Skip to content
All case studies

Decision brief 02 · Illustrative worked example

Before you own compute,
understand the commitment.

API, rented GPUs or owned hardware? A financial comparison for a hypothetical AI application with stable, repeatable inference demand.

Hypothetical inputs · No customer savings claimed

The decision

Would moving off an API
improve the economics?

A useful review answers three questions: what does the workload require, what does each feasible option cost, and what would change the choice?

This example holds the workload constant, documents assumed prices and separates upfront cash from monthly cost. It illustrates the analytical approach; it does not establish that a GPU meets the workload.

Default scenario · Assumptions, not vendor quotes

1Mrequests per month
1,000 / 500input / output tokens per request
36 monthshardware allocation horizon
720 hoursrental billing per 30-day month

API rates: $0.50 per million input tokens and $1.50 per million output tokens. Rental: $2/hour plus $300/month operations. Ownership: $24,000 upfront plus $850/month operations, comprising $500 hosting/power and $350 support. Hardware residual value is assumed to be zero.

Explore the assumptions

What changes your answer?

All default prices are hypothetical. These inputs are held in your browser; this tool does not submit them.

MONTHLY EQUIVALENT · CONSTANT DEMAND

API$1,250

Usage charges only; add any service or integration fees.

Rented GPU$1,740

720 billed hours plus $300 monthly operations.

Owned hardware$1,517

36-month hardware allocation plus $850 monthly operations.

API has the lowest modeled cost.

The first fixed-cost alternative matches API spending at approximately 1,213,333 requests per month. This is a cost screen. Capacity, quality, service requirements and cash availability still determine feasibility.

Capital commitment and modeled cost over 36 months
OptionUpfront hardware36-month total
API$0$45,000
Rented GPU$0$62,640
Owned hardware$24,000$54,600

Ownership is a monthly allocation, not a monthly payment. The purchase is paid upfront; its cost is counted once in the 36-month total. API and rental deposits, setup fees and prepayments are excluded.

Discuss a comparison using your actual quotes

Interpretation of the default inputs

Start with flexibility.
Revisit the fixed commitment.

01 / CURRENT VOLUME

API wins the initial cost screen.

At one million requests per month, modeled API spending is $1,250, versus $1,740 for rented capacity and $1,517 for owned hardware. Ownership also requires $24,000 upfront.

02 / DECISION THRESHOLD

Demand can change the answer.

Under these assumptions, ownership matches API cost near 1.21 million monthly requests; rental near 1.39 million. Above those points, fixed costs may be attractive if the assumed system can actually serve demand.

03 / NEXT ACTION

Validate before committing.

Request dated quotes and test the same model against representative traffic. Compare model quality, latency percentiles, peak capacity, failover and migration effort before treating a cost advantage as actionable.

How the model works

Make the reasoning
reviewable.

API cost

Requests × [(input tokens × input price + output tokens × output price) ÷ 1,000,000].

Rented cost

Hourly price × 720 billed hours + $300 monthly operating costs.

Owned monthly equivalent

Hardware purchase ÷ 36 months + $850 monthly operating costs. This straight-line allocation is a comparison method, not an accounting or financing schedule.

Comparison boundary

Same model quality, adequate capacity and service fit are assumed for every option and remain unvalidated. Demand is constant for 36 months. Financing, tax, discounting, API integration fees, migration, staff opportunity costs, networking, storage, redundancy, additional capacity and hardware resale are excluded. Actual quotes must define these items before a purchasing decision.

All numbers come from the stated hypothetical assumptions. Performance benchmarks from our separate study are not used to assert that this configuration works.

From an example to your decision

Bring the workload.
Bring the quotes.

A focused review can organize comparable costs, identify missing evidence and define the next validation step.

Request a clarity call