Cost versus cloud benchmark¶

OpenMed can turn a measured local PerfReport into a dated cost comparison against published paid cloud tiers. The report is deterministic, contains no source text, and keeps every cloud row linked to the price captured for it. It is planning evidence, not a vendor quote or a guarantee of savings.

What the report computes¶

The input performance report must provide docs_per_second and chars_per_document (either at the top level or in metadata). The hardware model supplies:

{
  "purchase_price_usd": 1200.0,
  "useful_life_hours": 12000.0,
  "power_watts": 45.0,
  "electricity_usd_per_kwh": 0.2
}

For each paid cloud tier, the benchmark reports:

• cloud USD per million characters;
• local USD per million characters, amortizing purchase price over useful life and adding electricity;
• savings per million characters; and
• the character volume at which purchase price is recovered versus the cloud tier after marginal electricity cost, or never when that crossover would fall beyond the hardware's declared useful life.

The formulas use measured characters per second. Free tiers, taxes, support plans, negotiated discounts, storage, data transfer, operator time, request rounding, and financing are deliberately excluded. Those exclusions make the result reproducible but require human review before a business decision.

Run the comparison¶

Store the hardware model in the performance report as hardware_cost_model, or pass it separately with --hardware. Then run:

openmed benchmark cost \
  --perf reports/mobile/perf.json \
  --prices openmed/eval/data/cloud_prices.json \
  --output-dir reports/cost

The command writes cost-vs-cloud.json and cost-vs-cloud.md atomically. Every normalized semantic input contributes to the report's SHA-256 fingerprint; unknown fields and raw text do not. Local model paths are replaced by stable digests. Input files, price rows, strings, numeric ranges, and report sizes are bounded before they are materialized. A price row is rejected when it lacks a safe HTTPS source, capture date, positive normalized price, or explicit verify: true marker, or when its monthly band overlaps another row for the same provider, service, and region.

Paid rows represent marginal monthly bands. The reported cloud USD per million characters and breakeven value are sensitivity calculations at one row's marginal rate; they do not integrate a progressive monthly invoice across all preceding bands. Monthly tier boundaries and lifetime breakeven volume are therefore shown separately.

Committed price snapshot¶

The bundled USD table was captured on 2026-08-21:

AWS's page states that requests use 100-character units and its worked production example publishes the three NERe prices. Microsoft's retail meter publishes $20, $15, $6, and $5 per 1,000 health text records across the paid volume bands; the normalized values divide by the 1,000 characters covered by each record. Microsoft's public pricing page also directs customers above 10 million monthly records to contact sales, so the final retail-API row must not be treated as a negotiated quote.

Re-capture and review the table before a release, pricing statement, or hardware purchase. Set verify to true only after confirming the source, unit, region, tier boundary, and effective date.