a16z raises $1.1B Machine Age Fund to rebuild AI infrastructure

The next phase of AI depends less on new models than on the chips, power, and machines needed to run them.

a16z's $1.1 billion Machine Age Fund targets the physical layer of AI — chips, memory, power, and data centers — as hyperscaler capital spending heads past $1 trillion and component orders stretch to 2028.

"Any problem you have can be solved with enough infrastructure — essentially GPUs plus money. As long as problems remain unsolved, demand won't stop," Ben Horowitz, co-founder of Andreessen Horowitz, said in a video conversation announcing the fund.

The fund will invest across semiconductors, memory, networking, storage, data centers, robotics, and edge devices. Partners Martin Casado and Raghu Raghuram joined Horowitz to detail the supply-demand gap driving the strategy: hyperscaler capital expenditure is running at roughly $700 billion this year and projected to exceed $1 trillion in 2027, while GPU and memory orders are already booked through 2028. A leading memory vendor told the Hot Chips conference that existing orders alone would require three years of production capacity.

The fund formalizes a shift already visible in a16z's deal flow — hardware startups now represent more than 20 percent of its investments, up from 3 to 5 percent a few years ago. With token demand growing near 1,000 percent annually and new data centers requiring 44 gigawatts of power by 2028 against only 25 gigawatts of expected grid additions, the firm is betting the compute stack must be rebuilt from first principles.

The supply-demand gap is not a cyclical squeeze but a structural mismatch. Hardware supply has historically grown 20 to 30 percent per year, while AI compute demand is expanding at triple-digit rates. Casado described the situation as unprecedented: "Every GPU being produced is already pre-sold." He noted instances of thousands of GPUs going to multi-day auctions with scalper prices at four times retail. Raghuram added that the impact extends all the way down the supply chain — "to the copper mines."

Rack power requirements are jumping from 5 to 10 kilowatts to 100 to 150 kilowatts, with compute density rising roughly 70-fold. Cooling has shifted from air to liquid as a mandatory requirement for frontier data centers. Rack voltage is climbing to 800 volts DC — a level that only about 2 percent of U.S. electricians are certified to handle, according to Horowitz. Concrete prices for data center construction are also rising sharply, Raghuram noted.

A $2 billion case for custom silicon

The capital intensity of frontier AI has reached a point where bespoke hardware becomes economically rational. Training a frontier model now costs $3 billion to $5 billion, Casado said. Inference needs to generate roughly $10 billion to recoup that investment, meaning a 20 percent efficiency gain in inference — worth about $2 billion — is enough to justify developing a custom ASIC.

"We've never spent $5 billion on a single digital artifact before," Casado said. "This will place the highest demands on hardware we've ever seen."

Hardware deals jump from 5% to 20% of a16z's flow

The shift is already visible in the startup community. Casado estimates that hardware deals from top-tier founders have risen from roughly 3 to 5 percent of a16z's deal flow to more than 20 percent. Recent hardware investments include Unconventional AI, Nexthop, Volta, Atoms, Heron Power, and Mind Robotics, alongside earlier positions in SpaceX, Anduril, and Waymo.

Raghuram noted that these founders must be "system-level" operators — able to architect chips, design supply chains, and navigate manufacturing constraints that software founders never face. AI labs are also signing agreements with early-stage hardware companies before products exist, a shift that reduces commercialization risk for startups. First-round funding for hardware startups has reached hundreds of millions of dollars, a scale rarely seen in previous hardware cycles.

The fund's name reflects a deliberate framing. "Artificial intelligence was always the wrong term — it should be machine intelligence," Casado said. "There's a deep irony that the people who said 'software eats the world' have arrived at a place where the real constraint is the physical machines underneath."

For investors, the fund's launch points to sustained capital deployment across the AI hardware stack. Nvidia, whose H100 and upcoming Rubin platforms anchor the current compute generation, faces a growing field of challengers enabled by the economics Casado described. Data center operators, power infrastructure providers, cooling system vendors, and memory manufacturers all stand to benefit from the multi-year buildout. The 44-gigawatt power gap by 2028 suggests the constraint may shift from silicon to electricity — and to the transformers, turbines, and grid connections needed to deliver it.

This article is for informational purposes only and does not constitute investment advice.