The Hidden Cost of Thinking Machines
Agentic AI’s Toll on Energy, the Environment, and Human Well-being
By Dhanada Kanta Mishra*
In early September 2026, an artificial intelligence firm reported a proposed solution to one of the seven Millennium Prize Problems of mathematics. Roughly 10,000 AI agents, running through OpenAI, worked in parallel for 88 hours to produce a 165-page proof resolving key parts of the 90-year-old Navier–Stokes problem.
They exchanged some 2.7 million messages and consumed around 130 billion output tokens—basic units of data processed by artificial intelligence models—at an estimated cost of $10 million to $40 million. Days earlier, the same company released GPT-6 “Astra,” a model trained on more than 100,000 GPUs, with the company’s president declaring that “the AGI era” had begun—a GPU being a graphics processing unit, an electronic circuit that adds speed and efficiency to computers, and AGI being artificial general intelligence, a digital-based intelligence that can not only just answer questions but that can reason, learn, and apply knowledge creatively from one field to another to independently solve complex problems.
Agentic AI—artificial intelligence that employs individual AI “agents” that each manage a part of an overall massive task and plans and pursues goals on its own—is no longer a research curiosity or a science-fiction fantasy. It is here, it is working, and it is consuming resources at a scale that’s only beginning to be understood.
A single complex agentic query can draw 136.5 times more electricity than a standard chatbot response, and agentic token usage, which overtook human-driven queries in February 2026, grew 14-fold to 7.3 trillion tokens by August.
As these autonomous systems reshape how we work and live, the question is no longer whether they are coming, but at what cost—to the planet and to ourselves.
From Chatbots to Agents
The differences between traditional AI and agentic AI are significant.
An AI chatbot answers a question and stops. An agent plans, calls tools, checks its own work, and keeps going until a goal is met, with far less human oversight.
Astra itself is marketed in the terms of agentic AI: It fills online forms, updates customer records, organizes calendars, conducts research, and builds websites.
The shift matters environmentally because agentic workloads involve far more tool calls, more tokens, and more idle GPU cycles per task than a simple query—every planning loop, every moment of reflection, every “thinking token” can multiply compute cost a hundredfold or more per prompt.
A human asking a chatbot for directions is cheap; a teammate walking beside you, checking every step, is not.
A human asking a chatbot for directions is cheap; a teammate walking beside you, checking every step, is not.
The Hidden Cost of ‘Thinking’
The Navier–Stokes milestone turns an abstraction into a concrete number: A single research result burned through 130 billion output tokens and millions of dollars of computation. It is an extreme case of a general pattern, not an anomaly.
An audit by the South Korean public research university KAIST found agentic tasks can consume 136 times more energy than a comparable chatbot response, largely because GPUs idle while waiting for tools to respond.
A Stanford study of agentic coding workloads found they can use up to 1,000 times more tokens than simple code chat, with input tokens—context and tool definitions—driving most of the cost.
Navveen Balani, executive director of the Green Software Foundation, describes how planning loops, tool calls, retries, and compounding context can escalate a 2,000-token interaction into a 2 million-token workflow.
OpenAI, for its part, declined to claim the $1 million prize associated with the Navier–Stokes challenge, and the proof has yet to be independently verified. The “milestone” was as much a demonstration of scale as of mathematics.
But the impact is already clear: Real-world data from OpenRouter, a major AI gateway, shows agentic token usage has already overtaken human-driven queries and is growing exponentially. The same Stanford study, coauthored by Erik Brynjolfsson and Sandy Pentland, maps how unpredictably agentic tasks spend tokens, arguing that token economics deserve the same scrutiny as model accuracy.
“The token is the unit we bill on,” writes Balani of the Green Software Foundation. “The kilowatt-hour, the gram of CO₂e, and the liter of water are the units that cost the planet.”
Every token means GPU time, grid energy, emissions, and cooling water. The International Energy Agency expects data centers’ roughly 1.5% share of global electricity to more than double by 2030, while their water use climbs toward 1.2 trillion liters a year.
The International Energy Agency expects data centers’ roughly 1.5% share of global electricity to more than double by 2030, while their water use climbs toward 1.2 trillion liters a year.
The Future of Work
What agentic AI means for employment is still unknown—rigorous longitudinal data does not yet exist. However, the early evidence suggests the shift may reshape work more than replace it outright.
A new report from the Boston Consulting Group estimates AI could reshape roughly half of jobs in the coming years, while tech researcher IDC projects that 40% of roles at the world’s 2,000 largest companies will involve direct engagement with AI agents.
The likelier near-term results are agentic AI programs absorbing tasks requiring routine cognitive skills while humans shift toward oversight, judgment, and relationship work—even as the boundary between an “always-on” digital colleague and personal time begins to blur.
The practical question is less whether jobs disappear than who is equipped to oversee, audit, and correct the agents doing them—a gap that may prove hard to close.
The “AGI era” framing deserves a closer look. Nvidia’s chief executive posted that “AGI has arrived,” but the ARC Prize team, whose benchmark OpenAI cited, declined to endorse the label. Capability is real; the rhetoric around it is contested.
Well-being vs. Anxiety
How do people react to advanced AI in their workplaces?
AI-based tools show early promise for supporting workers’ mental health. A recent systematic review in the British Medical Bulletin finds encouraging but still maturing evidence. Yet surveys now show AI-driven change as one of workers’ top concerns.
Gallup research across more than 30,000 employees finds the difference comes down to trust and clarity. Frequent AI use without a clear strategy is associated with higher anxiety and burnout, while in high-trust workplaces, the same tools yield larger productivity gains.
On the positive side, delegating drudgery to agentic AI may ease some forms of stress—the relief of a task finally done.
But this must be weighed against new anxieties around obsolescence, oversight, and the always-on pull of a digital colleague that never logs off. A May 2026 report by CES-IFO found that 14% to 19% of workers believe their job is “somewhat” or “very likely” to be eliminated within five years. Frequent AI use itself raises that perceived risk.
Christos Makridis, a senior researcher at Gallup, puts the human factor plainly: “The success of any AI initiative depends on the people, starting with whether employees feel that management creates a trusting work environment that is safe to experiment in. AI can help with tasks, generate insights, and sometimes validate results, but humans have a competitive edge in connecting with customers through empathy, applying creativity to shape and use AI-driven ideas, and developing the tasks of the future.”
“AI can help with tasks, generate insights, and sometimes validate results, but humans have a competitive edge in connecting with customers through empathy, applying creativity to shape and use AI-driven ideas, and developing the tasks of the future.”
Aligning Agentic AI with People and Planet
The cost of agentic AI is not fixed; it is a design choice. Balani calls energy “the lever—the dimension that practitioners directly control but rarely measure. Reducing the tokens reduces the GPU-seconds and reduces the kilowatt-hours.”
Three approaches to cut token and energy waste are to adopt “smarter reasoning budgets,” have fewer redundant tool calls, and cache information promptly. Asked which single design change matters most, Balani suggests eliminating needless “repeated reasoning,” e.g., designing agents to “reuse validated outcomes where appropriate, and invoke additional reasoning only when the situation genuinely requires it.”
Organizations like the Green Software Foundation are codifying this into “Lean Agentic AI,” an efficiency-first approach that pairs token-aware design with carbon-intensity measurement.
Two more elements—transparency about an agent’s environmental footprint and workplace policies that protect well-being alongside productivity—complete the responsible path.
Systemic Considerations with Agentic AI
The environmental question is not whether to use the agent, but whether the whole system reduces life cycle impacts.
Consider a building-maintenance agentic AI that gathers inspection photographs, checks them against the building’s digital model, queries maintenance histories, drafts a work order, and routes it for approval.
Used well, this AI can cut repeat site visits, catch defects earlier, and save the material and travel emissions of preventive maintenance.
But used carelessly—with redundant scans, oversized models, and constant reanalysis—it simply adds a new layer of digital waste on top of the old one.
Local inference—executing AI models directly on a user’s device, rather than relying on remote servers—is the other side of that coin, and it is coming faster than it once seemed. Distilled and quantized models already run on hardware that would have been unthinkable a few years ago—a five-year-old laptop with no dedicated GPU can handle lightweight models, if slowly and short of what agentic workloads demand.
The trajectory points toward frontier-class capability on ordinary devices. And the gains are real: less strain on water supplies and energy-hungry data centers described above, lower per-query cost, and data that never leaves the user’s machine. The honest caveat is that the energy is not eliminated, only relocated—the tokens still draw power, now on a desk instead of a rack, and the training behind the model happened in a data center regardless.
Local inference is best read as a complement to the efficiency levers already described rather than a substitute: It offloads the edges, while the heaviest agentic workloads continue to lean on centralized compute.
There is a cautionary thread in recent news, too. OpenAI’s own chief scientist has acknowledged that more capable models are harder to audit, while Nvidia’s chief executive celebrated that 400,000 more GPUs are coming online. The trajectory is steep in both directions: capability and consumption rising together.
Agentic AI has begun to do the work. Whether it becomes a burden on the planet and its people—or a tool we learn to lead wisely—depends on the choices being made now, not on the technology itself.
*Dhanada Kanta Mishra is a PhD in civil engineering from the University of Michigan and is currently working as the managing director of a Hong Kong-based AI startup for building technology for the sustainability of built infrastructure (www.raspect.ai). He writes on environmental issues, sustainability, the climate crisis, and built infrastructure.
Comments