As startups scramble for AI servers, a growing number of venture capital firms are stepping in to help their portfolio companies lock down computing resources. People familiar with the matter said that Andreessen Horowitz (a16z), which stockpiled more than 20,000 GPUs for its portfolio companies two years ago, is in talks to expand its computing supply capacity, potentially through two data center companies it has already invested in.
Meanwhile, two other sources said that General Catalyst is building its own data center to provide computing power for its portfolio companies, and several other VC firms are using intermediaries to snap up scarce chips for their portfolio companies. Jeff Jewett, chief operating officer of Shadeform, said: "More and more VC firms want to allocate computing resources across their basket of portfolio companies." Shadeform's business involves leasing servers from various cloud providers and reselling them to AI startups and other customers. "To attract the world's top startup teams, computing supply will inevitably become an essential bargaining chip," he said.
These moves reflect a shortage in the supply of specialized chips used for AI training and inference. Constrained by computing capacity, earlier this year cloud providers like Microsoft prioritized GPUs for internal projects or major customers, further squeezing smaller startups' access to chips. Even for more mature startups, computing costs can account for a significant portion of operating expenses. Traditional VC firms typically only provide soft support such as funding, networking resources, and hiring advice, but now they are increasingly venturing into hardware resources.
Daniel Gross and Nat Friedman were among the first investors to build computing clusters. In 2023, they bulk-purchased Nvidia's high-end chips and assembled a computing pool to open up to their portfolio companies, known as the Andromeda Cluster. The cluster has now transformed into a computing supply-and-demand matching platform.
Talks with Switch and Volta
a16z had previously expanded a computing cluster codenamed "Oxygen," which includes chip resources leased from multiple cloud providers. People familiar with the negotiations said that in recent months, the VC firm has discussed partnering with data center developer and operator Switch and infrastructure platform Volta to further expand the cluster's scale. a16z has capital ties with both companies: in August this year, it led Volta's $300 million funding round at a valuation of $2.4 billion; in July, it was reportedly in talks to lead Switch's $2 billion funding round. These investments are part of a16z's broader push into the underlying hardware and infrastructure of the AI wave. The specific scale of a16z's chip expansion is not yet clear. If portfolio companies do not use up all the computing power, the firm can resell the surplus externally. A spokesperson for a16z said the fund helps founders "find computing power, negotiate contracts, and secure favorable and special pricing."
General Catalyst, which has a broad investment portfolio and previously acquired hospital systems, has now also entered the chip computing space. In May, the firm recruited former Nvidia cloud computing executive Alexis Black Bjorlin as chief strategy officer, tasked with leading infrastructure projects to build "the efficiency frontier of large-scale AI systems" for its portfolio companies. Earlier this year, General Catalyst CEO Hemant Taneja sent a survey to portfolio founders to understand their computing access challenges. Reports in April indicated that the firm was already conceiving support plans, including building a shared computing pool and negotiating on behalf of startups. According to people familiar with the project, the firm is recently advancing the construction of its own data center specifically to output computing power. Sources said this computing cluster will serve all portfolio companies from early-stage to growth-stage, with a scale sufficient to support mature AI companies with massive token consumption.
These two firms are not alone. Coatue Management established data center development company Next Frontier last year and has already launched at least two joint venture projects with cloud service provider Fluidstack. People familiar with the matter at the firm's senior levels said that one purpose of the project is to ensure the portfolio gets access to scarce computing power. Large VC firms are making significant forays into chips and data centers, though the source of funding remains unclear, the business itself has profit potential. Many large funds have affiliated special-purpose vehicles that can be used to fund such projects; they can also provide loan guarantees and sign long-term leases to help partners complete financing.
The Logic Behind VC Computing Returns
Early-stage fund Crucible Capital tested the waters of building a computing business last year. People familiar with the matter said the firm owns a cluster of 576 Nvidia chips, with part of the computing power sold to frontier AI labs and the rest reserved for portfolio companies for product testing and project pilots. San Francisco-based computing startup San Francisco Compute, founded three years ago, builds AI chip clusters and operates a computing trading marketplace where enterprises can quickly rent GPUs without long-term contracts. Its founder Evan Conrad said: "From what we've observed, almost every large fund is doing this. VC firms want to directly control computing clusters to guarantee supply for their portfolio companies." Conrad is currently working with multiple VC firms to help them build their own clusters and allocate them on demand to portfolio companies; he also plans to sell idle computing power from VC firms externally.
Sriram Krishnan, co-founder and general partner of early-stage VC firm Kearny Jackson, wrote in early October that the firm has built a cluster with AI server suppliers and holds hundreds of AMD MI350 high-end server chips available for lease or purchase. Krishnan said that once the news broke, a large number of founders hungry for computing power came knocking, "the scene was like a storm." Some founders described to him that getting GPUs was equivalent to holding crude oil during an oil shortage era. Krishnan prioritizes GPUs in the cluster for his own portfolio companies; the business itself is not intended to be profitable, but he believes that the availability of computing resources will attract more quality entrepreneurs to come knocking. "If there's a match and an investment opportunity exists, then all the better."
Comments