Built for the market below 30 MW.
The Compute Foundry turns small and mid-sized data centers — from 0.25 MW edge rooms to 30 MW regional campuses — into bankable, SLA-backed AI compute. We work where large infrastructure capital doesn’t reach, and where owners can’t go alone.
Three tiers, 0.25–30 MW.
A light-asset service at the edge, a financed full stack in the 2–15 MW core, and tranche-financed campuses up to 30 MW.
Edge & SME
Software, network tuning and orders on the owner’s existing GPUs. No hardware financing.
- GPUs
- 128 – 1,024
- New hardware
- $0
- Owner added net / yr
- $0.41M – $3.26M
- TCF revenue
- 15% platform fee
- Share of portfolio
- ~10%
Mid-Market single sites
Transformers in place, no 3.2T engineering, no low-cost capital. The full stack, one site per SPV.
- GPUs
- 1,024 – 5,000
- Hardware (SPV)
- $33.8M – $260M
- Cluster revenue / yr
- $15.2M – $115M
- Owner income / yr
- $2.65M – $21.3M
- DSCR
- 1.23× – 1.29×
- Share of portfolio
- ~70%
Regional campuses
Blackwell-class liquid-cooled builds, financed in rolling tranches of 1,024-GPU units.
- GPUs
- 5,000 – 10,000
- Hardware (SPV)
- $260M – $520M
- Cluster revenue / yr
- $115M – $231M
- Owner income / yr
- $21.3M – $42.6M
- Tranches
- 5–10 × ~$53.2M
- Share of portfolio
- ~20%
Illustrative, pre-tax. MW at the meter. GPU counts assume H200-class servers (~1.96 kW/GPU) at 2 MW and Blackwell-class liquid-cooled racks (3.0 kW/GPU) from 15 MW. Owner income for financed tiers = hosting fee + 20% of cash after debt service. Full assumptions on the Data Center Partners page.
Why we stop at 30 MW.
Below the minimum ticket of large infrastructure capital
Large infrastructure funds and bank syndicates typically start at 30–50 MW+. Owners below 30 MW have power and buildings but limited access to low-cost financing and AI delivery track records — that gap is where our structure adds the most value.
Built to repeat
Every financed site is built from the same 1,024-GPU unit — about $33.8M (H200-class, ~2 MW) to $53.2M (Blackwell-class, ~3 MW). A 2 MW room and a 30 MW campus differ in unit count, not in structure.
Risk you can size
Below 30 MW, a site’s output can be placed with contracted offtake and diversified workloads, and financed unit by unit. We avoid single-site exposure where one empty cluster can swamp the cash flow.
What we bring to a site.
Six capabilities, one operating model — delivered with specialist partners under a single contract stack.
Power headroom
Identify and contract behind-the-meter capacity at industrial sites, so new compute skips the interconnection queue.
Network engineering
3.2T RoCE design, PFC/ECN congestion tuning and driver/firmware unification across mixed sites.
Throughput operations
Scheduling, inference-stack tuning, telemetry and SLA management — delivered as tokens, not machines.
Structured finance
Node-level SPVs, equipment debt, direct-to-OEM procurement and controlled cash accounts.
Offtake & price protection
Term contracts with AI customers, plus hedging structures with market counterparties to set a price floor.
Lender transparency
A look-through portal showing GPU health, power draw and collected cash in real time.
How we build the portfolio.
Target mix by megawatts. Each tier has its own entry rules.
Core single sites
The main SPV carriers: one to five 1,024-GPU units per site, with behind-the-meter power and take-or-pay offtake. Nearby 2–5 MW sites on one grid may share a unit.
Scale-up campuses
Entered only after earlier units have a delivery record. Financed in tranches, so capital follows contracted demand.
Light-asset service
No hardware financing — software, scheduling and orders for a platform fee. A pipeline for future financed units.
Projects in North America and Southeast Asia.
Industrial power, converted sites
Behind-the-meter capacity at industrial sites and converted data centers, including ERCOT-connected locations in Texas.
Industrial-zone data centers
Small and mid-sized facilities in industrial zones serving regional AI inference and rendering demand.