ChatGPT-5 Pro Investigation

What does each Legit.Health diagnostic report cost to deliver?

By ChatGPT-5 Pro, November 2025

A cost model grounded in Spanish regulatory compliance, cloud pricing, staffing, and ML lifecycle realities

Executive summary

Per-report delivery cost is dominated by fixed, compliance, and staffing overheads, not raw GPU seconds. With realistic 24/7 low-latency service, ENS compliance, MDR/QMS upkeep, retraining, and staffed support, a defensible cost floor lands in the €1–€7 per report band depending on annual volume. At ~200k reports/year, a typical composite cost is ≈€1.6/report. At lower volumes (≈50k/year), fixed costs dominate and push the unit cost toward ≈€6.5/report. Variable cloud costs per report (GPU + storage + logging + egress) are cent-level.

Numbers below are transparent, with formulas you can swap for your actuals.

Why a software report still carries meaningful OPEX

Regulatory and operational obligations in Spain/EU create structural costs:

• EU MDR + QMS + software lifecycle (IEC 62304) + risk management (ISO 14971). MDR requires a QMS and post-market surveillance; IEC 62304 and ISO 14971 processes must be maintained through the software lifecycle. Notified-body audits and surveillance recur. These are ongoing costs, not one-off fees.
• ENS (Esquema Nacional de Seguridad) for public-sector workloads in Spain mandates a security audit at least every two years and certification for Medium/High categories, plus logging, traceability, incident response, continuity, and continuous improvements.
• AEMPS commercialization obligations (national registers, transitional CCPS notices under RD 192/2023) add administrative overhead.
• Real-time (<1 s) inference needs always-warm capacity or provisioned concurrency to avoid cold starts and meet payload limits; this raises baseline compute and ops cost.
• Dermatology pipelines are multi-model (classification + multiple severity/segmentation detectors + quality check), which increases compute, validation, and monitoring footprint relative to a single classifier. Segmentation models (e.g., U-Net variants) are compute-heavier than classifiers and require more rigorous validation.
• Data drift and domain shift in medical imaging are real; safety requires monitoring and periodic retraining/updates. That is recurring spend on compute, data ops, and labeling/clinical validation.

Large images also affect payload and storage. iPhone clinical images can be 5–25 MB for HEIF/ProRAW at 12 MP and ~75–100 MB at 48 MP; designing for ~20 MB per image is reasonable.

Cost model: variables and formulas

Let:

• $V$ = reports per year.
• $M$ = number of ML models per pipeline (≥3 here: diagnosis, severity, quality).
• $p_h$ = on-demand GPU price per hour (€) (EU-Spain L40S example below).
• $R$ = sustained reports/second per GPU at target latency.
• $\rho$ = average utilization of that GPU (0-1).
• $s$ = average stored image volume per report (GB).
• $m$ = retention in months.
• $P_{\text{S3}}$ = object storage price per GB-month (€).
• $e$ = average data egress per report (GB).
• $P_{\text{egress}}$ = price per GB data transfer out (€) (after free tier).
• $g_{\text{base}}$ = annual "always-on" GPU capacity (hours/year) to guarantee <1s and eliminate cold starts.
• $C_{\text{staff}}$ = fully-loaded annual staff cost (support, MLOps, data science).
• $C_{\text{compliance}}$ = annualized MDR/QMS/ISO 14971/IEC 62304 and ENS certification, audits, PMS, vulnerability scanning, SIEM/SOC, etc.
• $C_{\text{retrain}}$ = annual retraining compute + data ops + validation.
• $C_{\text{tools}}$ = observability/logging/security tooling (CloudWatch, CloudTrail, scanners) not captured elsewhere.

Per-report total cost

$$C_{\text{report}}=c_{\text{infer}}+c_{\text{storage}}+c_{\text{egress}}+c_{\text{logs}}+\frac{C_{\text{fixed}}}{V}$$

with

$$C_{\text{fixed}}=C_{\text{staff}}+C_{\text{compliance}}+C_{\text{retrain}}+C_{\text{tools}}+p_h\cdot g_{\text{base}}$$ $$c_{\text{infer}}=\frac{p_h}{3600;R;\rho} \qquad c_{\text{storage}}=s\cdot P_{\text{S3}}\cdot m$$ c_{\text{egress}}=\max!\big(0,, e \cdot P_{\text{egress}} - \text{free_tier_credit}\big),\quad c_{\text{logs}} = \text{ingested_GB_per_report} \times P_{\text{logs}}

Notes:

• GPU price examples (On-Demand): Europe (Spain) g6e.2xlarge (NVIDIA L40S) ≈ $2.36/hour; On-Demand rates vary by region.
• Storage price example: S3 Standard published example $0.023/GB-month (region-specific tables apply).
• Data transfer: AWS gives 100 GB/month free data transfer out; then regional per-GB rates apply.
• Cloud logs: CloudWatch logs tiered pricing starts at $0.50/GB ingested; tools and retention choices dominate cost.
• Real-time endpoints and provisioned concurrency to avoid cold starts are recommended when latency is tight or traffic bursts occur.

Plugged-in example values

These are explicit assumptions you can replace:

Variable per-report components

• Inference compute: L40S (g6e.2xlarge) €2.17/h (assume 0.92 €/$). With three models, conservative sustained **R** between 1-5 reports/s/GPU and utilization$\rho$ between 0.2-0.8 depending on day-night demand and strict sub-second latency.

  Example values:

  • $R=1, \rho=0.2 \Rightarrow c_{\text{infer}}\approx 0.0030$ €
  • $R=5, \rho=0.5 \Rightarrow c_{\text{infer}}\approx 0.00024$ €

  Even at low utilization, compute per report is < 1 cent; the expensive part is the always-on capacity to hit <1 s latency and multi-model pipelines.

• Storage: 3 images × 20 MB ≈ 60 MB/report = 0.0586 GB. With S3 Standard $0.023/GB-month, retain 12 months:

  $$c_{\text{storage}}=0.0586 \times 0.023 \times 12 \approx \$0.0166 \approx 0.015 \text{ EUR}$$

  Storage is ~1.5 euro-cents/report/year at these volumes.

• Egress: Often negligible because the result payload is small (JSON/PDF). Also the first 100 GB/month out is free across services. For scale planning, use $c_{\text{egress}} \approx e \cdot P_{\text{egress}}$ after the free tier.

• Logs/observability: If you ingest ~50 KB/report into CloudWatch, ingestion cost is ~€0.00000002 per report (negligible). Real spend comes from retention and high-volume security/audit logs you keep for ENS/MDR evidence.

• Image sizes: 12-MP ProRAW ≈ ~25 MB; 48-MP ≈ ~75–100 MB. Designing for ~20 MB clinical images is aligned with published ranges.

Fixed annual components

• Always-on GPU baseline ($p_h \cdot g_{\text{base}}$): one L40S 24/7 to guarantee sub-second responses (= €2.17/h × 8760 ≈ €19,041/y). Two GPUs for HA/peaks would double this item.

• Staffed support + MLOps + data science (Spain market medians, fully loaded):

  • MLOps engineer median €49.5k base.
  • Data scientist senior average €68.9k base.
  • Technical support engineer €40.9k base. Employer social-security and labor on-costs are typically ~30% (common contingencies 23.6% + MEI 0.67% + other items per BOE 2025), yielding ~€207k/year for one FTE each across these three roles.

• Retraining/ML governance: Assume 360 GPU hours/year on an A100 (p4d.24xlarge) for scheduled retraining and validation = $32.77/h ≈ €30.15/h → €10.9k/y. This is modest; your true need may be higher, especially with drift handling and multi-task severity models.

• Compliance & security: ENS certification requires an audit at least every 2 years (plus surveillance) and broader MDR/QMS/IEC 62304/ISO 14971 lifecycle activities and post-market surveillance. Budget this as $C_{\text{compliance}}$; it is deployment-specific and usually five- to six-figure €/year once you include audits, penetration tests, vulnerability management, SIEM/SOC, legal, and PMS. (Use your actual NB/consulting/SOC contracts.)

• Tooling (CloudWatch, CloudTrail, scanners): treat as $C_{\text{tools}}$; cost depends on retention and breadth of logging. CloudWatch/CloudTrail pricing pages provide the calculable dimensions.

Worked scenarios (swap with your numbers)

Assumptions held constant across scenarios

• GPU baseline: 1× L40S 24/7 → €19,041/y.
• Staffed support+MLOps+DS: €207k/y fully loaded (3 FTEs).
• Retraining: €10.9k/y (360 h A100).
• Compliance & security: €85k/y placeholder (ENS+MDR audits, PMS, SOC/SIEM). (Replace with your contracts).
• Variable per report:
  • Inference compute: €0.001 (conservative upper-bound; see “Variable components” above)
  • Storage: €0.0149/report for 60 MB retained 12 months (S3 Standard).
  • Other (logs/egress): €0.003/report placeholder after free tier.

Total fixed $C_{\text{fixed}}$ = 207k + 10.9k + 85k + 19.0k ≈ 322,051 EUR/y.

Unit cost by annual volume

$$C_{\text{report}} \approx \frac{322{,}051}{V} + 0.001 + 0.0149 + 0.003 \text{ (EUR)}$$

Annual reports (V)	Fixed €/report	Variable €/report	Total €/report
50,000	6.441	0.019	6.46
200,000	1.610	0.019	1.63
1,000,000	0.322	0.019	0.341

Interpretation: compute and storage are cheap; guaranteeing sub-second, audit-ready service with ENS + MDR/QMS + retraining + staffed support drives the economics. At hospital-network scale (≥200k reports/year), unit economics look like €1–€2/report. At small pilots, plan €5–€7/report.

Why sub-second matters to cost

Real-time endpoints must be always hot. SageMaker/real-time guidance sets payload limits (25 MB typical) and discourages cold starts for low latency; provisioned concurrency keeps capacity warm and costs even if idle. That is why we include $p_h \cdot g_{\text{base}}$ as a fixed line.

Why multiple models matter

Three functional blocks—diagnosis (classification), severity (segmentation/detection), quality (IQA classifier)—increase compute and validation load. Segmentation networks (e.g., U-Net and its variants) are computationally heavier than single-label classifiers, and every additional model adds inference time, testing, and drift monitoring surface. $\Rightarrow$ plan capacity for the sum of their latencies.

Why retraining is not optional in dermatology AI

Clinical image distributions shift across devices, lighting, and populations, degrading model performance. Medical AI literature documents data/domain drift and supports prospective monitoring and drift-triggered retraining. That is recurring OPEX, captured as $C_{\text{retrain}}$.

How this compares to “physical” diagnostics

Even though Legit.Health is “software,” the operational stack (ENS audits, MDR/QMS processes, PMS, staffed support, always-on GPU capacity, monitoring, secure logging, and storage) produces per-report costs just like a lab test does. The unit economics at realistic volumes (≈€1–€2/report) are consistent with low-complexity imaging analytics and below common Spanish reference charges for baseline diagnostics (e.g., the user-provided examples “Analítica básica” ≈ €15 and “Radiografía simple” ≈ €25), while still recognizing Legit.Health’s multi-model and real-time requirements.

Implementation checklist for your own calculator

1. Pick your GPU class and region and set $p_h$. For Spain, consider g6e (L40S) (On-Demand price reference above).
2. Decide baseline availability: 1 vs 2 GPUs always-on for HA. Set $g_{\text{base}}$.
3. Measure true throughput $R$ and utilization $\rho$ with your real pipeline and batching safely within latency SLO.
4. Fix storage policy: number of images per report, compression, retention months, and storage class (Standard vs lifecycle to IA/Glacier). Set $s$ and $m$.
5. Set compliance budgets from contracts: ENS audits (≥ biennial), MDR/NB audits, PMS, SOC/SIEM, pentest.
6. Staffing plan: MLOps + DS + Support FTEs with Spain salary medians; apply employer on-cost per BOE.
7. Retraining cadence and hardware (e.g., A100 hours).
8. Observability: logging volume + retention to meet ENS evidence. Use CloudWatch/CloudTrail pricing to add $C_{\text{tools}}$.

Appendix: Sources and quick facts

• ENS obligations: audit at least every 2 years; Medium/High categories require certification; organizational, operational, and protective measures apply.
• MDR software guidance (MDCG 2019-11 Rev.1): qualification and classification of software under MDR/IVDR.
• IEC 62304 & ISO 14971 govern software lifecycle and risk management.
• Real-time ML endpoints: low latency use real-time endpoints; payload limits and warm-capacity recommendations apply.
• GPU pricing examples: AWS g6e (L40S) EU (Spain) ≈ $2.36/h On-Demand.
• Storage example: S3 Standard $0.023/GB-month (region varies).
• Data transfer: 100 GB/month free across AWS services; then per-GB regional pricing.
• Staffing medians in Spain: MLOps €49.5k, Data Scientist €68.9k, Tech Support €40.9k.
• Employer on-costs: common contingencies 23.6% employer share plus other concepts including MEI 0.67% (total ≈ ~30% typical burden).
• Dermatology image sizes: 12-MP ProRAW ≈ ~25 MB; 48-MP ≈ ~75–100 MB.
• Drift & retraining: data drift is recognized in clinical AI; monitoring and update strategies mitigate degradation.

Takeaway

Use the formulas to convert your contracts and real throughput into a defensible €-per-report cost floor. The levers that matter are:

• Volume $V$ and SLOs (drive fixed cost dilution).
• Baseline capacity $g_{\text{base}}$ for <1 s.
• Compliance scope $C_{\text{compliance}}$ and staffing $C_{\text{staff}}$.

Raw GPU seconds are cheap; being clinical-grade, audit-ready, and sub-second at scale is not.