Health Economics Metrics

Topic 63 of 76 · AI acceleration

Inference Unit Economics

Inference unit economics price AI features by their marginal compute: cost per token, rolled up to cost per transaction, per user, per clinical episode. The defining dynamic: LLM prices have fallen roughly an order of magnitude every 1–2 years at constant capability — a deflation rate with no precedent in health-technology costing.

Why it matters

Two consequences follow from the price collapse. Commercially, an AI feature that is marginal today may be trivially profitable in 18 months — and a competitor priced on today's costs will be undercut. For economic evaluation, any cost-effectiveness model for an AI-enabled clinical service that freezes 2024 inference prices materially overstates ongoing cost — the analysis needs price-decline scenarios the way drug models handle patent expiry and generic entry. (Reference points from the research: frontier output tokens ~$15–75/M in mid-2026, mid-tier models an order cheaper, GPT-4-level capability down from ~$20/M in 2022 to ~$0.40/M; Epoch AI measured 9×–900×/year declines depending on the capability milestone.)

The math

Cost per call     = input tokens × input rate + output tokens × output rate
Cost per unit     = Σ calls per unit of business output (per triage episode,
                    per drafted letter, per consultation summary)

Blended reality   = base call + retries + RAG context (input-heavy)
                    + evaluation/guardrail calls (often 20–50% overhead)

Price-decline scenario for multi-year models:
  cost_t = cost_0 × d^t, test d ∈ {0.3, 0.5, 0.7}/year in sensitivity analysis

Worked example

An AI discharge-summary service: average summary uses 12,000 input tokens (record context) + 1,200 output, plus a verification pass (6,000 in / 300 out). At $3/M in, $15/M out:

Draft:   12,000 × 3/1M + 1,200 × 15/1M  = $0.036 + $0.018 = $0.054
Verify:   6,000 × 3/1M +   300 × 15/1M  = $0.018 + $0.0045 ≈ $0.023
Per summary ≈ $0.077 → per 100,000 summaries/year ≈ $7,700

Against ~20 clinician-minutes saved per summary (≈ £25), inference is
0.25% of the value created — the economics are dominated by everything
EXCEPT the tokens: integration, evaluation, governance, adoption.

That conclusion — inference cost is rarely the binding constraint, at current prices, for high-value clinical tasks — is itself the finding worth carrying into pricing meetings.

Software engineering connection

This is cloud unit economics specialized for AI, with three practice notes: meter per business unit, not per API call, so the number plugs into ICER/budget impact models directly; watch the input/output asymmetry (output typically ~4× input price; RAG architectures are input-heavy — architecture choices are pricing choices); and route by task tier — matching model capability to task difficulty (cheap models for classification, frontier for synthesis) routinely cuts blended cost 5–10× at equal quality, the software version of using the cheapest effective intervention (cost-minimization, equivalence evidenced).

Pitfalls

  • Frozen-price multi-year models — overstates cost; but also assumed-deflation revenue models — a price war is not a contract; scenario both.
  • Ignoring evaluation overhead: guardrails, judges, and retries are real tokens, often the majority in regulated settings.
  • Per-token myopia: latency, rate limits, and context-window constraints carry costs no token price captures.

Sources