PortfolioC3-2

Business of AI · Gallery

Build vs Buy vs Fine-Tune

SIMULATEDVerified Jul 2, 2026

Three paths, one 3-year number each, scored on more than cost. The recommendation is the easy part — the flip condition is the one worth remembering, because you'll re-run this in 18 months.

Same instrument · three industries — pick a use-case to reconfigure the run

1.0M calls

Data sensitivity

Differentiation need

Latency requirement

Team skill

API (usage-based)

Recommended

Pay per call; fastest to value; no control of the model.

3-yr TCO

$164k

Score

70

  • Integration (one-time)$20k
  • Usage · 36 mo$144k

Fine-tune / self-host

Train + host + maintain; most control and differentiation; needs the team.

3-yr TCO

$361k

Score

49

  • Training (one-time)$60k
  • Eval-harness build$40k
  • Hosting · 36 mo$216k
  • Eval maintenance · 3 yr$45k

Buy (license)

COTS product; fast, but lock-in and little differentiation.

3-yr TCO

$440k

Score

40

  • License · 3 yr$360k
  • Integration$50k
  • Lock-in premium (risk)$30k
API (usage-based)at 1.0M calls/mo, this data sensitivity, and this team.

Flip condition — Fine-tune / self-host wins if volume roughly triples (self-host amortizes) or differentiation need rises.

Know the flip, not just the answer

Cross the volume slider slowly and watch API and fine-tune trade places — the crossover is the whole decision. Buy wins on speed when differentiation is low; it loses the moment the capability becomes your edge.

Steering-committee takeaway: This decision is re-made every 18 months. The evaluator matters less than knowing your flip conditions.

How this is built & assumptions

TCO (3 yr): API = integration + volume × 36 × $0.004/call. Fine-tune = training + eval-harness + hosting (scales with volume) + eval maintenance. Buy = license + integration + a lock-in risk premium.

Score = weighted blend of cost (inverse TCO), speed, control, differentiation, and risk (cost 35% · diff 20% · others 15% each). Data sensitivity and latency shift control/risk; differentiation need scales the diff weight; team skill gates fine-tune feasibility.

Stack: Next.js (static) + shared design system; deterministic client-side.

Limitations: rates and line items are illustrative defaults; real TCO needs your negotiated pricing and utilization. It frames the decision and its sensitivity, not a procurement quote.