Ben Thompson: we are not in an AI bubble — agents make the CapEx case ironclad

Summary

Ben Thompson argues that the AI industry is not in a bubble. Writing on Stratechery, he stakes this position on the economics of agentic systems and identifies three inflection points in AI adoption—language models, reasoning models, and agents—each exponentially increasing compute demand in ways that justify hyperscalers' massive capital expenditure.

Language models like ChatGPT required heavy training compute but efficient inference. Users received model output directly. Reasoning models like o1 and Claude Opus flipped the equation by generating longer chains of thought, which multiplied token usage through longer outputs and increased adoption. Agents tip the scale entirely.

Agents drive demand through three mechanisms. They make multiple calls to reasoning models per task. They themselves consume more compute, often better served by CPUs and GPUs. And they create a step function increase in usefulness that drives higher utilization than chatbots. Thompson notes that far more people use chatbots than agents today, but most chatbot users underutilize the technology because it requires intentional agency. Users must actively choose to deploy AI. Agents are autonomous and persistent, removing that friction entirely.

Thompson applies this logic to enterprise software. Companies demonstrate willingness to pay for tools that make employees more productive. What excites executives is not job elimination but doing more work with the same headcount—growing faster while cutting costs. Agents will tilt toward pure acceleration. In large organizations, a small number of people drive real value, but their impact flows through costly coordination apparatus with humans who both accelerate and retard progress. Agents replace that friction-laden machinery with tireless, continuous executors. The net effect is not primarily headcount reduction but multiplication of output from high-leverage individuals.

Thompson's conclusion rests on three observations. Exponential compute increases are addressing every language model weakness. The number of people who need to wield AI for demand to skyrocket is shrinking—agents work autonomously without requiring skilled operators. Economic returns from agents flow both to the bottom line through cost savings and the top line through revenue growth. That dual impact explains why every hyperscaler reports compute demand exceeding supply and why all are announcing CapEx plans that exceed expectations.

Thompson inverts the traditional bubble narrative. Most bubble warnings assume speculative excess requires capitulation to trigger collapse. But capitulation itself is evidence of a bubble because it means everyone agrees the bubble exists and has already priced in the downside. His argument is that March 2026 shows the opposite. Hyperscalers are unconstrained in their conviction about capacity needs because the demand signal is real and growing.