Methodology & data provenance

1. Measured inputs

These are recorded, not estimated. Re-runnable from the scripts in the source repository.

Gemini pipeline (real billed totals)

Source: gemini_usage_daily MongoDB collection. Aggregated by scripts/gemini-usage-aggregate.mjs.

Claude Code (sampled)

Source: ~/.claude/projects/-Users-dereklomas-sourcelibrary/ JSONL session logs. Each contains assistant message usage blocks emitted by the Anthropic API.

2. Bottom-up energy model

Roofline analysis of transformer inference. Output tokens are memory-bandwidth-bound; input tokens are compute-bound; cache reads use ~10% of fresh input.^[6][7]

Validated against MLPerf Inference v5.1 (Llama 3.1-70B on 8×H200 = 31,391 tok/s at ~5.6 kW = 0.178 J/output-tok).^[5] Our model predicts 0.114 J/output-tok for that workload — within 36%, so we apply a 1.56× correction.

Source: scripts/co2-footprint-model.py

3. The six estimation lines

Each is independent. Each makes different assumptions. The geometric mean of all centrals is 1805 kg.

Source: scripts/co2-estimation-lines.py

4. Equivalents — conversion factors

Every "1805 kg equals X burgers" statement in the post uses the conversion factors below. Each is cited.

5. All sources

Every ^[n] footnote in the post points here.

1. 8.0 million Gemini API calls, 30.6 billion tokens, $19,588 spend, 22.6 million page-credits processed [measured]
   Source: gemini_usage_daily MongoDB collection, aggregated 2026-01-02 to 2026-05-23. page_count backfilled 2026-05-25 for ~748K legacy records (PR #2006 + #2009) that were missing the field due to pre-April 2026 direct-Mongo writers; translation page-credits ratio went from 0.45 to 1.83 pages/call after the fix. Token and cost totals were unaffected by the bug — only the per-call page-count attribution.↻ scripts/gemini-usage-aggregate.mjs
2. Claude Code: 139M output tokens, 43.5B cache-read tokens, 1.6B cache-create tokens [measured]
   Source: ~/.claude/projects/-Users-dereklomas-sourcelibrary/ session JSONL files; 11.9% byte-sample extrapolated to full corpus↻ scripts/sample_tokens.py
3. Median Gemini Apps prompt: 0.24 Wh energy, 0.03 g CO2e [cited]
   Source: Elsworth et al. (Aug 2025), 'Measuring the environmental impact of delivering AI at Google scale' ↗
4. Mistral Large 2: 1.14 g CO2e per 400-token response (2.85 mg/token full LCA) [cited]
   Source: Mistral AI (Jul 2025), 'Our contribution to a global environmental standard for AI' ↗
5. Llama 3.1-70B on 8×H200 = 31,391 tok/s offline = 0.18 J/output-tok [cited]
   Source: MLPerf Inference v5.1 results (Sep 2025) ↗
6. Decode is memory-bandwidth-bound; prefill is compute-bound [cited]
   Source: Roofline analysis of transformer inference; arXiv 2507.14397 'Efficient LLM Inference' ↗
7. Cache hits use ~90% less energy than full inference [cited]
   Source: Li et al. (2024), 'A Survey on Large Language Model Acceleration based on KV Cache Management' ↗
8. Claude Opus runs on AWS Trainium2 (Project Rainier, ~500K chips) [cited]
   Source: Data Center Dynamics (Oct 2025), 'AWS activates Project Rainier cluster' ↗
9. Anthropic has not disclosed Scope 1, 2, or 3 emissions [cited]
   Source: Stanford CRFM Foundation Model Transparency Index (Dec 2025) ↗
10. Geometric mean of six estimation methods: ~1,800 kg CO2e total [computed]
    Source: Computation across all six methods (see lines 1-6 in /methodology)↻ scripts/co2-estimation-lines.py
11. Beef: 60 kg CO2e per kg (global LCA median) [cited]
    Source: Poore & Nemecek (2018), Science ↗
12. EU passenger car fleet average: ~180 g CO2/km well-to-wheel [cited]
    Source: European Environment Agency, CO2 performance of new passenger cars (2024) ↗
13. AMS-LHR economy RT: ~245 kg CO2e with radiative forcing factor [cited]
    Source: atmosfair flight calculator + DEFRA 2024 conversion factors ↗
14. AMS-NYC economy RT: ~2,950 kg CO2e with radiative forcing factor [cited]
    Source: atmosfair flight calculator + DEFRA 2024 long-haul factor (0.149 kg/passenger-km) ↗
15. Dutch per-capita emissions: ~8.4 t CO2e/year [cited]
    Source: CBS (Centraal Bureau voor de Statistiek), 2024 ↗

6. What we don't know

• Anthropic per-token energy. Not published. Required disclosure under California SB 253 begins in 2026.
• Active parameter counts for Gemini Pro and Claude Opus. Best estimates ±50%.
• Production batch sizes for major providers. Estimated 12-128 depending on workload type and latency target.
• Grid intensity for Anthropic's deployment regions. We model 60% Project Rainier (Indiana, ~650 g/kWh) + 40% older AWS clusters (us-west-2 etc.). Mix may shift over time.
• Training amortization beyond what is implicit in Mistral's LCA. No comparable disclosure from Anthropic / Google / OpenAI.
• Long-context overhead factor for Claude Code. We apply ×2.0 based on TokenPowerBench (Dec 2025) which measured 3× per-token energy from 2K → 10K context. Source Library's ~155K-token contexts are extrapolated, not measured directly.

7. Known data-quality issues

During this audit, we found several issues in the upstream pipeline data. None invalidate the headline number, but worth knowing if you're reading deeply.

• Translation page-count under-logging — FIXED 2026-05-25. The long-form translation endpoint had been logging ~0.45 page-credits per call (vs ~3 for OCR). Root cause: the batch-collector wrote page_count: successCount, which is the count of pages whose DB row was matched at result-collection time — zero for batches whose results were already collected on a prior pass. Separately, ~748K legacy gemini_usage records written before the April 2026 Supabase migration had page_ids populated but no page_count field. Both fixed in PRs #2006 and #2009, plus a backfill of the historical records. Translation ratio is now 1.83 pages/call, total page-credits 413K → 2.68M, consistent with the ~5.4M pages with translation text in the pages collection.
• Two translation type labels in history. Older records use type: "translate" (per-page, manual scripts) and newer records use type: "translation" (long-form, batched context — current default). A 2026-03 normalization consolidated the active code paths but ~250K legacy translate rows remain. Both are summed in this post's totals.
• Book-level status fields are abandoned (and now confirmed safe to delete). books.ocr_status and books.translation_status have 4 and 10 entries respectively across 46K books. Status is tracked per-page now (via the pages collection) and per-book progress via pages_ocr / pages_translated counters. The legacy fields are not in the active Book TypeScript schema and not read by any public route; safe to drop in a future cleanup PR.
• Re-OCR rate is 1.18×, not 1.4×. Earlier drafts assumed more re-OCR than is actually happening. Now corrected. The page_revisions collection (164K rows) stores pre-overwrite snapshots, but only for pages that were re-OCR'd, so it can't derive an absolute rate by itself — the API-log ratio of 7.24M OCR page-credits ÷ 6.14M unique OCR'd pages is the authoritative source.
• Per-book denominator: 14,295 books with index/summary generated in window. The carbon-per-book averages divide window-total tokens by this count. Alternatives — 28K visible books, 15K with any OCR, 17K with anypipeline_auto.last_updated in window — would shift the per-book number 10–40%. We use index-generation because that's the phase that fires the full enrichment pass (summary + chapters + quality + collection assignment).
• Visual + Unknown language books may be OCR'd unnecessarily. 14,544 books with language=Visual (image-only manuscripts, alchemical plates) and 10,703 with language=Unknown have no OCR filter upstream — the pipeline runs Gemini on them and saves whatever auto-detect returns. Net impact on this post's totals is small because batch API discount is 50%, but it's wasted compute. Pending fix as of 2026-05-25.
• Claude Code token totals are sampled, not exhaustive. 11.9% by bytes of the JSONL session logs, extrapolated linearly. Sampling error ~±20%. A full scan was attempted but stalled at 98% due to slow I/O on the project directory.