a16z has raised $1.1bn to invest in the physical layer of AI

Andreessen Horowitz has closed a $1.1bn fund that will invest exclusively in hardware. The Machine Age Fund, announced on Friday, covers the physical layer artificial intelligence runs on, from chips to the buildings that house them.

The mandate takes in chips, memory, networking and storage, and then keeps going. Complete systems qualify too, defined broadly enough to stretch from data centres to robotics to AI appliances for the home.

Five partners put their names to the launch: Ben Horowitz, Martin Casado, Raghu Raghuram, David Ulevitch and David George. That is a lot of senior attention for a single vehicle, and a fair indication of how central the firm considers the thesis.

The case for it rests on physics rather than market sizing. Compute density has risen 28-fold between Nvidia’s H100 generation and its Rubin racks, and a rack that once drew 5 to 10 kilowatts now draws between 100 and 250.

Within three years, that is expected to reach a megawatt per rack. Individual data centres are already moving from tens of megawatts to hundreds, with some campuses approaching gigawatt scale, the trend now sending builders to their banks for financing on a different order.

Every layer of that stack is running into the limits of what the supply chain can produce, and in places into the limits of physics and computer science. The bottleneck, on this reading, has moved out of the models and into the world.

a16z’s own deal flow has shifted with it. Hardware went from a marginal share of the deals the firm sees to more than 20% of them.

Hardware has been the harder sell in venture for two decades, on the grounds that it takes longer, costs more, and scales worse than software. A fifth of a firm’s pipeline is the point at which that objection stops being decisive.

The portfolio the fund builds on spans a good deal more than semiconductors. Unconventional AI, Nexthop, Volta, Atoms and Mind Robotics sit alongside Skydio, SpaceX, Anduril and Waymo.

That list is what defines how broadly systems is being read here. Drones, launch vehicles, defence hardware and autonomous vehicles all qualify, which puts the fund some distance from a conventional deeptech mandate.

On the demand side, the argument turns on token intensity. Both the volume of AI work and the compute each unit of it consumes are rising by orders of magnitude, against growth the firm puts in triple digits.

The specific areas of interest are narrower and rather more revealing. Memory and interconnect improvements, power-efficient edge devices, and the cooling, materials, electrical, and real estate infrastructure that surrounds a modern facility.

Memory and interconnect are where the constraint has quietly settled. A rack full of accelerators that cannot be fed data fast enough is an expensive way to generate heat, and the industry has spent two years discovering how often that is the actual limit.

That last category is the one worth watching in Europe. Real estate, power distribution and cooling have not historically been venture categories, and siting is already the binding constraint here, with 63% of new capacity now going somewhere other than the five established hubs.

The fund arrives on top of an already unusual year. a16z announced more than $15bn across new funds in January, among them a $1.7bn Infrastructure Fund 2 and a $1.18bn American Dynamism Fund 2.

How the Machine Age Fund relates to those has not been made clear, and there is no word yet on limited partners, cheque sizes, or stage focus. Recent cheques point the same way regardless, among them a Series A into Netris, which automates the networking that slows down GPU clouds.

The pitch, stripped of the language around it, is that the scarce thing has changed. For most of the past decade, it was talent and distribution, and a16z is now betting $1.1bn that it is transformers, substations, and thermal design.