1/ Everyone talks about what AI can do. Almost nobody talks about what AI costs — and who actually pays. A thread on the hidden ledger. 🧵
2/ COMPUTE: Training GPT-4 cost an estimated $100M+. Training frontier models now runs $500M-$1B. That money comes from venture capital and cloud revenue — which means it comes from your software subscriptions.
3/ ELECTRICITY: A single ChatGPT query uses roughly 10x the energy of a Google search. US data centers now consume more electricity than most countries. Someone pays that power bill. Usually local ratepayers.
4/ WATER: Data centers in The Dalles, Oregon used over 25% of the city's water supply for cooling in 2022. Microsoft's global water consumption jumped 34% in one year. Dry communities subsidize your AI outputs.
5/ LABOR: Behind every "intelligent" system are human labelers. Kenyan workers rated harmful content for OpenAI at roughly $2/hour through outsourcer Sama. The AI industry runs on invisible labor arbitrage.
6/ DATA: Your Reddit posts, your photos, your medical records. AI training data was scraped without consent from billions of people. Kate Crawford calls this an extractive industry. The raw material is us.
7/ MINERALS: Lithium from Chile. Cobalt from the DRC. Rare earth elements from China. The physical supply chain of AI hardware maps onto centuries-old patterns of resource extraction from the Global South.
8/ TALENT: The AI industry has vacuumed up researchers from universities and governments. Public institutions trained the talent; private companies captured the returns. A subsidy nobody voted for.
9/ ATTENTION: Every AI-generated article, image, and video competes for your attention. The cost is harder to measure but real: a flood of synthetic content degrades the information commons we all share.
10/ Add it up: compute, electricity, water, labor, data, minerals, talent, attention. AI isn't free. The costs are just distributed to people with no seat at the table.
11/ The redistribution question isn't "Will AI create value?" It clearly does. The question is: Who pays the costs, who captures the value, and is that arrangement just?
12/ If the people bearing AI's costs — click workers, ratepayers, data subjects, mining communities — had a vote in how it's built, it would look very different. That's the conversation worth having. /end
LinkedIn version:
Everyone talks about what AI can do. Almost nobody talks about what it costs — or who pays.
The hidden ledger is long. Training a single frontier model now costs $500M-$1B. A ChatGPT query uses 10x the energy of a Google search. Data centers in The Dalles, Oregon consumed over 25% of the city's water for cooling. Kenyan workers labeled harmful content for $2/hour. Billions of people had their data scraped without consent.
Then there's the material layer: lithium from Chile, cobalt from the DRC, rare earth elements from China. The supply chain of AI hardware maps onto centuries-old extraction patterns.
Kate Crawford documented this in Atlas of AI — the "cloud" is a physical industry with real environmental and labor costs. Add the talent drain from public universities to private labs, and you have a massive unacknowledged subsidy flowing from public to private.
The redistribution question isn't whether AI creates value. It does. The question is whether the people bearing the costs — click workers, ratepayers, data subjects, mining communities — have any say in how the value gets distributed.
If they did, AI would look very different. That's the conversation Redistributed exists to have.
What hidden AI cost surprises you most?