Inside the Trillion-Dollar AI Buildout | Dylan Patel Interview

Dylan Patel argues the OpenAI-Nvidia deal is a high-stakes equity-for-compute swap, and that traditional SaaS is structurally broken by AI economics.

• Nvidia’s OpenAI deal: $10B equity investment per gigawatt returns ~$30B gross profit to Nvidia via GPU sales at 75% margin — not round-tripping.
• Oracle signed a $300B contract with OpenAI, whose annualized revenue is roughly $15-20B — the math only works if OpenAI dominates consumer AI at scale.
• Anthropic revenue grew from under $1B to $7-8B faster than any company in history; nearly all of it is code-related via Claude, Cursor, Copilot, Windsurf.
• Token demand doubles every two months but compute hardware is not; algorithmic efficiency gains (2,000x cost drop for GPT-3-class inference) are bridging the gap.
• GPT-5 was kept the same size as 4o deliberately — a bigger model can’t be served at acceptable latency, stalling the adoption curve.
• Post-training via RL and environments is in the first inning; pre-training on text is early-to-mid innings, with video/audio/multimodal barely started.
• Traditional SaaS faces a structural squeeze: AI inflates COGS, lowers barriers for competitors to build rather than buy, and customer acquisition cost stays high — escape velocity becomes harder.
• China has invested $450B+ into its semiconductor ecosystem over a decade and is prioritizing insular supply chains over peak performance; Dylan believes the US literally needs AI to avoid economic decline.

2025-09-30 · Watch on YouTube