Voltage Park
Last reviewed
May 21, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 ยท 4,140 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 21, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 ยท 4,140 words
Add missing citations, update stale details, or suggest a clearer explanation.
Voltage Park is a United States cloud computing company that operates a fleet of NVIDIA H100 graphics processing units for artificial intelligence training and inference workloads. The company launched publicly on October 30, 2023 with roughly 24,000 H100 GPUs financed by a $500 million investment from the Navigation Fund, a 501(c)(3) public charity endowed by Stellar and Ripple co-founder Jed McCaleb.[^1][^2] Voltage Park initially operated as a for-profit subsidiary of the Navigation Fund, with profits flowing back to the parent charity to subsidize grants in scientific research, animal welfare, climate, criminal justice, and AI safety.[^3][^4] By 2026, the company operated six data centers across four U.S. states, expanded through the March 2025 acquisition of GPU marketplace TensorDock, and in January 2026 completed a merger with developer platform Lightning AI.[^5][^6][^7]
| Field | Detail |
|---|---|
| Company name | Voltage Park |
| Type | For-profit subsidiary of a 501(c)(3) public charity |
| Public launch | October 30, 2023 |
| Parent | The Navigation Fund |
| Founder/Donor | Jed McCaleb (via Navigation Fund) |
| Initial CEO | Eric Park |
| 2024 to 2026 CEO | Ozan Kaya |
| Headquarters | San Francisco, California |
| Initial fleet | About 24,000 NVIDIA H100 SXM5 (80 GB) GPUs |
| Initial capital | About $500 million for GPUs; $1 billion long-term capital base |
| Locations | Texas, Virginia, Washington, Utah (six Tier 3+ facilities) |
| Managed services partner | Penguin Solutions |
| Storage partner | VAST Data |
| Notable transactions | Acquired TensorDock (March 2025); merged with Lightning AI (January 2026) |
Voltage Park emerged from a personal philanthropic initiative by Jed McCaleb, a serial entrepreneur best known as the original creator of the Mt. Gox bitcoin exchange in 2010, a co-founder of the Ripple payments protocol in 2011, and a co-founder of the nonprofit Stellar Development Foundation in 2014.[^8] McCaleb left active operation of Mt. Gox before its 2014 collapse and continued to serve as chief technology officer of Stellar. In parallel he founded the aerospace company Vast in 2021 to develop artificial-gravity space stations. Forbes lists him as a multibillionaire whose net worth rose with the appreciation of XLM and other cryptocurrency holdings.[^8]
In 2023 McCaleb and his spouse Seemay Chou, a microbiologist and chief executive of the open science nonprofit Arcadia Science, organized a new grant-making vehicle called the Navigation Fund. Inside Philanthropy and a fund disclosure described the Navigation Fund as a public charity (rather than a private foundation) with an endowment in the low billions of dollars at launch, and an unusual asset composition: a substantial majority of its assets consisted of a single operating company, namely Voltage Park itself.[^4] David Coman-Hidy, formerly president of The Humane League, joined as Navigation Fund president in August 2023 and Seemay Chou chairs the board.[^1][^4]
The Navigation Fund publicly unveiled Voltage Park on October 30, 2023. Coverage by TechCrunch the next day, October 31, 2023, framed the launch as a new nonprofit-affiliated entrant to the H100 cloud market: Navigation Fund had purchased about 24,000 NVIDIA H100 accelerators for roughly $500 million and transferred ownership to a for-profit operating subsidiary, Voltage Park, that would lease the GPUs to AI developers.[^1] The Register characterized the deal as one of the larger Hopper-class GPU fleets globally at that time, noting that the Navigation Fund paid full sales and use taxes on the purchase despite its 501(c)(3) status.[^2] Data Center Dynamics reported the same purchase volume and noted the deployment would be spread across data centers in Texas, Virginia, and Washington.[^9]
At launch the initial chief executive of Voltage Park was Eric Park, who told reporters 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."[^1] In separate comments during the launch press cycle, Park said that in his conversations with AI companies, "pretty much every single one tells me they can't get enough H100s to train their models," and observed that "ML teams and AI founders have to wait months or pay exorbitant sums" for access to current-generation accelerators.[^22] McCaleb himself took a deliberately hands-off posture, neither operating Voltage Park nor sitting on its board; Decrypt reported that he wanted the operating company to function independently of him while channeling its profits back to the parent charity.[^10]
The launch landed during a period of acute supply pressure for NVIDIA Hopper-class accelerators. From mid-2023 through early 2024, lead times on direct H100 orders from NVIDIA reportedly stretched to many months, and capacity at established hyperscalers was frequently reserved by a small number of large frontier model labs. By inserting roughly 24,000 H100s into the market under a transparent on-demand pricing model, Voltage Park positioned itself explicitly against the pattern of multi-year, multi-million-dollar capacity reservations that dominated the rest of the segment.[^1][^9]
Voltage Park's own launch blog post described the cluster as approximately 24,000 H100 SXM5 80 GB GPUs interconnected with 3.2 Tb/s InfiniBand, placing it among the largest single-operator H100 deployments announced in 2023. The company stated that some racks were already serving customers at launch and that full deployment was expected to complete by February 2024.[^11] The initial software offering was deliberately narrow: bare-metal access only, suitable for organizations that wanted to manage their own scheduler, container runtime, and storage layer. The Voltage Park launch post listed future support for Slurm, Kubernetes, and Mosaic-style training tooling as roadmap items rather than initial capabilities.[^11]
In July 2024 Voltage Park selected Penguin Solutions as its managed services partner. Penguin, an NVIDIA DGX-ready Managed Services Provider, was contracted to integrate and operate the majority of the 24,000 GPU fleet alongside the associated InfiniBand and Ethernet fabrics.[^12] The arrangement effectively outsourced day-to-day cluster operations (cabling, firmware, NVIDIA Collective Communications Library tuning, RDMA fabric validation, on-site break-fix) to a specialist with prior experience operating NVIDIA DGX SuperPOD-class systems, allowing Voltage Park's own staff to remain focused on customer-facing software and commercial operations.[^12] Tacoma and Pierce County economic development authorities confirmed that one of Voltage Park's western facilities was sited at the Centeris data center in Puyallup, Washington, with additional capacity in Quincy, Washington; Sterling, Virginia; Salt Lake City, Utah; and Fort Worth and Allen, Texas, for a total of six Tier 3+ facilities across four states by 2026.[^13][^14] VAST Data supplied storage for the clusters.[^14][^15]
On March 26, 2025 Voltage Park announced the acquisition of TensorDock, a GPU cloud marketplace offering on-demand consumer and professional GPUs. The companies did not disclose the financial terms. TensorDock founder Jonathon Lei joined Voltage Park as general manager of on-demand services, while Melissa Du, previously director of customer experience at Voltage Park, took over as general manager of the TensorDock marketplace, which continued to operate under its own brand.[^5][^16] The acquisition broadened Voltage Park's offering from premium H100 clusters into a broader marketplace spanning consumer RTX cards, HGX H100 SXM5 servers, and Blackwell-class systems.[^5]
Ozan Kaya succeeded Eric Park as chief executive of Voltage Park in 2024 and continued in the role through 2025.[^17][^14] On May 15, 2025 the company announced the appointment of former AWS leader Saurabh Giri as chief product and technology officer, along with Jesica Church as vice president of marketing and Cameron Huang as vice president of finance.[^17] During the same period the company reported growth from approximately $18 million in annual recurring revenue in 2024 to more than $500 million by late 2025, according to disclosures issued at the time of the Lightning AI merger.[^6]
In January 2026 Voltage Park and Lightning AI, the company behind the PyTorch Lightning training framework, announced that they had completed a merger that combined Lightning AI's developer platform and Voltage Park's GPU fleet. After the transaction the combined entity operated under the Lightning AI brand and was led by Lightning AI founder William Falcon as chief executive, with Ozan Kaya transitioning into a president role and Saurabh Giri remaining as chief product and technology officer.[^6] The merger disclosure described an owned fleet of more than 36,000 H100, B200, and GB300 GPUs across the combined data center footprint.[^6] Existing Voltage Park customer contracts and deployments were stated to continue without disruption.[^6]
Voltage Park's organizational structure was unusual for a GPU cloud provider. The Navigation Fund, a 501(c)(3) public charity registered in 2023, directly owned Voltage Park as a for-profit operating subsidiary. Inside Philanthropy reported that Voltage Park was carried on the Navigation Fund's books at approximately $883 million in 2023, which represented roughly three-quarters of the charity's assets.[^4] Profits earned by Voltage Park from GPU rentals flowed up to the Navigation Fund and were redeployed as grants in five program areas: scientific research, farm animal welfare, criminal justice, climate change, and digital sentience (research and policy related to the moral status of digital minds, including AI safety).[^3][^4]
The charity reported that its grant-making scaled rapidly from $4.1 million in 2023 to roughly $36.8 million in 2024, with continued growth projected in 2025. Inside Philanthropy reported that the AI safety and digital sentience portfolio was budgeted at roughly $20 million annually.[^4] The structure differs from comparable for-profit GPU clouds in that none of Voltage Park's net income could be distributed to McCaleb or other private parties: it had to be retained in the operating company or distributed to the parent charity. In tax-policy terms the arrangement is a variant of the program-related investment, in which a charity holds an operating asset whose mission directly furthers its charitable purpose (in this case, providing affordable compute to AI researchers) and whose net income funds further grant-making.
Press coverage flagged tax aspects of the launch: while the Navigation Fund's 501(c)(3) status could in principle exempt certain transactions from sales and use tax, both TechCrunch and The Register reported that the fund paid the full applicable taxes on the chip purchase, framing this as a deliberate choice to avoid the appearance of tax arbitrage.[^1][^2]
The Navigation Fund's governance was organized around domain-expert program officers rather than career philanthropy professionals. Inside Philanthropy reported that the fund hired David Coman-Hidy, the former president of the animal welfare organization The Humane League, as its president, and that Seemay Chou, an Arcadia Science executive and McCaleb's spouse, chaired the board.[^4] The fund's published practice was to deploy capital through a mix of open requests for proposals and invitation-only funding cycles, with substantial discretion delegated to program officers within each thematic area.[^4]
Voltage Park's launch fleet consisted of approximately 24,000 NVIDIA H100 SXM5 accelerators, each with 80 GB of HBM3 memory. The H100 is the flagship product of NVIDIA's Hopper architecture launched in 2022 and was, at the time of Voltage Park's announcement, the dominant chip for large-scale transformer training.[^11][^9] The fleet was deployed as 8-GPU server nodes (the standard HGX H100 reference design) with 1 TB of system memory per node, fully interconnected by 3.2 Tb/s NVIDIA Quantum-2 InfiniBand for low-latency collective communication across nodes, and parallel 100 GbE Ethernet for control-plane and storage traffic.[^11][^14]
After the Lightning AI merger in January 2026 the combined fleet was described as 36,000+ GPUs and explicitly included Blackwell-generation B200 and GB300 systems in addition to H100 and H200 capacity.[^6]
VAST Data supplied the all-flash storage system used as the primary high-throughput layer for training data and checkpoints.[^14][^15] Voltage Park offered customers bare-metal access (where the customer manages the operating system and scheduler directly), with managed services provided by Penguin Solutions for integration and production readiness on most of the fleet.[^12] Higher-level offerings published on the company website by 2025 included managed Kubernetes, virtual machines, and observability services, with a 15 minute provisioning target advertised for bare-metal H100 nodes.[^14]
For training workloads the standard customer pattern was to launch a contiguous slice of the InfiniBand fabric (typically a multiple of 8 GPUs corresponding to one HGX server, but extending to 248 GPUs under short-term lease and up to 4,088 GPUs under year-long commitments) and then run a customer-supplied training framework directly against the hardware. Voltage Park did not at first offer a managed model-training control plane or a managed inference endpoint of the kind sold by Together AI and similar firms; that gap was a primary stated rationale for the eventual merger with Lightning AI.[^6][^11]
Voltage Park did not initially own data center buildings. The H100 capacity was deployed in colocation facilities across six sites in Puyallup and Quincy, Washington; Sterling, Virginia; Salt Lake City, Utah; and Fort Worth and Allen, Texas, all reported as Tier 3+ facilities.[^13][^14]
A central piece of Voltage Park's positioning was undercutting incumbent cloud GPU pricing for H100 capacity. At launch the company quoted as low as $1.89 per GPU-hour for H100 capacity on its on-demand tier, increasing to $1.99 in subsequent published price lists.[^18][^19] The 2024 on-demand structure offered 1 to 8 GPUs at the lowest hourly rate, 8 to 248 GPUs under short-term leases, and capacity up to 4,088 H100s under year-long commitments.[^9]
Independent reporting at launch contrasted these prices with hyperscaler and specialist competitors. Coverage in The Batch reported that Amazon Web Services charged about $98 per hour for an on-demand p5.48xlarge instance containing eight H100s (which works out to roughly $12.29 per GPU-hour), or about $43 per hour for the same configuration under a three-year reserved commitment, while CoreWeave then listed H100 capacity starting at $4.76 per GPU-hour.[^18] Lambda Labs published on-demand H100 instances at roughly $2.99 per GPU-hour in 2024 to 2025.[^19]
| Provider | Reported H100 on-demand list price (per GPU-hour) | Source date |
|---|---|---|
| Voltage Park | $1.89 launch; later $1.99 | Oct 2023; later 2024 to 2026 |
| Lambda Labs | About $2.99 | 2024 to 2025 |
| CoreWeave | $4.76 list at launch period; lower for committed reliability tiers | 2023; 2026 |
| Amazon Web Services (p5.48xlarge) | About $12.29 per GPU-hour on-demand; $5.40 under 3-year reservation | 2023 to 2024 |
The pricing differential reflected structural choices. Voltage Park did not bundle a managed Kubernetes or model-serving stack at launch and relied on colocation facilities rather than owned-and-operated buildings; the parent Navigation Fund supplied the GPUs with non-dilutive philanthropic capital rather than venture or debt financing, which removed a required return on equity from the unit economics of each GPU-hour.[^1][^4]
At launch Voltage Park identified three early customers in the AI startup ecosystem: Imbue (formerly Generally Intelligent), a research lab building general-purpose agents; Character.AI, the consumer chatbot company; and Atomic AI, a biotechnology startup applying machine learning to RNA drug discovery.[^11][^2] Imbue chief executive Kanjun Qiu was quoted in the launch materials saying that Voltage Park had "helped us access critical compute much more quickly than other providers could have."[^11] The three named launch customers were representative of the targeted segment: a small AGI-oriented research lab, a consumer-facing model company, and an applied life-science startup, all with significant H100 needs but without sufficient scale to negotiate multi-year hyperscaler reservations.
Subsequent disclosures by the company described "over 100 global customers" across research institutions, AI startups, and enterprises by 2025.[^14] The customer profile, oriented toward research labs and venture-backed startups rather than Fortune 500 enterprises, reflected the founding mission of subsidizing access for organizations less able to negotiate hyperscaler reserved-capacity contracts. Sacra's company analysis described Voltage Park's commercial model as a hybrid of on-demand GPU rentals and dedicated capacity contracts of typically six to twelve months, with transparent usage-based billing.[^23]
On July 16, 2025 the U.S. National Science Foundation announced that Voltage Park had joined the National Artificial Intelligence Research Resource (NAIRR) pilot, a two-year proof-of-concept launched in January 2024 to coordinate federal and private compute for academic AI research. Voltage Park committed to donate the equivalent of one million H100 GPU-hours to NAIRR-allocated projects in science, engineering, health, and climate, alongside in-kind support from its operations staff to help match researchers with appropriate cluster configurations.[^20][^21] The pilot at the time involved twelve federal agencies and twenty-seven private, nonprofit, and philanthropic partners.[^20]
Voltage Park entered a market segment that journalists and analysts in 2023 to 2024 referred to as the "neocloud" or specialty GPU cloud segment, characterized by providers that offered dense H100 and Hopper-class capacity at prices below the major hyperscalers. The most prominent peer companies included CoreWeave, Lambda Labs, Crusoe Cloud, and Together AI for managed inference. Independent surveys by Spheron, IntuitionLabs, and gpuleaseindex in 2025 to 2026 consistently placed Voltage Park among the lowest-cost providers of on-demand H100 capacity for short-duration workloads.[^19]
Distinctive features relative to peers included:
Voltage Park drew attention beyond the GPU cloud market for several reasons. It represented one of the largest single philanthropic deployments of physical AI infrastructure since the launch of large H100 clusters became feasible in 2023. The structure (a for-profit infrastructure operator wholly owned by a public charity) is unusual in cloud computing and provided a template for using physical infrastructure as a perpetual funding source for grant-making. Inside Philanthropy described the model as a deliberate alternative to traditional foundation investing, in which an operating asset rather than a financial portfolio underwrites the philanthropic budget.[^4]
The company also contributed to the 2023 to 2024 narrative of GPU supply scarcity. By making published, transparent on-demand pricing available at below-hyperscaler rates for the same NVIDIA H100 SKU, Voltage Park provided a public reference point that other providers and customers used in negotiating committed-capacity contracts. The deeplearning.ai newsletter The Batch credited Voltage Park, alongside specialty providers such as Lambda Labs and CoreWeave, with widening short-term H100 access for academic and startup users during a period when hyperscaler queues stretched into 2024.[^18]
Finally, Voltage Park's participation in the NAIRR pilot represented one of the larger in-kind compute contributions to U.S. federal AI research infrastructure planning during the pilot's two-year window, alongside donations from NVIDIA, Anthropic, and other commercial providers.[^20]
From a broader standpoint, Voltage Park can be read as one of several attempts in the 2023 to 2026 period to assemble physical AI infrastructure outside the dominant hyperscaler stack. Where Crusoe focused on energy-aware colocation of GPU capacity near stranded renewable or flared-gas resources, and where CoreWeave sought to scale conventional venture-backed GPU cloud capacity to hyperscaler-rivaling size, Voltage Park's particular contribution was a capital structure that combined non-dilutive philanthropic financing with a transparent on-demand pricing model. The merger with Lightning AI in January 2026 partially closed the unusual capital-structure chapter by re-integrating the operating company with a more conventional venture-backed software and platform business, while the underlying compute assets continued to support the Navigation Fund's grant-making indirectly through the combined entity.[^6][^4]
Several aspects of Voltage Park drew skepticism in press coverage.
| Company | Relationship | Notes |
|---|---|---|
| CoreWeave | Competitor | Larger specialty GPU cloud, NYSE/Nasdaq-listed in 2025; began in cryptocurrency mining |
| Lambda Labs | Competitor | Specialty GPU cloud popular with researchers; published on-demand H100 instances |
| Crusoe | Competitor | Stranded-energy GPU cloud focused on flared-gas and renewable colocation |
| Together AI | Adjacent | Focuses on managed inference and fine-tuning rather than bare-metal H100 rental |
| Amazon Web Services | Hyperscaler benchmark | p5.48xlarge instance used as the public price comparison for H100 capacity |
| Character.AI | Customer | Identified as launch-period customer for training capacity |
| Imbue | Customer | Early customer; CEO endorsement appeared in launch materials |
| PyTorch Lightning | Adjacent (Lightning AI) | Open-source training framework owned by Lightning AI, with which Voltage Park merged in January 2026 |