Why Top Founders Are Racing Into AI Infrastructure

Video summary

a16z’s Machine Age Fund targets AI infrastructure as GPU supply is booked to 2028 and memory demand needs three years of capacity.

Ben Horowitz, Martin Casado, Raghu Raghuram, and Erik Torenberg introduce a16z’s Machine Age Fund, arguing that AI’s bottleneck is shifting from models to the infrastructure beneath them: chips, memory, networking, power, cooling, and data centers. Hyperscaler capital spending is expected to reach $1 trillion next year, GPU supply is booked through 2028, and a leading memory vendor says current demand alone would take three years of capacity to meet. As reasoning and agents consume far more tokens, the speakers say compute demand has no natural engineering limit and could keep growing for decades. They discuss agents such as GrokBot becoming employee-like tools, while warning that integration brings cost, reliability, and security challenges. Building AI-ready infrastructure will require redesigned systems, liquid cooling, new power solutions, and skilled workers; by 2028, data centers may need 44 gigawatts of additional power against 25 gigawatts of expected grid additions. The fund will back founders rethinking the full computing stack, with the ambition of helping America lead in infrastructure.

Chapters

1. 0:00Introducing the Machine Age Fund: AI Requires New Chips, Power, and Copper Alternatives
2. 2:00Why Founder Interest in Hardware Just 4x'd: Top-Founder Hardware Deals Rose from 5% to Over 20%
3. 4:00How Do We Know Demand Isn't a Hype Cycle? Hyperscaler CapEx Nears $1 Trillion and GPUs Are Booked to 2028
4. 7:00Sold Out to 2028: GPUs Are Pre-Sold, Resell at 4x, and Memory Demand Takes 3 Years of Capacity
5. 10:00What's Actually Bottlenecked Right Now: 3–4-Year Chip Cycles, 4–5-Year Data Centers, and Physical Limits
6. 14:00Tokens, Scaling & Why There's No Natural Regulator: Inference and $3 Billion Compute Clusters
7. 19:00Agents as a New Kind of Employee: GrokBot Handles Computer Tasks, but Teams Must Manage Risks
8. 25:00What “AI-Designed” Infrastructure Actually Looks Like: Jointly Optimize the Stack and Consider Model-Specific ASICs
9. 28:00Rack Power, Liquid Cooling & Data Center Redesign: 50–100 kW Racks and DC Power
10. 34:0044 Gigawatts by 2028: Grid Shortfalls, Equipment Delays, and Projects Moving Abroad
11. 38:00Why “Machine Age” Fits: Machine Intelligence and Hardware-Driven Breakthroughs
12. 40:00Why Nvidia Won’t Take Everything: Fragmenting Markets and New Hardware Startups
13. 48:00The Founder Profile: Hardware Experience, SpaceX Alumni, and Systems Teams