Andreessen Horowitz, General Catalyst Race to Secure the Chips Their Startups Need

As startups scramble to get access to AI servers, more venture capital firms are finding ways to secure compute on their behalf.

Andreessen Horowitz, which two years ago set up a cluster of 20,000 graphics processing units for its startups, is in talks to expand its offering, potentially via two data center companies it has backed, according to people familiar with its plans. Meanwhile, General Catalyst is building a data center to supply compute for its startups, said two other people. And several other VC funds have used brokers to find sought-after chips for their companies.

“More and more we are working with venture capital firms who are looking to deploy” compute across their portfolios of startups, said Jeff Jewett, chief operating officer at Shadeform, which rents server capacity from various cloud providers and provides it to customers including AI startups.

“It’s absolutely going to be a requirement to be able to attract the hottest startups in the world,” he said.

The efforts reflect a shortage of the chips designed to handle the complex workloads of training and running AI. These constraints earlier this year led Microsoft and other large cloud providers to divert graphics processing units to their internal teams or bigger customers, making it harder for smaller startups to find a supply.

Even for mature startups, compute can be a significant portion of their operating costs. That’s led more venture firms, which typically stick to supplying cash and more ephemeral assets—such as connections to skilled executives and advice on hiring—to wade into hardware.

Daniel Gross and Nat Friedman in 2023 were among the first investors to organize a compute cluster, buying advanced Nvidia chips they pooled together and made available to portfolio companies at a discounted price. Their Andromeda Cluster has now pivoted to matching buyers with sellers of compute.

Switch, Volta Talks

Andreessen Horowitz last year expanded its cluster, initially called Oxygen, into multiple private clusters, including groups of chips rented from different cloud providers.

In recent months, the VC firm has discussed partnering with companies, such as Switch, a data center developer and operator, and Volta, a platform that helps build, finance and operate data centers, to expand the cluster, according to people familiar with the talks.

Andreessen Horowitz has ties to both. It co-led a $300 million funding round in Volta in August at a $2.4 billion valuation, and it reportedly was in talks to lead a $2 billion round in Switch in July. The investments are part of the firm’s bigger swing at hardware and the infrastructure companies that underpin the AI boom, spearheaded by general partner Martin Casado.

It’s unclear how big Andreessen Horowitz’s chip expansion will be. If portfolio companies don’t end up using 100% of the capacity, the firm could resell the capacity to others who want the compute.

An Andreessen Horowitz spokesperson said the firm helps founders “navigate finding compute, negotiating contracts and obtaining desirable or special pricing.”

General Catalyst, a sprawling investment firm that previously bought a hospital system, is now also delving into chips. In May, the VC firm hired Alexis Black Björlin, who formerly helped run Nvidia’s cloud computing division, as its chief strategy officer. She is focused on infrastructure projects at the firm, or “the efficiency frontier of large-scale AI systems” for the firm’s portfolio companies.

Earlier this year, General Catalyst CEO Hemant Taneja sent a survey to founders inquiring about their ability to access compute. At the time, the firm was considering ways to help its portfolio companies with supply, potentially by facilitating a shared pool of capacity or negotiating on a startup’s behalf, The Information reported in April.

Lately, the VC firm has been working on securing a data center to supply this compute, according to someone with knowledge of the project. It will serve companies from early to growth stage across its portfolio, another person said, suggesting that the cluster could be sizable enough to meet the needs of mature companies that process more tokens than younger startups.

These add to other initiatives such as Coatue Management’s launch last year of a data center development company, Next Frontier, which has at least two joint ventures currently underway with neocloud Fluidstack. That effort is in part designed to secure access to scarce compute for its investments, according to someone who spoke with a senior member of the firm.

It’s not clear how the biggest VC firms will finance more aggressive forays into chips and data center development, and they certainly could find ways to generate profits from the endeavors. Many of the larger funds have side vehicles they can use to finance such a project, or they could guarantee a loan or sign a lease commitment that would allow a partner to raise the money.

Venture Payback

Early-stage investor Crucible Capital began building its own cluster last year mostly as an experiment. Now it sells some of the capacity from the 576-chip cluster of Nvidia B300s to a frontier lab, while its portfolio companies use it to test products or run pilot projects, according to a person with knowledge of the firm’s activities.

“Our impression is [that] every major fund is attempting to do this,” said Evan Conrad, founder of San Francisco Compute, a three-year-old startup that builds clusters of AI chips and operates a compute marketplace where companies can quickly rent GPUs without requiring long-term contracts. “They’re trying to own the clusters to secure access for their portfolio.”

Conrad is currently working with a few VC firms to set up clusters they will control and parcel out to portfolio companies when the startups need the compute. Before that happens, SF Compute plans to sell the VCs’ excess access to other buyers, he said.

Sriram Krishnan, a co-founder and general partner at early-stage VC firm Kearny Jackson, in early October posted on X that he had access to “hundreds of AMD mi350s”—AMD’s high-end server chips, which can train and run AI—for lease or purchase, in a cluster he recently set up with a supplier of AI servers.

Krishnan said he immediately received an onslaught of outreach from founders hungry for compute: “It was a bit of a hurricane,” he said in an interview. Some founders who messaged him likened having access to GPUs to “having oil during an oil shortage,” Krishnan said.

Krishnan will dole out the GPUs in Kearny Jackson’s cluster to portfolio companies and doesn’t plan to turn a profit on the effort. Though he surmises that the prospect of quick access to compute will attract new founders to the firm. “If there’s a mutual fit and an opportunity to invest, then great,” he said.