All case studiesAI Infrastructure · Compute Economics

[e.g., "One asset, two economies: training vs inference on a GPU cluster"]

[one line: the same fleet underwritten two ways — capex-heavy training vs throughput-driven inference]

[e.g., "One asset, two economies: training vs inference on a GPU cluster"]
01

Project Snapshot

Asset
[GPU cluster, e.g., 10k accelerators]
Mode
[Training / Inference toggle]
Capex
[$/GPU]
Power
[MW, PUE]
Useful life
[obsolescence assumption]
02

The Challenge

[The same fleet has two different economic profiles — training is capex- and utilization-driven with obsolescence risk; inference is throughput- and cost-per-token-driven. How do you underwrite an asset whose returns depend on its use case?]

03

The Approach

  1. 01

    [Build the shared cost base: capex, power, PUE, depreciation/obsolescence]

  2. 02

    [Training case: cluster utilization, time-to-train, cost per training run]

  3. 03

    [Inference case: throughput, cost per token, utilization, opex]

  4. 04

    [Add a toggle/scenario switch]

  5. 05

    [Sensitize to the GPU obsolescence curve]

04

Inside the Model

[Describe the toggle switching training/inference outputs.]

[ Image placeholder — Toggle view: training vs inference unit economics ]
05

Results at a Glance

[$/training run][Cost / 1M tokens][Fleet utilization][Payback][IRR by case]
06

What This Demonstrates

  • GPU/accelerator economics
  • Capex vs opex underwriting
  • Scenario/toggle modeling
  • Depreciation & obsolescence handling
07

Key Takeaway

[Idea + mechanism, two sentences.]

Illustrative case built on representative data; not based on any confidential or client transaction.