This is a submission for the Google I/O Writing Challenge
I actually appointed myself as AGI Police some days back. On account of the number of times I've found myself sipping AI slop milkshake.
As a reformed individual (with regulated slop intake), I keep listening to Yann LeCun's insights through several notable podcasts. He constantly reminds us that “LLMs in general cannot predict the consequences of their actions.”
Google I/O 2026 made an impact, with lots of exciting announcements. Training across the largest clusters in the world. Over 7x more tokens processed every month. Bigger infrastructure. Faster inference. More intelligence delivered instantly to billions of people.
But somewhere in the middle of all the demos and applause, I found myself thinking less about the models and more about the machinery underneath them.
So I asked Google AI Mode to help calculate the energy consumption behind large-scale token processing and compare it to something human-sized.
Here is what we found out in less than 30 seconds of processing.
We could power nearly 3 million light bulbs continuously, 24 hours a day, for an entire year. Let that sink in.
What struck me wasn’t just the raw number itself. It was the inversion of intuition.
AI feels weightless.
You type words into a chat box and receive intelligence back in seconds. No smoke. No factory floor. No visible machinery. Just text appearing instantly on a glowing rectangle in your hand.
But underneath that interface sits an industrial system consuming electricity, water, cooling infrastructure, and global semiconductor supply chains at unprecedented scale.
Before I looked away, Gemini made another suggestion that made me even more curious. It suggested comparing the energy consumption with Nvidia infrastructure and also estimating the amount of water required to cool the servers powering the inference workloads.
I indulged.
And in less than 5 seconds (which means I was exaggerating when I said 30 seconds earlier), this happened:
457 million litres matches the total annual water footprint of roughly 1,200 average household families.
At that point, the conversation stopped feeling like a fun experiment and started feeling like a glimpse into the physical economics of intelligence itself.
The real takeaway from Google I/O wasn’t simply that models are getting smarter.
It was that intelligence is becoming infrastructure.
Every prompt now has a physical cost attached to it:
- Electricity generation
- Water cooling systems
- Data center expansion
- Semiconductor fabrication
- TPU supply chains
- Thermal management at planetary scale
And the strange part is that users rarely see any of it.
What fascinated me most was how Gemini framed the answers. Instead of treating the numbers like an alarming revelation, it immediately contextualized them against broader industry infrastructure. The response was technically useful, but it also revealed something subtle about AI systems: they do not merely answer questions; they shape how scale is emotionally interpreted.
The answer, although accurate, still felt slightly biased — almost like the system instinctively softened the psychological impact of the numbers by normalizing them within the broader AI race.
And honestly, I understand why.
Because the benefit of knowledge being delivered at our fingertips is genuinely incredible.
A student can learn quantum mechanics from a village with weak infrastructure. A founder can prototype an idea in hours instead of months. A developer can debug systems faster than ever before. The productivity gains are real.
But so is the cost.
For years, software scaled mostly through abstraction. AI may be the first mainstream computing paradigm where scaling intelligence also means scaling physical consumption in the real world.
That may ultimately become the defining tradeoff of this era.
The question after Google I/O is no longer just:
“How intelligent can these systems become?”
But also:
“What will it cost to sustain them?”