AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Is There an AI Bubble? Baker & George on the AI Capex Cycle

· The a16z Show

Gavin Baker (Atreides Management) and David George (a16z) argue that AI is not in a bubble, contrasting today's buildout with the 2000 telecom bust. The key distinction: 97% of fiber laid in 2000 was dark, whereas every GPU deployed today is fully utilized, with major hyperscaler spenders seeing roughly a 10-point increase in ROIC since ramping CapEx. They note that the largest spenders collectively generate around $300B in annual free cash flow and hold about $500B in cash, creating a large buffer against the roughly $40–50B cost of lighting up a gigawatt of Nvidia capacity. On competitive dynamics, the discussion frames the real contest as Nvidia versus Google's TPU (with Broadcom and AMD as supporting players), argues that Nvidia's "round-tripping" deals are rational competitive responses to Google's TPU ecosystem, and predicts that most custom ASIC programs will fail over the next three years. They expect frontier AI labs to have structurally lower gross margins than traditional SaaS due to scaling laws and test-time compute, and recommend that incumbents lean in even at the expense of near-term margins, drawing parallels to the cloud transition for Microsoft and Adobe. At the application layer, they see opportunity in fragmented SMB SaaS, outcome-based pricing (e.g., customer support), and consumer AI that closes the affiliate loop. Reasoning models, they argue, have fundamentally changed frontier lab economics by creating a user-base flywheel similar to early consumer internet. They close with bullish short timelines on AGI and view humanoids as the likely winning robotics form factor, pitting Tesla against Chinese competitors.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.