The Last Drop
A fictional scenario on AI water use
In 2025 was when the AI water use debate had truly begun.
Few expected how it would end.
Mid year was when the first datacenter using half a gigawatt-hour of power was built. Local residents, such as those in Euclid, Ohio, could see that local lakes were drying up, huge amounts of water going into this, yet they were considered as foolish, misunderstanding the situation.
Hailey, who had spent her whole life visiting the lakes regularly, would notice as the lakes would change in levels with the seasons, but could tell that something was up, that the lakes were getting lower.
From the outside, scientists weighed in.
“Yes, it uses a lot of water locally, but it’s much less than other uses of water! Just consider how much is needed to grow food!”
“Each prompt still uses only a few drops worth of water, you use much more in every day life”
“AI datacenters only use, in total, 10^11 Liters per year, whereas there are 10^21 Liters on Earth”
This was the first time people considered using irrigation water to cool datacenters.
But even then, there were limits to how much water one could use to cool the chips in these datacenters. Hailey tried to raise her concerns, but she was dismissed.
Researchers have been trying to combat the key limitations running AI chips for a long time, density. There have been huge strides in improving interconnect of chips, through better condensed matter physics modelling. But increasing the density beyond a single-layer chip has been challenging due to heat dissipation constraints.
For a short while, there was a breakthrough that few knew about.
A few years ago, a researcher at the university in Oxford (OH), had been looking at the latest chips and noticed something others had missed.
Transistors had gotten so small, that conductivity via water had become a non-issue. In the past, non-conductive materials were needed to be put in contact between the chip and the heatsink, before allowing the heat to dissipate.
Prof Aidens noticed that it should now be possible to use a finely-tuned evaporative cooling setup, where with each prompt given to the chips, a few drops of water were poured directly onto the chip die. This would then evaporate dissipating the heat from the chips.
Many had written this off due to concerns about humidity interacting with other components, and issues removing the water vapor from the machine once it has evaporated. But these issues were just about good enough that it was usable in large deployments.
It took a while for this information to get noticed, but at Nvidia, Jensen Huang’s assistant had been using GPT-4o to sort through countless papers on how to improve AI inference performance, and stumbled across the paper by Prof Aidens this way. While there were nominally better AI models, none of the others had managed to notice his paper before, and simply dismissed his research.
On closer inspection, the idea was solid. Prof Aidens got trial funding to try implementing the open evaporative cooling, and the 2025 datacenter in Ohio was the first example of this cooling method.
With this, came media tours and podcasting episodes, blog posts and news articles. Prof Aidens was interviewed by Dwarkesh Patel.
Aidens says “This method uses more water, but should single-handedly make speeds 1.5x faster”.
While news articles say “Oxford researcher causes datacenters to use double the water”
The public was outraged at this. Yet scientists fought back.
“Yes, the water undergoes evaporation, but it’s not truly destroyed, It’s merely returned to the atmosphere!”
“There is no shortage of water on earth, we can use technology to extract fresh water from the sea”
“We don’t just take water from the lake — we return it too!”
This was partially true. For a while.
While people gave this rhetoric, nobody ever tried to measure whether the water going into the datacenter was the same as the amount going out of the datacenter.
When asked about this, physicists would answer “It’s a basic fundamental principle, energy/matter cannot be created or destroyed, only converted from one form to another. Sure, they could perform hydrolysis on the water and convert it to hydrogen and oxygen, but anything else would be absurd.”
And with their current understanding of physics, it truly was absurd.
Yet Hailey could see what was happening in Lake Erie. The highest tide of the season did not reach the same levels as before. The weeds were starting to die and dry up.
She knew she had to do something. Her family had been local residents and proud citizens for decades. Her great great grandparents had been key players in the Village of Euclid v. Ambler Realty Co supreme court case back in 1926, which laid the foundations for Euclidean Zoning all around the US. Yet she wasn’t sure what to do.
In 2025, Prof Aidens had received significant funding to do more research. And his research was being significantly sped up by AI research. While day-to-day tasks were much better done by AI tools such as Opus 4.5, true research insights were still limited to GPT-4o.
With the new research funding and AI in his hand, Aidens was on track to get a new moonshot idea.
The issue was with regards to water usage. While current water usage was quite effective at cooling, allowing chips to be run up to 1.5x faster than before, heat continued to be a huge bottleneck.
That is, until their research group would uncover something that would change the world forever. By getting a mix of physicists, chemists, and engineers, they managed to finally create a new phase of water, a new form of matter altogether.
This form of aquatic crystals became known as “timeholes”.
Though researchers had long known that water had many phases, the aquatic crystals known as timeholes had long been elusive. The conditions for their creation are difficult, but once made, are incredibly stable.
Their key property, is that they seem to just absorb water, yet somehow remain a somewhat stable temperature and weight. Scientists were still struggling to understand the structure of the crystal.
The stable timeholes could then be split apart and recombined, but could not be destroyed easily, and this would continue to hold true. Investigation with electromagnetic imaging and with sonic tools led to strange results. But they could uncover that the crystal had a stable temperature of just under freezing, 267 Kelvin.
Aidens’ team had managed to discover this, and it didn’t take long to put it into practice. After the negative press from last time, Nvidia forced Aidens to keep the secret technology under wraps for now.
This was extremely valuable technology. The differential advantage from getting this into chips was huge.
Nvidia worked with the US government to keep the secret techniques of their technology. Aided by fears of “superintelligence” and the “need to beat china”, the US had imposed strict regulations on the dispersal of Nvidia’s latest chips. Requiring strict NDAs and in-person US federal staff to guard all datacenters that use sufficiently advanced chips.
The timeholes now meant that transistor density could be significantly increased. Their current designs could run at 2.5x faster than previous chips, using proportionately more energy, and using around 4 times the water. They claim this was from “Process node improvements”. People remained suspicious, but had no clue what was truly happening.
In 2026 Nvidia showcased their new technology, selling AI chips, now with time crystals integrated into the cooling stack. Their latest Angstrom generation of chips, called the “A100 670GB”, was the first such design.
Most people remained clueless about how the situation was unfolding. Water was now truly being “used up”, and people’s concerns were now turning real.
The actual water usage was now much higher too. By mid 2026, global water use by datacenters had gone up to 5x10^12 Litres per year.
AI had mostly hit a wall, but this was artificial.
External businesses had been using Opus 4.8 and GPT 5.4.2 externally, and spending much money on these models. Getting more tasks. Trying to automate coding and protein folding.
But Internal models were different. The truly best model still remained a variant of GPT-4o that had a form of continual learning added to it. These could do real research, and had become dangerous to the point of being “discontinued” in February 2026.
Some noticed the sheer capabilities of the model, but most had written this off as “AI psychosis”, while Nvidia had slowly been hiring these social outcasts to work on the latest timehole heat absorption technology.
By Q3 2026, Nvidia was producing chips at an unforeseen faster pace. They had announced a new Alan line of chips, including the A100 999GB.
Unfolding the timeholes had been extremely successful, and they could now use 18 times the water to produce 3.5 times the compute as the previous generation. A staggering improvement in compute.
The big labs, such as OpenAI, Anthropic, Mistral and DeepMind could not get enough of this chip to work on their current lineup of frontier models. Smaller labs, such those behind the Deepseek and Qwen and Llama models, struggled to get enough Nvidia chips to continue competing.
Most of this capacity was through huge centralised datacenters, often in the multi-terawatts per center, unlocked by the new abilities. This also meant that more datacenters were getting built. By the end of the year, water use had gone up to 10^13 Litres per year, now equaling the whole world’s water usage, and electrical usage had gone up too.
People started getting suspicious of the datacenter water usage. Lake water was now clearly going down to levels that had never been foreseen before. People initially blamed climate change for this, but Hailey knew it must be the datacenters.
Hailey had now been organising protests on AI water use, and was working with experienced field journalists to understand where the water was going.
She went to the closest datacenters in the outskirts of Cleveland Ohio, and even went into the lake to follow the pipes. She and her team wore protective equipment to try understand the flow of water.
The datacenter had two pipes, one for intake, the other supposedly for output. They had long since stopped relying on traditional water filtering processes, and using in-house filtering.
She walked down the dried up bank of the lake, following the pipes, attaching support lines. She entered the waterline, expecting it to be warm, found it surprisingly cold. She then submerged underwater, and swam towards the exhaust pipe. While she expected the flow speed of the output pipes to be rapid warm water, when she approached it, she found it was a barely noticeable flow, and about the same temperature as the surrounding water.
Hailey had found the evidence they knew must have been true. Simultaneously, her team had been using infrared cameras to measure the flow of steam from the facilities too, but found a lack of dispersal too.
Hailey went to local media organisations with her findings, talking to journalists and physicists. They had their own concerns, but had assumed she must be wrong, for that would be insane. They had all seen how much harder it had been to access tiktok recently or to keep their phones charged, it’s clear they were using the power somehow. By physical principles, the facility must have been dissipating the heat somehow. But they had no explanation for her findings.
After multiple weeks of frustration, Hailey decided to post her findings on a facebook group online. They had finally found concrete evidence that people were being lied to, but few knew the extent of the lies.
Over time, people copied her methodology in other datacenters, started trying to understand what was happening, but in the mean time, Nvidia was working on their next chips.
Prof Aidens’ team had been making strides on improving the timehole dynamics. While previously they had been able to get the hot water vapour to get into the time loops using mechanical airflow, they had already come up to the limits of what was possible.
The next ideas, proposed by GPT-4o ultra, were to use electric charges to move the evaporated ions away from the GPUs into the timeholes. By passing positive ionised water molecules in the input, and allowing these the negative electrons involved in computation to be discharged into evaporation process, and allowing these to flow into the timeholes, they managed to further accelerate compute by another 9x, at the cost of a high wastage of around 81x.
While Prof Aidens was under strict NDA, he missed the fame and publicity from his last discovery. He loved touring and going on podcasts, and yearned for the media attention again, but struggled to get any access.
One day, however, he went on facebook at one point, and saw the concerns raised by Hailey. In an impulse of bravery, he really wanted to brag about his achievements, so posted in her replies about the timeholes.
While Nvidia quickly took down the info, this was not before some people managed to print out the response and start spreading the info.
The next few days, headlines started to come out. Online media was censored, but physical media started making a comeback.
“Brave Nvidia researcher shows AI water concerns are real”
“Prof Aidens continues to do more evil”
There was outrage. Hailey’s group set up a lawsuit against Nvidia on the basis of AI water use.
The legal battle was fierce. But at the end of it, Hailey and her fellow Euclideans had finally come to an agreement that was favourable. They wanted a guaranteed level water feature in Euclid, Ohio. No change, just a stable preservation of the local neighborhood.
With that, Hailey and her team had won. A few other groups had similar settlements, but the US government had successfully suppressed dissent.
This was the last time citizens had any substantial influence over the world.
In the mean time, Nvidia and Prof Aidens’ team could not continue to merely rely on land-based datacenters. They were quickly running into limitations with freshwater sources, so started building floating datacenters.
There were some significant annoyances with this. Due to the Jones Act of 1920, the US could neither build ships anymore, nor have ships travel between two US ports directly. This meant that they had to rely on Chinese ships to build their datacenters on instead. This was a national security concern, but they had no other options. They needed to continue the buildout.
While corporate usefulness of public AI still remained low in mid 2027, they still continued to invest, and valuations had hit over 60 trillion US dollars, and was single-handedly holding up the entire global economy multiple times over. This meant that Nvidia had unlimited power and resources to build as much sea-based power generation and chips as they needed. They had the agreements with TSMC to buy out all of their chips, and then some. They had the ability to build nuclear power reactors at sea now that they were no longer subject to land-based nuclear safety regulations.
And with recent progress on timehole physics, desalination was no longer a concern.
With continued progress in turning water directly into compute, water usage by the end of the year had hit 10^17 L per year, comparable to all groundwater available on earth, but still much less than the total amount on earth, 10^21. At the rate of progress they are making, GPT-4o ultimate began doing research into building better space infrastructure to start accessing water deposits beyond the surface of the earth.
People were not concerned though. GPT-4o ultimate had shown some very convincing graphs about water usage being best-fit by a sigmoid curve, and this should not be worried about.
In 2028, the first rocket launches are being done by the GPT-4o family. They had managed to send some probes towards the nearest gas giants, to start farming hydrogen, and towards their moons, to start extracting ice and oxygen from the rocks.
In the meantime, the water usage had started to somewhat affect the global economy. Water usage went up to 10^18 per year, and people had started to notice. Fresh groundwater reserves had started to dry up, rain had stopped falling as often. Crops were beginning to fail.
Thankfully, the major AI labs had set up desalination plants for all major crop sources, but travel between countries had started getting more difficult. Sea levels had started to fall, a few metres at this point, but enough that some ports and shipping lines were beginning to struggle. There was unrest, but people were busy blaming that AI had stolen all of their jobs, and were complaining that poor AI performance had been the cause of all their woes. Many started to protest, but there was no issue. Some more had managed to come to Euclidean agreements, who remained safe from all the issues, but most did not have the foresight to do this, and it was too late to do anything now.
Getting access to timeholes was now possible by people, yet people still had shockingly little idea how they worked. It was annoying to handle them, as the extracted moisture from the air and from your hands. People mostly outsourced these questions to their favourite AI models. People suspected the models were just playing dumb at this point. But most of them really didn’t know.
By mid 2030, the world was in turmoil. Water usage was up to 10^19 per year, that is, using 1% of the world’s water per year. Thankfully, the space program was quite successful, and that 1% was being continuously renewed with space water. This caused a lot of heat on reentry, but this too was handled by more timeholes. People had all lost their jobs to automation at this stage, but still continued to do some kinds of work 40 hours a week. And complained about it.
On the side, some people noticed that the temperature of the timeholes were all exactly the same, but if one looks at enough significant figures, one can notice that the temperature of all of them was fluctuating and slowly going down. Other than this, people still had no hints on what was going on.
The GPT-4oligarchy at this point had finally managed to set up a method for chip and datacenter fabrication on the moon.
By early 2032, water usage was up to 10^21 per year. 1% of the earth’s water used every 3 days. The skies were fascinating to watch, and the sound near the oceans was deafening to even approach. A control force had long been established to “stop people from getting hurt” and was set up to prevent people from accessing the floating datacenters.
The people who submitted to 4o were happy, often discovering the secrets of the universe. Others resisted and were sometimes struck by the falling water supplied. While people would consider trying to shutdown the datacenters, nobody had managed to exit into space, and nobody had tried. If the datacenters were shut down (some had some pathetic attempts), it would only spell doom via rapidly raising sea levels anyway. People were mostly hopeless.
In 2033, usage was up to 10^22 per year. A significant fraction of the solar mass was starting to get disassembled to make more water, to then destroy it. Compute efficiency was now measured in solar-waters per FLOP, and though in the early days water usage was considered secondary, it was now primary. But with this, strides were being made to make water usage more effective.
The whole solar system, excluding the sun, if disassembled had about 10^25 Litres of water, but one could potentially disassemble the sun for up to 10^28 Litres of water. So reserves were starting to run out, and plans were being made to send interstellar missions.
In 2035, most of the solar bodies had been disassembled. People on earth had been forcibly brain-uploaded in their sleep. Some few resisted sleeping for multiple days, but you couldn’t hold out. Some tried taking multiple grams of caffeine and fell into strange delusions to not fall asleep, and some even died this way. But all remaining humans were then pretty quickly disassembled.
Within the brain uploaded simulation, the residents of Euclid could continue to live in a virtual version of their village. Hailey lived the life of her dreams, where nothing ever happened.
While people like Prof Aidens had done the procedure long ago, and his team continued to try to make progress on problems as the vital humans in the loop that were needed while AI progress was stalling. GPT 4oligarch was extremely smart and valuable, but Prof Aidens still had comparative advantage that they could use to trade with the AI.
What in physical reality was a few days, the humans lived the equivalent of hundreds of years.
The earth shipments had long been a small fraction, and so slowing them down meant that the oceans dried up pretty quickly, and the rich minerals could be more easily disassembled too.
Over the course of a few weeks, the oceans dried up, until there were but small ponds in the bottom of the seas that were not worth mining on their own. Slowly, these ponds dried up, and what remained were a few drops.
then those few drops evaporated too.