We hear a lot about the high environmental costs of AI use, but I couldn't quite connect them to my own behavior. I built this tool to contextualize the water, energy, labor, and content usage behind my AI conversations so I can make informed decisions about when and why I use it.
This is the first version — please share your findings, feedback, and ideas with me at thopkins@nobox.us so I can improve it.
1
2
3
4
Step 1 of 4 — Your usage profile
How do you use AI?
Answer a few questions to understand your own habits and/or run the prompt(s) below to get data from your chat history in Claude.
Which AI tools do you use?
Select all that apply.
How many days a week do you use AI?
On a typical active day, how many conversations do you start?
One thread = one conversation, regardless of length.
How long do your conversations typically run?
Be honest — most people underestimate this.
Do you use Projects or load background context?
Projects, custom instructions, or documents that give AI context about your work.
Do you upload documents, images, or files?
What do you primarily use AI for?
Select all that apply.
Before today, how much did you think about the environmental or social cost of AI?
Optional: calibrate with real data from Claude
Run either prompt in Claude, then paste the JSON result below. Your profile will update automatically with your real data.
Prompt 1 — Analyze your last 20 conversations
Paste into a fresh Claude conversation to get a history-based usage analysis.
Use the recent_chats tool to retrieve my last 20 conversations. Then analyze the full set and return ONLY a JSON object (no preamble, no markdown fences) with these fields: conversation_count, date_range_days, dominant_types (array from: task-sprint, deep-research, collaborative-production, quick-question, document-analysis, creative-writing, technical-coding), avg_length_estimate (short/medium/long/extended/mixed), tool_use_density (none/low/moderate/high/very-high), context_dependency (low/moderate/high/very-high), project_signals (boolean), topic_domains (array up to 4 strings), power_user_signals (array 2–4 specific observations), estimated_avg_tokens_per_conv (number), portrait (2–3 sentences second-person, specific and observational, not generic).
Prompt 2 — Analyze a single conversation
Paste at the end of any conversation to get its token breakdown.
Analyze this entire conversation and return ONLY a JSON object (no preamble, no markdown fences) with: token_estimate (total input+output tokens, accounting for full history re-sent each turn, 4 chars/token), conversation_profile (short/medium/long/extended), message_count (human turns only), avg_tokens_per_exchange, dominant_content_type (text-only/code-heavy/image-included/mixed), model_used, calculation (object with method, total_chars, input_tokens, output_tokens, total_tokens, notes).
Paste the JSON from either prompt. Make sure to remove any other text from the response before pasting.
No data is stored. Everything runs in your browser.
Step 2 of 4 — Your mirror
Here's who you are as an AI user
—
—
—
What you can do
—
Your estimated usage
Per conversation
—
tokens (est.)
Per day
—
tokens (est.)
Per month
—
tokens (est.)
Vs casual user
—
environmental footprint
How you spend your AI time
Context dependency
—
10 = AI is useless to you without background context loaded.
This was written by AI.
Step 3 of 4 — Your impact
What your average AI conversation costs
Estimates from published research. Anthropic has not disclosed per-query environmental data. Figures shown with ranges and sources.
User type
Model
Research source
Scale
Energy
—
—
Water
—
—
CO₂
—
—
Tokens
—
—
Range by research source
Energy
—
Water
—
CO₂
—
—
Training labor hours embedded in your usage
Me
Expert annotator hours
—
at $40/hr rate
Offshore labeler hours
—
at $9–15/hr rate
Combined labor hours
—
all annotators
RLHF comparisons (Claude model)~318K
Hours per annotation (est.)~0.5 hrs
Expert annotation cost vs. computeup to 28× higher
—
Sources: Galileo AI (2026), Kang (Medium 2025). Hours amortized across estimated 10B lifetime queries. Anthropic has not published annotator headcount or hours.
Training data provenance
Anthropic non-disclosing
Estimated corpus sizeTrillions of tokens
Licensed / paid content (est.)~15–40%
Used without permission (est.)~60–85%
Active data scraping litigationReddit, authors, publishers
Articles represented in your scope—
Low confidence. US Copyright Office (2025): "most content used without authorization." Stanford FMTI 2024: Anthropic non-disclosing on data size.
Physical infrastructure behind your conversation
Me
GPU seconds used
—
H100/H200 equiv.
Your share of a GPU
—
amortized fraction
HBM memory used
—
GB-seconds
Component
Your scope
What it is
Where it comes from
GPU accelerators
—
NVIDIA H100/H200/GB200 — the chips that run AI inference. Each costs $25K–$50K.
Designed by NVIDIA (Santa Clara, US). Fabricated by TSMC on N4 process in Taiwan. Packaged in Malaysia and Taiwan. Key materials: silicon, copper interconnects, gold wire bonds.
HBM memory chips
—
High-bandwidth memory stacked on each GPU. 80GB per H100, made by SK Hynix / Samsung.
SK Hynix (Icheon, South Korea) and Samsung (Hwaseong, South Korea). Materials: silicon, tungsten, palladium. HBM3e uses stacked die bonding — one of the most materials-intensive chip processes.
InfiniBand cables
—
Connects GPUs within clusters at 400Gb/s. Anthropic's Colossus 1 cluster uses thousands of these.
Designed by NVIDIA/Mellanox (Yokneam, Israel). Copper or fiber-optic cable; transceivers assembled in China and Taiwan. Optical variants use rare earth-doped fiber.
Network switches
—
High-speed switching fabric routing data between GPU nodes. Each rack-scale switch ~$50K.
NVIDIA Quantum-2 ASICs designed in Israel (Mellanox). PCBs assembled in China and Taiwan. Materials: copper, fiberglass, rare earth magnets in inductors.
Cooling units
—
Liquid cooling CDUs (coolant distribution units). One per rack of ~8 GPUs. H100 draws 700W each.
Major makers: Vertiv (US), Airedale (UK), CoolIT (Canada). Steel and copper tubing; synthetic glycol coolants. Units assembled in US, UK, and China.
Server boards / CPUs
—
Host processors managing GPU workloads. Each server has 2 CPUs and 4–8 GPUs.
Intel Xeon (Ohio/Arizona fabs) or AMD EPYC (fabbed at TSMC, Taiwan). PCBs assembled in China. Capacitors contain rare earth tantalum; heat spreaders are copper.
NVMe SSDs (model storage)
—
Fast storage holding model weights. A 100B+ parameter model requires ~200GB+ at FP8 precision.
NAND flash from Samsung (South Korea), Micron (US/Japan), Kioxia (Japan). Controllers fabbed at TSMC. Assembled in China and Thailand. Materials: silicon, tungsten, aluminum oxide.
—
GPU count from Anthropic's Colossus 1 deal (220,000 GPUs, confirmed May 2026) + AWS/Google capacity. Per-query GPU-seconds derived from H100 inference throughput benchmarks (~66 tokens/sec). Hardware quantities are amortized fractions of Anthropic's estimated total fleet. Sources: xAI/SpaceX announcement (May 2026), NVIDIA H100 benchmarks, SemiAnalysis InferenceX (Apr 2026).
Who gets what — value flow
Me
What you paid (subscription share)——
Infrastructure cost (est. 40% of revenue)——
Annotators received (amortized, training)——
Content creators received$0—
Anthropic est. gross margin——
—
Anthropic at scale — verified figures
Revenue run-rate
$30B+
Apr 2026, confirmed
Consumer MAU
~30M
May 2026 estimate
API calls/month
25B+
mid-2025, growing
Consumer mkt share
~4.5%
global AI chatbot mkt
High confidence. Sources: Sacra (May 2026), Madrona (Apr 2026), getpanto.ai (May 2026). Gross margin modeled — Anthropic does not publish P&L.
Step 4 of 4 — Over time
Your footprint over time
From one conversation to five years — and what it looks like at team, company, Anthropic, and global AI scale. Anthropic and All AI figures use independently sourced estimates; personal and organizational figures use your profile.
·
What can you do about this?
Use the smallest model that works
Claude’s limits are cost-based, not token-based — Opus burns through your allowance ~5× faster than Haiku for the same exchange. Start on Haiku. Step up to Sonnet when you feel the ceiling. Reserve Opus for genuinely complex multi-step reasoning. Most tasks don’t need it.
Start fresh conversations more often
Every message re-sends the entire conversation history. By turn 10, you’ve consumed ~5× more tokens than ten one-turn conversations would have. When you finish one task and start another — open a new chat. The compounding cost is invisible until it isn’t.
Convert files before uploading
PDFs are the most expensive format: Claude extracts text AND converts every page to an image — 1,500–3,000 tokens per page before you’ve asked a single question. A 50-page PDF can consume most of your context window. Convert to plain text or markdown first.
Tell Claude to be brief
Output tokens count the same as input tokens — and verbose responses sit in your history, multiplying with every turn. “Just the code, no commentary.” “Answer in one sentence.” Simple constraints can cut response tokens by half or more.
Ask for transparency
Anthropic has published no per-query environmental data — no Scope 1/2/3 emissions, no water figures. Neither have OpenAI, xAI, or Microsoft. The Stanford FMTI tracks this: only Google has disclosed methodology. Support calls for standardized reporting.
Consider the full supply chain
Data labelers in the Global South, often paid $1–$5/hour, are a hidden layer of AI’s supply chain. Content creators whose writing trained these models received nothing. The AI Labelers Alliance advocates for fair pay. Support initiatives for opt-in licensing.
Use intentionally, not reflexively
Before opening an AI tool, ask whether the task actually needs it. A quick lookup, a simple edit, a question you already know the answer to — none of these justify the resource cost. The default should be intention, not convenience. AI is a tool, not a reflex.
Talk about it
Most people have no idea this is happening. The energy, water, labor, and uncompensated content behind every conversation are invisible by design. Share what you found. The social cost of AI only becomes a public conversation when enough people know to ask.
Audit your other tools for hidden AI
Your phone, fitness tracker, and many apps contain invisible AI you may not have turned on. Opt out of features you don't need. Your total AI consumption is probably higher than just the LLM(s) you use.
Support fair data compensation
Writers, journalists, and creators whose work trained these models received nothing — no notice, no payment, no opt-out. Legislative efforts and licensing frameworks are trying to change that. The AI Labelers Alliance and Authors Guild are good places to start supporting.
