Beneath the glamorous surface, OpenAI’s “four major dilemmas”

Former a16z partner and renowned tech analyst Benedict Evans recently published an in-depth article directly pointing out four fundamental strategic dilemmas OpenAI faces behind its apparent prosperity. He believes that despite OpenAI's enormous user base and ample capital, issues like the lack of technological moat, low user stickiness, fast-following competitors, and product strategy being restricted by lab-led R&D are threatening its long-term competitiveness.

Evans notes that OpenAI's current business model lacks clear competitive advantages. The company has neither unique technology nor established network effects. Of its 900 million weekly active users, only 5% are paid, and 80% of users will send fewer than 1000 messages in 2025—amounting to less than three prompts on average per day. This "one mile wide, one inch deep" usage pattern suggests ChatGPT hasn't become a daily habit for users.

Meanwhile, tech giants like Google and Meta have caught up technologically with OpenAI and are leveraging their distribution advantages to seize market share. Evans believes that real value in the AI field will stem from novel experiences and use cases yet to be invented, and OpenAI cannot create all this innovation alone. This forces the company to fight on multiple fronts, laying out from infrastructure all the way to the application layer.

Evans’s analysis reveals a core contradiction: OpenAI seeks to establish competitive barriers through massive capital investment and full-stack platform strategy, but in the absence of network effects and user locking mechanisms, whether this strategy can succeed remains dubious. For investors, this means they need to reassess OpenAI’s long-term value proposition and its actual standing in the AI competitive landscape.

Disappearance of Technological Advantage: Intensifying Model Homogenization

Evans points out in his analysis that currently about six institutions can launch competitive frontier models with largely similar performance. The companies leapfrog each other every few weeks, but none can establish a technical lead that others can't match. This contrasts sharply with platforms like Windows, Google Search, or Instagram—which use network effects to self-reinforce their market share, making it nearly impossible for rivals to break the monopoly regardless of how much capital and effort they invest.

This technological leveling might change with certain breakthroughs, most notably the realization of continuous learning. But Evans believes OpenAI cannot plan for this at present. Another possible differentiating factor is the scale effect of proprietary data, including user data or vertical industry data, but existing platform companies also hold advantages here.

With model performance converging, competition is shifting toward brand and distribution channels. The rapid growth of Gemini and Meta AI’s market share confirms this trend—for ordinary users, these products appear almost identical, while Google and Meta possess formidable distribution abilities. In contrast, Anthropic’s Claude model frequently tops benchmark tests but, lacking a consumer strategy and product, has near-zero consumer awareness.

Evans likens ChatGPT to Netscape, which once held early advantage in the browser market but was eventually defeated by Microsoft’s distribution edge. He believes chatbots face the same differentiation challenge as browsers: essentially little more than an input and output box, with very limited room for product innovation.

Fragile User Base: Scale Masks Low Stickiness

Despite OpenAI’s apparent lead with 800-900 million weekly active users, Evans notes this hides severe user engagement problems. Most users who know about and can use ChatGPT have not made it a daily habit.

Data shows only 5% of ChatGPT users are paid. Even among American teenagers, those using ChatGPT a few times a week or less far outnumber those using it multiple times daily. OpenAI’s “2025 Year in Review” event disclosed that 80% of users will send fewer than 1000 messages in 2025—averaging less than three prompts per day, with actual chats even fewer.

Such shallow use means most users don’t perceive differences between models in personality or focus, nor can they benefit from “memory” features designed to build stickiness. Evans notes that memory features can only bring stickiness, not network effects. Meanwhile, user base data could be an advantage, but when 80% of users use it only a few times a week at most, the value of this advantage is questionable.

OpenAI itself admits the issue, proposing a “capability gap” exists between model ability and actual user usage. Evans sees this as avoiding the issue of unclear product-market fit. If users can’t think of what to do with it in their daily life, then it hasn’t changed their lives.

The company launched advertising projects, partly to cover service costs for the 90%+ non-paying users, but more strategically, to deliver the newest and most powerful (and expensive) models to these users in hopes of deepening engagement. Still, Evans questions whether giving users a better model will change things if they can’t think of what to use ChatGPT for today or this week.

Platform Strategy in Doubt: Lack of True Flywheel Effect

Last year, OpenAI CEO Sam Altman tried to integrate the company’s various initiatives into a coherent strategy, showcasing a chart and quoting Bill Gates: the definition of a platform is that the value it creates for partners exceeds that it creates for itself. Meanwhile, the CFO released another chart illustrating the “flywheel effect.”

Evans considers the flywheel effect an elegant and coherent strategy--capital expenditure forms a virtuous cycle and becomes the foundation for building a full-stack platform company. Starting from chips and infrastructure, building every layer upward, the higher you go, the more you help others use your tools to create their own products. Everyone uses your cloud, chips, and models, and at higher layers, the tech stack reinforces itself, forming network effects and an ecosystem.

However, Evans bluntly states he does not believe this is the right analogy. OpenAI does not possess the platform/ecosystem dynamics Microsoft or Apple once had, and that flywheel chart does not actually show a true flywheel effect.

In capital spending, the four major cloud computing companies invested about $400 billion in infrastructure last year and announced at least $650 billion this year. OpenAI claimed a $1.4 trillion and 30 gigawatts commitment for computing power (no clear timeline), while actual usage in late 2025 is 1.9 gigawatts. Lacking large-scale cash flow from existing business, the company resorts to fundraising and leveraging others’ balance sheets (partly involving “round-trip revenue”) to achieve these goals.

Evans argues massive capital investment may only get you a seat at the table, not a competitive edge. He likens AI infrastructure costs to airplane manufacturing or the semiconductor industry: no network effects, but every generation becomes more difficult and expensive, leaving a handful of companies able to stay on the frontier. Yet, TSMC, though holding a de facto monopoly in cutting-edge chips, does not gain leverage or value extraction power upstream in the tech stack.

Evans notes developers must build for Windows because it has almost all users—users must buy Windows PC because it has nearly all developers—this is network effect. But if you invent a great new app or product using generative AI, you can just call the foundational model via API in the cloud, and users neither know nor care which model you used.

Loss of Product Leadership: Strategy Constrained by the Lab

Evans quotes OpenAI product head Fidji Simo in 2026: “Jakub and Mark set the long-term research direction. After months of work, amazing results emerge, then the researchers reach out and say: ‘I have something cool. How will you use it in chat? How will you use it in our enterprise product?’”

This contrasts sharply with Steve Jobs’s famous line from 1997: “You must start with the customer experience and work backward to the technology. You can’t start with the technology and try to figure out where to sell it.”

Evans observes that when you’re a product lead at an AI lab, you don’t control your roadmap, and your ability to set product strategy is very limited. You open your email in the morning, see what the lab has developed, and your job is to turn it into a button. Strategy happens elsewhere—but where?

This problem highlights the fundamental challenge OpenAI faces: unlike Google in the 2000s or Apple in the 2010s, OpenAI’s smart and ambitious employees do not have a genuinely effective product others cannot replicate. Evans suggests one way to interpret OpenAI’s activity over the past 12 months is that Sam Altman deeply realizes this, and before the music stops, tries to convert the company’s valuation into more durable strategic positioning.

Most of last year, OpenAI’s answer seemed to be: “do everything, all at once, immediately.” Application platforms, browser, social video app, partnership with Jony Ive, medical research, advertising, etc. Evans thinks some of this looks like “all-out assault,” or simply the result of quickly hiring many aggressive people. Sometimes it feels like people copy the forms of past successful platforms, but don’t fully understand their purpose or dynamic mechanisms.

Evans repeatedly uses terms like platform, ecosystem, leverage, and network effect, but admits these are widely used in tech and quite vague. He quotes his college medieval history professor Roger Lovatt: power is the ability to make people do things they don’t want to—that is the real issue: does OpenAI have the ability to get consumers, developers, and enterprises to use its systems more regardless of what the system actually does? Microsoft, Apple, and Facebook once had this ability, and Amazon has it too.

Evans suggests a good way to interpret the Bill Gates quote is: what a platform fundamentally achieves is leveraging the creativity of the entire tech industry, so you don’t have to invent everything yourself, and can build much more at scale, but all done on your system, under your control. Foundational models are indeed multipliers; a lot of new things will be built on them. But do you have a reason for everyone to use your product even if competitors have built the same thing? Do you have a reason to keep your product superior regardless of how much money and effort rivals invest?

Evans concludes that without these advantages, all you have is daily execution. Out-executing everyone else is certainly an aspiration; some companies do manage it for extended periods, even convince themselves they’ve institutionalized it—but it is not a strategy.

Risk warning and disclaimerThe market carries risks, and investing should be done cautiously. This article does not constitute personal investment advice and does not take into account the unique investment goals, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article apply to their specific situation. Invest accordingly at your own risk.