, some of them, some of which will survive…”
Evans: User data does not currently make AI models better
“At the moment, they don't. At the moment, the companies all say, well, indeed, we don't train, we don't use your data to make the models better, partly because the amount of user data involved isn't enough to make the, relative to the broader scale of the trai…”
Evans: AI product heads cannot set roadmaps because research breakthroughs dictate capabilities
“What that's saying is the head of product isn't setting the product strategy and has no idea what the product is going to be in a month's time. Because there's going to be stuff coming out of the research lab that they don't know about, that the research lab d…”
Evans: No AI lab has unique models because everyone builds identical tech
“What's happened so far is that people leapfrog each other every couple of weeks or every month or two, but because everyone is basically building the same stuff nobody has anything unique in the models.”
Evans: Survey data shows Claude holds between 0% and 1% consumer usage
“If you look at the, you know, survey data on usage, it's,
It's ChatGPT, and then it's sort of two-thirds of that to a half of that is Gemini and Meta AI.
Even though Meta AI is a fiasco, Llama IV was a fiasco, the new models haven't launched yet, but consume…”
Evans: Undifferentiated tech with divergent market shares creates market instability
“When the technology is basically undifferentiated, the product is undifferentiated, and you've got radically different market shares, that tends to be an unstable situation.”
Evans: OpenAI lacks a technology lead, lock-in, or network effects
“If you're open AI, you've got, don't fundamentally have a technology lead. You're first among sequels at best. You've got this giant user base. That's not, that has very shallow usage and engagement. And it isn't really locked. It isn't locked in. It doesn't h…”