Voltage Park
GPU-as-a-service: on-demand bare-metal GPU rental plus reserved-capacity contracts on an owned fleet; post-merger, bundled with Lightning AI's developer software layer (training, inference, model serving)
Voltage Park was never independently priced: the 2023 point is grant capital deployed (a capitalization proxy, not a negotiated equity valuation), and the 2026 point is the merged Lightning AI entity, not Voltage Park standalone. No primary or secondary priced rounds exist between the two.
Earnings, margins, COGS & capex
Voltage Park is private and, unusually, was capitalized by philanthropy: Jed McCaleb's Navigation Fund (a nonprofit; McCaleb co-founded Ripple and Stellar and built Mt. Gox) bought ~$500M of NVIDIA H100s at founding (Oct 2023) and transferred them to for-profit subsidiary Voltage Park, with total grant funding of $900M (Forbes, Jan 2026; the fund's assets are ~$1.25B). Profits were designed to flow back to the nonprofit. Standalone revenue, margins, and cash flow were never disclosed. The only hard financial anchor is the Jan 2026 merger with Lightning AI: combined valuation over $2.5B, combined ARR over $500M (including GPU rentals booked through Voltage Park), up from $18M in 2024 - though the Voltage Park rental vs Lightning software split is not broken out. Navigation Fund retains a significant (unspecified) equity stake in the merged company, which operates under the Lightning AI name.
Revenue trend
Margins
Structurally advantaged: the founding GPU fleet was grant-funded, so depreciation/financing drag is lower than debt-financed neocloud peers; company has marketed below-market pricing (H100s from $1.89/GPU-hr at 2023 launch)
Likely negative or reinvested given Blackwell (B200/GB300) fleet expansion
COGS structure
Dominated by data center operations: power (~60MW active across 6 sites in 4 US states at merger), colocation/facility costs (Centeris among named colo providers), GPU depreciation (H100 fleet aging toward the Blackwell cycle), 3.2 Tbps InfiniBand networking, VAST Data storage, and deployment/support contracts (Dell hardware; Penguin Solutions cluster management). Grant-funded original capex means COGS carries less financing cost than leveraged peers like CoreWeave.
Capex
Initial ~$500M H100 purchase (24,000 units, Oct 2023) funded by the Navigation Fund grant ($900M total); fleet since expanded to 36,000+ GPUs including B200 and GB300 Blackwell units - implying substantial ongoing capex of undisclosed size and funding source. TensorDock acquired (March 2025, undisclosed price) for marketplace reach rather than capacity.
Latest earnings
None issued; the only forward-looking anchor is the merger's stated ambition to build a full-stack 'first cloud built for AI'
- Combined ARR at merger
- $500M+ (Jan 2026)
- Combined valuation
- $2.5B+ (Jan 2026)
- GPU fleet
- 36,000+ owned H100/B200/GB300 (merger PR, Jan 2026; Forbes cites 35,000+)
- Active data center capacity
- ~60MW, 6 sites, 4 US states (Forbes, Jan 2026)
- Neocloud rank
- #3 by chips deployed, behind CoreWeave and Nebius (Forbes, Jan 2026)
Growth drivers
- AI training and inference compute demand outstripping hyperscaler supply, sustaining the neocloud category
- Merger with Lightning AI — 400,000+ developers, startups, and enterprises on the Lightning platform (PyTorch Lightning, 400M+ downloads per Forbes) convert into GPU consumption on Voltage Park capacity
- Blackwell (B200/GB300) fleet refresh commanding higher $/GPU-hour than aging H100s
- Move up-stack from bare metal to managed Kubernetes (launched June 2025) and post-merger inference/model-serving software - higher-margin attach
- Below-market pricing enabled by grant-funded capital base, winning price-sensitive startups and labs
- TensorDock marketplace acquisition (March 2025) extending reach into long-tail developer demand
Bull & bear
The cleanest balance sheet in the neocloud sector meets a 400K-developer software funnel. A grant-funded fleet means Voltage Park can price below leveraged rivals and still generate cash, and the Lightning AI merger converts commodity GPU-hours into a differentiated full-stack AI cloud growing from $18M to $500M+ ARR in two years.
- Capital-structure edge: the founding ~$500M H100 fleet carries no debt service, so at market-clearing rental prices Voltage Park earns spread where debt-financed peers earn zero - a structural cost moat as GPU prices deflate
- $500M+ combined ARR against a $2.5B+ merger valuation is roughly 5x ARR - a fraction of CoreWeave's and Nebius's public revenue multiples, implying substantial re-rating room if growth persists
- The merger fixes each side's weakness: Lightning had software and developers but rented compute (its CEO was a Voltage Park customer from March 2024); Voltage Park had chips but no up-stack differentiation - together they mimic the full-stack model CoreWeave paid for via Weights & Biases
- 36,000+ owned GPUs including GB300s ranks #3 among neoclouds by chips deployed (Forbes), with frontier customers (Cursor, Reflection) validating infrastructure quality
- Inference era favors owned, well-utilized mid-scale fleets over hyperscale training monoliths: steady, diversified serving workloads improve utilization economics
A subscale, H100-heavy landlord in a brutally deflating commodity market, with opaque financials, an awkward nonprofit-linked cap table, and a merger whose $500M ARR headline cannot be decomposed or verified. The structural cost advantage is a one-time grant, not a repeatable capital engine.
- 60MW active is orders of magnitude below where the frontier is going; CoreWeave and hyperscalers contract gigawatts - Voltage Park cannot serve the largest training deals that anchor the sector's economics
- The grant advantage does not compound: the next 36,000 GPUs must be bought with earned cash or raised capital at market cost, at exactly the moment GPU rental prices are deflating industry-wide
- H100 fleet obsolescence: much of the installed base is 2023-vintage H100s whose market rate has collapsed; depreciation on ~$500M of aging silicon may consume the apparent pricing edge
- $500M+ 'combined ARR including GPU rentals' is an unaudited private figure mixing rental commitments with software subscriptions - neocloud 'ARR' has historically overstated durable revenue, and the Voltage Park/Lightning split is undisclosed
- Nonprofit-anchored governance (Navigation Fund's significant stake, profits designed to flow to philanthropy) plus VC investors (Coatue, Index, Bain Capital, NVIDIA on the Lightning side) creates mixed incentives for the aggressive capital raising this market demands
- No disclosed margins, utilization, or customer concentration - the absence of every number that matters is itself a bear signal in a sector where public comps disclose all three
What it is worth
Merger-negotiated valuation cross-checked against public neocloud revenue multiples (CoreWeave CRWV, Nebius NBIS as the comp set) and private marks (Lambda at $5.9B, late 2025)
sub-$1.5B: GPU rental deflation compresses the legacy fleet's earning power, ARR churns at reserved-contract renewals, and expansion stalls at ~60MW for lack of repeatable capital
~$2.5-3.5B
merger valuation holds; steady rental growth offsets H100 price deflation; software attach improving but unproven
$5-7B+ (merged Lightning AI)
ARR compounds past $1B with software-weighted mix and new power capacity; re-rates toward public neocloud multiples in an IPO or late-stage round
The only price is the Jan 2026 merger: $2.5B+ combined value on $500M+ combined ARR, roughly 5x ARR. Public neoclouds have traded materially above that on revenue multiples, but with audited financials, gigawatt pipelines, and disclosed backlogs - the private discount reflects opacity, subscale power (~60MW), and H100 fleet age. Reverse-read: ~5x ARR implies the merger parties priced meaningful ARR durability risk; sustaining high growth with a software-shifting mix would justify re-rating toward public comps. Standalone Voltage Park financials were never disclosed - all valuation math is at the merged-entity level. Not financial advice.
SWOT
Strengths
- Owned, grant-funded GPU fleet (36,000+ H100/B200/GB300) — little to no financing drag vs debt-financed peers, enabling durable price leadership
- Third-largest neocloud by chips deployed (Forbes) with 3.2 Tbps InfiniBand bare-metal clusters across Tier 3+ data centers
- Post-merger software layer (Lightning AI, 400K+ developers, PyTorch Lightning ecosystem) differentiates from pure capacity resellers
- Real enterprise and AI-native logos — Cursor, Higgsfield, Reflection, Infineon, Monks (per Forbes); earlier Imbue and Character.AI (2023 launch cohort)
Weaknesses
- No disclosed standalone financials - revenue quality, margins, and utilization are unverifiable from outside
- 60MW active capacity is small vs CoreWeave (gigawatt-scale contracted) and hyperscalers; scale ceiling without new capital
- H100-heavy legacy fleet faces rapid depreciation and $/hr price erosion as Blackwell/Rubin supply grows
- Unusual nonprofit-linked governance (Navigation Fund significant stake, profits designed to flow to philanthropy) may complicate future capital raises or an IPO path
- Post-merger integration risk: hardware operator culture merging with an open-source software company
Opportunities
- Inference demand shift — serving workloads on owned capacity via Lightning's model-serving stack is a higher-margin, stickier business than raw rental
- Enterprise AI adoption expanding the buyer pool beyond AI-native labs (Infineon, Monks are early evidence)
- Blackwell/GB300 refresh lets the company reprice upward and win frontier training jobs
- Consolidation of the fragmented long-tail GPU market (the TensorDock playbook)
Threats
- GPU rental price deflation as hyperscaler and neocloud supply catches up to demand; per-hour H100 rates have already fallen sharply industry-wide
- CoreWeave (CRWV), Nebius (NBIS), Crusoe, and Lambda all scaling faster with public-market or debt capital access
- Hyperscalers (AWS, Azure, Google Cloud) bundling GPU capacity with enterprise agreements and custom silicon (Trainium, TPU) undercutting merchant GPU clouds
- Customer concentration typical of neoclouds - loss of one large reserved-capacity tenant materially dents ARR
- NVIDIA allocation dependence: access to GB300-class supply is gated by one vendor's priorities
Moats, dependencies & bottlenecks
Moats
Real but one-time; does not extend to Blackwell-era expansion capex
400M+ downloads, 400K+ platform users) Post-merger; open-source funnel to paid compute is the strongest long-term asset
InfiniBand bare-metal cluster engineering at multi-thousand-GPU scale Table stakes among top-10 neoclouds; differentiates only vs long-tail providers
Switching costs on reserved contracts and managed Kubernetes environments Rental workloads are portable; software attach post-merger is what would deepen this
Dependencies
Sole GPU supplier (H100/B200/GB300) plus InfiniBand networking; allocation and pricing power sit entirely with NVIDIA - which is also an investor in Lightning AI
GPU server hardware for the H100 clusters
Deployed and manages the original large H100 cluster
Multi-tenant storage layer across the fleet
capital/governance Founding $900M grant (fund assets ~$1.25B); retains significant equity in merged Lightning AI - future capital strategy depends on its posture
4 US states at merger; incl. Centeris) ~60MW active; expansion gated by grid interconnect availability, the sector's binding constraint
Neocloud revenue is typically concentrated; standalone concentration undisclosed
Advantages
- Debt-free founding fleet enables below-market pricing with positive unit economics
- Third-largest neocloud GPU count (Forbes) with frontier-grade 3.2 Tbps InfiniBand clusters
- Merged software layer (Lightning) converts open-source developer mindshare into compute demand
- Bare-metal access with a managed-Kubernetes option - genuinely developer-friendly vs hyperscaler friction
- US-only footprint simplifies compliance for domestic enterprise and government-adjacent workloads
Weaknesses
- No disclosed standalone financials - revenue quality, margins, and utilization are unverifiable from outside
- 60MW active capacity is small vs CoreWeave (gigawatt-scale contracted) and hyperscalers; scale ceiling without new capital
- H100-heavy legacy fleet faces rapid depreciation and $/hr price erosion as Blackwell/Rubin supply grows
- Unusual nonprofit-linked governance (Navigation Fund significant stake, profits designed to flow to philanthropy) may complicate future capital raises or an IPO path
- Post-merger integration risk: hardware operator culture merging with an open-source software company
Bottlenecks
- Power and data center capacity: ~60MW active is the hard ceiling on revenue until new sites energize
- Access to next-gen NVIDIA allocation (GB300/Rubin) without hyperscaler-scale purchase commitments
- Capital for fleet refresh now that the grant is deployed - must earn or raise at market cost
- Post-merger integration bandwidth: unifying Voltage Park bare metal with Lightning's software control plane
- Sales motion transition from price-led rentals to enterprise full-stack contracts
Top signals & trends
Top signals
First test of whether $500M+ combined ARR is durable and software-weighted
Expansion is the single biggest unlock; silence implies capital constraint
Confirms NVIDIA treats the merged entity as a priority neocloud partner
Continued deflation compresses the legacy fleet's earning power
Enterprise mix means longer contracts and lower churn
Would mark the real market value beyond the merger's negotiated $2.5B+
Sets the comp band for any future Lightning AI raise or IPO
Trends
Favors utilization-optimized owned fleets with serving software - exactly the merged stack's pitch
Compresses commodity rental margins; pushes value up-stack to software
Voltage Park has been an acquirer (TensorDock, March 2025) and a merger participant; scale via M&A beats organic in a capital-intensive sector
60MW active is small; grid interconnect queues slow expansion for everyone below hyperscale
The merger is a direct bet on this trend, mirroring CoreWeave's Weights & Biases acquisition
Long-run demand risk for all NVIDIA-only neoclouds
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
H100/B200/GB300 GPUs and InfiniBand networking - sole compute supplier; also a Lightning AI investor
GPU server hardware
Multi-tenant storage platform
Cluster deployment and management services
AI coding startup; named merged-entity customer (Forbes)
Open-source frontier lab; its CEO publicly endorsed the merger
AI video generation startup (Forbes)
Enterprise semiconductor customer (US ADR; primary listing Frankfurt: IFX)
Marketing-services arm of LSE-listed S4 Capital
Anchor tenant on the original H100 cluster at the Oct 2023 launch
Early tenant - cluster being finalized at the Oct 2023 launch, alongside Atomic AI
Largest pure-play neocloud; gigawatt-scale, debt-financed, public since March 2025; also went full-stack via Weights & Biases
Public (Nasdaq) AI cloud, #2 neocloud by chips per Forbes; large Blackwell fleet and multi-billion-dollar Microsoft and Meta capacity contracts
Private GPU cloud with developer focus; $5.9B valuation (late 2025), NVIDIA-backed, reported IPO prep for H2 2026; closest model overlap
Private; energy-first data center builder (Abilene/Stargate); competes for the same AI-native tenants
Hyperscaler GPU capacity plus Trainium custom silicon; wins enterprise bundles
Largest AI cloud by revenue; OpenAI anchor; enterprise distribution Voltage Park lacks
TPU + GPU capacity; aggressive AI-startup credits program
Major merchant GPU capacity supplier to AI labs (incl. Stargate participation)
Private; inference/fine-tuning platform - competes at the software layer post-merger
Long-tail developer GPU clouds competing with the TensorDock marketplace segment