The AI Selloff Doesn't Match the Data | Top AI Investor Explains

Summary

Gavin Baker dissects the AI and semiconductor drawdown of July 2026, arguing that public market panic sharply contradicts accelerating on-the-ground fundamentals. He explains why GPU spot prices are still surging, how open source models actually boost infrastructure demand, and why hyperscaler operating cash flow acceleration will fund the buildout without a debt crisis. The episode also covers NVIDIA’s new revenue model, the game theory of memory LTAs, SpaceX’s underappreciated compute business, and the looming threat of AI regulation.

• AI names fell 40–60% in July 2026, but every quantitative demand metric (GPU prices, DRAM spot, token growth) is accelerating.
• Open source model adoption takes margin from frontier labs but shifts token share to cheaper tokens, which still consume the same compute—bullish for AI infrastructure.
• Hyperscale operating cash flow accelerated from 28% to 35% last quarter, with contracted compute set to reprice much higher, making the buildout self-financing.
• Memory is the single most important lever for AI throughput; LTAs lock in hyperscalers and give memory suppliers durable pricing power.
• NVIDIA’s credit wrapper/revenue share model effectively creates a cloud royalty stream, reinforcing its competitive position while the stock trades at a 10-year low P/E.
• SpaceX’s public market debut is misunderstood: its compute capacity and ability to bring on power are far ahead of consensus, with significant upside.
• China’s DUV progress is a long-term risk for semicap equipment but an extreme near-term overreaction likely created a watching setup for ASML.
• Regulatory pushback—data center moratoriums and misinformation about water/power drag—is viewed as the biggest potential risk to the AI capex cycle.