Analyst Benedict Evans Questions OpenAI's Moat, Citing Commoditization and Low User Engagement

Leading technology analyst Benedict Evans has posited that OpenAI's large language models are essentially commodities, suggesting the company could face a similar fate to Netscape in the 1990s. Evans's analysis, shared during an episode of The MAD Podcast with Matt Turck, highlights the high cost of foundational models alongside their commoditized nature.

Abstract visualization of interconnected data nodes, representing the high cost and commoditization of AI models.

Evans noted that OpenAI, despite having 900 million weekly active users, sees 80% of these users interacting with its platforms fewer than 1,000 times annually. This low engagement, coupled with significant capital expenditure in the AI industry—such as Meta dedicating 50% of its revenue to chips—indicates a challenging financial landscape for AI companies.

The Erosion of Competitive Advantage

Historically, technological monopolies have been built on network effects, such as the ecosystem of software developers for Windows or the continuous optimization of Google's search results through user clicks. However, Evans argues that this dynamic does not apply to large language models.

Abstract representation of shifting peaks, symbolizing the rapidly changing leadership in AI model performance.

He contends there is no current evidence that a leading model can sustainably prevent competitors from developing equally capable alternatives. Models like Claude, Gemini, and Llama are rapidly advancing, with leadership in performance often shifting within weeks. Evans stated that OpenAI, lacking its own infrastructure or differentiated network effects, primarily holds "mind share," which he likened to Netscape's position in 1995. Netscape, despite having the leading browser, was quickly overtaken once the technology became an infrastructure and larger players entered the market. OpenAI, according to Evans, is now attempting to convert this mind share into tangible assets.

ChatGPT's Usage Challenges

Analysis of user data, including annual summaries shared on Reddit, revealed that most users interact with AI less than three times daily on average. Users who issued 1,000 commands to ChatGPT last year are considered among the top 20% of global users. For many, ChatGPT serves as an occasional search enhancer or a tool for rewriting text, rather than a central productivity platform.

Person casually using an AI chatbot on a laptop, illustrating low user engagement with AI tools.

This low-frequency usage results in less than 5% of users being willing to pay for the service. Evans emphasized that simply improving model accuracy may not resolve this issue. Even if a model's accuracy improves from 90% to 95%, human verification remains necessary, preventing full automation and value liberation from human workflows.

Financial Pressures in the AI Sector

The AI industry's rapid expansion is marked by substantial financial outlays. Meta's recent financial report projects capital expenditures to exceed 50% of its total revenue, an investment scale Evans described as unprecedented for a company of its size. Giants like Google and Microsoft are also increasing their capital expenditures at double-digit rates.

Vast, futuristic data center interior, symbolizing massive capital expenditures in the AI industry.

This investment, according to Evans, goes beyond typical factory construction, resembling the development of national infrastructure. He questioned whether such investments would yield proportional returns, pointing to a revenue structure within the AI industry that includes significant circular revenue and vendor financing. This involves Nvidia selling chips to cloud providers, who then lease computing power to AI startups, often funded by the same large corporations, potentially masking underlying weaknesses in market demand.

The Rise of Improvised Software

Evans introduced a distinction between "improvised software" and "institutionalized software" when discussing AI's impact on the software industry. He dismissed the notion that every company would develop its own enterprise resource planning (ERP) system, noting that large organizations require stable, compliant systems like SAP and Oracle, which cannot be entrusted to AI agents for improvisation.

However, AI has significantly reduced the cost of code, enabling the transformation of numerous non-standard tasks into software. This does not imply a contraction of the software industry; rather, in line with Jevons' paradox, increased resource efficiency often leads to greater overall demand. Evans predicts an increase in software development, but with a shift in value from code writing to a deep understanding of business requirements. Companies that merely wrap databases, he suggests, may struggle in an environment saturated with AI-generated code.

Abstract visualization contrasting fluid, AI-generated code with stable, institutionalized software systems.

AI's "Netscape Moment"

Reflecting on history, Evans drew parallels between the current AI landscape and the internet boom of 1997. While predictions at the time included e-commerce and video conferencing, unforeseen innovations like Uber emerged. He suggested that while AI's importance is clear and investments are substantial, current applications might still be in an early, "PDF-online" stage.

Abstract depiction of an early internet browser window against a backdrop of emerging futuristic digital city, symbolizing AI's 'Netscape Moment'.

True AI-native applications have yet to fully materialize. OpenAI's efforts to develop agent tools like OpenClaw could position it in direct competition with major players like Google and Apple. As AI attempts to integrate with user desktops and inboxes, challenges related to privacy, permissions, and system control may become more significant than algorithmic advancements. Evans concluded that while some companies will undoubtedly profit from this wave, success may not hinge solely on superior algorithms.