# AI infrastructure purchase and ROI worksheet

Original Ryuc worksheet, prepared 31 October 2026. Numbers entered here are decision assumptions until validated.

## 1. Specify useful work

- Workflow and business owner:
- Eligible task population and observation period:
- Quality and safety acceptance criteria:
- Baseline completion rate, latency, review and repair effort:
- Business benefit: cash saving, useful capacity, revenue or service improvement:
- How finance and the operational owner will validate that benefit:

## 2. Locate spending and constraints

- Owned, rented, reserved or provider-hosted capacity:
- Commitment period, cancellation rights and alternative uses:
- Model/API subscriptions and usage-based bills:
- Data ingestion, retrieval, storage and network costs:
- Engineering, operations, review and repair:
- Actual utilization metric, sampling interval and hardware scope:
- Observed bottleneck and evidence supporting it:

## 3. Compare options

For each option, document accepted outcomes, quality by task category, service levels, full operating cost, one-time transition cost and incremental cost over the actual decision horizon.

Cost per accepted outcome = all allocated costs for attempted work / accepted outcomes.

Include failed attempts and retries. Show an unavailable or zero denominator as undefined, not zero cost.

Options to assess:

- Improve existing workflow:
- Hosted model/API:
- Self-host on available capacity:
- Replace or consolidate platform:
- Stop or defer:

## 4. Stress the business case

- Lower adoption:
- Lower acceptance rate:
- More human review:
- Longer migration:
- Higher usage or engineering cost:
- Commitment renewal versus a temporarily sunk reservation:

Do not compare a marginal cost with a full cost without labeling the difference.

## 5. Record the decision

- Chosen option and reason:
- Evidence that could reverse the decision:
- Pilot scope and spending limit:
- Responsible engineering and business owners:
- Stop conditions:
- Review date:

## Headline audit

For any utilization statistic, record the exact numerator, denominator, sample, dates, screening rules and whether it measures organizations, hardware, allocation, engine activity or accepted work. A respondent share is not a share of global GPU capacity.