Why AI Demand Is Outrunning Compute Supply

Video summary

Gavin Baker argues AI compute demand will outpace supply, with sub-one-year infrastructure paybacks and orbital data centers approaching.

Gavin Baker tells a16z’s David George he cannot find a worsening business metric: OpenAI, open-source models and AI applications are accelerating, and the boom need not have a single winner. He argues compute investments can pay back in under a year, while the number of heavy AI users may be below 10 million against 1.5 billion knowledge workers; one company’s token spending rose 100x from March to August. They discuss bubble risks but see a greater danger in underbuilding capacity through 2028, with data centers bringing tax revenue and industrial jobs to American towns. Baker also makes the case for orbital compute: Starship reusability could shift launch economics, and a SpaceX-NVIDIA Rubin rack could launch as early as late 2027. They close on hybrid AI models, NVIDIA’s financing and supply-chain advantages, and why companies should plug into its ecosystem rather than pick a head-on fight.

Chapters

1. 0:00Intro: Elon and Jensen as defining figures of the AI era
2. 1:06Finding the Bear Case: AI Accelerates and Multiple Players Could Win
3. 8:06How's This All Gonna Go Wrong? Frontier Models, Open Source, and NVIDIA Can All Win
4. 9:09Will Labs Reinvest Profits Into Training? Nebius Payback Could Be 9–10 Months
5. 14:44The Demand Side: Fewer Than 30 Million Heavy Users and Token Spend Rising 100×
6. 19:16From Reactive Coding to Autonomous Agents: Action-Taking Bots and the Risk of Overbuilding
7. 21:34From Reactive Coding to Autonomous Agents: Compute Constraints and Data Center Benefits
8. 26:12From Reactive Coding to Autonomous Agents: Underbuilt Compute Could Push Token Costs Up 10×
9. 30:18What Happens If There's a Massive Supply Shortage? Compute Inequality and Starlink Access
10. 34:21Orbital Data Centers: Airplane-Sized Racks and Starship Launch Costs
11. 40:00Starlink’s Nearly $2 Trillion Market and SpaceX’s Compute Upside
12. 44:04The Most Futuristic SpaceX Idea: Asteroid Mining and a Robot-Built Mars Presence
13. 48:30The Most Futuristic SpaceX Idea: Asteroid Mining Chapter Turns to Model Ensembles and Fireworks Nexus
14. 54:01Who Becomes the Abstraction Layer of Intelligence? Cursor, Microsoft, and Enterprise Complexity
15. 57:57Harvey, Cursor & Vertical AI Winners: Legal Agents and the $500 Million Build
16. 1:00:26Jensen, Nvidia & the Central Bank of AI: $50B Data Centers Need $15B in Equity
17. 1:03:56Jensen, Nvidia & the Central Bank of AI: Supply-Chain Scale and the Risks of Chip Failure
18. 1:08:47Jensen, Nvidia & the Central Bank of AI: Jalapeno’s ASIC and Nvidia’s Nine-Chip System
19. 1:12:16How Chip Deal Structures Reveal True Customer Preference: H100 Shortages, RVGs, and Warrants