F.F

F.F Research · AI · 31 July 2026

NeoCloud Bare-Metal Unit Economics

A three-year rental model: owned GPUs, leased power, sold hours—and cash payback that ignores depreciation mythology.

Thesis

NeoCloud sells financed GPU-hours, not chips

Buy accelerators on a balance sheet Park them on contracted kW Sell billable hours under SLA leakage
  • Product: bare-metal capacity with fabric, not a miniature Azure.
  • P&L: hardware + colo + team versus rental + attach.
  • Trap: high EBITDA while cumulative free cash stays negative.
  • Adult metrics: peak funding, cash payback, life = lease.
  • Tool: interactive 36-month workbook on the Blog view.

Article architecture

Define → decompose → model → stress

  1. 01

    Define the product

    Boundary, stack position, bare metal versus managed GPU cloud.

  2. 02

    Decompose the P&L

    Hardware, hosting, team, revenue—then the adult metrics.

  3. 03

    Model and stress

    36-month workbook, EBITDA trap, four ramps, market edge.

Chapter 01

What a NeoCloud actually sells

Separate the chip marketing from the hour that appears on the invoice.

  1. 1.1 Definition and boundary
  2. 1.2 Position in the AI stack
  3. 1.3 Bare metal versus managed

Chapter 1 · 1.1 Definition

Four traits of the NeoCloud form

Owned / financed GPUs

Not pure software brokerage; the asset sits on a balance sheet or SPV.

Leased power shell

Wholesale AI colo or JV campuses supply kW; few own the full stack.

Training-grade fabric

IB / high-radix Ethernet sized for multi-node jobs, not commodity VMs.

Capacity contracts

Reserved clusters and take-or-pay blocks beat infinite on-demand myths.

Chapter 1 · 1.2 Stack position

NeoCloud is the hour layer between kW and tokens

AI infrastructure layers Five stacked layers from power at the bottom to model platforms at the top, with NeoCloud bare metal highlighted in the middle. POWER · LAND · INTERCONNECTMW · WATER · FIBER WHOLESALE AI COLO / POWERED SHELL$/kW-MONTH NEOCLOUD BARE METAL / GPU CLOUD$/GPU-HOUR · CLUSTER-MONTH · GPU + FABRIC + COLO + OPS HYPERSCALE GPU SKUsMANAGED INSTANCE FAMILIES MODEL / AGENT PLATFORMSTOKENS · APPS · OUTCOMES

Chapter 1 · 1.3 Product shape

Bare metal is the clean unit-economic object

Bare metal

Nodes or NVL racks to the customer. Customer brings the runtime. Price densest and most underwritable. Attach is storage, egress, support.

Managed GPU cloud

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

Three cost layers, one revenue engine

Force every line into hardware, hosting, team, or revenue—or you will double-count power.

  1. 2.1 Hardware balance sheet
  2. 2.2 Data-center hosting
  3. 2.3 Team and G&A
  4. 2.4 Revenue path
  5. 2.5 Metrics not to miss

Chapter 2 · Structure

P&L is a spread after four blocks

Four P and L blocks Hardware, colo, and team feed cost; revenue feeds income; the spread is rental price per hour minus cash cost of the hour. HARDWARECAPEX · AFR · DEPR AIDC HOSTINGRENT · ENERGY · ADMIN TEAM / G&APAYROLL · COMMISSION CASH COSTOF ONE GPU-HOUR REVENUE$/HR · ATTACH · LOCK/DECAY SPREADPRICE − CASH COST − HAIRCUTS DEPRECIATION IS NOT OPTIONAL INTELLECTUALLY—EVEN WHEN LENDERS BUY EBITDA

Chapter 2 · 2.5 Metrics

Adult metrics beyond slogan ROI

Capex Initial capital = GPU + ancillary + NRC
Peak $ Deepest cumulative free-cash trough
Cash PB Payback ignoring depreciation
Acct PB Cum. EBIT covers invested capital
IRR Annualized monthly free cash
Fixed opex Rent, admin, payroll that lag occupancy

Not modeled: interest, advance rates, take-or-pay floors, residual sale, taxes.

Chapter 03

Interactive three-year calculator

The Blog view hosts the live workbook. Presentation freezes the structure and default story.

  1. Assumption panels vs fixed quotes
  2. KPI strip and annual roll-up
  3. M1–M36 revenue and cash opex grid

Chapter 3 · Monthly workbook

One month, top to bottom

Billable hours

Deployed × hours × occupancy × commercial ramp.

Revenue

GPU rental + attach; locked price or annual decay.

Haircuts

SLA credits and bad debt on gross, not opex fiction.

Colo stack

Base rent, metered energy, admin & connectivity.

Ops cash

Spares, insurance, commission, payroll, G&A.

EBITDA → EBIT

Then depreciation and NRC amortisation.

Memo fixed opex

Occupancy-independent cash drag.

Free cash

EBITDA − capex − NRC in purchase months.

Chapter 3 · Default case

The base case is intentionally tight

Inputs that set the mood

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.

What often prints

Strong mid-period EBITDA margin once full—while cumulative free cash can stay negative inside 36 months.

Why ship it anyway

That tension is the NeoCloud problem statement, not a calculator defect. Open Blog → section 3 to edit every assumption live.

Chapter 04

How to read the model without fooling yourself

Path dependence beats terminal ROI slogans.

  1. 4.1 The EBITDA trap
  2. 4.2 Four ramps, one sheet
  3. 4.3 Sensitivity order

Chapter 4 · 4.2 Path dependence

Four clocks that rarely align

Four ramps over 36 months Purchase ramp, deployment gap, occupancy ramp, and price path drawn as staggered schedules on a 36-month axis. M0M18M36 BUY RAMP DEPLOY GAP OCCUPANCY RAMP PRICE PATH STRESS ONE RAMP AT A TIME · WATCH PEAK FUNDING FIRST

Chapter 4 · 4.3 Sensitivities

Levers that usually dominate cash

LeverHelpful directionWhy
GPU capexLower / secondaryDominates invested capital
$/GPU-hour & lockHigher, longerRevenue density + decay shield
Occupancy & fillHigher, fasterFixed colo and payroll leverage
$/kW rent & energyLowerLargest cash opex block
Deploy gapShorterLess pre-revenue burn
Depr. life vs leaseAlignedKills fake accounting returns

Chapter 05

Why NeoClouds still exist

Speed and reserved fabric still clear—until the generation and the financing disagree.

  1. Where demand still routes to NeoClouds
  2. Structural risk: depreciating semis on leased buildings

Chapter 5 · Edge versus fragility

Edge is delivery; risk is the residual

Why buyers still call

Faster multi-node cluster delivery. Reserved training fabrics. Contract sizes hyperscalers ignore. Power not yet absorbed upstream.

What breaks the story

List-price compression from hyperscalers. Customer concentration on take-or-pay. Advance rates that lag peak funding. Accounting life longer than economic life.

Conclusion

Underwrite the hour, not the logo on the GPU

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.