Andreessen Horowitz has raised $1.1 billion for its new Machine Age Fund, a16z said in its August 28 announcement. The fund targets the physical infrastructure needed to build and run AI, from chips and memory to networking, storage, data centers, robotics, and home AI appliances.
This is a broader hardware bet than a conventional accelerator fund. a16z says the common thread is that every layer of the stack is running into supply-chain limits, as well as constraints imposed by physics and computer science. The firm is making hardware an official investment focus while AI demand expands beyond chat into coding, reasoning, and other compute-heavy work.
A fund aimed below the model layer
The announcement describes a stack that reaches well below the model itself. Faster systems need more capable processors, but they also need cheaper and higher-bandwidth memory, larger interconnects, power-efficient edge devices, and the supporting cooling, materials, electrical equipment, and real estate. That emphasis on moving data is the same problem behind our explainer on memory bandwidth in AI hardware: a faster compute block is only useful when the rest of the system can keep it fed.
a16z frames the moment as another shift in computing architecture, comparing it with the move from mainframes to client-server systems, the internet, cloud, and mobile. Its argument is that this cycle is arriving with more breadth and less time to spare. The fund can invest in individual components, complete systems, and the facilities that connect them, rather than treating hardware as a narrow category beneath software.
That scope also covers physical AI. Robotics and home appliances bring the model out of the data center and into devices that need sensors, motors, reliable local compute, and a way to operate within real power and thermal limits. It gives the fund a route into edge systems as well as the largest cloud installations.
The rack is becoming an infrastructure project
The scale of the buildout is clearest in the figures a16z chose to publish. The firm says compute density per rack has increased 28 times from an H100 rack to a Rubin rack, while in-rack networking is reaching the limits of copper cabling. It also says rack power has moved from roughly 5 to 10 kilowatts to 100 to 250 kilowatts for current systems, with 1 megawatt racks possible within the next three years.
Those are a16z’s estimates and framing, not a promise that every future rack will use the same configuration. They do show why the chip race is also a power race. As accelerator density rises, the electrical distribution, cooling loops, network fabric, and building around the rack become part of the product.
a16z says data centers are moving from tens of megawatts toward hundreds of megawatts, with some campuses reaching gigawatt scale. It also points to a shift from grid-only supply toward combinations of grid power, behind-the-meter generation, and captive sources. The firm’s linked NVIDIA reference puts the 100 to 250 kilowatt rack range in the context of high-voltage direct-current power architecture, but the fund announcement is the source for a16z’s wider investment thesis.
The firm says hardware startups now represent more than 20% of its deal flow, up from a small share, and names recent investments including Unconventional AI, Nexthop, Volta, Atoms, and Mind Robotics. a16z also points to earlier investments in Skydio, SpaceX, Anduril, and Waymo as evidence that the firm’s hardware interest predates this fund.
The Machine Age Fund does not make new chips or solve the power bottleneck by itself. Its significance is that one of the most visible software-focused venture firms is now treating manufacturing capacity, energy, and physical systems as a central AI investment lane. The next test will be whether that capital helps turn an increasingly dense rack into infrastructure that can actually be built, powered, cooled, and maintained.
