Dylan Patel - Inside the Trillion-Dollar AI Buildout - [Invest Like the Best, EP.442]

What It Covers

Dylan Patel maps the trillion-dollar AI infrastructure buildout, explaining OpenAI's strategic partnerships with NVIDIA and Oracle, the economics of gigawatt-scale data centers costing $50 billion each, reinforcement learning's early innings, and why America's competitive position depends on AI success.

Key Questions Answered

• •Infrastructure Economics: One gigawatt of data center capacity costs $50-75 billion over five years in rental payments, with $10-15 billion annual operating costs. NVIDIA captures roughly $35 billion of initial $50 billion CapEx per gigawatt, maintaining 75% gross margins while effectively lowering prices through equity investments in customers like OpenAI.
• •Scaling Law Reality: Model improvement follows log-log scaling where 10x more compute yields one tier of capability increase. This resembles progression from six-year-old to thirteen-year-old intelligence levels. Pre-training on text data reaches late innings, but multimodal pre-training and reinforcement learning remain in second inning, with vast unexplored territory in environment-based learning.
• •Tokenomics Trade-offs: Companies face critical decisions between serving larger, slower models with higher intelligence versus smaller, faster models with broader adoption. OpenAI chose GPT-5 at similar size to GPT-4 rather than scaling up because user experience degrades with latency, limiting revenue despite superior capabilities in larger models like Claude Opus.
• •Reinforcement Learning Paradigm: Post-training through synthetic environments enables models to learn tasks absent from internet data, like spreadsheet manipulation or physical object recognition. This approach generates training data through iterative trial-and-error in simulated environments, teaching models to reason through problems rather than memorize answers, fundamentally changing capability development.
• •Value Capture Dynamics: Gross profit currently flows to hardware layer (NVIDIA, Broadcom) while application companies like Cursor send most revenue to model providers (Anthropic), who reinvest in training compute. Power shifts as application companies accumulate proprietary user interaction data and can train specialized models, creating frenemy relationships throughout the stack.