There are two stories about AI and the environment running at the same time. AI infrastructure is straining power grids and water systems at a pace that local communities are absorbing first. The same technology is helping climate scientists and grid operators do work that was not possible at scale before.
Our position AI's environmental impact is two stories happening at the same time, one requires leaders to exercise caution, the other requires our support. The buildout is creating measurable strain on electricity grids, water systems, and local communities. The same technology is producing real advances in climate modeling and renewable energy forecasting. The potential harms require regulatory care.
We are skeptical of catastrophic framing that treats every AI workload as a moral failure. We are equally skeptical of techno-optimism that treats environmental cost as a rounding error. The numbers on both sides deserve to be looked at clearly. The benefits are important, but they must not be positioned as more important than the environment we all rely on.
AI is a physical technology
AI is not a cloud. It is electricity drawn from a substation, water evaporating from a cooling tower, and fiber-optic cable running into a warehouse-sized building somewhere. Whatever happens on the screen rests on that physical infrastructure, and that infrastructure has a footprint that grows with demand. That footprint is the part of the story public conversation has been catching up to. It is also not the whole story. Honest reporting looks at the whole ledger.
Real and measurable
Rising electricity demand from data centers, concentrated regional impacts, water consumption for cooling, e-waste from rapid hardware turnover, and the carbon intensity of the energy mix powering it all.
Also real and measurable
Climate modeling at unprecedented resolution, renewable energy forecasting that reduces grid waste, materials science breakthroughs in batteries and carbon capture, and rapid efficiency gains in the technology itself.
What the numbers show
Two of these numbers are growing. The third is shrinking faster than almost any other metric in energy history. Which trajectory dominates the next decade depends on choices being made right now about where data centers get built, what they get cooled with, and how the underlying technology gets engineered.
What the costs actually look like
Demand is growing faster than the grid can absorb it
Global data center electricity consumption was about 415 terawatt-hours in 2024, roughly 1.5 percent of global use. The IEA projects it doubles to 945 TWh by 2030 with AI as the dominant driver, and AI-focused data center electricity grew 50 percent in 2025 alone. In the U.S., data centers are projected to account for roughly half of all electricity demand growth out to 2030.
The aggregate number understates the local problem. Virginia's data centers consumed 26 percent of state electricity in 2023, and Ireland's are projected to reach 32 percent of national consumption by 2026. Where data centers cluster, transmission lines get reinforced at ratepayer expense, electricity rates rise, and other electrification work (transit, building heat, manufacturing) competes for the same constrained capacity. The source mix compounds it: renewables are growing fast, but natural gas and coal together still account for over 40 percent of the additional electricity needed for data centers through 2030. Renewable generation that could be displacing fossil fuels elsewhere is being absorbed by AI instead.
The footprint is present, and the disclosure is uneven
A single large hyperscale data center can consume around 1.1 million gallons of water per day for cooling, roughly what a town of 10,000 people uses. U.S. data centers directly consumed about 17 billion gallons in 2023, and Lawrence Berkeley National Laboratory projects that figure could double or quadruple by 2028. Indirect water consumption from electricity generation typically runs three to four times higher than direct on-site use.
Where this lands matters more than the aggregate. Friction has already emerged in Dublin, Uruguay, and parts of the U.S. Southwest, where data centers compete with agricultural and residential users for stressed water tables. Disclosure is inadequate. Per-query water numbers from companies and researchers vary by more than a factor of a thousand depending on what is counted, which makes the picture impossible to verify from outside.
The cost the conversation gets wrong
E-waste from rapid hardware turnover is the part of the footprint that gets counted least. GPU generations are now turning over every two to three years rather than the five to seven typical for prior data center equipment. Volumes of obsolete accelerators and supporting gear heading to disposal are rising sharply, and recycling infrastructure has not kept pace. The mining footprint for the rare earth elements that go into the next generation belongs on the same ledger.
"The speed of the AI revolution is increasingly contrasting with the speed of the physical, social and economic systems that underpin it."International Energy Agency, 2026
What AI is contributing on the other side
AI is doing environmental work, even as it consumes
The biggest engineering challenge of the renewable transition is forecasting. Solar and wind depend on weather, and a grid operator who cannot predict next-day generation accurately has to keep fossil-fuel plants running as backup. Open Climate Fix's partnership with Google DeepMind improved solar forecast accuracy by 40 percent for the UK's national grid operator, directly reducing fossil-fuel backup. NOAA's AI-based ensemble forecast matches the skill of its traditional model while using 9 percent of the computing resources.
Beyond grid work, AI is accelerating climate science previously bottlenecked by computational cost. Materials science applications are identifying candidate compounds for batteries and carbon capture faster than traditional methods. Satellite analysis tracks deforestation and illegal mining close to real time. Microclimate models that used to require government supercomputers now run on individual GPUs. None of this displaces the underlying problem of climate change, and none of it should be treated as offsetting the cost side of the ledger. It does change what is operationally possible.
Per-task cost is dropping, even as aggregate use rises
Energy required for an individual AI task is dropping by at least an order of magnitude annually, a rate the IEA calls unprecedented in energy history. A simple text query now typically uses less electricity than running a television for the same length of time. The gains come from new chip architectures, smaller distilled models that handle many tasks at a fraction of the cost, and closed-loop cooling designs that recycle water and heat rather than evaporating fresh supply.
This does not solve the problem. Aggregate consumption is still growing because demand is growing faster than per-task efficiency improves. The trajectory matters, though, and any honest forecast has to count it alongside the demand growth, not against it.
Small ops, by design
Every Passons AI pilot is built without training a new model. We use existing infrastructure from established providers and design for the smallest computational footprint that gets the job done. Our target two-week scope means we are not primarily running large-scale agentic systems that loop indefinitely. We put a useful tool in front of a worker who needs it for a specific, bounded task.
That is a small operating choice in a large industry. It does not change the trajectory of global data center demand. It does let us be honest about our role: a thin layer of practitioner work sitting on top of infrastructure built and operated by others. We ask our clients to make their own choices about that infrastructure clearly, and we make ours clear in return.