CI cost in a monorepo: Turborepo vs Nx vs Bazel

The zero-vendor approach using git diff and workflow paths filters. Works well for small monorepos (under 10 packages) where the dependency graph is simple enough to manage manually.

# Detect changed packages with git diff
- id: changes
  uses: dorny/paths-filter@v3
  with:
    filters: |
      api:
        - 'packages/api/**'
      web:
        - 'packages/web/**'

# Run only affected jobs
jobs:
  api-tests:
    if: needs.changes.outputs.api == 'true'
    # ...
  web-tests:
    if: needs.changes.outputs.web == 'true'
    # ...

Limitation: does not handle transitive dependencies. If package A depends on package B and you change B, this approach does not know to re-run A's tests without explicit configuration.

Turborepo understands your package dependency graph and runs only the tasks affected by a change, including transitive dependencies. Remote cache stores task outputs: on a cache hit, the task is skipped entirely and the output replayed.

# Run only affected packages since origin/main
- run: npx turbo run build test --filter='...[origin/main]'
  env:
    TURBO_TOKEN: ${{ secrets.TURBO_TOKEN }}
    TURBO_TEAM: your-org

Pricing: Vercel hosts Turborepo remote cache free for small teams. Self-host on S3/R2 with turborepo-remote-cache package for free. No per-minute charge from Turborepo itself.

Cost impact: 30-60% CI time reduction on a typical 10-package monorepo with a warm remote cache.

Nx provides nx affected commands that scope tasks to changed packages, plus Nx Cloud remote cache and distributed task execution.

Nx Cloud pricing (2026): Free tier: 500 CI pipeline hours/month. Pro: $300/month unlimited. Enterprise: custom pricing.

For large monorepos, Nx Cloud's distributed execution (running individual tasks across multiple agents) can reduce wall-time by 60-80% independently of the cost reduction from running only affected tasks.

Bazel is the maximum-correctness option. It builds a precise dependency graph at the file level (not package level), stores all outputs in a content-addressed remote cache, and only rebuilds targets whose inputs have changed.

At scale (100+ engineers, deep C++/Java/Python polyglot codebases), Bazel's precision pays dividends that no other tool matches. Dropbox's published migration showed 10x build speed improvement and ~50% CI compute reduction at their scale. See case studies.

The honest counterpoint: Bazel setup cost is high. Plan 2-6 months of migration for a large existing codebase. Maintenance requires dedicated build-engineering capacity. Below 100 engineers, Turborepo or Nx almost always gives better ROI per engineering hour invested.

Scenario: 25-dev team, 10-package monorepo, 50 PRs/day, GH Actions hosted Linux. April 2026 rates.