Andreessen Horowitz boosts AI infrastructure fund to $3B

Andreessen Horowitz expanded its artificial intelligence infrastructure fund to $3 billion, committing an additional $1.7 billion to the vehicle launched in 2024 to back software built for developers. The firm confirmed the new capital was committed this month, Bloomberg reports.

The Silicon Valley firm, known as a16z, applies a broad definition to AI infrastructure. In its framework, the category covers AI software sold to technical buyers inside companies, including coding applications, foundational models and network security tools. The fund targets the systems and software layers that power AI products.

Raghu Raghuram, a managing partner and former VMware chief executive, framed the thesis simply: “Some of the most important companies of tomorrow will be infrastructure companies.”

The portfolio has recorded several transactions in recent months. Stripe agreed to buy a16z-backed billing platform Metronome for a reported $1 billion. Salesforce acquired Regrello, an AI provider for manufacturers, and Meta Platforms purchased AI audio company WaveForms.

Valuations have climbed. AI coding startup Cursor raised money in November at a $29.3 billion valuation, up from about $400 million when a16z first invested in 2024. LMArena, a platform for ranking AI models, is valued at $1.7 billion less than a year after it spun out of an academic project. Unconventional AI, an AI computer startup a few months old, raised a $475 million seed round and was valued at $4.5 billion. World Labs, founded with AI pioneer Fei-Fei Li, raised $230 million to develop “world models” that simulate physical spaces.

Ben Horowitz, the firm’s co-founder, emphasized that it remains early to judge returns, which are usually measured over a decade. In his words: “It’s one of the best funds, like, I’ve ever seen,” while noting it is too soon to draw conclusions about performance.

General partner Martin Casado, a former computational physicist who joined a16z after selling Nicira to VMware, described today’s private market bluntly: “Private valuations are crazy.” He added: “The users are real. The demand is real. The GPU usage is real.”

The firm has avoided directly backing the AI data center buildout. Casado acknowledged second thoughts about skipping so-called neocloud providers, pointing to CoreWeave’s rise to around a $50 billion market capitalization. By his account, “We just talked ourselves out of it stupidly.”

A16z’s approach is to invest early with relatively small checks and work closely with founders. Casado, who sits on Cursor’s board, has helped the startup recruit, secure computing capacity and resolve outages, according to Cursor President Oskar Schulz. Fei-Fei Li recounted that he regularly spends time at World Labs’ offices working with the team.

Casado has also worked behind the scenes on outcomes for companies under pressure. Cloudflare acquired Replicate, which helps deploy AI models to the cloud, after a matchmaking effort to find a partner and a price the startup would accept, according to a person familiar with the process. In a separate deal, he worked for months to help a16z-backed Tabular sell to Databricks for more than it had been discussing with Snowflake, according to people familiar with that transaction.

The latest capital arrives alongside personnel changes. A16z is promoting longtime infrastructure investor Matt Bornstein to general partner on the fund, joining Casado and Jennifer Li, who advanced from analyst to general partner. The fund concentrates on products sold to technical buyers inside enterprises.

Casado expects uneven outcomes as the market develops. In his view: “There’s going to be a bunch of companies that don’t work. There’s always fewer winners than people assume, but they are much larger than people assume.”

As we reported earlier, StandardHash CEO and founder Leon Lyu wrote on X that some miners are shifting power capacity to AI computing because margins can be more predictable. He noted that large mining sites already have assets that are costly and slow to build, including grid access, reliable cooling, and supporting infrastructure. Those facilities can be repurposed for the high-performance computing required to train and run neural networks.