# AI Factory Economics — Tokens, Utilization, PUE/WUE

Source: https://savrn.com/institute/courses/aio-2905
Markdown alternate of the page above; the site index is https://savrn.com/llms.txt

---

Advanced · AIO

Proposed course specification. Delivery, enrollment, and assessment arrangements require a partner program. Credential references describe intended alignment or pathways; external credentials and academic awards are issued only by the responsible organizations.

60Total hours

28Lecture

32Hands-on lab

AIO 2904Prerequisite

—Proposed credential / alignment

32 lab hours 28 lecture hours

This course covers the economic and efficiency metrics used to evaluate an AI factory's operating performance. Students calculate Power Usage Effectiveness (PUE), defined by The Green Grid as total facility energy divided by IT equipment energy, and Water Usage Effectiveness (WUE), defined by The Green Grid as annual site water usage in liters divided by annual IT equipment energy in kilowatt-hours, from facility metering data. The course covers GPU utilization and idle-time analysis, capacity-factor calculation for a cluster over a billing period, and tokens-per-second and cost-per-million-tokens as the primary unit economics of an inference-serving business. Students build a rack-level and site-level cost model incorporating power cost, PUE overhead, depreciation, and labor to produce a fully loaded cost-per-token figure, and compare that figure against representative market pricing to assess margin. The course closes with a capacity-planning exercise sizing GPU fleet additions against forecast demand and power availability. This course is quantitative and spreadsheet-based rather than hands-on hardware, extending the technical operations skills built in prior AIO courses into the financial language used by site operators and investors.

## What you'll be able to do

1. Calculate PUE from facility metering data over a defined billing period.
2. Calculate WUE from annual site water usage and IT equipment energy data.
3. Calculate GPU cluster capacity factor and idle-time percentage from DCGM utilization telemetry over a billing period.
4. Calculate tokens-per-second throughput and fully loaded cost-per-million-tokens incorporating power, PUE overhead, depreciation, and labor.
5. Build a rack-level cost model comparing GB300 NVL72 and VR200 NVL72 deployments on a fully loaded cost-per-token basis.
6. Compare a calculated cost-per-token figure against representative market inference pricing to assess gross margin.
7. Size a GPU fleet capacity addition against forecast demand and available site power.

← Previous course

Model Serving, Inference Ops & Observability

Next course →

Responsible AI & Sovereign Compute Controls

### Department

Applied AI & Compute Operations

9 courses · 860 hours →

## Train the team that runs the factory.

Shape a program around your facility, your learners, and the competencies employers need.

[Plan a Program →](https://savrn.com/institute/plan) [Explore programs](https://savrn.com/institute/programs) [Read catalog](https://savrn.com/institute/catalog)