{"slug": "show-hn-estimate-your-ai-co2-footprint", "title": "Show HN: Estimate your AI CO2 footprint", "summary": "A new web tool lets users estimate the energy and CO2 emissions of their AI usage, using modeled coefficients from public AgentX benchmarks on NVIDIA B300 hardware. The calculator, posted on Hacker News, allows input of token counts and grid intensity to compute footprint, with presets for chat, agent assistants, and sourced projects like Anthropic's compiler experiment and Bun's Zig-to-Rust rewrite. The tool's methodology notes that energy estimates are modeled, not measured, and that carbon calculations depend on grid scenarios.", "body_md": "# What’s your *AI footprint?*\n\nEstimate the energy and emissions behind your AI use. No token counts? Start with an example.\n\n## Your usage\n\nK = thousand · M = million · B = billion · T = trillion\n\nExamples replace all settings; energy estimates are modeled.\n\n## Advanced settings\n\n## Your estimated footprint\n\nEnter token usage to calculate.\n\n**Energy breakdown**\n\n**—** Cached input\n\n**—** Output\n\n**—**\n\n## In everyday terms\n\nSelect a comparison for its assumptions and source.\n\nComparisons are alternative yardsticks, not amounts to add together. Energy-only benchmarks are separate from emissions.\n\n## Methodology & benchmark assumptions\n\n**Example energy estimates are modeled, not measured.**\n                The chat and agent assistant presets use illustrative token\n                counts, not measured averages. “This page” uses a rounded\n                estimate of development usage, not recorded token totals. “Major\n                project” uses the reported token totals from\n                [Anthropic’s compiler experiment](https://www.anthropic.com/engineering/building-c-compiler)\n                —2B input and 140M output tokens across nearly 2,000 sessions.\n                That does not establish a typical feature size, cache rate, or\n                input/output ratio.\n                [Claude Code’s usage documentation](https://code.claude.com/docs/en/costs)\n                separates fresh input, cache reads, and cache writes; include\n                all three in this calculator’s input total. Example cache rates\n                and model classes are assumptions, not measured characteristics\n                of those workloads.\n              \n\n                “Agent team project” uses\n                [Bun’s reported pre-merge Zig-to-Rust rewrite usage](https://bun.com/blog/bun-in-rust): 5.9B uncached input + 72B cached input = 77.9B total input,\n                and 690M output. Its 92.4% cache share is rounded to 92% to\n                match the whole-percent control. Both sourced projects are\n                exceptional examples, not typical workloads; their reported\n                token totals do not validate this calculator’s energy\n                coefficients.\n              \n\nThe estimator uses three contemporary public model proxies on modern NVIDIA Blackwell hardware. Its core equation is:\n\n```\nE = fresh_input × Jfresh + cached_input × Jcache + output ×\n                  Joutput\n```\n\nCurrent per-token coefficients are modeled from public AgentX benchmarks on B300 hardware. The model assumes cache hits use ~56% of fresh-input energy and output costs ~10× per token.\n\n**Flash** Qwen3.8-Flash-Next · 176B total / 6B active · B300-class AgentX proxy\n\n**Frontier** DeepSeek V4 Pro · 1.6T total / 49B active · B300 AgentX proxy\n\n**SOTA** Kimi K3 · 2.8T total / ~104B active · B300 AgentX proxy\n\n**Serving profile matters.** Efficient serving\n                batches more work and reduces energy per token; low-latency\n                serving sacrifices utilization. The displayed range reflects\n                different serving profiles, not a statistical confidence\n                interval. Architecture, quantization, context length, cache\n                placement, hardware generation, speculative decoding, and\n                datacenter routing can move the true value outside it.\n              \n\n**Carbon is calculated separately:**\n`CO₂e = kWh × grid intensity`. Changing the grid does\n                not change estimated inference energy.\n              \n\n**Grid presets are scenarios, not routing claims.**\n                Presets use EPA eGRID2023 revision 2 annual total-output CO₂e\n                rates, not marginal or consumption-based emissions. These\n                generation averages do not adjust for electricity imports,\n                transmission losses, or upstream fuel emissions. West Virginia\n                is a coal-heavy benchmark; selecting it does not imply that a\n                request was served there. Dedicated generation or contractual\n                clean-energy purchases can differ from these averages.\n              \n\n**Benchmark/source links:**\n[Qwen3.8 / InferenceX](https://inferencex.semianalysis.com/model/qwen-3-8-flash-next)\n                ·\n                [DeepSeek V4 Pro / InferenceX](https://inferencex.semianalysis.com/compare/deepseek-v4-b300-vs-gb300)\n                ·\n                [Kimi K3 / InferenceX](https://inferencex.semianalysis.com/run/kimi-k3-on-b300)\n                ·\n                [EPA eGRID2023 tables (XLSX)](https://www.epa.gov/system/files/documents/2025-06/summary_tables_rev2.xlsx).", "url": "https://wpnews.pro/news/show-hn-estimate-your-ai-co2-footprint", "canonical_source": "https://llmfootprint.fyi/", "published_at": "2026-09-09 05:02:44+00:00", "updated_at": "2026-09-09 05:20:29.870882+00:00", "lang": "en", "topics": ["ai-tools", "ai-infrastructure", "ai-ethics"], "entities": ["NVIDIA", "Anthropic", "Bun", "Qwen3.8-Flash-Next", "DeepSeek V4 Pro", "Kimi K3", "EPA eGRID2023"], "alternates": {"html": "https://wpnews.pro/news/show-hn-estimate-your-ai-co2-footprint", "markdown": "https://wpnews.pro/news/show-hn-estimate-your-ai-co2-footprint.md", "text": "https://wpnews.pro/news/show-hn-estimate-your-ai-co2-footprint.txt", "jsonld": "https://wpnews.pro/news/show-hn-estimate-your-ai-co2-footprint.jsonld"}}