Owned / financed GPUs
Not pure software brokerage; the asset sits on a balance sheet or SPV.
F.F Research · AI · 31 July 2026
A three-year rental model: owned GPUs, leased power, sold hours—and cash payback that ignores depreciation mythology.
Thesis
Article architecture
Boundary, stack position, bare metal versus managed GPU cloud.
Hardware, hosting, team, revenue—then the adult metrics.
36-month workbook, EBITDA trap, four ramps, market edge.
Chapter 01
Separate the chip marketing from the hour that appears on the invoice.
Chapter 1 · 1.1 Definition
Not pure software brokerage; the asset sits on a balance sheet or SPV.
Wholesale AI colo or JV campuses supply kW; few own the full stack.
IB / high-radix Ethernet sized for multi-node jobs, not commodity VMs.
Reserved clusters and take-or-pay blocks beat infinite on-demand myths.
Chapter 1 · 1.2 Stack position
Chapter 1 · 1.3 Product shape
Nodes or NVL racks to the customer. Customer brings the runtime. Price densest and most underwritable. Attach is storage, egress, support.
K8s / slurm / serverless wrappers. Higher attach—and support cost. Packing gains are easy to overclaim. Model occupancy explicitly, not in a black box.
Chapter 02
Force every line into hardware, hosting, team, or revenue—or you will double-count power.
Chapter 2 · Structure
Chapter 2 · 2.5 Metrics
Not modeled: interest, advance rates, take-or-pay floors, residual sale, taxes.
Chapter 03
The Blog view hosts the live workbook. Presentation freezes the structure and default story.
Chapter 3 · Monthly workbook
Deployed × hours × occupancy × commercial ramp.
GPU rental + attach; locked price or annual decay.
SLA credits and bad debt on gross, not opex fiction.
Base rent, metered energy, admin & connectivity.
Spares, insurance, commission, payroll, G&A.
Then depreciation and NRC amortisation.
Occupancy-independent cash drag.
EBITDA − capex − NRC in purchase months.
Chapter 3 · Default case
H100-class ~$30k, $1.68/GPU-hour locked, 256 GPUs, 36-month life, 15% ancillary, ordinary colo at ~$120/kW-month and $0.06/kWh, three-person ops pod.
Strong mid-period EBITDA margin once full—while cumulative free cash can stay negative inside 36 months.
That tension is the NeoCloud problem statement, not a calculator defect. Open Blog → section 3 to edit every assumption live.
Chapter 04
Path dependence beats terminal ROI slogans.
Chapter 4 · 4.2 Path dependence
Chapter 4 · 4.3 Sensitivities
| Lever | Helpful direction | Why |
|---|---|---|
| GPU capex | Lower / secondary | Dominates invested capital |
| $/GPU-hour & lock | Higher, longer | Revenue density + decay shield |
| Occupancy & fill | Higher, faster | Fixed colo and payroll leverage |
| $/kW rent & energy | Lower | Largest cash opex block |
| Deploy gap | Shorter | Less pre-revenue burn |
| Depr. life vs lease | Aligned | Kills fake accounting returns |
Chapter 05
Speed and reserved fabric still clear—until the generation and the financing disagree.
Chapter 5 · Edge versus fragility
Faster multi-node cluster delivery. Reserved training fabrics. Contract sizes hyperscalers ignore. Power not yet absorbed upstream.
List-price compression from hyperscalers. Customer concentration on take-or-pay. Advance rates that lag peak funding. Accounting life longer than economic life.
Conclusion
NeoCloud bare metal is financed hardware rental with a power landlord and a short technological half-life. Prefer invested capital, peak funding, cash payback, and accounting payback when life equals lease—then open the Blog calculator and break your own quote.