It’s strange times we’re living in when a blockchain billionaire, not the usual Big Tech suspects, is the one supplying the compute capacity needed to develop generative AI.
It was revealed yesterday that Jed McCaleb, the co-founder of blockchain startups Stellar, Ripple and Mt Gox and aerospace company Vast, launched a 501(c)(3) nonprofit that purchased 24,000 Nvidia H100 GPUs to build data centers that’ll lease capacity to AI projects.
Already, the cluster of GPUs — worth an estimated half a billion dollars and among the largest in the world — is being used by startups Imbue and Character.ai for AI model experimentation, claims Eric Park, the CEO of a newly-formed organization, Voltage Park, that’s charged with running and managing the data centers.
“The goal was to unlock access to startups, scale-ups and research organizations that are currently blocked from this space due to restrictive contracts, the scarcity of GPUs and high minimum purchase thresholds so that they can access the vital resources they need to innovate,” Park told TechCrunch in an email interview. “We’re continuing to speak with folks in the industry to understand their needs and are using what we learn to inform how we build out our remaining clusters so that they’re useful to as many customers as possible.”
Most companies training models, particularly generative AI models like ChatGPT and Stable Diffusion, heavily rely on GPU-based hardware. GPUs’ ability to perform many computations in parallel make them well-suited to training — and serving — today’s most capable AI.
But there aren’t enough chips to go around.
Microsoft is facing a shortage of the server hardware needed to run AI so severe that it might lead to service disruptions, the company warned in a summer earnings report. And some of Nvidia’s best-performing AI cards are reportedly sold out until 2024.
“The shortage for cutting-edge compute is dire,” Park said. “I have conversations with companies of all sizes — startups, scale-ups and big labs — and pretty much every single one tells me they can’t get enough H100s to train their models. There’s a compounding difficulty for startups and scale-ups who can’t sign the large contracts many clouds require to access these chips, and that’s really limiting AI innovation.”
Voltage Park was established under a somewhat unusual structure.
McCaleb started a nonprofit called the Navigation Fund to whose endowment he donated. The Navigation Fund then purchased the aforementioned H100s — paying “full sales and use taxes,” despite the Navigation Fund’s ostensible nonprofit status — and transferred ownership of the GPUs to for-profit Voltage Park, which is technically a subsidiary of the Navigation Fund.
“Navigation’s board determined that a for-profit subsidiary would be better able to set up specialized operations to manage a cluster of this scale,” a Navigation Fund spokesperson told TechCrunch in an email. “They created Voltage Park to do that, and transferred ownership of the GPUs as part of the initial capital contribution. There was good strategic reason for doing this. A subsidiary would be better placed to pursue the market opportunity around cutting-edge compute, while the Navigation Fund would be able to focus exclusively on its mission and charitable grant-making.”
The spokesperson went on to say that McCaleb doesn’t own, run or earn profits from the Brooklyn, New York-based Navigation Fund or Voltage Park, which are run by separate executive teams and have independent boards of directors. Presumably, there’s a considerable tax benefit to McCaleb’s donation — even if the nonprofit recipient immediately re-gifted the assets to Voltage Park.
Voltage Park won’t be the Navigation Fund’s only project. The spokesperson described it as a “long-term” nonprofit foundation with interests in a range of sectors, including farmed animal welfare, criminal justice reform, open science, climate and AI safety, that’ll support “organizations, activists, advocates and entrepreneurs.”
The Navigation Fund’s got a ways to go before it embarks on any of those endeavors, though. It doesn’t appear to be formally registered as a nonprofit yet; when TechCrunch consulted an expert last week to perform a search for the Navigation Fund on two industry-standard databases, Charity Navigator and GuideStar, the expert couldn’t find a listing. And the Navigation Fund’s president, David Coman-Hidy, only joined in August.
That’s not to suggest impropriety. The Navigation Fund might well have its paperwork in order and filed, but the IRS simply hasn’t gotten around to processing it yet.
Whatever the Navigation Fund’s tax status, a percentage of Voltage Park’s profits will go to the Navigation Fund to support its philanthropic mission, the spokesperson said.
Voltage Park’s full cluster isn’t online. But once it is closer to the end of the year (assuming all goes according to plan), it’ll be live across three states — Texas, Virginia and Washington. Park says that a “significant portion” will be reserved for early stage startups and developers and that capacity will be available both for short-term leases and hourly billing.
“We set up Voltage Park to help ensure a broad range of companies have a seat at the [AI] table,” Park said. “We currently offer bare-metal machine learning training infrastructure, but our mission is broader – we want to make machine learning accessible to a wider audience by lowering the barrier to entry. This may include additional services built on top of our infrastructure. We’re still in the early days of this industry and we’ll adapt as the field develops and matures.”