AI cost: cloud versus owned environment

Measure the work that will happen

Estimate monthly requests, input and output size, operating hours, peaks and concurrency. Include retrieval, storage and human review where they belong in the workflow.

Count people and continuity

An API brings consumption, integration, transfers and supplier terms. An owned environment adds purchase or rental, energy, maintenance, updates, support and idle capacity.

Keep assumptions visible

Model low, likely and high use. Record date, currency, quoted prices and sensitivity to volume or model changes.

Compare quality, latency, residency and operating capability too. The cheapest spreadsheet line may be the wrong architecture.

A comparison sheet without preset prices

For each option, separate monthly fixed cost, usage-based cost and human work. A simple scenario calculation is: period total = reserved capacity + variable consumption + operations + integration + contingency. Show which items are amortised over years and which move when volume doubles. Do not apply the same unit price to models with different task quality.

Add non-financial assumptions: where data is processed, acceptable response time, recovery needs and ease of switching supplier. A useful estimate shows which choice would change if an assumption proved wrong. Without a measured workload and dated prices, any published break-even point would only be illustrative.

Fictional worked example, not market pricing

Imagine two alternatives that have already passed the same quality test, using invented euro amounts solely to show the calculation. The API has €550 in monthly platform and operating fixed costs, plus €0.03 per request. The owned environment has €1,350 monthly fixed costs: €500 amortisation, €700 operations and €150 energy; add €0.01 per request. These are not Hesperia prices, a supplier quote or a client estimate.

Under these fictional assumptions, the lines cross at 40,000 monthly requests: (1,350 − 550) / (0.03 − 0.01). More support, uneven demand or inadequate local model quality moves the crossing point or makes it irrelevant. Model a range of demand and idle capacity as well as an average. AWS’s component cost guidance includes operations and management; Google Cloud’s cost framework distinguishes capital and operating expenditure. These are supplier frameworks, not evidence that either option wins here.

For a real case, replace every number with a dated source and owner, account for taxes or currency where relevant and test against the actual workload. The architecture comparison sets out what to verify beyond cost.