Atlasresearch

Workspace

WorkspacePokedexHPC MapPortfolioCalendarTaiwan RevenueMemory
Calculators
  • 01AI Factory
  • 02CPO Stack
  • 03HBM Bridge
  • 04GPU Cloud TCO
  • 05Inference Tokens
  • 06Power Bottleneck
  • 07Custom ASIC
  • 08OCS TAM
  • 09N3 Allocation
  • 10Optical TAM
Agent APISubscribe
Sign in
Calculator
Atlasresearch

Workspace

WorkspacePokedexHPC MapPortfolioCalendarTaiwan RevenueMemory
Calculators
  • 01AI Factory
  • 02CPO Stack
  • 03HBM Bridge
  • 04GPU Cloud TCO
  • 05Inference Tokens
  • 06Power Bottleneck
  • 07Custom ASIC
  • 08OCS TAM
  • 09N3 Allocation
  • 10Optical TAM
Agent APISubscribe
Sign in
Calculators/GPU Cloud TCO

Jason's Chips

Subscribe or sign in with your Substack email for live inputs, outputs, and saved scenarios.

Subscribe to Jason's ChipsSign in

Inputs

Revenue

Hardware capex

Opex & capital

Last calculated: Not yet

Calculator 04AI cloud unit economics

GPU Cloud Unit Economics Calculator

Convert active AI cloud power into GPUs, GPU-hour revenue, gross margin, economic TCO, payback, and breakeven utilization. The model separates realized contract pricing from public spot pricing and separates billable utilization from customer goodput.

Sensitivity

Fleet GPUs

48.1K

2.08 kW / GPU

Annual revenue

$1.07B

$3.000 realized x 85% billable

Revenue / MW

$10.7M

100 MW active IT base

GAAP gross margin

31.2%

$2.063 GAAP cost / GPU-hour

Economic TCO

$2.748

18% capital recovery over 5.0 yrs

Payback

2.7 yrs

76% economic breakeven util.

GPU-hour bridge

Market price to useful compute

Public anchor

$8.600

65% discount to realized

Realized price

$3.000

Contract / modeled revenue price

Useful-hour price

$3.333

90% goodput efficiency

Economic cost

$2.748

Opex plus capital recovery

Capacity conversion

Critical IT power

100 MW

Facility power

120 MW

Racks

752

Hourly spread

Realized price$3.000
GAAP cost$2.063
Economic cost$2.748

Cost bridge

Annual revenue to gross profit

$336M GAAP gross profit

14%
9%
7%
38%
Line itemAnnual dollars$/GPU-hour% revenue
Revenue$1.07B$3.000100%
Colocation hosting$150M$0.41914.0%
Power pass-through$0.00$0.0000.0%
Support and operations$100M$0.2809.3%
Network, storage, software$78.8M$0.2207.3%
Hardware depreciation$410M$1.14438.1%
Cash gross profit$746M$2.08169.4%
Economic profit after capital recovery$90.2M$0.2528.4%

Sensitivity analysis

Realized GPU-hour price x utilization

$/GPU-hr \ Util.70%78%85% base93%100%
$2.250-6.6%1.6%8.3%13.9%18.7%
$2.6258.7%15.7%21.4%26.2%30.3%
$3.000 base20.1%26.2%31.2%35.5%39.1%
$3.37529.0%34.4%38.9%42.6%45.8%
$3.75036.1%41.0%45.0%48.4%51.2%

This grid keeps hardware, hosting, goodput, useful life, and cost of capital fixed. It isolates the two questions that dominate neocloud return debates: what realized price clears, and how much of the fleet is billable.

Preset comparison

Same active power, different GPU cohort

PresetkW / GPUPublic $/hrRealized $/hrRevenue / MWEconomic TCO
H100/H2001.25$6.160$2.250$13.4M$2.341
B200 HGX2.08$8.600$3.000$10.7M$2.748
GB200 NVL722.08$10.50$3.350$12.0M$2.949
GB300 NVL722.22$12.60$3.900$13.1M$3.437
Rubin NVL722.50$18.00$5.250$15.6M$3.671

Assumptions

B200 HGX

GenerationBlackwell
Provider caseFrontier HGX cohort
Preset noteBlackwell HGX case assumes higher GPU-hour price compresses when translated into revenue per MW.
Hardware capex$2.05B
Capex / GPU$42.6K
Capex / MW$20.5M
Billable GPU-hours358.3M
Goodput GPU-hours322.5M
Public market run-rate$3.08B

Saved scenarios

B200 realized base

42% GM / 3.2 yrs

$16.8M / MW

GB200 scarcity cohort

36% GM / 3.5 yrs

$13.2M / MW

H100 renewal defense

33% GM / 3.8 yrs

$13.4M / MW