What your CI pipeline and AI inference are actually costing you — and how to cut both
Anand Srinivasan
22 Jul 2026 · 7 min read
Most engineering teams have three invisible cost lines: CI compute waste, AI inference waste, and compliance preparation. Here is what each actually costs — and how fixing the first two makes the third significantly cheaper.
The CI compute bill
GitHub charges $0.008/min for a 2-core Linux runner. At scale:
TEAM SIZE RUN FREQ AVG DURATION MONTHLY COST ──────────────────────────────────────────────────────────── 10 engineers 3 pushes/day 8 minutes $576/mo 30 engineers 3 pushes/day 12 minutes $2,592/mo 80 engineers 4 pushes/day 15 minutes $11,520/mo ──────────────────────────────────────────────────────────── macOS runners cost 10× more: $0.08/min Windows runners cost 2× more: $0.016/min
Most teams overpay in two places: full test suites triggered on changes that don't touch tested code, and concurrent PR runs that aren't cancelled when a new commit arrives. EmitCI surfaces both patterns from your workflow history.
The AI inference bill
Teams running AI in CI typically use the model that was convenient when the agent was built, not the one appropriate for the task.
MODEL INPUT OUTPUT TYPICAL TASK ─────────────────────────────────────────────────────────────── gpt-4o $0.0025/1k $0.010/1k code review, reasoning gpt-4o-mini $0.00015/1k $0.0006/1k classification, short outputs gpt-4 $0.030/1k $0.060/1k legacy integrations (avoid) claude-3-5-haiku $0.00015/1k $0.0006/1k fast extraction, summarisation gemini-1.5-flash $0.000075/1k $0.0003/1k lowest cost per token ─────────────────────────────────────────────────────────────── gpt-4o → gpt-4o-mini on short outputs: ~17× cheaper per output token gpt-4 → gpt-4o: ~6× cheaper, similar quality
The bigger problem is repeated inference on unchanged content. A CI bot that re-embeds the same documentation on every run instead of caching on a content hash spends ~75% of its budget on zero-value calls.
Where cost and carbon align — and where they don't
EFFICIENCY SAVINGS — cost AND carbon fall together ────────────────────────────────────────────────────────────── Path filtering (skip full suite on docs/config changes) → 30–50% fewer runner minutes, 30–50% less CO₂ Concurrency cancel on PR push → 40–65% fewer runs for AI-assisted teams (higher commit freq) AI model right-sizing (gpt-4o → gpt-4o-mini for short outputs) → ~90% less CO₂ per inference call, ~17× cost saving Embedding cache (content-hash skip for unchanged files) → ~75% fewer API calls on stable documentation REGION SELECTION — where cost and carbon can diverge ────────────────────────────────────────────────────────────── ap-south-1 (Mumbai) cheapest APAC compute, 713 gCO₂/kWh polandcentral cheapest EU compute, 640 gCO₂/kWh eu-north-1 comparable cost, 7 gCO₂/kWh ────────────────────────────────────────────────────────────── A FinOps migration to cheaper but dirtier compute lowers the invoice and raises Scope 3 by 80–120%. Both show up in your CSRD filing. Cost and carbon need to be read together.
What a 30-engineer team actually saves
Scenario modeller outputs — your numbers depend on run volume, model mix, and current configuration.
LEVER CO₂ SAVING EST. COST SAVING/MO ─────────────────────────────────────────────────────────────── Path filtering on 6 workflows ~35% ~$900 Cancel-on-push for PR runs ~20% ~$520 AI model right-sizing ~90%* ~$600 Embedding cache for doc bot ~75%* ~$200 ─────────────────────────────────────────────────────────────── Combined (moderate scenario) ~40% ~$2,200/mo * Applies to the inference portion of Scope 3 only. Verify against your actual inference volume in the EmitCI run dashboard — token counts vary by agent design.
$26,400 per year. Before the Green Scheduler (10–25% further CI carbon reduction at no additional compute cost) and before any macOS/Windows runner rationalisation.
The compliance cost
CSRD and SB 253 are not optional above the revenue thresholds. The fine is a percentage of revenue, not your CI bill.
REGULATION FINE WHO IS IN SCOPE
──────────────────────────────────────────────────────────────
CSRD Up to 10% of global EU: >500 staff or >€150M
annual turnover revenue (filing now)
Public non-compliant EU: >250 staff or >€40M
company registry (filing from FY2025)
SB 253 Up to $500,000/day CA nexus + >$1B revenue
per violation Scope 3 due Jan 2027
──────────────────────────────────────────────────────────────
A €15M ARR SaaS company: CSRD exposure up to €1.5M/filing yearThe day-to-day cost is preparation. Without a pipeline, collecting emissions data spread across GitHub, AWS, Stripe, and HR systems is a 3–6 month manual project, repeated annually.
COMPLIANCE PREPARATION: MANUAL VS. AUTOMATED
──────────────────────────────────────────────────────────────
Manual With EmitCI
─────────── ────────────
Data collection 3–4 months Continuous (real-time)
CI/CD carbon Estimated Measured per run
Cloud carbon Billing export Enriched per region
Methodology doc Consultant Auto-generated PDF
CSRD ESRS E1 CSV Custom build One-click export
Audit trail None typically SHA-256 per report
External consultant cost (typical, mid-size tech co)
ESG data gap analysis: $15,000–$40,000
Methodology documentation: $8,000–$20,000
Annual GHG assurance (limited): $20,000–$60,000
──────────────────────────────────────────────────────────────
First-year compliance without tooling: $45,000–$120,000The full cost picture
COST EXPOSURE WITHOUT EMITCI WITH EMITCI (TEAM) ────────────────────────────────────────────────────────────────── CI compute waste $26,400+/yr $0 (surfaced + fixed) AI inference waste $7,200+/yr $0 (surfaced + fixed) Annual compliance prep $45K–$120K/yr ~$2,400/yr (Team plan) GHG assurance premium $20K–$40K extra Included in prep CSRD/SB 253 fine exposure Up to 10% revenue Defensible disclosure ────────────────────────────────────────────────────────────────── Estimates based on 30-engineer team with mixed AI CI tooling. Compliance costs are market-rate estimates — verify with your advisors.
One investment addresses four separate cost exposures. That is the case for measurement.