Name: Why Top Founders Are Racing Into AI Infrastructure
Uploaded: 2026-08-28T13:11:59+00:00
Duration: 3239 s

Speakers

Ben Horowitz — Co-Founder, Andreessen Horowitz (a16z)

Martin Casado — a16z General Partner

Raghu Raghuram — Operating Partner, Andreessen Horowitz

Summary

Ben Horowitz, Martin Casado, Raghu Raghuram, and Erik Torenberg discuss a16z’s new Machine Age Fund and the AI infrastructure buildout. They argue AI demand is outpacing supply across chips, memory, networking, power, cooling, and data centers, creating a broad investment opportunity. The group also covers custom ASICs, data center robotics, and the need for new infrastructure software.

a16z launches the Machine Age Fund focused on AI infrastructure.
Model progress is no longer the main bottleneck; infrastructure south of the model is.
Hyperscaler capex and sold-out supply indicate demand is not a hype cycle.
Memory, power, cooling, networking, and data center capacity are constrained.
Custom per-model ASICs may become economically rational.
Data centers require power, liquid cooling, physical rebuilds, and automation.
Embodied AI and agents could multiply token and compute demand.
U.S. policy and permitting affect where data centers get built.

Ideas

Ben Horowitz Co-Founder, Andreessen Horowitz (a16z) 1:12

AI requires massive new infrastructure buildout

AI is the most important new technology and requires an entirely new infrastructure stack—new chips, system software, power, cooling, networking, and even materials. The buildout is widespread, supply is tight, and the new Machine Age Fund is dedicated to this infrastructure shift.

Raghu Raghuram Operating Partner, Andreessen Horowitz 2:19

Semiconductor supply chain is bottlenecked

The bottleneck has moved south of the model into the semiconductor supply chain: chips, memory, networking, and power chips are constrained, components are booked out to 2028, and a leading memory vendor says current demand would take three years of capacity to supply. Improving tokens per dollar, watt, or rack requires new full-system innovations.

Martin Casado a16z General Partner 14:05

AI compute demand persists for decades

AI scaling has shifted from an engineering problem to a resource problem. Reasoning, agents, computer use, and embodied AI multiply token consumption, and there is no natural engineering regulator like before, so compute and token demand should persist for decades.

Raghu Raghuram Operating Partner, Andreessen Horowitz 19:21

Robots and embodied AI grow

Embodied AI and robots are another major source of compute demand. Large cloud data center operators are already experimenting with robots to assemble and install servers, and adoption should increase as AI evolves.

Martin Casado a16z General Partner 26:44

Custom ASICs make sense for models

Training a frontier model costs $3–5 billion, and inference must pay that back several times over. A 20% inference-efficiency gain can be worth about $2 billion, enough to justify a custom ASIC, and fixed model weights make per-model custom silicon economically plausible.

Martin Casado a16z General Partner 28:24

Data center power is bottlenecked

Rack power is rising from roughly 5–10 kilowatts to 100–150 kilowatts, making AC power obsolete and forcing DC power, on-site generation, and grid upgrades. Permitting, turbine and transformer shortages, and a projected 44 GW data-center power need versus 25 GW of grid additions make power a major bottleneck and investment opportunity.

Martin Casado a16z General Partner 29:20

AI racks require liquid cooling

Cooling is moving from air to liquid as a requirement for state-of-the-art AI data centers. Liquid cooling alone may not be enough, as political and environmental constraints push demand toward eco-friendly and water-efficient cooling solutions.

Martin Casado a16z General Partner 30:37

Data centers need physical rebuild

Dense AI racks require stronger floors, thicker walls for noise mitigation, and large amounts of reinforced concrete. Many existing data center buildings are becoming obsolete, driving demand for new physical data center infrastructure and materials.

Raghu Raghuram Operating Partner, Andreessen Horowitz 45:21

AI infrastructure needs management software

As AI infrastructure scales, a software layer is needed to automate and manage fleets of chips, systems, and data center resources. This infrastructure software layer is another important investment area.

Why Top Founders Are Racing Into AI Infrastructure

Summary

Ideas

Up Next