A Relational Intelligence prototype by Nobox Inc.
What does your AI use
actually cost?
The environmental and social costs of AI use are real but invisible by design. Answer three questions to see what your usage costs in energy, water, carbon, labor, and physical resources — and what you can do about it.
Nothing is stored. Everything runs in your browser.
Step 1 — Get your real data
Run this prompt in Claude first
Paste the prompt below into a fresh Claude conversation. It uses Claude's built-in tools to pull your last 20 conversations and return a JSON summary of your actual usage patterns. Then paste the result here.
This gives you real numbers instead of estimates — and it only takes about 30 seconds.
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, avg_length_estimate (short/medium/long/extended/mixed), context_dependency (low/moderate/high/very-high), project_signals (boolean), estimated_avg_tokens_per_conv (number), portrait (2–3 sentences second-person, specific and observational, not generic).
Don't have Claude open right now?
Step 1 — Quick estimate
Answer three questions for a rough estimate. Run the prompt above in Claude later for real numbers.
How many days a week do you use AI?
How long do your conversations typically run?
Do you load background context — projects, instructions, or documents?
Estimates use Claude Sonnet 4.6 and arXiv 2505.09598 as defaults.
Your profile
—
—
Per conversation
—
tokens (est.)
Per day
—
tokens (est.)
Per month
—
tokens (est.)
Your impact
Energy
—
—
Water
—
—
CO₂
—
—
Research range
—
Source: arXiv 2505.09598 direct benchmark (30 models). WUE 0.5 L/kWh. Grid 0.41 kg CO₂/kWh (US avg). ±40% range. Anthropic has not disclosed per-query environmental data.
What you paid (subscription share)—
Infrastructure cost (~40% of revenue)—
Annotators received (amortized from training)—
Content creators received$0
Anthropic est. gross margin—
$20/mo Pro ÷ ~60 convs/mo = your per-conversation subscription share. Content creator figure is structural — not an estimate. US Copyright Office (2025): "most content used without authorization." Annotator wages amortized across 10B estimated lifetime queries.
At scale
Your use, multiplied
The debate around AI's footprint focuses on individual behavior. But individual use only becomes meaningful at the scale of hundreds of millions of people — and the vast majority of the world hasn't started yet.
One avg AI conversation
× 500M users, once today
× 500M users, every day for a year
Water per conversation (~500 mL): Li et al., "Making AI Less 'Thirsty'", UC Riverside, arXiv:2304.03271 (2023) — includes both direct cooling and indirect water from electricity generation. 500M users: Statista/various, 2024–2025. LA residential water use: LA DWP Annual Report 2023.
Who's using it
Each dot is 1% of the global population. Filled dots are regular AI users today.
~6% — current regular AI users (~500M of 8.2B)
~94% — not yet regular users
The infrastructure being built now — datacenters, power contracts, hardware orders — is sized for far higher adoption than this.
Global AI electricity demand
2024
~200 TWh/yr
AI-attributed share of global data center energy. Comparable to the electricity use of Thailand.
2027 projected
~500 TWh/yr
Approaching the total annual electricity consumption of Germany (~550 TWh).
2030 projected
~1,000 TWh/yr
Equivalent to Japan's entire annual electricity consumption. More than 3× global aviation's energy use today.
These projections assume ~50% adoption and current efficiency trajectories. Efficiency gains per query are real — but volume growth has outpaced them in every prior technology transition.
AI electricity projections: IEA World Energy Outlook 2024; Goldman Sachs Research, "AI Infrastructure: Too Much Spend, Too Little Benefit?", Jun 2024. Germany consumption: IEA Country Profile 2023. Japan consumption: IEA Country Profile 2023. Aviation energy: IEA "Net Zero by 2050" (2023), ~300 TWh electric-equivalent. AI user count: Statista, various 2024–2025 sources. Global population: UN, 2024 estimate.
What these numbers don't show
The rest of the bill
Energy, water, and carbon are the measurable tip. Underneath every conversation is a global supply chain of materials, content, and human labor that never appears in any environmental report.
Physical resources
~500 miles of copper cable connect a single large GPU cluster — before the building, power systems, or cooling loops are counted.
In large quantities
Copper — 300–600 miles of cabling per large cluster, plus thousands of tons in racks and power distribution. Steel — 5,000–15,000 tons of structural steel per hyperscale datacenter building. Silicon — each H100 GPU die is ~814mm²; across an estimated fleet of 220,000+ GPUs, that's roughly 2,800 sq ft of fabricated silicon.
In small quantities — from conflict zones
Tantalum (coltan) — dozens of capacitors per circuit board; ~70% mined in the DRC. Cobalt — battery backups and magnetic components throughout; ~70% from DRC. Tungsten — wire bonds inside every chip; primarily China and Rwanda. Palladium — connector plating on every server; primarily Russia and South Africa.
Plastic & e-waste
Every cable is jacketed in PVC or LSZH plastic. Every circuit board is built on FR4 fiberglass-epoxy substrate. Every chip is encased in epoxy mold compound. AI server racks exceed 4,000 lbs — a significant portion is plastic and composite materials. Fewer than 20% of discarded electronics are properly recycled globally, and AI hardware cycles out in 3–5 years.
Content resources
Training data isn't borrowed — it's extracted at scale. The web alone yields 400 terabytes of raw text every month. Most of the people who wrote it received nothing.
Web content
Common Crawl collects 3–5 billion web pages per month (~400 TB). Its archive exceeds 250 billion pages over 17 years — every news article, forum post, blog, and product page ever indexed. Compensation: none.
Books & research
LibGen (7.5M books, 81M papers) trained Meta's Llama 3. Books3 (183,000 pirated books) trained Meta, EleutherAI, and Bloomberg. Anthropic settled for $1.5B after a ruling found it used 7M+ pirated books — ~$3,000 per identified title. Compensation: none prior to settlement.
Code
GitHub's public repositories — 100M+ users, hundreds of millions of codebases — form a major component of most models' training data. Compensation: none.
Images
LAION-5B contains 5.85 billion image-text pairs scraped from the web — photos, illustrations, medical images, artist portfolios. Stable Diffusion trained on 2.3 billion of them; Pinterest alone supplied 8.5% of one sample. The dataset was later found to contain CSAM, prompting a partial re-release in 2024. Compensation: none.
Whose knowledge counts
Common Crawl is 42% English — yet English speakers are ~16% of the global population. AI models answer 79% of US culture questions correctly vs. 12% for Ethiopian culture. Of 7,000+ languages, only a few hundred appear in usable corpora. The blind spots encoded in training data don't disappear — they compound.
Human resources
Tens of millions of people contributed labor to make AI systems work — across mining, fabrication, annotation, and research. Most of it is invisible to the people using the product.
Data annotators
Scale AI (240,000+ contractors) and Appen (1M+ registered workers) supply the human feedback that shapes model behavior. A single training run may require millions of annotation hours. Compensation: $2–25/hr depending on location and task.
ML researchers & engineers
~1,000–1,500 at Anthropic; tens of thousands across the industry. Compensation at frontier labs: $200K–$600K+/yr.
Chip fabricators
TSMC (73,000+), Samsung, and SK Hynix employ hundreds of thousands; assembly and packaging facilities across Malaysia, China, and Taiwan add millions more. Compensation: TSMC avg ~$116K/yr; fab technicians $59K–$81K/yr. Downstream assembly wages substantially lower and undisclosed.
Miners
150,000–255,000 artisanal cobalt miners in the DRC — often informal and unregulated. Tantalum and tungsten operations in Central Africa employ hundreds of thousands more. Compensation: avg ~$8/day; most earn below the DRC minimum wage of ~$5/day and well below the Kolwezi living wage of $480/month.
What you can do
Use the smallest model that works
Opus burns through resources ~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.
Start fresh conversations more often
Every message re-sends your 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.
Convert files before uploading
PDFs extract text AND convert 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 first.
Ask for transparency
Anthropic has published no per-query environmental data. Neither have OpenAI, xAI, or Microsoft. Only Google has disclosed methodology (Stanford FMTI). Support calls for standardized reporting — and ask your AI vendor directly.
Tell Claude to be brief
Output tokens cost the same as input — and verbose responses compound with every turn. "Just the code." "One sentence." Simple constraints can cut response tokens by half or more.
Use intentionally, not reflexively
Before opening an AI tool, ask whether the task actually needs it. A quick lookup, a simple edit — these don't justify the resource cost. The default should be intention, not convenience.
Audit your other tools for hidden AI
Your phone, fitness tracker, and many apps contain AI you may not have turned on. Opt out of features you don't use. Your total AI consumption is probably higher than just the LLMs you open deliberately.
Consider the full supply chain
Data labelers — often paid $1–$5/hour — are a hidden layer of AI's supply chain. Content creators whose writing trained these models received nothing. Support the AI Labelers Alliance and opt-in licensing initiatives.
Support fair data compensation
Writers, journalists, and creators whose work trained these models received no notice, no payment, no opt-out. Legislative efforts are trying to change that. The Authors Guild and AI Labelers Alliance are good places to start.
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 here.