The most expensive part of building a large language model isn't the talent or the data. It's the electricity and the silicon. For the AI labs and startups that need to train frontier-scale models, the bottleneck is securing thousands of high-end GPUs for months at a time. Andromeda AI, founded by Nat Friedman and Daniel Gross, is betting that the market for dedicated, large-scale compute is underserved by the hyperscalers. The company, which recently raised a $60 million Series A at a $1.5 billion valuation led by Paradigm, is building a business around renting out dedicated clusters of Nvidia H100 GPUs, with a current fleet reported at over 4,000 units [Forbes, February 2024] [SiliconANGLE, March 2026].
The dedicated cluster wedge
The pitch is straightforward: if you are training a model that requires weeks of uninterrupted time on a cluster of hundreds or thousands of GPUs, you don't want to be competing for spare capacity on a public cloud. Andromeda's core offering is access to these large-scale clusters, which it sources, operates, and rents out, primarily from a data center in Santa Clara, California [AI Business]. The company also operates gpulist.ai, a marketplace that lists both its own inventory and third-party GPU capacity [Tom's Hardware].
The founder-led capital advantage
Nat Friedman, the former CEO of GitHub and current Co-Chief AI Officer at Meta, and Daniel Gross, a former Y Combinator partner, are not just operators but also investors through their billion-dollar AI venture fund, NFDG [The Information, 2026] [Forbes, 2024]. This gives Andromeda a unique position. The founders are simultaneously building the infrastructure and investing in the companies that will need it. The recent $60 million round from Paradigm suggests confidence in the thesis beyond the founders' own networks [Upstarts Media, January 2026]. The leadership structure is clarified by Wil Moushey serving as CEO [Portal of Robotics and Artificial Intelligence].
Where the wheels could come off
- Capital intensity and hardware cycles. Building a fleet of thousands of the latest GPUs requires hundreds of millions of dollars. The company is competing with well-funded specialists like CoreWeave and Lambda.
- The commodity trap. While dedicated clusters offer performance guarantees, the underlying compute is still a commodity. Andromeda's differentiation must extend beyond mere access to include software or orchestration.
- Founder bandwidth. Friedman's role at Meta and the pair's VC activities are assets, but they also represent significant commitments outside of Andromeda.
The company's early traction, including the deployment of a 2,512 H100 cluster in 2023, shows it can execute on the hardware side [Wikipedia, Daniel Gross, retrieved 2026].
The next twelve months
The Series A capital provides a substantial runway to scale the fleet and prove the model. The key metrics to watch will be customer concentration and contract duration. A handful of large, long-term contracts with named AI labs would validate the dedicated cluster thesis. Andromeda's ideal customer profile is the AI lab or well-funded startup that has a model architecture ready to train and needs guaranteed access to a specific, large GPU footprint for a known period of time.