Stratechery · Ben Thompson · March 2026 ★★★★★

Agents Over Bubbles

Why we're NOT in an AI bubble — and why that matters more than anyone realizes

AI Investment Strategy Agents

The Three LLM Paradigms

Ben Thompson identifies three distinct inflection points in AI development that explain why massive capex spending is justified, not speculative:

Paradigm 1 — ChatGPT (Nov 2022)

LLMs were impressive but had two fatal flaws: they hallucinated and required users to know exactly what to ask for. Useful for experts, useless for most.

Paradigm 2 — o1 (Sep 2024)

Reasoning models self-evaluate. Instead of committing to a wrong answer immediately, they iterate internally. The brilliance of ChatGPT was making LLMs readable; the brilliance of o1 was making them reliable.

Paradigm 3 — Opus 4.5 + Agentic Code (Nov 2025)

Claude Code with Opus 4.5 and GPT-5.2-Codex can accomplish multi-hour tasks correctly. The critical difference: agents abstract the human away from the model entirely. A coding agent generates code, checks if it works, and retries — all without human involvement.

Core insight: "In paradigm one, an LLM would generate code. In paradigm two, an LLM would think about the code it was generating and iterate. In this paradigm, an agent directs a model to generate code, then checks to see if the code actually works — and if it doesn't, tries again — all without the user needing to be involved."

Why Agents Narrow the Agency Gap

Previously, AI required human agency to be useful — you had to know what to ask, verify the output, and manage the process. Agents change this equation fundamentally.

One person can control multiple agents simultaneously
The number of humans with "agency" needed for massive economic impact is dramatically smaller
AI still needs agency — just from far fewer people
Result: a step-function increase in usefulness without a corresponding step-function in human adoption

"How many industries are not media, in that they still need a team to implement the vision of one person? How many apps haven't been built, not because one person can't imagine them, but because they haven't had the resources or team to actually ship them?"

The Enterprise Economic Imperative

Thompson's most provocative claim: companies will use AI not just to save costs, but to shrink their workforce permanently — because AI makes smaller teams with agents more effective than large human organizations.

Companies become bloated because that's been the only way to scale
AI gives both the excuse and the capability to "rightsize" to much smaller teams
The most forward-looking companies will cut more rather than less, betting that remaining humans must rebuild scale with agents
Dramatically smaller competitors built with AI from the beginning will challenge incumbents with both lower costs and more capabilities that structurally increase over time

Why we're not in a bubble (three reasons):

All LLM weaknesses are being addressed by exponential compute increases

The number of people needed to wield AI effectively is decreasing

Economic returns from agents impact both the bottom line AND the top line

On Apple's "Brilliant Move"

Addressing Horace Dediu's argument that Apple won by not building its own model: Thompson counters that in the agentic paradigm, this reasoning breaks down. When AI drives enterprise productivity, owning the customer's workflow (like Apple does) matters less than building the most capable agents. "The question is who has the best agent, not who has the best device."

Key Quote

"You can see them coming and yet still be amazed when they arrive — and, as one must say with everything related to AI, in a form that is the worst they will ever be."