Written by: Zhao Ying

Source: Wall Street Journal

Benedict Evans, a former partner at a16z and renowned technology analyst, recently published an in-depth analysis article, pointing out the four fundamental strategic dilemmas faced by OpenAI behind its apparent prosperity. He believes that although OpenAI has a vast user base and ample capital, the lack of a technological moat, insufficient user stickiness, rapid competition, and issues with product strategy constrained by laboratory research directions are threatening its long-term competitiveness.

Evans pointed out that OpenAI's current business model lacks a clear competitive advantage. The company has neither unique technology nor established network effects, with only 5% of its 900 million weekly active users paying, and 80% of users sending fewer than 1,000 messages in 2025—equivalent to an average of less than three prompts per day. 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 with OpenAI technologically and are utilizing their distribution advantages to seize market share. Evans believes that the real value in the AI field 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 comprehensively from infrastructure to application layers.

Evans' analysis reveals a core contradiction: OpenAI tries to establish competitive barriers through massive capital investment and a full-stack platform strategy, but whether this strategy can work without network effects and user lock-in mechanisms remains questionable. For investors, this means a need to reassess OpenAI's long-term value proposition and its real position in the AI competitive landscape.

Disappearing Technological Advantage: Model Homogenization Intensifies

In his analysis, Evans noted that currently about six organizations can launch competitive leading-edge models, and their performances are roughly equivalent. Companies surpass each other every few weeks, but none can establish a technological lead that others cannot match. This contrasts sharply with platforms like Windows, Google Search, or Instagram, which have achieved self-reinforcing market share through network effects, making it difficult for competitors to break their monopolies regardless of how much funding and effort they invest.

This equalization of technology may change with certain breakthroughs, the most notable being the realization of continual learning capabilities, but Evans believes that OpenAI cannot currently plan for this. Another possible differentiating factor is the scale effect of proprietary data, including user data or vertical industry data, but existing platform companies also have advantages in this area.

In the context of similar model performance, competition is shifting towards branding and distribution channels. The rapid growth in market share of Gemini and Meta AI confirms this trend—these products appear very similar to ordinary users, while Google and Meta possess strong distribution capabilities. In contrast, Anthropic's Claude model, although frequently topping benchmark tests, has close to zero consumer awareness due to a lack of consumer strategy and products.

Evans draws a parallel between ChatGPT and Netscape, which held early advantages in the browser market but was eventually defeated by Microsoft’s distribution advantages. He believes that chatbots face the same differentiation dilemma as browsers: they are essentially just an input box and output box, with very limited product innovation space.

Weak User Base: Scale Cannot Mask Insufficient Stickiness

Despite OpenAI's apparent lead with 800 to 900 million weekly active users, Evans points out that this figure obscures serious user engagement problems. The vast majority of users who are aware of and know how to use ChatGPT have not developed it into a daily habit.

Data shows that only 5% of ChatGPT users pay, and even among American teenagers, the ratio using it a few times a week or less is much higher than the ratio using it multiple times a day. OpenAI disclosed during its "2025 Annual Summary" event that 80% of users sent fewer than 1,000 messages in 2025, which, at face value, translates to an average of less than three prompts per day, with actual chatting being even less.

This shallow usage means that most users do not see differences in personality or focus between different models and cannot benefit from features like "memory," which are aimed at establishing stickiness. Evans emphasizes that memory features can only bring stickiness, not network effects. While usage data from a larger user base can be an advantage, it is questionable how much of an advantage this is when 80% of users use it only a few times a week.

OpenAI itself has acknowledged the problem, suggesting that there is a "capability gap" between the model's capabilities and actual user usage. Evans believes this is an evasion of the fact that product-market fit is unclear. If users cannot think of what to do with it in everyday situations, it indicates that it has not changed their lives.

The company has launched an advertising program partly to cover the service costs of over 90% of unpaid users, but more strategically, this allows the company to offer the latest, most powerful (and expensive) models to these users in hopes of deepening user engagement. However, Evans questions whether providing better models would change the situation if users cannot think of what to do with ChatGPT today or this week.

Questionable Platform Strategy: Lack of True Flywheel Effect

Last year, OpenAI CEO Sam Altman attempted to integrate the company's various initiatives into a coherent strategy, presenting a chart and quoting Bill Gates' saying: the definition of a platform is that the value created for partners exceeds the value created for itself. Meanwhile, the CFO released another chart illustrating the "flywheel effect."

Evans believes that the flywheel effect is a clever, coherent strategy: capital expenditure forms a virtuous cycle in itself and becomes the basis for building a full-stack platform company. Starting from chips and infrastructure, building upward each layer of the tech stack, the higher you go, the more you help others create their products using your tools. Everyone uses your cloud, chips, and models, and then at higher levels, the layers of the tech stack reinforce each other, forming network effects and ecosystems.

However, Evans candidly stated that he does not believe this is the correct analogy; OpenAI does not possess the kind of platform and ecosystem dynamics that Microsoft or Apple once had, and that flywheel chart does not actually demonstrate a true flywheel effect.

In terms of capital expenditure, the four major cloud computing companies invested about $400 billion in infrastructure last year and announced plans to invest at least $650 billion this year. A few months ago, OpenAI claimed future commitments of $1.4 trillion and 30 gigawatts of computing power (without a clear timeline), whereas the actual usage by the end of 2025 was 1.9 gigawatts. Due to the lack of large cash flow from existing business, the company is achieving these goals through financing and utilizing others' balance sheets (partly involving "recurring revenue").

Evans believes that large-scale capital investment may only secure a seat, not a competitive advantage. He likens AI infrastructure costs to those of aircraft manufacturing or the semiconductor industry: there are no network effects, but the craftsmanship of each generation of products becomes increasingly difficult and expensive, ultimately allowing only a few companies to maintain the necessary investment at the forefront. However, while TSMC possesses a de facto monopoly in cutting-edge chips, it has not gained leverage or value extraction capability in the upstream tech stack.

Evans points out that developers must build applications for Windows because it has nearly all users, and users must buy Windows PCs because it has nearly all developers—this is the network effect. But if you invent an outstanding new application or product using generative AI, you simply call the cloud-based foundational model via API; users do not know or care what model you used.

Missing Product Leadership: Strategy Constrained by Laboratories

Evans opened the article by quoting OpenAI's product head Fidji Simo from 2026: "Jakub and Mark set long-term research directions. After months of work, amazing results emerge, and then researchers contact me and say: 'I have something cool. How do you plan to use it in chat? How is it used in our enterprise products?'"

This statement starkly contrasts with Steve Jobs' famous quote from 1997: "You have to start with the customer experience and work backward to the technology. You can't start with the technology and then figure out where to sell it."

Evans believes that when you are the product head of an AI lab, you cannot control your roadmap, and your ability to set product strategy is very limited. You open your email in the morning and see what results the lab has produced, and your job is to make it into a button. Strategy occurs elsewhere, but where?

This question highlights a fundamental challenge facing OpenAI: unlike Google in the 2000s or Apple in the 2010s, OpenAI's smart and ambitious employees do not have a truly effective product that others cannot replicate. Evans believes that one interpretation of OpenAI's activities over the past 12 months is that Sam Altman has become acutely aware of this and is trying to convert the company's valuation into a more lasting strategic position before the music stops.

For most of last year, OpenAI's answer seemingly was "doing everything, all at once, immediately." Application platforms, browsers, social video apps, collaborations with Jony Ive, medical research, advertising, and so forth. Evans believes that some of this looks like a "full-scale attack," or simply the result of quickly hiring many aggressive individuals. At times, it also seems that people are replicating the forms of previously successful platforms without fully understanding their purposes or dynamic mechanisms.

Evans repeatedly uses terms such as platform, ecosystem, leverage, and network effects, yet he acknowledges that these terms are widely used in the tech industry but remain quite vague in meaning. He quotes his medieval history professor from college, Roger Lovatt: power is the ability to make people do what they do not want to do. This 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 itself actually does? Microsoft, Apple, and Facebook once had that capability, as did Amazon.

Evans believes a good way to interpret Bill Gates' statement is that platforms truly leverage the creativity of the entire tech industry, so you do not have to invent everything yourself, allowing you to build more at scale, but all of this is done within your system and under your control. Foundational models are indeed multipliers, and many new things will be built using them. But do you have a reason to make everyone use your product, even if competitors have built the same thing? Is there a reason to ensure your product always outperforms competitors, regardless of how much they invest in funding and effort?

Evans concludes that without these advantages, the only thing you have is the execution power each day. Executing better than others is certainly a desire; some companies have done this over extended periods and even convinced themselves they have institutionalized it, but this is not a strategy.