Building a Bottom-Up GPU-as-a-Service Returns Model

Kashish built this model at EY-Parthenon to evaluate the returns a company could earn from operating in the GPU-as-a-service space. The sector was moving fast, and the model needed to be rigorous enough to hold up in front of the engagement partner.

He structured it from the ground up. On the revenue side, he started with chip-level assumptions: which GPU was being modeled, the revenue per chip per hour, and the utilization hours. On the cost side, he worked through the acquisition cost of the GPU, the cost of hosting it in a data center, and the depreciation schedule, which required careful treatment given the high rate of technology obsolescence in the space.

From those inputs, he built through to EBITDA, project IRR, and equity IRR, giving a complete picture of the economics at each level of the capital structure. Every assumption was triangulated against multiple sources before it went into the model.

Presenting and Defending the Model

When he presented to the partner, the cross-examination was thorough. Questions came at every step: why a specific GPU cost was taken at one figure and not another, what sources backed each assumption, and whether the triangulation was sufficient to give comfort. Kashish had prepared for this by sourcing and cross-checking every input two to three times, so he could answer each question with a clear reference rather than a judgment call.

The model held up under questioning and was used to inform the engagement's view on the GPU-as-a-service opportunity.