Beneath the surface, OpenAI's "Four Major Dilemmas"

a16z former partner and renowned tech analyst Benedict Evans recently published an in-depth analysis article, directly pointing out four fundamental strategic dilemmas faced by OpenAI behind its surface prosperity. He believes that despite OpenAI’s large user base and ample capital, issues such as lack of technological moat, insufficient user stickiness, rapid catch-up by competitors, and product strategy being constrained by laboratory R&D directions threaten its long-term competitiveness.

Evans notes that OpenAI’s current business model does not have a clear competitive advantage. The company neither possesses unique technology nor has formed network effects. Only 5% of its 900 million weekly active users are paying, and 80% of users sent fewer than 1,000 messages in 2025—equivalent to less than three prompts per day on average. This “one mile wide, one inch deep” usage pattern indicates that ChatGPT has not yet become a daily habit for users.

Meanwhile, tech giants like Google and Meta have caught up technically with OpenAI and are leveraging their distribution advantages to capture market share. Evans believes that the true value in AI will come from new experiences and application scenarios that have yet to be invented, and OpenAI cannot create all these innovations alone. This forces the company to fight on multiple fronts, deploying across infrastructure and application layers comprehensively.

Evans’s analysis reveals a core contradiction: OpenAI attempts to build competitive barriers through large-scale capital investment and a full-stack platform strategy, but in the absence of network effects and user lock-in mechanisms, the effectiveness of this approach remains uncertain. For investors, this means reevaluating OpenAI’s long-term value proposition and its true position in the AI competitive landscape.

Disappearance of Technological Advantages: Increasing Model Homogeneity

Evans points out that currently about six organizations can launch competitive cutting-edge models with performance that is largely comparable. Companies frequently surpass each other every few weeks, but no one can establish a technological lead that others cannot match. This contrasts sharply with platforms like Windows, Google Search, or Instagram—where network effects reinforce market share, making it difficult for competitors to break monopolies regardless of how much they invest.

This technological equalization could change with certain breakthroughs, most notably the realization of continuous learning capabilities, but Evans believes OpenAI currently has no plans in this direction. Another potential differentiator is the scale effect of proprietary data, including user data or vertical industry data, but existing platform companies also hold advantages here.

As model performance converges, competition shifts toward branding and distribution channels. The rapid growth of Gemini and Meta AI’s market share confirms this trend—these products appear similar to ordinary users, while Google and Meta possess strong distribution capabilities. In contrast, although Anthropic’s Claude model often ranks top in benchmarks, its consumer awareness is nearly zero due to lack of consumer strategy and product focus.

Evans compares ChatGPT to Netscape, which once held early dominance in the browser market but was ultimately defeated by Microsoft leveraging distribution advantages. He believes, chatbots face the same differentiation challenge as browsers: fundamentally, they are just an input box and an output box, with very limited space for product innovation.

User Base is Fragile: Scale Cannot Cover Lack of Stickiness

Despite OpenAI’s obvious lead with 800 to 900 million weekly active users, Evans points out that this data masks serious user engagement issues. The vast majority of users who are aware of and know how to use ChatGPT have not cultivated it into a daily habit.

Data shows that only 5% of ChatGPT users pay, and even among American teenagers, the proportion who use it several times a week or less far exceeds those who use it multiple times daily. OpenAI disclosed in its “2025 Annual Summary” that 80% of users sent fewer than 1,000 messages in 2025, which at face value is less than three prompts per day, with actual chat frequency likely even lower.

This shallow usage means most users do not see differences in personality or focus among models, nor benefit from features like “memory” designed to build stickiness. Evans emphasizes that memory functions can only create stickiness, not network effects. While a larger user base’s usage data might be an advantage, when 80% of users use it only a few times per week, the actual benefit of that advantage is questionable.

OpenAI itself admits there is a problem, citing a “capability gap” between model abilities and actual user usage. Evans believes this is an evasion of the unclear product-market fit. If users cannot think of what to do with it in everyday life, it indicates that it has not yet changed their routines.

The company has launched advertising projects, partly to cover the costs of serving over 90% of non-paying users, but more strategically, it allows the company to offer the latest, most powerful (and most expensive) models to these users, hoping to deepen engagement. However, Evans questions whether giving users better models today or this week will truly change their inability to think of uses.

Platform Strategy is Questionable: Lacking True Flywheel Effect

Last year, OpenAI CEO Sam Altman attempted to unify the company’s initiatives into a coherent strategy, presenting a diagram and quoting Bill Gates: “A platform is defined as creating more value for partners than for itself.” Meanwhile, the CFO released another diagram illustrating the “flywheel effect.”

Evans sees the flywheel as a clever, coherent strategy: capital expenditures form a virtuous cycle, serving as the foundation for building a full-stack platform company. Starting from chips and infrastructure, each layer of the tech stack is built upward, and the higher you go, the more it helps others use your tools to create their own products. Everyone uses your cloud, chips, and models, and at higher levels, the layers of the tech stack reinforce each other, forming network effects and ecosystems.

However, Evans bluntly states that he does not believe this is the correct analogy. OpenAI does not possess the platform and ecosystem dynamics that Microsoft or Apple once had, and that diagram does not truly depict a flywheel effect.

In terms of capital expenditure, the four major cloud providers invested about $400 billion in infrastructure last year and announced at least $650 billion for this year. OpenAI claimed a few months ago that it would have a future commitment of $1.4 trillion and 30 gigawatts of compute (without a clear timeline), but actual usage by the end of 2025 was only 1.9 gigawatts. Due to the lack of large-scale cash flow from existing business, the company relies on financing and leveraging others’ balance sheets (partly involving “recurring revenue”) to achieve these goals.

Evans believes large-scale capital investment may only buy a seat at the table, not a competitive advantage. He compares the costs of AI infrastructure to aircraft manufacturing or semiconductor industries: no network effects, but each generation’s craftsmanship becomes more difficult and expensive, ultimately only a few companies can sustain the necessary investments at the frontier. TSMC, despite its de facto monopoly in advanced chips, has not gained leverage or value extraction in upstream technology layers.

Evans points out that developers build applications for Windows because it has nearly all users, and users buy Windows PCs because they have nearly all developers—this is a network effect. But if you invent a great new application or product using generative AI, you only need to call an API to run the underlying model in the cloud; users don’t know or care what model you use.

Lack of Product Dominance: Strategy is Lab-Dependent

At the start of the article, Evans quotes OpenAI product lead Fidji Simo from 2026: “Jakub and Mark set the long-term research directions. After months of work, amazing results emerge, and then researchers contact me: ‘I have some cool stuff. How do you plan to use it in chat? How about for our enterprise products?’”

This contrasts sharply with Steve Jobs’s famous 1997 statement: “You have to start with the customer experience and then work backwards to the technology. You can’t start with the technology and try to figure out where to sell it.”

Evans believes that when you are an AI lab product leader, you cannot control your own roadmap, and your ability to set product strategy is very limited. You open your email in the morning and see what the lab has researched, and your job is to turn it into a button. The strategy is happening elsewhere—where?

This highlights the fundamental challenge OpenAI faces: unlike Google in the 2000s or Apple in the 2010s, OpenAI’s talented and ambitious staff do not have a truly effective product that others cannot replicate. Evans interprets the past 12 months of OpenAI’s activities as a realization that Sam Altman is deeply aware of this and is trying to turn the company’s valuation into a more durable strategic position before the music stops.

Most of last year, OpenAI’s answer seemed to be “doing everything simultaneously, executing immediately.” Application platforms, browsers, social video apps, collaborations with Jony Ive, medical research, advertising, and more. Evans believes some of these look like “full-scale assaults” or are simply the result of quickly hiring a large number of proactive people. Sometimes it feels like copying the form of previously successful platforms without fully understanding their purpose or dynamic mechanisms.

Evans repeatedly uses terms like platform, ecosystem, leverage, and network effects, but admits these terms are widely used in tech industry with quite vague meanings. He quotes medieval history professor Roger Lovatt from his university days: Power is the ability to make people do what they don’t want to do. That’s 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, Facebook once had this ability, and so did Amazon.

Evans suggests that a good way to interpret Bill Gates’s statement is that a platform truly leverages the entire tech industry’s creativity, so you don’t have to invent everything yourself—just build more at scale, but all within your system, under your control. Foundation models are indeed multipliers; many new things will be built with them. But is there a reason to force everyone to use your product, even if competitors have built similar things? Is there a reason to always keep your product superior, regardless of how much competitors invest?

Evans concludes that without these advantages, the only thing you have left is daily execution. Excelling at execution than others is of course desirable; some companies have achieved this over long periods and even institutionalized it, but that is not a strategy.

Risk Disclaimer and Disclaimers

Market risks are inherent; investments should be cautious. This article does not constitute personal investment advice and does not consider individual users’ specific investment goals, financial situations, or needs. Users should consider whether any opinions, viewpoints, or conclusions in this article are suitable for their particular circumstances. Invest accordingly at your own risk.