Nvidia Sold $194 Billion In Chips. The AI Bubble Story Is A Lie

TL;DR

• 1

  The key question is not whether AI is a bubble: Nate argues the useful split is between speculative froth in stocks and private valuations, versus real physical demand already showing up in revenue, chip sales, and capacity shortages.

• 2

  OpenAI, Anthropic, and Nvidia are showing paid demand at extreme scale: OpenAI reportedly went from about $2 billion in 2023 to more than $20 billion in 2025, Anthropic grew even faster from a smaller base, and Nvidia's fiscal 2026 data center revenue hit about $193.7 billion.

• 3

  Inference, not just training, explains the giant capex wave: Agentic workloads loop through prompts, tools, files, retries, and verification steps, turning what used to look like a chat product into a token-hungry production system.

• 4

  The market is now shifting from narrative to sorting: Jones says the next phase will separate companies with real paid usage from those with AI branding, and separate bottlenecks and workflow owners from commodity suppliers.

• 5

  Enterprise ROI looks messy because AI is uneven across workflows: Coding agents, legal review, and customer support can justify expensive inference, while shallow website chatbots and weak pilots often cannot.

• 6

  History says a real platform shift can still destroy investors: He compares AI to railroads, fiber, cloud, and the dot-com era, where the technology was real but many adjacent companies were overpriced, mistimed, or structurally weak.