CFO WORKSHEET · 36-MONTH TCO · 5 min
Local vs Cloud AI TCO Worksheet
Three-year total-cost-of-ownership comparison for a hospital AI scribe deployment — cloud vendor fees versus on-prem CapEx plus operating cost (power, ops labor, maintenance, depreciation). Anchored on the same evidence the WalledCare directory uses; nothing leaves your browser. Adjust inputs to match your specific deployment.
The worksheet
| Year | Cloud annual | On-prem annual | Cloud cumulative | On-prem cumulative |
|---|
What's in the model
The cloud side is straightforward: fee × clinicians × 12, escalated annually. The on-prem side has more components and that is the substance of the exercise — most procurement committees underestimate one or two of them and the comparison comes out wrong as a result:
- checkCapEx, amortized. The $120K hospital-scale build divided over a 4-year depreciation schedule equals $30K/year. The capex shows up on the balance sheet on day one; the operating cost smoothes it out for the comparison.
- checkImplementation, year one only. Integration, EHR connectors, security review, clinician training — typically $30–80K for the first deployment. Excluded from year 2-3.
- checkPower and cooling. ~$300/month for a 4× H100 node at $0.12/kWh, plus cooling overhead. Often invisible on the IT budget; very visible on the facilities budget.
- checkOps FTE allocation. The line most pilots underestimate. A working private AI stack needs ~0.25–1.0 FTE depending on scale — capacity planning, model approvals, audit log review. Make this explicit and the comparison is honest.
- checkHardware maintenance. Vendor support, refresh budget, spare-parts reserve. Underestimated when the procurement team treats the GPU like a typical 7-10 year hospital server.
What the model deliberately excludes
Three categories that move the answer further toward on-prem but are harder to quantify and easier to overcount, so the worksheet leaves them out:
- closeMulti-workflow leverage. One on-prem stack can serve scribe, document Q&A, discharge drafter, handoff, and private medical search on the same hardware. Each additional workflow on a cloud vendor adds another subscription. Include this manually if the deployment will host more than one workflow.
- closeCompliance overhead reduction. The PIA, cross-border-transfer documentation, and audit choreography for on-prem is materially smaller than for cloud-vendor processing under Quebec Law 25, PHIPA, and HIA. Real cost; hard to quantify cleanly.
- closeVendor risk and switching cost. Cloud vendor pricing, defaults, and product strategy change. On-prem hardware does what it did on day one. Optionality has value; the model doesn't price it.
How to read the result
Three honest framings:
- checkIf the 3-year delta is small in either direction: the architecture decision is being made on the non-TCO criteria — residency, audit control, multi-workflow leverage, vendor risk. The TCO is a tie-breaker, not the answer.
- checkIf on-prem wins by a wide margin: stress-test the ops FTE and maintenance assumptions. Most over-confident on-prem TCO projections under-allocate operational labor.
- checkIf cloud wins by a wide margin: the deployment is below the multi-workflow / multi-year threshold where on-prem starts to compound. A 50-clinician single-workflow deployment running for 24 months will not pay back $120K capex; a 200-clinician multi-workflow deployment for 60 months almost certainly will.
Where this fits in the WalledCare directory
This worksheet pairs with the ROI calculator (labor-time value side of the equation), the hardware sizing guide (capex anchor), and the local vs cloud checklist (the qualitative decision framework). Use them together when the conversation is "what does this cost and which side wins on the math."