A new computing provider says it is building a way for startups and researchers to access AI hardware that has become difficult to secure. Voltage Park is managing a cluster of 24,000 Nvidia H100 GPUs, purchased by the nonprofit Navigation Fund founded by blockchain entrepreneur Jed McCaleb.
The cluster is already being used by Imbue and Character.ai for AI model experimentation, according to Voltage Park CEO Eric Park. The project pairs a large supply of specialized chips with a stated aim of lowering the cost and contract barriers that can keep smaller organizations from training models.
Why access to GPUs matters
Generative AI systems rely heavily on GPU-based computing. These chips can perform many calculations in parallel, making them useful for training and serving capable AI models, including systems such as ChatGPT and Stable Diffusion.
But demand has outpaced the available supply of leading hardware. The source reports that Microsoft warned of a shortage of server hardware needed to run AI, while some of Nvidia’s strongest AI cards were reportedly sold out until 2024.
Park said companies of different sizes have told him they cannot obtain enough H100s to train their models. He described an added hurdle for startups and scale-ups: some cloud providers require large contracts to access chips, which smaller companies may not be able to sign.
Voltage Park’s proposed answer is to offer computing in ways that fit a wider range of customers. Park said a significant portion of the capacity will be reserved for early-stage startups and developers, with both short-term leases and hourly billing planned.
How the nonprofit and company are connected
The Navigation Fund purchased the GPUs using its endowment and transferred ownership to Voltage Park as an initial capital contribution. Voltage Park is a for-profit subsidiary of the fund, created to operate data centers and manage a cluster on this scale.
A Navigation Fund spokesperson told TechCrunch that the board believed a for-profit subsidiary would be better positioned to run specialized operations and pursue the market opportunity for cutting-edge computing. The fund, in turn, could focus on its charitable grant-making and broader mission.
The spokesperson said McCaleb does not own, run or earn profits from either organization. The Navigation Fund and Voltage Park have separate executive teams and independent boards of directors. A percentage of Voltage Park’s profits is expected to go to the fund to support its philanthropic work.
The GPU purchase also involved a detail about taxes: the fund paid full sales and use taxes despite its stated nonprofit status. The source raises the possibility of a tax benefit to McCaleb from donating to the fund, but does not establish that such a benefit occurred.
A wider philanthropic mission is planned
Voltage Park is intended to be one project of the Navigation Fund, rather than the fund’s sole focus. The spokesperson described a long-term foundation with interests that include farmed animal welfare, criminal justice reform, open science, climate and AI safety.
The fund plans to support organizations, activists, advocates and entrepreneurs working in those areas. Its broader plans are still at an early stage: TechCrunch reported that an expert could not find the fund in searches of Charity Navigator and GuideStar the prior week, and that its president, David Coman-Hidy, had joined in August.
That absence from those databases does not by itself establish that anything is improper. The article notes that the fund may have filed its paperwork and that the IRS may not yet have processed it.
Capacity is still being built out
The full Voltage Park cluster was not online at the time of the report. The company expected it to become operational across Texas, Virginia and Washington closer to the end of the year, assuming the rollout went according to plan.
For now, Voltage Park offers bare-metal machine learning training infrastructure. Park said the company may add services built on top of that infrastructure as the industry develops, and described the mission as making machine learning accessible to a wider audience by lowering the barrier to entry.
The plan’s effect will depend on how much capacity becomes available, how the pricing works for smaller customers and how the service develops as more GPUs come online. The central promise is straightforward: provide AI computing through shorter commitments and hourly access, so organizations that cannot meet large cloud contract thresholds can still experiment with models.