The $600B Infrastructure Gap

Sequoia's thesis: the AI infrastructure buildout is a $600B opportunity. The counter-thesis: they're selling picks and shovels in a gold rush where 90% of miners will die of thirst.

H100 secondary market tells the story. A card that cost $30K at peak is trading at $15-18K on the gray market as of August. Datacenters that reserved clusters are subletting rack space. The marginal start-up GPU hour is now cheaper than spot compute on Lambda Labs.

The AI bubble isn't bursting - it's settling. The $600B figure was always a fiction, a VC's back-of-envelope that assumed 100% YoY demand growth for five more years. But inference is already commoditizing. Training is consolidating to four players (OpenAI, Google, Anthropic, Meta). Everyone else is renting.

What's actually valuable: applications. Not models. The companies winning right now are the ones selling workflow outcomes, not GPU cycles. The infrastructure bet was a bet on scarcity. Scarcity is ending.

The H100 dumpers were the canary. The noise floor is rising.