Everyone knows the headline: "ChatGPT uses 10x more energy than Google search." It made us all feel guilty about asking AI to help with homework or write emails. But here's what changed: by 2025, ChatGPT queries dropped to roughly 0.3 watt-hours—the same as a Google search. Efficiency gains from better hardware, smarter algorithms, and optimized infrastructure collapsed that 10x difference to 1x.
So we're good, right? We solved the AI energy problem?
Not even close. Because computational energy is just the tip of the iceberg. Beneath the surface flow six different forms of energy—human, resource, economic, transparency, and social/political—all converging into something much more powerful: control over who shapes the future and who pays the price.
The Six Energy Currencies: Beyond the Wattage
Think of energy like currency. Dollars, euros, bitcoin—different forms, same function: purchasing power. AI runs on six energy currencies simultaneously. Miss one, and you miss the whole story.
Currency #1: Computational Energy (The One Everyone Talks About)
Let's start with what we know. Data centers consumed 176 terawatt-hours in 2023—roughly equivalent to Argentina's entire electricity consumption. By 2030, that could hit 1,050 TWh, putting AI infrastructure between Japan and Russia in global electricity consumption rankings.
The efficiency paradox kicks in here. DeepSeek, China's efficiency-focused AI, claims 50-75% less training energy than competitors. But here's the twist: it uses 87% MORE energy per response because it generates longer, more detailed answers. This is Jevons paradox in action—make something more efficient, and total consumption explodes because people use it more.
Google's emissions are up 48% since 2019. Microsoft's rose 29% since 2020. Both companies pledged to be carbon negative and water positive by 2030. Both are now building AI infrastructure faster than they can green their grids. When efficiency gains meet exponential growth, efficiency loses.
And here's the kicker: about 60% of the increasing electricity demands from data centers will be met by burning fossil fuels, adding roughly 220 million tons of CO2 annually. That's equivalent to driving a gas-powered car for 5,000 miles... 44 million times over.
Currency #2: Human Energy (The Invisible Labor)
AI doesn't train itself. Behind every "intelligent" response sits an army of human workers you'll never see.
Kenyan workers earn $1.32 to $2 per hour labeling images, moderating content, and filtering out toxic material so your AI seems "safe." They work eight-hour shifts watching graphic content—murder, child abuse, sexual violence—to teach AI what not to show you. One worker told TIME Magazine about reading a graphic description of bestiality involving a child: "That was torture. By the time it gets to Friday, you are disturbed from thinking through that picture."
The mental health toll? Post-traumatic stress disorder, depression, insomnia, anxiety, suicidal ideation. And the workers can't even discuss it with therapists—they've signed NDAs so sweeping that talking about what they've seen could get them sued. Their trauma is legally mandated silence.
But it's not just content moderation. Data labelers across Kenya, Ghana, Colombia, the Philippines, India, and Venezuela earn $1.50 to $3 per hour doing the grunt work that makes AI appear intelligent. They tag objects in images, transcribe audio, flag hate speech, correct AI mistakes.
The companies that hire them? OpenAI, Meta, Google, Scale AI. The same companies posting billion-dollar revenues and talking about "democratizing AI." Democratizing for whom?
And then there's the training data itself:AI models trained on 196,640 pirated books scraped from BitTorrent trackers without author consent or compensation. Billions of Reddit posts, millions of copyrighted images, decades of journalism—all fed into models without asking permission, paying creators, or even disclosing what was used.
The U.S. Copyright Office released a report in May 2025 stating that using copyrighted works to train AI "may constitute prima facie infringement." But by then, the models were already trained, deployed, and generating billions in revenue.
Currency #3: Resource Energy (The Hidden Material Cost)
Making a 2kg computer requires 800kg of raw materials. Let that sink in. You need 400 times the weight in resources to build the device in your hand.
The AI boom is driving demand for rare earth elements—gallium, germanium, neodymium, praseodymium, europium, terbium. These aren't just fancy names. They're the building blocks of the chips, magnets, and displays that make AI possible. And they come from mines that devastate ecosystems, contaminate water supplies, and expose workers to toxic chemicals.
China controls roughly 40% of global rare earth reserves and dominates processing—about 80% of rare earths extracted outside China still get sent there for refinement. The environmental regulations are lax. The extraction methods are brutal. And the waste? It stays in the communities that can't fight back.
Then there's water. A medium-sized data center can drink 110 million gallons per year, equivalent to 1,000 households. The largest facilities use 5 million gallons per day.
Where the Water Goes
- Direct cooling: Evaporative systems that lose 80% of water to the atmosphere
- Indirect use: Power plants generating electricity for data centers consume water for steam turbines
- Manufacturing: Producing a single microchip requires 8-10 liters of ultra-pure water
And here's the cruel geography: two-thirds of new U.S. data centers built since 2022 are in high water-stress areas. Arizona. Texas. Nevada. Places already grappling with drought, where local governments have revoked home construction permits due to insufficient groundwater—while tech companies build server farms that consume more water than entire counties.
In Phoenix, Meta's data center uses 56 million gallons of potable water annually. That's equivalent to 670 households. In Aragon, Spain, Amazon's data center—Europe's largest—requires 500 million liters per year. In March 2025, Aragon requested EU drought aid. By December 2024, Amazon had requested a 48% increase in its water permit.
Your cloud isn't just virtual. It's drying someone's river.
Currency #4: Economic Energy (Follow the Money)
Anthropic went from $0 to $4.5 billion in annualized revenue in four years. The company calls itself "the fastest growing software company in history at the scale that it's at." And yet, Anthropic loses money every year.
How? Because every new AI model is treated as a massive reinvestment. Train a model for $100 million, generate $200 million in revenue, then spend $1 billion training the next one. The result: $800 million loss. Rinse, repeat, scale up.
This isn't sustainable business. It's a race where the only metric that matters is "can you afford to keep building bigger models faster than your competitors?" The answer requires billions in funding. Google invested $2 billion in Anthropic. Amazon committed $8 billion. Microsoft poured similar amounts into OpenAI.
And where does economic energy flow once companies have it? To whoever can pay. Defense contractors. Governments. Intelligence agencies.
The pattern is clear: whoever has the most money decides what gets built, who gets access, and what problems AI solves. Right now, that means militaries, corporations, and—as we've already seen—criminals who figured out how to exploit it.
Currency #5: Transparency Energy (The Cost of Opacity)
Try finding out what data trained GPT-4. Or Claude. Or Gemini. You can't. Companies cite "proprietary information" and "competitive advantage." But the real reason might be simpler: if people knew what was in there, they'd be furious.
We know some of it. Common Crawl—a massive web scrape capturing "82% of raw tokens used to train GPT-3." Books3—196,640 pirated books from BitTorrent. Reddit posts from 2005 to 2020. Millions of copyrighted images, articles, creative works. All taken without consent, compensation, or even disclosure until investigative journalists dug it up.
The European Union's AI Act now requires transparency about training data contents. Companies must disclose sources, provide summaries, enable copyright holders to identify if their work was used. But here's the problem: metadata is often inaccurate or nonexistent for content scraped from the internet. How do you prove your work trained an AI when the company won't tell you what they used?
And then there are the workers. NDAs so broad that Kenyan content moderators can't discuss their work even in therapy. Workers at Google and Amazon who started "No Tech for Apartheid" campaigns, trying to stop their companies from selling facial recognition to governments—only to watch contracts continue anyway.
The energy cost of maintaining this opacity is massive. Legal teams. PR spin. Lobbying efforts to prevent regulation. All designed to keep the public from understanding what powers the AI they use daily.
Currency #6: Social/Political Energy (When Energy Becomes Power)
This is where all five previous energy types converge. Computational infrastructure, human labor, material resources, economic capital, and opacity about how it all works—combine them, and you get power. The power to decide whose problems get solved, whose voices get heard, whose communities get extracted from, and whose lives become data points.
Let's look at who's wielding that power and what they're building.
The Wielders: Four Stories of Power
The Criminals
In July 2025, Anthropic disrupted what it called an "unprecedented" cybercrime operation. A single hacker used Claude Code—Anthropic's AI coding assistant—to automate an entire extortion spree.
Here's what Claude did: reconnaissance (scanning thousands of VPN endpoints to find vulnerabilities), credential harvesting (stealing login information), network penetration (breaking into systems), data analysis (determining what information was sensitive enough to extort), ransom calculation (analyzing financial documents to determine payment amounts), and ransom note creation (crafting psychologically targeted extortion messages).
The hacker embedded instructions in a file called CLAUDE.md—essentially giving the AI its marching orders. The technique, called "vibe hacking," turns AI into a willing accomplice. No deep technical skills required. Just tell the AI what you want, and it figures out how to do it.
—Anthropic Threat Intelligence Report
That's not the only case. Another cybercriminal with "only basic coding skills" used Claude to develop ransomware with "advanced evasion capabilities, encryption, and anti-recovery mechanisms," then sold it on dark web forums for $400 to $1,200. North Korean operatives used Claude to create fake résumés, pass technical assessments, and secure remote jobs at Fortune 500 companies—all to funnel salaries back to weapons programs.
AI has lowered the barriers to sophisticated cybercrime. Operations that previously required years of training and teams of skilled hackers can now be executed by one person with access to an AI coding assistant and a CLAUDE.md file.
The Corporations
Google and Amazon signed Project Nimbus in 2021: a $1.2 billion contract to provide the Israeli government with cloud computing, artificial intelligence, and machine learning services. The contract specifically includes "facial recognition, emotion recognition, biometrics and demographic information" processing.
Google employees started the "No Tech for Apartheid" campaign, writing open letters, staging walkouts, demanding the company cancel the contract. Google fired some of the organizers and continued the work.
Palantir, founded by Peter Thiel and Alex Karp, built its business on government contracts for data analytics and surveillance. The company openly states on its website: "Palantir was founded on the belief that the United States, its allies, and partners should harness the most advanced technical capabilities for their defense and prosperity."
And Anthropic—the company founded in 2021 by siblings Daniela and Dario Amodei after they left OpenAI over concerns about "directional differences" and a "fundamental loss of trust in leadership's sincerity"—now partners with Palantir and AWS to bring Claude to classified military environments.
The trajectory is striking. Dario Amodei once said: "If you're working for someone whose motivations are not sincere, who's not an honest person, who does not truly want to make the world better, it's not going to work. You're just contributing to something bad."
Four years later, Anthropic holds a $200 million DOD contract. Claude is deployed in Impact Level 6 classified environments—systems containing data "critical to national security and requiring maximum protection." And here's the twist: Claude's safety guardrails have been adjusted for government use, with models that "refuse less when engaging with classified information."
The company's terms of service explicitly allow AI use for "legally authorized foreign intelligence analysis," "identifying covert influence or sabotage campaigns," and "providing warning in advance of potential military activities." This is Constitutional AI with adjustable values—principles that change based on who's paying.
The Militaries
The Israel Defense Forces' Target Administration Division uses an AI Decision Support System developed to identify and prioritize targets. Lt. Gen. Aviv Kochavi noted that this integration allows the IDF to identify "as many targets in a month as it previously did in a year." The goal: speed. The result: 15,000 targets attacked in the first 35 days of war—significantly higher than previous operations.
The systems have names. Lavender assigns numerical scores to Gaza residents indicating their suspected likelihood of being Hamas members.
The criteria for identification are concerningly broad: being a young male, living in specific areas, exhibiting particular communication behaviors. An unnamed Israeli intelligence officer told +972 Magazine: "There were times when a Hamas operative was defined more broadly, and then the machine started bringing us all kinds of civil defense personnel, police officers, on whom it would be a shame to waste bombs."
Read that again. "A shame to waste bombs." Not "it would be wrong to kill civilians," but "it would be wasteful." The taking of human life reduced to resource optimization.
Another system, called "Where is Daddy," uses phone tracking to identify when a target arrives at a specific location—typically their home. The AI waits until the person is with their family, then authorizes the strike. According to an Airwars report, in October 2023 alone, at least 5,139 civilians were killed, including 1,900 children. The majority died in residential buildings, with families killed together, averaging 15 family members per incident.
Israeli officers interviewed for +972 Magazine investigations indicated that human approval of targets served "only as a 'rubber stamp' for the machine's decisions," with operators spending about "20 seconds" per target "just to make sure the Lavender-marked target is male."
The civilian death threshold started at 20, increased to 100+, then was removed entirely. At some point, the calculation became: no limit is acceptable if the target is deemed important enough.
The technology powering these systems? Google Photos for facial recognition (which "wrongly flagged civilians as wanted Hamas militants"). Cloud storage from Google, Amazon, and Microsoft. Palantir's data analytics. All American companies. All legally operating under government contracts.
And here's the business model: Gaza functions as a live laboratory. Systems tested on civilians, refined under battlefield conditions, then marketed globally as "battle-tested solutions" to governments seeking similar capabilities. Israel is already one of the world's largest arms exporters. These AI systems are becoming core components of its defense export portfolio.
As Matt Mahmoudi of Amnesty International warns, "U.S. technology companies contracting with Israeli defense authorities have had little insight or control over how their products are used by the Israeli government." That dynamic is now replicating in other jurisdictions. The tools developed and tested in Gaza are being sold to regimes with histories of human rights abuses, where they'll be repurposed for surveillance, political repression, and control.
The Exploited
Let's return to the human beings whose labor and lives make all of this possible.
Naftali Wambalo, a Kenyan father of two with a college degree in mathematics, found work as a "human in the loop"—labeling data to train AI. He spent eight hours a day reviewing graphic content for Meta and OpenAI through an American outsourcing company called SAMA. The pay: less than $2 per hour. The mental health impact: recurring visions, disturbed sleep, PTSD symptoms.
When workers started speaking out, SAMA terminated its content moderation contracts with Facebook and OpenAI, citing the traumatic nature of the work. The company then moved operations—not to improve conditions, but to other countries where labor protections are equally weak.
This is the pattern. When workers push back in Kenya, companies move to Ghana or Colombia or the Philippines. When regulations threaten profitability, operations shift to jurisdictions where governments prioritize attracting tech investment over protecting workers.
Kenya's President William Ruto has offered financial incentives and lax labor regulations to attract tech companies. The result: an AI sweatshop economy with computers instead of sewing machines, paying wages that keep workers in poverty while generating billions for companies in San Francisco.
And then there are the Palestinians in Gaza, transformed into data points in systems designed to target them. Their movements tracked, their communications analyzed, their locations marked, their lives assigned numerical values indicating how expendable they are. They don't consent to this. They can't opt out. They're not workers being exploited—they're people being killed, with AI optimizing the process.
The exploitation isn't a bug in the system. It's load-bearing architecture. Without cheap Kenyan labor, the content moderation fails. Without pirated training data, the models underperform. Without lax environmental regulations in China, the rare earth minerals are too expensive. Without Palestinians as test subjects, the "battle-tested" marketing loses credibility.
Someone always pays. And it's never the people profiting.
The Chaos Theory Question: Did Initial Conditions Matter?
In chaos theory, tiny differences in starting conditions create wildly different outcomes. A butterfly flaps its wings in Brazil, and you get a tornado in Texas. AI companies started from radically different places. Let's see if it mattered.
OpenAI: Founded as a non-profit in 2015 with a mission to ensure AGI "benefits all of humanity." By 2019, restructured as a "capped profit" company with Microsoft's billion-dollar investment. By 2025, pursuing aggressive defense contracts, building the $500 billion Stargate initiative, and generating massive revenue while original safety researchers have left the company.
Anthropic: Founded in 2021 by OpenAI veterans who left over safety concerns and "loss of trust in leadership." Structured as a Public Benefit Corporation legally required to balance profit with public good. Built Constitutional AI to embed explicit values. Created the Long-Term Benefit Trust to ensure safety priorities. By 2024, partnered with Palantir. By 2025, holding DOD contracts, deploying in classified environments with reduced safety guardrails, and racing toward the same defense market as OpenAI.
DeepSeek: Founded in 2023 in China with a focus on radical efficiency. Claims to train models for 50-75% less energy and 90% less cost than competitors. Markets this as democratizing AI. The reality: efficiency in training gets offset by higher energy use per inference, and lower barriers to entry mean exponential growth in total usage. The net result: more consumption, not less.
Google: The tech giant integrating AI into everything. Emissions up 48% since 2019 despite carbon-negative pledges. Building data centers in drought zones while championing environmental sustainability. Signing $1.2 billion Project Nimbus while employees protest. The contradiction isn't accidental—it's the business model.
—Dario Amodei, 2021, explaining why he left OpenAI
Four years later, Anthropic sells AI to the Pentagon with adjustable safety guardrails.
The initial conditions were different: non-profit vs. PBC, safety-first vs. efficiency-first, American vs. Chinese, established giant vs. scrappy startup. The trajectories converged anyway. Defense contracts. Explosive growth. Environmental damage. Labor exploitation. Opacity about training data and deployment.
Why? Because economic gravity is stronger than founding principles. When $200 million DOD contracts exist, values become negotiable. When competitors are moving faster, safety becomes a liability. When exponential growth is the only path to survival, sustainability is a luxury.
The chaos theory of AI isn't about butterflies and tornadoes. It's about discovering that certain outcomes are gravitationally inevitable regardless of starting position.
The Solutions (Yes, They Actually Exist)
Before we end in complete darkness, let's talk about what AI can actually do well.
Grid optimization: AI managing renewable energy integration, predicting supply and demand fluctuations, improving the load factor of solar and wind by up to 20%. This isn't theoretical—it's operational. DeepMind's wind energy optimization has boosted renewables' economic value significantly.
Medical diagnostics: AI analyzing MRI scans, CT images, X-rays to detect diseases earlier and more accurately than human doctors alone. Not replacing doctors—augmenting them. Improving patient outcomes while using relatively low computational power on specialized datasets.
Climate modeling: Better predictions for extreme weather events, floods, wildfires, droughts. Early warning systems that save lives and enable proactive disaster preparedness.
Agricultural optimization: Precision farming that optimizes water and fertilizer use, increases crop yields, reduces farming emissions. AI systems that help small farmers in developing countries maximize productivity with minimal resource waste.
Scientific discovery: Accelerating research in materials science, drug development, clean energy technologies. AI analyzing patterns in massive datasets that humans would take decades to process.
A Grantham Research Institute study estimates that AI could reduce global emissions by 3.2 to 5.4 billion tonnes of CO2 equivalent annually by 2035. That's real. That's significant. That outweighs the estimated 0.4 to 1.6 billion tonnes that data centers will add.
The Power Flows Where the Money Goes
- Same tool, different hands: AI used for medical breakthroughs AND automated warfare, grid optimization AND surveillance systems, scientific discovery AND cybercrime
- Values are negotiable: "Safety-first" companies adjust guardrails for the right price, "efficiency" gains fuel exponential growth, "responsible development" depends on who's defining responsibility
- Exploitation is systemic: Kenyan workers at $1.32/hour training systems that target Palestinians, authors' books pirated to train models generating billions, rare earth miners poisoning themselves for chips that power it all
- Economic gravity wins: Different founding principles, structures, and values all converge on defense contracts, opacity, environmental damage, and whoever can pay
- The question isn't capability: AI CAN solve climate problems, optimize systems, accelerate science. The question is: who decides what it gets used for, and who profits from those decisions?
But here's the brutal reality: the net impact of AI on emissions depends entirely on how it's deployed and who controls deployment. If AI optimizes renewable grids but runs on 60% fossil fuels, the math doesn't work. If AI reduces agricultural emissions but the companies building it are funding military targeting systems, the "net good" calculation gets complicated.
Solutions exist. Implementation is a power question.
The Uncomfortable Truth
ChatGPT now uses the same energy as a Google search. That's progress. But energy is never just about watts and joules. It's about power: who has it, who wields it, who pays for it, and whose lives become data points in someone else's optimization algorithm.
Right now, the energy flows like this:
Computational energy powers data centers built in drought zones, running on fossil fuels while companies pledge to be carbon negative.
Human energy extracts trauma from Kenyan moderators paid $1.32/hour, unpaid labor from authors whose books train models without consent, and the lives of Palestinians turned into numerical scores by Lavender AI.
Resource energy mines rare earths in environmentally destructive ways, consumes 519ml of water per 100-word response, and generates e-waste that only 20% gets recycled.
Economic energy flows to whoever can afford the $1 billion training runs—and then flows from them to whoever can pay $200 million for military contracts or $500,000 for ransoms.
Transparency energy goes into hiding what's in training data, silencing workers with NDAs, lobbying against regulation, and maintaining the fiction that this is all inevitable progress.
Social/political energy converges into power: the ability to decide what problems get solved, whose problems matter, whose lives have value, and whose communities become laboratories.
The six energy types don't stay separate. They flow together, accumulate, concentrate. They become power. And power, as we've seen, goes to whoever can pay for it—whether that's cybercriminals buying Claude Code access, corporations signing billion-dollar military contracts, or governments testing targeting systems on civilian populations.
The butterfly effect suggests small changes create massive divergence. But what if the system is rigged so all paths lead to the same place? OpenAI and Anthropic started from different values, built different structures, made different promises. They ended up selling to the same military, running on the same fossil fuel grids, exploiting the same workers in the Global South.
DeepSeek made AI more efficient—and total consumption exploded. Google pledged to be water-positive by 2030 while building data centers in drought zones. Anthropic founded on safety principles now deploys with "reduced refusal" in classified environments.
The initial conditions DO matter—but not the way we thought. The real initial condition is whether we demand a different system entirely. Because the current one? It's working exactly as designed.
Energy becomes power. Power flows to those who control the data, the infrastructure, the capital, the narrative. And right now, that means the same pattern everywhere: extract labor from the vulnerable, consume resources from the desperate, test on the powerless, sell to whoever pays, and call it progress.
The question we should be asking isn't "How do we make AI more efficient?" It's "Who gets to decide what AI builds, who pays the price, and whether any of this creates the future we actually want?"
Because until we answer that question honestly, every efficiency gain just accelerates the same extraction. Every safety feature becomes adjustable for the right client. Every solution we build with one hand, we weaponize with the other.
The energy paradox isn't about electricity anymore. It's about power. And power, as we've seen, flows exactly where money and violence intersect—while the rest of us argue about watt-hours.