AI Eats the World: Navigating the Next Platform Shift

The history of technology is defined by platform shifts—fundamental resets that occur every 10 to 15 years. From mainframes to PCs, the web to smartphones, each shift reshapes not only the tech industry but the entire economic landscape. According to Benedict Evans, we are currently in the midst of the next great shift: Generative AI.

This transition is characterized by an unprecedented explosion in capital expenditure and a frantic scramble for infrastructure. However, as the initial hype settles, a critical question emerges: will the companies building the "brains" (the models) capture the value, or will the value move up the stack to the applications and workflows that actually solve human problems?

The Capex Explosion and the Infrastructure War

We are witnessing a capital deployment cycle that rivals the largest industrial investments in history. The "Big Four" (Microsoft, Alphabet, AWS, and Meta) are projected to spend roughly $700 billion in 2026 alone—a figure that dwarfs global telecom spending and approaches the scale of the oil and gas industry.

This surge is driven by a prevailing fear: the risk of under-investing is far greater than the risk of over-investing. As Sundar Pichai and Mark Zuckerberg have suggested, the worst-case scenario for over-investing is simply having "pre-built" capacity for a few years. This has led to a "pig in a python" effect, where demand for GPUs, memory, and power capacity far outstrips supply, creating bottlenecks across the entire global supply chain.

The Commodity Trap: Models vs. Infrastructure

One of the most provocative theses in Evans' analysis is the potential commoditization of Large Language Models (LLMs). He draws a parallel to the mobile networks of 2010: while telcos spent trillions building the infrastructure (3G, 4G, 5G), the actual value capture happened in the apps (Uber, Instagram, TikTok) built on top of that infrastructure.

The Case for Commoditization

• Lack of Network Effects: Unlike social networks or marketplaces, using a slightly better model doesn't inherently make the model better for other users.
• Convergence: Frontier models from OpenAI, Anthropic, Google, and Meta are increasingly similar in aggregate benchmark scores.
• Utility Pricing: Sam Altman has envisioned intelligence as a utility, like electricity or water, bought on a meter.

However, this view is contested. Some observers argue that frontier models are more like TSMC in the semiconductor world—a winner-take-most scenario where the smartest model attracts the most users, generates the most revenue, and funds the next leap in intelligence, creating a powerful flywheel effect that prevents commoditization.

From "Chat" to Deployment

While ChatGPT has hundreds of millions of weekly users, the depth of engagement remains shallow. For the vast majority of users, AI is an experimental tool rather than a daily essential. The "capacity gap" exists not in the technology's ability, but in its deployment.

The UX Problem

Evans argues that "Chat is a terrible UX." For AI to move from a novelty to a utility, it must move beyond the blank screen. The future of AI software isn't just a chatbot; it's the integration of AI into specific workflows, GUIs, and proprietary data sets.

The Coding Revolution

Coding is the first area seeing a generational shift. We are seeing instances where one or two people can build in a week what previously took dozens of people months. But as the cost of writing code drops toward zero, the hard part shifts from how to build to what to build. The value moves from the act of coding to the implicit knowledge, taste, and strategic vision required to define the product.

Automation: Task vs. Job

To understand how AI will change the world, we must distinguish between a task and a job.

• The Task: A specific action (e.g., writing a spec, building a spreadsheet, listening to a customer call).
• The Job: The overarching role that requires judgment, taste, and human relationship.

History shows that when a task becomes "free" through automation, the job doesn't always disappear; it evolves. When grocery barcodes were introduced in 1974, it didn't kill the grocery store; it allowed stores to stock five times more products (SKUs) because inventory management became cheap.

Similarly, AI provides "infinite interns." The question for businesses is not "will AI replace my employees?" but "what becomes possible now that the boring parts of the job are free?"

Conclusion: Radical Uncertainty

As we look toward the late 2020s, the overarching theme is radical uncertainty. Every platform shift feels "totally different" while it is happening, yet they all rhyme. We often forget how many ideas failed during the early internet and mobile eras—the AOLs and Netscapes of the world.

Whether AI "eats the world" or merely "nibbles" at it depends on our ability to move beyond the raw technology and invent new questions. The real winners of the AI era will likely not be those who build the biggest models, but those who use that intelligence to redefine the boundaries of what is possible in software and business.