The $1.5B Compute Startup Helping AI's Hottest Companies Find GPUs
Once a service spun up by AI investors Nat Friedman and Daniel Gross, Andromeda has now raised $60M from Paradigm to stand alone.
On the Friday before Thanksgiving in 2023, investor Daniel Gross called entrepreneur Wil Moushey with a job offer.
Over the previous few months, Gross and the VC firm he’d co-founded with Nat Friedman, NFDG, had half-accidentally ended up running what looked like a 1980s Wall Street stock trading desk, but for Silicon Valley’s most prized commodity: computing.
“I got a bid!” “I got an ask!”
A small team of engineers already worked on the project, called the Andromeda Cluster, but it was too much. Someone needed to step in and run the project. So Gross asked Moushey point-blank: “Do you want to jump in?”
Moushey, a former consultant who’d founded a sustainability startup backed by Gross that had recently shut down, was intrigued; he wanted some time to think about it. He didn’t appreciate the urgency yet, he remembers now. “I need your answer on Monday,” Gross added. He took the job.
Over the next few months, Andromeda gained publicity as one of the most notable examples of VC firms looking to win in the exploding AI category by offering a new kind of value-add, GPU access. NFDG committed more than $100 million upfront for Andromeda, per a Forbes report, reaching a stockpile of more than 4,000 GPUs by early 2024. Fast-growing companies in the portfolio like voice unicorn ElevenLabs and video generation startup Pika took advantage.
Then, last year, Friedman and Gross joined Meta as two of the tech giant’s splashy and expensive AI leadership hires, and stepped back from NFDG. Andromeda faded from focus as attention shifted to multi-billion dollar computing projects being signed – and sometimes abandoned – by the industry’s heavyweights like Meta, Elon Musk’s xAI, and OpenAI.
But tucked away in Miami, Moushey kept building. Spun out as its own startup with backing from NFDG and Paradigm, the VC firm notable for its crypto investments, Andromeda quietly passed a revenue run rate of $100 million in 2025, up from $50 million-plus the year before; operating profitably since its start, Andromeda now manages tens of thousands of GPUs at any given time, for a bunch of notable AI startups and labs that Moushey says he can’t name.
“There’s a certain amount of discretion in the business model, especially in regards to working with AI labs,” Moushey says. “These are very high-dollar, very secretive, very strategic kinds of things that we’re doing with them.”
But Andromeda does want the wider startup ecosystem to know that it’s open for business: with a team of about 20, the startup recently opened an office in San Francisco, and hopes to widen its customer base as its compute-under-management swells this year to as many as 100,000 GPUs.
To do that, the startup has raised new funding from Paradigm, the equivalent of a Series A, at a valuation of $1.5 billion. The funding brings Paradigm’s total investment in Andromeda to $60 million.
“This is a massive new market that really needs a lot of infrastructure to help it work smoothly,” Paradigm co-founder Matt Huang tells Upstarts.
The ‘if you know, you know’ startup is interesting for a few reasons beyond its surprise unicorn valuation. There’s the curiosity factor of what happens to such a project when its creators move on in AI’s musical chairs. There’s the intrinsic bet on computing remaining a pressing need among startups and big companies alike.
Most interesting, there’s the angle of why a firm like Paradigm was so interested: the promise of a real commodity market for GPUs and computing access developing, more akin to how we buy futures in oil or wheat.
“We have the aspiration to enable the global flow of compute, and we think that’s going to be a multi-trillion-dollar market,” Moushey says. “And to do that, we have to be a little bit more clear that we’re not this weird compute provider anymore. We’re actually building the market infrastructure to solve this dislocation of supply and demand.”
More on that opportunity — as well as a special video illustration of how all of this works by friend of Upstarts Natalie Fratto — below.
A MESSAGE FROM OUR SPONSOR
Radically different banking loved by over 300,000 entrepreneurs
When I launched Upstarts last year, I knew we needed a banking partner that could move at our speed, and grow with us, too.
Mercury is thoughtfully designed for the unique needs of entrepreneurs like me, offering an easy and intuitive banking experience so I can spend more time with the founders we like to cover.
Building a media startup comes with new challenges every day. But with Mercury, banking is one we don’t need to worry about.
Visit mercury.com to learn more and apply online in minutes, like I did.
Mercury is a fintech company, not an FDIC-insured bank. Banking services provided through Choice Financial Group and Column, N.A., Members FDIC.
Making a market
When Moushey, a former DI football player at the University of Albany who later worked stints helping with government datacenter projects and at PwC and Bridgewater, first plugged into Andromeda, it was a four person, remote team.
Many of Andromeda’s initial users were AI startups that Gross and Friedman had backed via a vehicle they’d set up called the AI Grant, which allocated $10 million to batches of AI companies. Besides Pika, notable batchmates included Browserbase, Cursor, Granola and Perplexity.
Andromeda had sold out of its own capacity quickly, forcing staffers like Hersh Desai, who now leads investment activities for NFDG, to work the phones themselves. They’d call anyone: telecom companies, crypto miners, big longtime companies like Dell and HPE.
For the startups, the pitch was simple: rather than commit to a multi-year contract with a hyperscaler like Amazon or Google, or get in line with neocloud computing providers only to miss deadlines as bigger companies bumped them in priority, or access didn’t materialize, Andromeda would secure the compute upfront, then portion it out as needed. For the startups, such usage would often come in spikes as they ran training runs for new models or tools; it would be cheaper and faster to only access it a la carte.
Andromeda, meanwhile, could make a small spread and create a win-win through leveraging its buying power, as well as the ability to stockpile GPU access when the market dipped, or demand temporarily tapered off — like when the release of Chinese open-source model DeepSeek spooked the market last January.
Because it leased the access from CoreWeave and lots of other providers, Andromeda also didn’t take on the liability and depreciation risks of building and maintaining its hardware or own data centers.
The Compute Commodity Market, by Natalie Fratto
The biggest challenge: finding GPUs and compute in lots of different places, testing it and verifying it, then sorting it and matching it to the startups in what resembled a more homogeneous, easy-to-use product.
“The conventional wisdom that no capacity exists out there was actually wrong,” says Moushey.
And while lots of neoclouds and exchanges were popping up and growing, from large-scale infra providers like Crusoe and Nebius to neoclouds like Lambda and Together AI, and marketplaces like Mithril and Runpod, Andromeda found a sweet spot in the mid-market, serving companies in the $250 million to $500 million annual compute spend range.
Net neutral
In a chaotic and growing GPU market, Andromeda sees itself as a Switzerland-like, neutral matchmaker.
NFDG portfolio companies don’t get preferential treatment, though many remain customers. Gross and Friedman, now at Meta, have no active relationship with Andromeda anymore and didn’t respond to requests for comment for this story.
When Gross posted on X in January that he was building out a compute desk for Meta, Moushey’s phone lit up, but he claims not to have inside knowledge: “I’m usually just like, you should reach out to Daniel.”
But Moushey says Andromeda is open to work with anyone, from Meta and the leading AI labs down to inference providers like Baseten and Fireworks AI.
That’s because he and investors hope that what’s most valuable about Andromeda isn’t just its GPU access, but its software and expertise in handling these contracts. It’s a technical challenge to vet and verify access from a wide range of sources, then quickly handle the deal structuring and contractual agreements.
“I tell the team we need to be able to provide the reliability and scale of a hyperscaler, with the speed and flexibility of the emerging market,” says Moushey. “Our competition is unutilized GPUs.”
The needs for those GPUs have shifted a bit, Desai at NFDG notes: from big, upfront training runs to a range of compute needs like pre-training, reinforcement learning, and long-running agents.
One AI startup founder who has worked with Andromeda for a while, and who asked to remain anonymous for competitive reasons, says that companies like theirs will continue to work with Andromeda so long as it can remove the headache of debugging the issues around GPU clusters with better customer service than a hyperscaler, while allowing a company like theirs to focus staffing resources elsewhere.
“They basically serve as this excellent engineering team for us,” the founder says. “It’s just more flexible.”
But if their startup grows big enough, at some point it might make sense to bring such capabilities in-house with their own fully-owned GPU clusters, the founder notes.
“You can’t run a $50 million training workload on a financial derivative.”
That’s part of the bet around Andromeda: that enough companies will continue to need GPU access and support working with a variety of providers to continue to grow its customer base; if only a few players stand tall in the long run, they won’t necessarily sustain a thriving marketplace.
“At the end of the day, everyone does not have a relationship with everyone,” argues Desai. “This isn’t like gold or coffee,” adds Moushey. “You can’t just deliver it.”
The winning trade
But when Upstarts talks to Huang at Paradigm, he brings up bushels of wheat, like we’re playing the boardgame Settlers Of Catan.
Not every bushel has the same quality, but the market has set standards for wheat and its grades in a way that it can be traded as a scaled commodity; that’s what Huang and Paradigm envision for compute with Andromeda.
Technical ability to test and standardize compute so that the types of GPUs involved are more interchangeable can help it to become such a commodity, Huang says.
(Paradigm is reportedly looking to expand deeper into AI and robotics with a new fund, per a Wall Street Journal report; the firm declined any comment about those plans.)
Last July, Upstarts wrote about OneChronos and the race to develop an auction market for GPUs and compute, which would allow for companies to invest in futures and hedge their risk. Andromeda can help such financial vehicles as the underlying marketplace, Moushey says. “You can’t run a $50 million training workload on a financial derivative,” he says.
Such a market could also help with what Moushey argues is the biggest hangup in AI right now: credit. Access to credit has reportedly held back OpenAI and other labs from investing as aggressively in new infrastructure projects as they’d like; even at the startup level, it’s no trivial task to spin up tens of thousands of GPUs today.
“There’s just a handful of balance sheets that can handle these investments,” Moushey says. “How to underwrite these investments is something that the market is going to figure out, and we’re going to be a key component.”
 | A guest post by
|
great read!