StarZero All shows Try the podcast clipper →

Dec 20, 2025 · 20m · another-podcast

How does OpenAI compete?

Benedict Evans · 15m spoken Toni Cowan-Brown · 2m spoken
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 dilemmas facing OpenAI and the broader AI ecosystem, examining whether defensible moats can be built across infrastructure, applications, or interfaces amid rapid model commoditization. Drawing historical parallels to the early web and mobile transitions, they explore the persistent gap between frontier benchmark performance and real-world adoption.

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 86.6% of the talking time here. How this is scored →

The hosts as informed peer 7.8 Guest teaching 1.3 Guest disagreement 1.3 The hosts pushing back 1.5
05100:0010:0020:000:00–3:44 · The hosts as informed peer 8/10 Model Parity and the Lack of AI Moats Benedict Evans opens by framing the structural dynamics of LLM competition, breaking down the lack of network effects, chip and data center economics, and model parity across vendors. Toni chimes in agreeably, noting the low-stakes usage patterns of consumers.3:45–7:10 · The hosts as informed peer 8/10 Up-the-Stack Strategies: Features, Applications, and APIs Evans systematically deconstructs up-the-stack strategies, comparing OpenAI's platform ambitions to historical models like Microsoft Windows proprietary APIs versus cloud infrastructure lock-in like AWS. Toni concurs that users neither know nor care about underlying infrastructure.7:10–9:15 · The hosts as informed peer 7/10 The Product Constraints of the Chatbot Interface Evans draws analogies to browser UI constraints, arguing that an open input/output chat box leaves very little surface area for defensible UI differentiation. The exchange is highly analytical and collaborative.9:15–13:55 · The hosts as informed peer 8/10 The Disconnect Between Frontier Benchmarks and User Needs Evans highlights reports of OpenAI engineers perplexed that users do not demand PhD-level reasoning benchmarks, while Toni remarks that users mainly want simple assistance like dating advice. Evans shares an anecdote about a Silicon Valley engineer being surprised by deterministic workflow needs.13:55–17:33 · The hosts as informed peer 8/10 Market Adoption Dynamics Across Competing AI Ecosystems Toni prompts Evans to consider who AI is being built for, leading Evans to detail consumer adoption data showing Meta AI and Gemini surging while Claude remains niche despite tech enthusiast sentiment.17:34–20:30 · The hosts as informed peer 8/10 Historical Parallels and Year-End Reflections Evans contextualizes the current AI era with deep historical parallels to the internet in 1997 and mobile in 2008, noting early assumptions about portals and TV information superhighways were completely wrong.0:00–3:44 · Guest teaching 1/10 Model Parity and the Lack of AI Moats Benedict Evans opens by framing the structural dynamics of LLM competition, breaking down the lack of network effects, chip and data center economics, and model parity across vendors. Toni chimes in agreeably, noting the low-stakes usage patterns of consumers.3:45–7:10 · Guest teaching 1/10 Up-the-Stack Strategies: Features, Applications, and APIs Evans systematically deconstructs up-the-stack strategies, comparing OpenAI's platform ambitions to historical models like Microsoft Windows proprietary APIs versus cloud infrastructure lock-in like AWS. Toni concurs that users neither know nor care about underlying infrastructure.7:10–9:15 · Guest teaching 1/10 The Product Constraints of the Chatbot Interface Evans draws analogies to browser UI constraints, arguing that an open input/output chat box leaves very little surface area for defensible UI differentiation. The exchange is highly analytical and collaborative.9:15–13:55 · Guest teaching 2/10 The Disconnect Between Frontier Benchmarks and User Needs Evans highlights reports of OpenAI engineers perplexed that users do not demand PhD-level reasoning benchmarks, while Toni remarks that users mainly want simple assistance like dating advice. Evans shares an anecdote about a Silicon Valley engineer being surprised by deterministic workflow needs.13:55–17:33 · Guest teaching 2/10 Market Adoption Dynamics Across Competing AI Ecosystems Toni prompts Evans to consider who AI is being built for, leading Evans to detail consumer adoption data showing Meta AI and Gemini surging while Claude remains niche despite tech enthusiast sentiment.17:34–20:30 · Guest teaching 1/10 Historical Parallels and Year-End Reflections Evans contextualizes the current AI era with deep historical parallels to the internet in 1997 and mobile in 2008, noting early assumptions about portals and TV information superhighways were completely wrong.0:00–3:44 · Guest disagreement 1/10 Model Parity and the Lack of AI Moats Benedict Evans opens by framing the structural dynamics of LLM competition, breaking down the lack of network effects, chip and data center economics, and model parity across vendors. Toni chimes in agreeably, noting the low-stakes usage patterns of consumers.3:45–7:10 · Guest disagreement 1/10 Up-the-Stack Strategies: Features, Applications, and APIs Evans systematically deconstructs up-the-stack strategies, comparing OpenAI's platform ambitions to historical models like Microsoft Windows proprietary APIs versus cloud infrastructure lock-in like AWS. Toni concurs that users neither know nor care about underlying infrastructure.7:10–9:15 · Guest disagreement 1/10 The Product Constraints of the Chatbot Interface Evans draws analogies to browser UI constraints, arguing that an open input/output chat box leaves very little surface area for defensible UI differentiation. The exchange is highly analytical and collaborative.9:15–13:55 · Guest disagreement 2/10 The Disconnect Between Frontier Benchmarks and User Needs Evans highlights reports of OpenAI engineers perplexed that users do not demand PhD-level reasoning benchmarks, while Toni remarks that users mainly want simple assistance like dating advice. Evans shares an anecdote about a Silicon Valley engineer being surprised by deterministic workflow needs.13:55–17:33 · Guest disagreement 2/10 Market Adoption Dynamics Across Competing AI Ecosystems Toni prompts Evans to consider who AI is being built for, leading Evans to detail consumer adoption data showing Meta AI and Gemini surging while Claude remains niche despite tech enthusiast sentiment.17:34–20:30 · Guest disagreement 1/10 Historical Parallels and Year-End Reflections Evans contextualizes the current AI era with deep historical parallels to the internet in 1997 and mobile in 2008, noting early assumptions about portals and TV information superhighways were completely wrong.0:00–3:44 · The hosts pushing back 2/10 Model Parity and the Lack of AI Moats Benedict Evans opens by framing the structural dynamics of LLM competition, breaking down the lack of network effects, chip and data center economics, and model parity across vendors. Toni chimes in agreeably, noting the low-stakes usage patterns of consumers.3:45–7:10 · The hosts pushing back 1/10 Up-the-Stack Strategies: Features, Applications, and APIs Evans systematically deconstructs up-the-stack strategies, comparing OpenAI's platform ambitions to historical models like Microsoft Windows proprietary APIs versus cloud infrastructure lock-in like AWS. Toni concurs that users neither know nor care about underlying infrastructure.7:10–9:15 · The hosts pushing back 1/10 The Product Constraints of the Chatbot Interface Evans draws analogies to browser UI constraints, arguing that an open input/output chat box leaves very little surface area for defensible UI differentiation. The exchange is highly analytical and collaborative.9:15–13:55 · The hosts pushing back 2/10 The Disconnect Between Frontier Benchmarks and User Needs Evans highlights reports of OpenAI engineers perplexed that users do not demand PhD-level reasoning benchmarks, while Toni remarks that users mainly want simple assistance like dating advice. Evans shares an anecdote about a Silicon Valley engineer being surprised by deterministic workflow needs.13:55–17:33 · The hosts pushing back 2/10 Market Adoption Dynamics Across Competing AI Ecosystems Toni prompts Evans to consider who AI is being built for, leading Evans to detail consumer adoption data showing Meta AI and Gemini surging while Claude remains niche despite tech enthusiast sentiment.17:34–20:30 · The hosts pushing back 1/10 Historical Parallels and Year-End Reflections Evans contextualizes the current AI era with deep historical parallels to the internet in 1997 and mobile in 2008, noting early assumptions about portals and TV information superhighways were completely wrong.

speaking balance: gold is the hosts, purple is the guest (3 minute bins)

0:00 · the hosts 90.1% · guest 9.9%0:00 · the hosts 90.1% · guest 9.9%3:00 · the hosts 96% · guest 4%3:00 · the hosts 96% · guest 4%6:00 · the hosts 99.2% · guest 0.8%6:00 · the hosts 99.2% · guest 0.8%9:00 · the hosts 78.3% · guest 21.7%9:00 · the hosts 78.3% · guest 21.7%12:00 · the hosts 75.2% · guest 24.8%12:00 · the hosts 75.2% · guest 24.8%15:00 · the hosts 78.6% · guest 21.4%15:00 · the hosts 78.6% · guest 21.4%18:00 · the hosts 89.3% · guest 10.7%18:00 · the hosts 89.3% · guest 10.7%
Sharpest disagreement ▶ 13:48 Questioning target audience and competition

Toni gently redirects Evans's framework by challenging whether the industry even understands who it is building for or competing against.

 Hardest push from the hosts ▶ 12:23 Pushing back on absolute lack of need

Evans clarifies that it is not that no one needs AI, but that discovering viable non-specialist workflows remains an unresolved friction.

 Biggest teaching moment ▶ 9:15 Highlighting builder-user disconnect

Toni notes that unlike web browsers which creators used natively, AI builders are fundamentally detached from how average consumers interact with software.

 The host holds their own ▶ 19:10 Detailed historical analysis of tech transitions

Evans demonstrates encyclopedic market history by walking through how Microsoft browser dominance failed to capture the web and how early mobile predictions missed Apple's ecosystem model.
the scores for every segment, with the reasoning behind each
ChapterTopicThe hosts as informed peerGuest teachingGuest disagreementThe hosts pushing backWhy
Model Parity and the Lack of AI Moats 8112 Benedict Evans opens by framing the structural dynamics of LLM competition, breaking down the lack of network effects, chip and data center economics, and model parity across vendors. Toni chimes in agreeably, noting the low-stakes usage patterns of consumers.
Up-the-Stack Strategies: Features, Applications, and APIs 8111 Evans systematically deconstructs up-the-stack strategies, comparing OpenAI's platform ambitions to historical models like Microsoft Windows proprietary APIs versus cloud infrastructure lock-in like AWS. Toni concurs that users neither know nor care about underlying infrastructure.
The Product Constraints of the Chatbot Interface 7111 Evans draws analogies to browser UI constraints, arguing that an open input/output chat box leaves very little surface area for defensible UI differentiation. The exchange is highly analytical and collaborative.
The Disconnect Between Frontier Benchmarks and User Needs 8222 Evans highlights reports of OpenAI engineers perplexed that users do not demand PhD-level reasoning benchmarks, while Toni remarks that users mainly want simple assistance like dating advice. Evans shares an anecdote about a Silicon Valley engineer being surprised by deterministic workflow needs.
Market Adoption Dynamics Across Competing AI Ecosystems 8222 Toni prompts Evans to consider who AI is being built for, leading Evans to detail consumer adoption data showing Meta AI and Gemini surging while Claude remains niche despite tech enthusiast sentiment.
Historical Parallels and Year-End Reflections 8111 Evans contextualizes the current AI era with deep historical parallels to the internet in 1997 and mobile in 2008, noting early assumptions about portals and TV information superhighways were completely wrong.

Statements from this episode (11)

Opinion
Evans: Frontier general-purpose AI models are at parity and indistinguishable
“The first of them is, like, the models are all basically the same. The general purpose model, the general purpose use cases, the models are all basically the same. And from week to week, the leader changes. They're all within a couple of percentage points of e…”
Benedict Evans Dec 20, 2025 ▶ 0:33
Insight
Evans: AI foundation models lack network effects and sustainable moats
“There's no network effect that we know of yet. And so unlike previous kinds of software platform shift, there's not any Obvious strategy in building the models where you can make yours better than everybody else's by doing that, following that plan. Well, you …”
Benedict Evans Dec 20, 2025 ▶ 1:07
Prediction Not checkable as stated
Evans: OpenAI will struggle to make data centers better than rivals
“In the end, A, it's going to be very hard for particularly OpenAI to, for their data center to be better than everybody else's or cheaper or more efficient than anybody else's.”
Benedict Evans Dec 20, 2025 ▶ 1:46
Prediction Not checkable as stated
Evans: AI infrastructure may become a high-cost oligopoly without network effects
“So you've got kind of a down the stack question, which might end up looking like semiconductors or aircraft or something where it's just, there's no network effects, it's just really expensive. And maybe, you know, there's only two people who make aircraft. Th…”
Benedict Evans Dec 20, 2025 ▶ 3:25
Assertion Supported
Evans: Sora usage dropped off the charts as the novelty faded
“Of course, it might also be Sora which is obviously an open AI attempt to solve this, or now Sora has now dropped off the charts as the novelty wore off.”
Benedict Evans Dec 20, 2025 ▶ 4:48
Insight
Evans: Building first-party AI apps means competing with all of Silicon Valley
“So if you're going to make your own experiences, then you are competing with the whole of Silicon Valley and the entire tech industry in trying to invent new experiences on this technology.”
Benedict Evans Dec 20, 2025 ▶ 5:17
Insight
Evans: Memory in AI chatbots creates stickiness, not network effects
“Memory is a feature and a stickiness, not network effect.”
Benedict Evans Dec 20, 2025 ▶ 8:26
Insight
Evans: Heavy LLM users are atypical, creating a paradox for AI developers
“If you're using this all day, every day, you're not a typical user. It's like the paradox. Normally you would say if you're only using this once a week, you don't, you shouldn't be working on it. You should only be working on it if you're using it every day. T…”
Benedict Evans Dec 20, 2025 ▶ 9:44
Opinion
Brown: Silicon Valley is building incredible AI capabilities that nobody needs
“And maybe this is the first time that the disconnect is just so obvious that we have this insane capability of building these things that are incredible and yet no one actually needs it right now.”
Toni Cowan-Brown Dec 20, 2025 ▶ 12:13
Disclosure
Evans: I only use LLMs for proofreading; they fail intern tasks
“I use it for proofreading sometimes, and that's about it. I, you know, I can't use it for most of the stuff that I have that I would give to an intern, because it won't be able to do it, or it won't get it right.”
Benedict Evans Dec 20, 2025 ▶ 12:57
Assertion Supported
Evans: Meta AI consumer usage rivals Gemini; Claude has zero adoption
“For consumers, Meta AI has got not actually very dissimilar usage to Gemini, and Claude is nowhere.”
Benedict Evans Dec 20, 2025 ▶ 16:40
Made with StarZero

Turn any episode into a week of clips.

This entire site, over 100 episodes transcribed, diarized, checked and made playable, runs on the StarZero media pipeline. Drop in your own episode and the podcast clipper finds the moments worth sharing, cuts them, captions them, and reframes them for every feed.

Try the podcast clipper Learn more