How will OpenAI compete? — Benedict Evans
Benedict Evans · Benedict Evans · 2026-02-19
OpenAI has scale but no durable moat: frontier models are converging, ChatGPT's huge user base is shallow and mostly unpaid, and owning the model or the cloud beneath it confers little leverage over what gets built on top.
The question isn't whether OpenAI is impressive — it's whether anything about the business compounds. Half a dozen labs ship roughly equivalent models and trade the lead every few weeks, and there's no known mechanic like Windows' or Google Search's network effects that would let one of them pull permanently ahead. Meanwhile the 800–900M weekly users mostly prompt a few times a week, only 5% pay, and the apps people actually use don't care which model is underneath. Strip out lock-in and what remains is daily execution against equally capable rivals, which is a hope rather than a strategy.
OpenAI's current business has no unique technology, no winner-takes-all dynamic, and no consumer products with real product-market fit on top of the models. The large user base hasn't translated into engagement, stickiness, or a network effect.
ChatGPT has 800–900M weekly active users, but only 5% pay, and 80% sent fewer than 1,000 messages in 2025 — under three prompts a day on average. Even US teens mostly use it a few times a week, not daily.
Half a dozen organizations ship roughly equivalent frontier models and leapfrog each other every few weeks. No known mechanic — analogous to the network effects of Windows, Google Search, or iOS — lets one lab open a gap others can't close.
Foundation models are multipliers, but if competitors can build the same thing and no mechanism forces customers to choose yours, you're reduced to out-executing everyone every day. That's an aspiration, not a strategy.
TSMC has a near-monopoly on cutting-edge chips but no leverage over what's built on them. Users don't know or care which cloud or which foundation model powers the app they're using, so being the underlying provider doesn't translate into power further up.
Open
- · What, if anything, could create a durable mechanic for a permanent lead among frontier labs?
- · Can a consumer product with real stickiness be built on top of these models, and by whom?
- · If being the model or cloud provider gives no upstack leverage, where in the stack does the value actually accrue?
Pipeline
- source kind
- url
- generated by
- anthropic+voyage
- candidates
- 27 (selected 5)
- embeddings
- voyage-3.5
Coverage
100% covered
Each block is one paragraph of the source. Darker means the decomposition captures it well; lighter means it was left out — the part of the document the summary doesn’t cover.
Considered candidates (22)
Below top-k · 19
- claimAltman's full-stack platform visionc 0.85
OpenAI's coherent strategy is to own every layer of the stack — chips, infrastructure, models, tooling, products — so the layers reinforce each other and developers and users get locked into the ecosystem, in the manner of Windows or iOS.
- claimThe 'capability gap' is really a missing product-market fitc 0.80
OpenAI's own framing of a gap between what models can do and what people do with them is a polite way of admitting that ChatGPT hasn't yet become essential to most users' daily lives.
- claimStandards and protocols could be a new source of powerc 0.80
An emerging stack of agent protocols lets websites, ads, e-commerce, and intent flows pipe into each other through chatbots. Whoever controls those APIs and uses a ChatGPT account as the glue could create a genuine network effect.
- claimThe next-generation AI experiences won't all be built by OpenAIc 0.75
The whole tech industry is racing to invent the second step of generative AI experiences, and there's no reason to assume OpenAI will be the one to invent them rather than one of thousands of competitors.
- caveatThe 'widget fallacy' undermines the agent-protocol visionc 0.75
Abstracting complex products behind a clean standard interface — 'APIs are the new BD' — has repeatedly failed because real workflows hit exceptions and because no one wants to be a dumb API call behind someone else's experience.
- implicationAltman is trading paper for durable position before the music stopsc 0.75
OpenAI sets the agenda and has talent, but unlike 2000s Google or 2010s Apple it lacks a thing only it can do. Altman's frenetic activity over the past year reads as an attempt to convert today's hype-fueled equity into something more durable.
- contextFour strategic problems frame the analysisc 0.70
Evans organizes the essay around four problems OpenAI faces: weak competitive position, a rapidly changing market, the need to cross the chasm without existing products or cashflows, and the fact that product strategy is dictated by what the labs invent rather than chosen.
- caveatBetter models may not fix the blank-screen problemc 0.70
If users can't think of anything to do with ChatGPT today, a more capable model may not change that. The bottleneck may be the chatbot form factor itself, not raw model quality.
- exampleChatbots resemble the browser warsc 0.70
Like Netscape, ChatGPT risks being crowbarred out by incumbents with distribution. Browsers and chatbots are both just an input box and an output box, with little room for UI differentiation.
- implicationMicrosoft won browsers, but that didn't matterc 0.70
Winning browsers was strategically empty because the value and experience of the consumer internet were captured elsewhere. The same may be true of chatbots — the real value will accrue to whoever builds the next-step experiences.
- contextAltman is willing OpenAI a seat at the capex tablec 0.70
OpenAI claims $1.4T and 30GW of future compute commitments against 1.9GW actually in use, funded by capital-raising and other people's balance sheets including 'circular revenue'. Altman is trying to self-fulfill a prophecy.
- caveatCross-platform agents don't create lock-inc 0.70
Even if the protocol vision works, developers can trivially support both ChatGPT and Gemini standards — the AI itself will write the glue. And users may not want to log into Tinder, Zillow, and Workday with their OpenAI account, nor will those services want it.
- mechanismAI infrastructure may follow the Rock's Law patternc 0.65
Unit costs fall while fixed costs rise generation over generation, as in airliners and semiconductor fabs, leaving only a handful of viable players. The result could be an oligopoly producing commodity infrastructure at marginal cost.
- contextProduct leaders don't set the AI roadmapc 0.60
At an AI lab, product heads don't choose what to build — they wait for research to ship something and then turn it into a button. Strategy happens upstream of product, which raises the question of where it actually happens.
- caveatPossible but unplannable sources of a future moatc 0.55
A breakthrough like continuous learning could create a network effect, and proprietary user or vertical data could create scale advantages, but none of this is something you can plan for today. So the working assumption has to be that models stay close.
- mechanismAds exist to subsidize giving free users the best modelsc 0.55
OpenAI's advertising play partly covers the cost of serving 90%+ non-paying users, but more strategically it lets the company push the most powerful models to free users in hopes of deepening engagement.
- implicationShallow engagement undermines the data-flywheel argumentc 0.50
If most users barely interact, they don't experience model personality differences or benefit from memory, and the data advantage of a larger user base shrinks correspondingly.
- contextOpenAI's scattershot 2024 responsec 0.50
Last year OpenAI launched at everything — app platforms, a browser, a social video app, Jony Ive, medical research, advertising, plus trillion-dollar capex announcements — which looked like flooding the zone or copying platform forms without their substance.
- exampleAnthropic shows that benchmark wins don't beat distributionc 0.45
Claude tops benchmarks regularly but has near-zero consumer awareness and no consumer strategy, illustrating that being best on the model isn't enough without distribution.
Redundant with selected · 3
- claimThe Windows/iOS analogy doesn't actually holdc 0.90 · sim 0.85
OpenAI's flywheel diagram isn't really a flywheel. It lacks the platform and ecosystem dynamics that made Microsoft and Apple defensible.
overlapped with: OpenAI lacks a durable competitive lead today
- claimThe real question is power, not 'platform'c 0.90 · sim 0.83
Platform, ecosystem, leverage, and network effect are loose words. The sharp question is whether OpenAI can make consumers, developers, and enterprises use its systems regardless of what the systems do — the way Microsoft, Apple, Facebook, and Amazon could.
overlapped with: OpenAI lacks a durable competitive lead today
- claimWith undifferentiated products, competition shifts to brand and distributionc 0.85 · sim 0.84
When products look the same to users, early adoption leads decay quickly. Gemini and Meta AI are gaining share precisely because Google and Meta can leverage distribution that OpenAI lacks.
overlapped with: OpenAI lacks a durable competitive lead today
Janitor
Non-content spans (acknowledgements, references, footnotes, headers, boilerplate) are dropped before the decomposition runs.
- total spans
- 52
- kept
- 46
- dropped
- 6
- outliers
- 3
- content · 46
- footnote · 3
- metadata · 2
- noise · 1