Generative AI: The Anatomy of a Platform Shift

Source Acknowledgment

This synthesis draws from the work of Benedict Evans, one of the most rigorous and level-headed analysts working in technology today. Evans spent twenty-five years analyzing mobile, media, and technology across equity research, strategy consulting, and venture capital – including six years as a partner at Andreessen Horowitz in Silicon Valley. He now operates as an independent analyst, writing a weekly newsletter that reaches nearly 200,000 subscribers and producing presentations for organizations including Alphabet, Amazon, L'Oréal, LVMH, Deutsche Telekom, and Vodafone.

Twice yearly, Evans publishes "AI Eats the World" – a comprehensive presentation exploring macro and strategic trends shaping the technology industry. These presentations have become essential reading for executives, investors, and strategists seeking clarity amid the noise. Where others traffic in hype or doom, Evans traces patterns. He asks questions rather than declaring certainties. He measures what can be measured and acknowledges what remains unknown.

This document synthesizes insights from three presentations: Autumn 2024, Spring 2025, and Autumn 2025. The analysis, reframing, and editorial voice are ours; the foundational research and data belong to Evans.

For the original presentations, newsletter archive, and ongoing analysis, visit ben-evans.com. The material there, not just these documents, is second to none.

The Core Truth

Generative AI represents something we witness once every decade or so – a platform shift that redirects the flow of capital, innovation, and company creation. This recognition has triggered an unprecedented capital expenditure surge, with Big Tech committing hundreds of billions annually to AI infrastructure. The logic is asymmetric: the cost of being wrong by under-investing dwarfs the cost of being wrong by over-investing.

Yet beneath this investment frenzy lies a paradox. The AI models themselves are becoming commodities. Performance converges across providers. Prices collapse. No defensible moat has emerged around the core technology. The battle shifts to scale, capital access, distribution, and product design – the unglamorous fundamentals that separate lasting enterprises from spectacular failures.

Meanwhile, a deployment gap persists between what the industry promises and what the market absorbs. Consumer awareness runs high; sustained daily use remains low. Enterprise adoption drowns in pilots. An August 2025 MIT analysis suggests that 95% of corporate generative AI pilots are failing (MIT, 2025). The questions that define this era are deceptively simple: How much further can the technology scale? What products will drive mass adoption? And in a world of commoditized models and massive capital burn, where will economic value ultimately settle?

Platform Shifts: The Recurring Pattern

Every ten to fifteen years, a platform shift reshapes the technology landscape. Mainframes gave way to PCs. PCs yielded to the web. The web surrendered to smartphones. Each shift redirected innovation, investment, and company creation – establishing new gatekeepers and new models for value capture.

Generative AI follows this pattern. Within technology, a new class of capability becomes the central focus. Outside technology, industries must determine whether this new platform represents a tool for efficiency, a source of new revenue, or an existential threat to their business models.

History teaches three uncomfortable truths about platform shifts. First, dominance is temporary. Microsoft commanded the PC era but became largely irrelevant when smartphones took center stage. Second, early leaders often fail. AOL in internet portals, Netscape in browsers, Nokia and RIM in smartphones – each led their moment and lost their future. Third, initial clarity is rare. The early days of every new platform overflow with uncertainty, failed experiments, and misguided investments. The current debate about whether AI's ultimate form will be browsers, agents, or wearables mirrors the chaotic early years of the web and mobile internet.

The Capital Expenditure Surge

Fear of missing out has ignited an investment boom without precedent in technology history. Leaders have stated the calculus explicitly: the risk of under-investing vastly exceeds the risk of over-investing.

"The risk of under-investing is significantly greater than the risk of over-investing," declared Sundar Pichai in Alphabet's Q2 2024 earnings call (Pichai, 2024). Mark Zuckerberg offered a similar framing in Meta's Q3 2025 report: "The very worst case would be that we have just pre-built for a couple of years" (Zuckerberg, 2025).

The numbers tell the story. The Big Four – Microsoft, AWS, Google, and Meta – spent approximately $220 billion on capital expenditure in 2024 and are projected to reach $400 billion in 2025. This nearly doubled from initial January 2025 estimates. Total projected generative AI capital expenditure across the industry reaches $500 to $750 billion annually. For context, global telecommunications capex runs approximately $300 billion; oil and gas upstream investment stands at roughly $540 billion.

This surge strains every link in the supply chain. Nvidia and TSMC cannot expand capacity fast enough to meet demand. In the United States, data center construction is on trajectory to overtake office construction. Access to utility power has become, in industry parlance, an "extremely significant" constraint – AI is expected to add substantial demand to a grid growing at only two percent annually.

The financing structures have grown correspondingly complex and often circular. OpenAI, lacking its own cash flow, has made massive commitments by partnering with capital-rich entities. One particularly striking arrangement: Nvidia agreed to invest up to $100 billion in OpenAI, which placed massive orders for Nvidia chips while linking a $300 billion cloud deal with Oracle, itself a major Nvidia customer. As one analyst observed, OpenAI is essentially "buying Nvidia chips with Nvidia's cashflow."

Model Commoditization: The Vanishing Moat

While capital flows at historic rates, the models themselves are losing defensibility. They are becoming commodities.

Performance has converged. After OpenAI's initial lead, model quality across providers has reached rough parity. By late 2025, on general benchmarks like LMArena, leading models from OpenAI, Google, and Anthropic trade positions weekly. Open-source models and those from Chinese labs like DeepSeek have achieved state-of-the-art results.

Prices have collapsed. The Stanford AI Index observes that an LLM is "the most expensive and fastest-depreciating asset in tech history" (Stanford AI Index, 2025). Meta has accelerated this dynamic through its strategy of open-sourcing Llama models – turning a potential competitor's moat into basic infrastructure.

The basis of competition shifts. Microsoft's capital expenditure as a percentage of sales has surged, suggesting a strategic move from its traditional software moat – built on network effects – to one based on access to capital. The question becomes where value will accrue. The potential moats are no longer the best model but access to capital, proprietary vertical data, distribution and go-to-market strategy, product and UX design, or simply building what was once called a "normal" software company atop commodity AI.

The Deployment Gap: Promise Versus Absorption

A stark contrast separates industry investment from actual adoption. The pattern is one of broad but shallow engagement.

Consumer usage reveals high awareness paired with low frequency. Data consistently shows that a large percentage of users have tried tools like ChatGPT only once or twice. The ratio of daily active users to weekly active users appears low. Surveys from Pew, Deloitte, Bain, and others throughout 2024 and 2025 confirm that far more people use chatbots occasionally than integrate them into daily life (Pew Research, 2025; Deloitte, 2025; Bain, 2025). OpenAI's reporting of 800 million weekly active users for ChatGPT, with only five percent paying and a conspicuous absence of daily active user figures, has drawn scrutiny.

Enterprise adoption drowns in pilots. A June 2025 McKinsey survey showed that in most business functions, pilot and experimental use of AI "agents" far outpaced actual deployment (McKinsey, 2025). CIO surveys reveal a slow, deliberate adoption curve. A September 2025 Morgan Stanley survey found that while a quarter of CIOs had deployed a project, forty percent had no plans to do so until 2026 or later (Morgan Stanley, 2025). This mirrors the slow but steady adoption of public cloud, which after more than a decade still accounts for only roughly thirty percent of enterprise workloads.

Enterprises cite significant barriers to full deployment: security and privacy concerns, lack of expertise, error rates, unclear return on investment, and inadequate data readiness.

Emerging Use Cases: Where Absorption Begins

Despite the deployment gap, clear use cases are crystallizing. Technology absorption follows a predictable pattern across three stages.

The first stage is absorption – automating obvious existing use-cases and adding AI as a feature to current products. The second stage is innovation – creating new products and business models, often as startups unbundle incumbents. The third stage is disruption – redefining entire markets or questions, a less predictable outcome that emerges from the first two.

Current success concentrates in absorption, particularly in domains where errors are easily identified and corrected by humans in the loop.

In coding, AI assistants represent what industry observers have called "a new step change reduction in software creation costs." Y Combinator CEO Garry Tan noted that for approximately a quarter of their Winter 2025 startups, ninety-five percent of the code was written by AI (Tan, 2025). Companies like Cursor have achieved significant valuations by focusing on this space.

In marketing and advertising, generative AI drives scaled content creation. Brands including L'Oréal, Mondelez, and Unilever use AI to generate thousands of on-brand assets, create ad variations, and reduce production costs by thirty to fifty percent.

In customer support and automation, these represent common early use-cases – though they carry risks if not properly managed.

A wave of AI-native startups is emerging, with Y Combinator batches becoming dominated by AI companies. Their primary function is unbundling – extracting workflows locked inside incumbents like Excel, email, and Oracle to create targeted, AI-powered solutions.

The Jevons Paradox offers a useful lens here. As automation makes a resource cheaper – in this case, intern-level tasks – consumption increases dramatically. Rather than doing the same work with fewer people, organizations may do vastly more work with the same number of people. The "infinite interns" unlock new possibilities, not just cost savings.

The Questions That Define This Moment

Generative AI's platform shift unfolds under fundamental uncertainty. The impact will be enormous. Its final shape remains unknown.

The scaling question asks whether the current approach – more data, more compute – will continue yielding better results, or whether we are approaching a plateau of diminishing returns. Reports in late 2024 of leading labs struggling to achieve desired performance from next-generation models fuel active debate.

The product question asks what user experience will drive mass adoption. The initial chatbot interface has struggled to achieve deep, daily engagement. The future may lie in AI absorbed as features within existing software – like a better spell-check. Or AI may become the overarching UX that controls other applications. Or entirely new product categories may emerge. As Steve Jobs stated, "You've got to start with the experience and work backwards to the technology" (Jobs, 1997).

The value capture question asks where profit pools will form if models are commodities. Possibilities include the capital-intensive infrastructure layer (like Nvidia), distribution channels (like Apple embedding AI in iOS), proprietary data sets, or the application layer where startups build specific workflow solutions.

The disruption question asks about ultimate impact. AI may simply automate tasks and add features. Or it may fundamentally unbundle and re-aggregate entire industries. AI-driven recommendation engines that move beyond correlation to genuine understanding of user intent could reshape the trillion-dollar global advertising market and the nature of online retail and media discovery.

The Destination: Invisible Infrastructure

History teaches that when automation works, it disappears. It becomes infrastructure. Automatic elevators eliminated a job category but became an invisible, essential component of modern architecture. No one "uses" an automatic elevator consciously anymore. They simply go up.

The long-term trajectory points toward a similar outcome. "AI" will cease to be a buzzword. It will become what it was always meant to be: software. Invisible. Essential. Unremarkable.

Between now and then lies the work of a generation – sorting the durable from the disposable, the transformative from the transient, the platforms from the features. The capital is committed. The technology advances. The questions remain open.

The answers will reveal themselves in construction, not contemplation. We build our way to clarity.

References

Bain & Company. (2025). Consumer AI Adoption Survey.

Deloitte. (2025). State of AI in the Enterprise.

Evans, B. (2024, Autumn). AI Eats the World. Retrieved from https://www.ben-evans.com/presentations

Evans, B. (2025, Spring). AI Eats the World. Retrieved from https://www.ben-evans.com/presentations

Evans, B. (2025, Autumn). AI Eats the World. Retrieved from https://www.ben-evans.com/presentations

Jobs, S. (1997). Apple Worldwide Developers Conference Keynote.

McKinsey & Company. (2025, June). The State of AI in 2025.

MIT Sloan Management Review. (2025, August). Why Most AI Pilots Fail.

Morgan Stanley. (2025, September). CIO Survey on AI Adoption.

Pew Research Center. (2025). Americans and AI Tools.

Pichai, S. (2024). Alphabet Q2 2024 Earnings Call.

Stanford Institute for Human-Centered AI. (2025). AI Index Report.

Tan, G. (2025). Y Combinator Winter 2025 Demo Day Remarks.

Zuckerberg, M. (2025). Meta Q3 2025 Earnings Call.