Episode Summary
Executive Summary: Patrick O'Shaughnessy interviews Dylan Patel of SemiAnalysis on the physical reality behind AI: compute, data centers, power, talent, and geopolitical competition. Patel argues AI is still early, demand is compounding, and value is shifting toward the layers that control compute, distribution, and infrastructure.
Main Topics: OpenAI, NVIDIA, and Oracle deal dynamics (Priority: 10/5): Large AI deals are really about funding compute upfront and securing future capacity. Tokenomics and inference demand (Priority: 9/5): AI economics depend on token cost, model quality, latency, and user adoption. Pre-training, reinforcement learning, and environments (Priority: 10/5): The next gains come from RL environments and synthetic data, not just web-scale text. Reasoning and memory (Priority: 8/5): Longer reasoning and better memory handling will make models more agentic and useful. Power, grids, and data-center buildout (Priority: 9/5): AI is forcing a new industrial wave in power, transformers, electricians, and generators. U.S.-China AI competition (Priority: 9/5): AI is framed as strategic necessity for the U.S. and a long-game industrial policy battle for China. Software business model disruption (Priority: 8/5): AI raises COGS and lowers build costs, pressuring classic SaaS economics.
Key Arguments: Compute comes first: models need clusters before revenue; capacity is now the bottleneck. OpenAI must secure allies because trillion-dollar rivals can outspend it on compute. Model quality still scales with compute, but serving cost and latency constrain deployment. Token demand is growing fast; economics hinge on how cheaply intelligence can be served. RL environments are early and will drive much of the next progress beyond pre-training. Power infrastructure is the hidden constraint: AI data centers are reshaping utilities and supply chains. Classic SaaS gets squeezed as AI lowers software build cost while raising inference COGS. The U.S. needs AI growth to preserve hegemony; China is building for self-sufficiency and speed.
Data Points: AI expense review automation: 85% - Ramp uses AI to automate expense reviews. Expense review accuracy: 99% - Ramp claims its AI expense review accuracy. Company savings from Ramp: 5% - Ramp says it saves companies about 5%. OpenAI users: 800 million - Patel cites OpenAI's user base while warning revenue and compute still matter. Data-center capacity pricing: $10 to $15 billion per gigawatt - Annual rental price for one gigawatt of capacity in the discussion. Five-year capacity deal cost: $50 to $75 billion - Estimated outlay for a five-year gigawatt deal. OpenAI-Oracle deal: $300 billion - Patel references a reported OpenAI contract with Oracle. NVIDIA equity investment into OpenAI: $100 billion - Described as part of the strategic compute-financing arrangement. One gigawatt initial tranche: $10 billion - The first chunk of the NVIDIA/OpenAI press release deal. Data-center capex captured by NVIDIA: $35 billion - Patel estimates NVIDIA captures most of the capex from a $50 billion build. Gross margin of NVIDIA: 75% - Used in Patel's simplified deal math. Anthropic revenue ramp: from a billion or less to seven to eight - Patel describes Anthropic's rapid revenue growth. Token demand growth: doubling every two months - Mentioned as a rough external stat on inference demand. Etsy traffic from GPT: more than 10% - Patel cites GPT-driven traffic as proof of shopping intent. U.S. power share used by data centers: 3-4% - He says data centers are still a small share of U.S. electricity. Texas/PJM notice window: 24 hours or 72 hours - Grids may cut large loads with advance notice. Electrician wage trend: doubled - Mobile electricians for data-center work are in high demand. Blackwell hourly cost assumption: $2 - Patel uses a simplified cost per hour for Blackwell GPUs. CoreWeave/contractor deal example: $19 billion - He references a Microsoft-backed deal with Nebius. China capital into semis: at least $400-500 billion - Patel estimates China's semiconductor ecosystem investment.
Pivotal Quotes: "The compute precedes the buildup of business." — Dylan Patel: Explaining why AI companies must secure infrastructure before monetizing products. "If you don't move fast enough... they will get beaten." — Dylan Patel: On OpenAI's need to keep pace with mega-cap tech rivals in compute. "The reason America's rich is because we've exported all the labor we've kept all the value." — Jensen Huang (quoted by Dylan Patel): Patel uses this to explain why high-value layers capture the profits.
Implications: The unresolved question is where AI's real profit pool lands as infrastructure, models, and apps all fight for margin; listeners should watch power, memory, and RL as the next bottlenecks.
From the Transcript
Money glitch here. No, no, no. That's not actually what's happening. If they pay each other, then their market caps all keep going up. What's really happening is OpenAI has an insatiable demand for compute. The compute precedes the buildup of business. You have to have the cluster before you can run models on it for inference. You have to have the cluster to train the model that's good enough that it unlocks new use cases, which then can be adopted, and there's an adoption curve there for any new use case. So you have to have all these things like sequence. Given this is a game of the richest people in the world, or rather the biggest tech giants in the world, right? It's it's Zuck, it's all the biggest people in the world. It's Elon, right? Google, Larry, and Sergey is like constantly in the business now again. There's very much a risk of OpenAI being too small to matter, which is crazy to say because they've got 800 million users, but where's the revenue? Where's the compute? They could easily get swamped in terms of how much compute they have if they don't move fast enough. And if they don't have the most compute or among the most compute,
But then there's all these like dynamics. Like if you don't push them a little bit away from having the chips because they have all this talent, probably half the AI engineers in the world are Chinese, whether they're immigrated to America or not. You look at the list of meta superintelligence, like it's like they push like 80% Chinese people, right? And like how many of them are from China versus being ABCs, right? American-born Chinese. Like you could look it up, but it's like push China too hard. They have the talent. They could go crazy. They could, if we no longer have Taiwan, actually China could build a way bigger cluster than us. And of course Compute is all that matters. They could do all of these things and they own the means of production for everything. There's this like challenging aspect of geopolitical risk. That's why people don't want to invest in TSMC, but it's like almost like you can't invest in Amazon or Apple or Google or like Microsoft if you have geopolitical risk. If you believe Taiwan has risk, and so it's like, yo, invest in TSMC. I know a lot of people with PMs are like, oh, you can't invest in TSMC because geopolitical risk. And it's like, no, dude, you can't invest in fucking Apple. Who is?
About Invest Like the Best with Patrick O'Shaughnessy
Conversations with the best investors and business builders in the world.
View all episodes from Invest Like the Best with Patrick O'Shaughnessy