"How will OpenAI compete?" is a February 19, 2026 essay by Benedict Evans published on ben-evans.com. It argues that OpenAI lacks a defensible competitive position and that the platform-and-ecosystem narrative the company advanced in late 2025 does not describe the dynamics it claims. It is an opinion piece and is treated here as an argument, not as evidence.
Four strategic questions
Evans opens by framing OpenAI's position as four problems.
- No clear competitive lead today. The business has no unique technology or product. The models have a very large user base but "very narrow engagement and stickiness, and no network effect or any other winner-takes-all effect so far." OpenAI also has no consumer products built on the models that have product-market fit.
- The market will change substantially. Experience, product, value capture, and strategic leverage will all shift over the next couple of years as incumbents and "thousands of entrepreneurs" try to turn foundation models into commodity infrastructure sold at marginal cost. Having started the LLM boom, OpenAI must now invent a further set of new things, or "fend off, co-opt and absorb" those who do.
- Crossing the chasm without distribution. OpenAI, like Anthropic, must cross the "messy middle" without existing products that act as distribution, and compete in "one of the most capital-intensive industries in history" without cashflows from an existing business. Evans adds that firms that do have those assets face their own problem of self-disruption, but that the field is "well past the point that people said Google couldn't do AI."
- Product does not control the roadmap. Quoting Fidji Simo, OpenAI's head of product, and citing earlier remarks by Mike Krieger and Kevin Weil, Evans argues that heads of product at AI labs have limited ability to set strategy: research sets the agenda and product's job "is to turn that into a button. The strategy happens somewhere else."
He reads OpenAI's activity over the preceding twelve months as Sam Altman being "deeply aware of this, and … trying above all to trade his paper for more durable strategic positions before the music stops."
The models
Evans counts "something like half a dozen organisations" shipping competitive frontier models with near-equivalent capabilities, leapfrogging each other every few weeks. He notes variation within that group — Meta having dropped off the curve "for now"; Apple, Amazon, and Microsoft having failed to get onto it; China remaining roughly six months behind the frontier and relying heavily on others' work — but concludes that "today there is no mechanic we know of for one company to get a lead that others in the field could never match," with no equivalent of the self-reinforcing network effects that ran from Windows to Google Search to iOS to Instagram.
He identifies two things that could change this and argues neither can be planned for: a breakthrough enabling a network effect, "most obviously continuous learning"; or practical scale effects around proprietary data, which he calls "the dark matter of AI." On the data path he distinguishes user data (where the existing platform companies already hold a great deal, and the relevant scale is unclear) from vertical data (foundation models "don't know what happens inside SAP or Salesforce," but such data can be bought or commissioned by any large player).
The user base
OpenAI's one clear lead, Evans writes, is 800–900 million users — but these are weekly actives with shallow engagement. He cites The Information's reporting that only 5% of ChatGPT users pay; Pew survey data showing US teens are much more likely to use chatbots a few times a week or less than multiple times a day; and OpenAI's own "2025 wrapped" promotion, from which he derives that 80% of users sent fewer than 1,000 messages in 2025 — "an average of less than three prompts per day, and many fewer individual chats." His summary: "Usage is a mile wide but an inch deep."
He draws two consequences. Most users never encounter the differences in model personality and emphasis that heavy users notice, and most do not benefit from memory or the other stickiness features labs copy from each other — and he distinguishes these explicitly: "memory is stickiness, not a network effect." Second, a larger user base may yield a usage-data advantage, but its size is unclear if most of that base uses the product a couple of times a week.
Evans reads OpenAI's advertising project as serving two purposes: covering the cost of serving the 90%-plus of users who do not pay while capturing an early position with advertisers, and, more strategically, making it possible to give non-paying users the newest and most expensive models in the hope of deepening engagement. He quotes Simo that "diffusion and scale is the most important thing," and treats OpenAI's own language about a "capability gap" between what models can do and what people do with them as "a way to avoid saying that you don't have clear product-market fit." He allows that better models may deepen engagement but argues it is "at least equally likely" that users are stuck on the blank-screen problem, or that the chatbot is the wrong product for their use cases regardless of model quality.
The browser analogy. Where a product is undifferentiated, Evans argues, early adoption leads tend not to be durable and competition shifts to brand and distribution. He points to share gains by Gemini and Meta AI — noting that although people in tech wrote off Llama 4 as a fiasco, "Meta's numbers seem to be good" — and to Anthropic regularly topping benchmarks while having "no consumer strategy or product … and close to zero consumer awareness," with an aside that Claude Cowork "asks you to install Git." He extends the widely made ChatGPT-to-Netscape comparison into a claim about product form: differentiating a chatbot is the same problem as differentiating a web browser, because "the browser itself, and the chatbot itself, are just an input box and an output box." The last successful browser product innovations, he notes, were tabs and merging search into the URL bar. He also observes that Microsoft won the first-generation browser war and it "turned out not to matter, because the experiences that did matter, and the value capture, were created elsewhere."
Capex and the platform narrative
Evans addresses the diagram Altman presented in late 2025 alongside the Bill Gates definition of a platform as something that creates more value for its partners than for itself, and a companion diagram published by OpenAI's CFO. His response: "a 1:1 relationship between capex and revenue is not a flywheel." He grants the strategy is coherent on its face — build every layer from chips upward, with the higher layers mutually reinforcing — but argues it is the wrong analogy, and that OpenAI has none of the platform and ecosystem dynamics Microsoft or Apple had.
On the capital numbers he cites roughly $400 billion of 2025 infrastructure spending by the big four cloud companies with at least $650 billion announced for the following year, and OpenAI's claim of $1.4 trillion and 30 gigawatts of compute commitment with no timeline against 1.9 gigawatts in use at the end of 2025 — financed through capital-raising, not all of it closed, and other companies' balance sheets, some of it the "circular revenue" arrangements. See AI Bubble vs. Buildout — Synthesis.
Rock's Law. Evans raises the possibility that AI infrastructure resembles airliners or semiconductors: no network effects, but each generation more difficult and more expensive, consolidating from dozens of firms to Boeing and Airbus, or to TSMC. He pairs Moore's Law with Rock's Law — that the cost of a state-of-the-art fab doubles every four years — and suggests generative AI may follow the same pattern, with unit costs falling and fixed costs rising until only a handful of firms can sustain competitive model investment. He notes the resulting oligopoly would have a price equilibrium at unknown margins, possibly "commodity infrastructure sold at marginal cost," particularly since some participants will use their models to power more differentiated businesses. In footnotes he offers two counter-scenarios: the ZIRP-era "capital as a weapon" pattern, which he says worked arguably for Uber and not for WeWork, "which, like OpenAI, had no network effect"; and the possibility of a large number of models of many sizes, some running free on the edge, making "an oligopoly of AI infrastructure" as odd a phrase as an oligopoly of SQL infrastructure.
On Altman's stated aspiration to build a gigawatt of compute every week — which Evans notes implies something on the order of a trillion dollars of annual capex — he declines both the dismissal ("braggawatts"; TSMC staff reportedly calling Altman "podcast bro") and the endorsement, reading it as an attempt to create a self-fulfilling prophecy that has so far worked. His question is whether it buys anything beyond a seat at the table: TSMC has "a de facto monopoly on cutting edge chips" and little to no leverage or value capture up the stack, because "people don't build TSMC apps or Intel apps." Developers had to build for Windows because it had the users and users had to buy Windows because it had the developers; by contrast, an application calling a foundation model through an API leaves users unaware of which model was used, as no Snap user cares whether it runs on AWS or GCP. Evans grants that the APIs are not interchangeable and that developers choose between clouds for real reasons, but maintains that "running a cloud doesn't give you leverage over third part products and services that are further up the stack." See Inference Economics and Token Pricing.
Standards, agents, and the "widget fallacy"
Evans acknowledges one structural difference from earlier cloud generations: the emerging set of standards and protocols letting models and websites talk to each other across advertising, e-commerce, and automation, so that a website can surface a subset of its capabilities inside ChatGPT. He cites the brief enthusiasm around OpenClaw as capturing some of this, and the example of instructing an agent to read a recipe on Instagram and order the ingredients on Instacart. Setting and controlling those APIs would confer power — "standards have been a basic competitive weapon in every generation of technology," recalling Microsoft's "embrace and extend" — and OpenAI's suggestion is that the ChatGPT account becomes the glue linking them, which the company frames as a network effect.
Evans doubts it on two grounds. The first he names the widget fallacy: the recurring belief in technology that many different complex products can be abstracted into a simple standard interface. He identifies the decade-old "APIs are the new BD" as the same idea, and says it mostly failed — partly because of the gap between demos and the interaction models and workflows of real products, where exception cases quickly require the actual product UI and a human decision, and partly because the incentives are misaligned: "no-one wants to be someone else's dumb API call." He notes the resulting tension between the distribution an abstraction layer offers (Google Shopping, Facebook shopping, now ChatGPT shopping) and a company's desire to own the experience and the customer relationship, observing that all of Instacart's profits come from showing ads.
The second is that plugged-together systems produce weak lock-in. If chatbot-feed apps work and OpenAI and Gemini use different standards, supporting both is far less work than shipping separate iOS and Android apps — and, he asks, "can't you get the AI to write the code for you? What does that do to developer lock-ins?" He raises the same doubt about a universal login: it is not obvious that users want to sign in to Tinder, Zillow, and Workday with the same account, or that those services want them to.
Power
Evans closes by noting that he has "returned again and again to terms like platform, ecosystem, leverage and network effect," which "get used a lot in tech, but they have pretty vague meanings" — Google Cloud, Apple's App Store, Amazon Marketplace, and TikTok are all called platforms and are all very different. The word he settles on is power, quoting his university medieval-history professor Roger Lovatt that power is "the ability to make people do something that they don't want to do."
Applied to OpenAI, the question becomes whether the company can get consumers, developers, and enterprises to use its systems more than anyone else's regardless of what those systems do. Microsoft, Apple, and Facebook had that, and so does Amazon — "this is a real flywheel," illustrated with the Amazon flywheel diagram. Evans reads the Gates definition accordingly: a platform harnesses the creative energy of the whole industry so that the platform owner need not invent everything, "but, it's all done on your system with you holding the reins."
His concluding question is whether large language models have that property. Foundation models are "certainly multipliers," and a great deal will be built with them. But absent a reason everyone must use one company's version when competitors have built the same thing, and absent a reason that version will stay better "no matter how much money and effort they throw at it," what remains is "execution, every single day" — which he calls an aspiration, and "not a strategy." In a footnote he notes the counter-argument that Microsoft, Google, Apple, and Meta lived with winner-takes-all effects but never believed they had won, citing Andy Grove's "only the paranoid survive" and Intel's loss of both its network effect and its technology lead.
Relation to other sources
The essay's central economic claim — that frontier-model capability is converging and competition therefore shifts to distribution, brand, and cost — is the same starting point Ben Thompson takes in Who's Afraid of Chinese Models? (Ben Thompson, Stratechery, July 2026), but the two reach different conclusions about who benefits. Thompson argues that whoever is on the frontier is best placed to dominate non-frontier markets because applied intelligence compounds and the frontier labs likely hold the lowest cost per unit of frontier-quality intelligence; Evans argues that converged capability leaves no durable advantage for a firm without distribution, and that infrastructure ownership confers no leverage up the stack. Evans's treatment of engagement depth extends the measurement skepticism of his earlier "AI metrics" essay (2025), summarized at Benedict Evans.
Evans's argument that OpenAI lacks the network effects of prior platform generations is a direct challenge to applying Aggregation Theory to foundation-model companies, and reaches that conclusion by a different route than Thompson's later observation that AI restores marginal costs and so reverses the theory's premise.
Provenance
Retrieved from the author's own site, ben-evans.com, the canonical host. The gap-identifier separately verified the essay on 2026-06-19 against the site's essays index (listing it under 19 February 2026), the page's datePublished metadata, and distinctive passages including the four strategic questions, the "widget fallacy," the Roger Lovatt definition of power, the Amazon flywheel, and Rock's Law. That pull contained a summary rather than the full text and was merged into this raw file, which carries the complete essay.
Figures in the essay are Evans's own citations to third parties (The Information on paying-user share; Pew on teen usage; OpenAI's "2025 wrapped" promotion; Bloomberg and the Wall Street Journal on financing) and are reported here as he reports them, not independently verified.
Relationships
- contradicts (partial): Who's Afraid of Chinese Models? (Ben Thompson, Stratechery, July 2026) — same premise of converged frontier capability, opposite conclusion about whether frontier position confers durable advantage
- depends-on: Aggregation Theory — argues the framework's platform dynamics do not apply to foundation-model companies
- related: OpenAI, Anthropic, Inference Economics and Token Pricing, AI Bubble vs. Buildout — Synthesis
- related: Benedict Evans — author