There is a strange contradiction in modern society that becomes harder to ignore the deeper artificial intelligence moves into everyday life. And no worries, this won’t become a ethical lecture about lifestyle.

We became incredibly disciplined about visible consumption. People replace old lightbulbs with LEDs because saving a few watts matters. The Food gets sourced locally to reduce transportation impact and short trips are done by bicycle instead of car. The plastic packaging gets questioned and straws became a geopolitical topic for a while.

And honestly, none of that is unreasonable. The most interesting part comes directly afterwards.

The same people who carefully optimize every visible part of their environmental footprint, and being very verbal about it, often sit down in front of systems powered by industrial-scale GPU infrastructure and spend hours arguing with a language model about nonsense inside a giant context window.

Not because they are solving medical research problems or because they are modeling climate systems. But because an invitation decline should sound „slightly warmer but still professional“. Or because an AI agent network is currently comparing productivity apps to avoid paying 5 dollars for an existing SaaS subscription.

That is the part I find fascinating.

Not the existence of AI itself, but the disconnect between how we emotionally perceive physical consumption versus computational consumption.

The Resource Nobody Associates With AI: Water

Electricity is only one side of the equation.

Modern AI infrastructure also consumes enormous amounts of water because GPU clusters generate extreme heat that must be cooled continuously. Large AI data centers therefore rely heavily on water-based cooling systems, especially at hyperscale.

And the numbers become surprisingly large very quickly.

Current estimates suggest that AI-related data centers could consume between roughly 279 and 765 billion liters of water globally in 2025 depending on infrastructure assumptions and cooling methods. (see Heinrich Böll Foundation)

Even individual AI requests are not completely negligible. Depending on the model and infrastructure, a single prompt may indirectly consume somewhere between 14 and 50 milliliters of water.

That sounds harmless until billions of people continuously regenerate images, retry prompts, extend conversations, or run autonomous agent workflows because the first result was “almost right.”

Nobody opens a chatbot thinking about cooling towers or water infrastructure. The interaction feels weightless because the complexity stays hidden behind a clean interface.

But operationally, modern AI infrastructure increasingly behaves less like software and more like industrial utility infrastructure.

AI Feels Weightless. That Is the Problem.

Most people still emotionally categorize AI as „software“. And software feels abstract, is mostly Invisible and almost environmentally neutral.

You type something into a chat window and words appear. The interaction feels harmless because the infrastructure remains invisible.

But modern AI systems are not lightweight software features quietly running on a laptop somewhere. They are industrial-scale compute systems that convert electricity into statistical prediction across globally distributed accelerator clusters.

Training modern frontier models already consumes astonishing amounts of electricity. But training is only one side of the equation. The larger long-term impact likely comes from inference, meaning the billions of interactions users have with these systems every single day.

Every generated image, every rewritten email, every „make this slightly more friendly“ and every retry because the answer „did not feel right“ consumes resources in terms of compute and therefore electricity and water. Even worse when you start working with autonomous agent loops that are continuously calling additional models in the background.

At small scale this sounds trivial, but at internet scale it becomes infrastructure.

One report discussing OpenAI’s long-term infrastructure ambitions estimated possible targets around 250 gigawatts of compute capacity by 2033. That would place AI infrastructure energy consumption into ranges comparable to entire industrialized nations. The same analysis estimated tens of millions of GPUs operating continuously to sustain such environments. (see Tom’s Hardware)

That is no longer „just tech“, it becomes a challage for the energy infrastructure.

Computation Became Emotionally Free

Technology history follows a very predictable pattern. The moment something becomes cheap enough that humans stop emotionally registering the cost, usage explodes.

When storage became cheap, we duplicated everything endlessly instead of cleaning data properly. When cloud infrastructure became flexible, entire industries normalized massive overprovisioning because the hardware disappeared behind APIs and monthly invoices.

Now AI entered the same phase.

People regenerate the same image hundreds of times because the fingers looked strange in version 34. Teams continuously rewrite presentation slides because the wording “still feels slightly off,” while developers increasingly burn inference calls to avoid reading documentation carefully for fifteen minutes.

The funny part is that the industry itself already started joking about the absurdity. When someone asked OpenAI CEO Sam Altman how much money the company loses because users constantly write “please” and “thank you” to ChatGPT, he joked that it was “tens of millions of dollars well spent.” Yahoo Finance coverage of Sam Altman’s comment

And while the statement was obviously playful, it accidentally highlights something important. Tiny human behaviors multiplied across billions of interactions suddenly become industrial-scale infrastructure demand. Politeness itself now has measurable compute cost, which may be the sentence that best summarizes both the absurdity and brilliance of the modern AI economy.

The Real Problem Is Not AI. It Is Scale Without Friction.

One AI request does not matter. One generated image does not matter, and one conversation with a chatbot is environmentally irrelevant.

The problem starts when billions of users collectively behave as if computational resources are infinite because the interface feels frictionless.

This is exactly what happened with streaming video. Nobody watches a single movie and thinks about data center infrastructure. But collectively, streaming platforms became one of the largest drivers of global internet traffic and infrastructure expansion because billions of people simultaneously consume high-bandwidth content every day.

AI follows the same economic law, except the workloads are often computationally far more expensive. And unlike traditional systems, generative AI scales exceptionally well with human indecisiveness.

Longer prompts require more compute, larger context windows require more memory, and every retry generates additional workloads. Agent systems make this even more interesting because they recursively create further model calls in the background while the user experiences everything as one simple interaction.

That is probably the most fascinating part. Modern interfaces abstract the complexity away almost perfectly. The user experiences “a conversation,” while behind the scenes enormous GPU clusters, cooling systems, networking fabrics, orchestration layers, storage infrastructure, and redundancy systems coordinate in real time to generate that experience.

The cloud always sounded soft and harmless. Operationally, modern AI increasingly resembles heavy industry.

AI Efficiency May Become the Real Competitive Advantage

Ironically, this is where the discussion becomes genuinely interesting from a technical perspective.

The companies likely to dominate the next phase of AI may not necessarily be the ones building the largest models. They may simply be the ones capable of delivering useful intelligence most efficiently, because physics eventually forces every hype cycle back into operational reality.

Electricity costs remain real, cooling infrastructure remains real, hardware supply chains remain real, and latency constraints remain real. At some point, “just add more GPUs” stops being strategy and starts becoming evidence that architectural efficiency was neglected.

This is exactly why techniques such as quantization, sparse architectures, retrieval-augmented generation, intelligent routing, and model compression matter so much. These are no longer niche optimization exercises for engineers who enjoy benchmarking hardware. They are rapidly becoming economic survival mechanisms for companies operating AI systems at scale.

For years, the industry normalized astonishing infrastructure burn rates because investor enthusiasm temporarily masked operational reality. But eventually every system meets physics, and physics does not particularly care whether electricity was consumed by a steel factory, a streaming platform, or somebody running a six-agent workflow to draft a two-sentence LinkedIn reply.

The Strange Blind Spot of Modern Sustainability

The point here is not that AI is bad. Quite the opposite. AI will create enormous value in medicine, logistics, science, accessibility, industrial optimization, safety systems, and countless other areas that genuinely improve human capability.

The more interesting question is whether we are currently applying the same computational seriousness to meaningful problems and digital noise generation, because right now society often treats both with nearly identical infrastructure intensity.

And perhaps that is the defining contradiction of this decade.

We became extremely aware of visible consumption while remaining surprisingly blind to invisible computation. A plastic straw feels wasteful because you can physically hold it in your hand, while a giant context window filled with endless back-and-forth reasoning feels free because it disappears behind a clean chat interface.

Unfortunately, the underlying infrastructure does not care whether the workload felt productive, philosophical, or completely unnecessary. Electricity was consumed either way.

My personal Thoughts

AI is not environmentally problematic because it exists. The real question is whether we are using industrial-scale infrastructure intentionally or simply because friction disappeared.

We learned to optimize visible consumption remarkably well, while invisible computation still feels emotionally free. But modern AI systems are no longer abstract software. They are physical infrastructure powered by electricity, cooling, water, and global supply chains.

The future winners in AI will likely not just build the smartest systems, but the most efficient ones.