What if the debate about data centers isn't really about data centers? What if it's about something much larger What if it's about how societies decide which problems deserve attention, which narratives become dominant, and ultimately which solutions become acceptable?
Recently, I found myself following an interesting thread. A research paper helped popularize the idea that artificial intelligence consumes significant amounts of water. The story spread quickly. Articles appeared. Social media amplified it. Communities began questioning data center projects. Activists raised concerns. Policymakers took notice.
The paper that appears to have launched the modern "AI is thirsty" discussion is "Making AI Less 'Thirsty': Uncovering and Addressing the Secret Water Footprint of AI Models" (2023) by Pengfei Li, Jianyi Yang, Mohammad A. Islam, and Shaolei Ren. The authors were affiliated with the University of California, Riverside.
The paper combines:
- On-site cooling water at data centers
- Indirect water used for generating the electricity used
- Power usage effectiveness (PUE) overhead
- Geographic and time-varying water intensity factors
These were used to estimate a total water footprint, not water physically consumed at the data center. The authors explicitly argue that off-site water used in electricity generation and so forth should be included in the number. When people say, ”AI consumed X liters of water”, they are referring to the broader lifecycle framework introduced in this research, not cooling water pumped into a facility.
What I found interesting is that the researchers themselves do not appear to be advocating for orbital computing or space-based AI infrastructure.
Looking at their broader body of work, the focus appears to be environmental accounting of systems. Other papers from the same group explore carbon footprint, water footprint, environmental equity, geographic load balancing, and ways to reduce environmental impacts through better scheduling and energy sourcing.
In other words, their proposed solutions involve making terrestrial computing more efficient, not moving computation into space.
That distinction matters , as the those creating an article and embracing a narrative and the people who benefit from a narrative are not always the same people.
First Glance
The idea that data centers are bad seems like a simple story: Researchers discovered a problem.
The public learned about it.
As I looked deeper, I realized I was asking the wrong question. The more interesting question was:
What happened next?
The answer appears to follow a pattern common to modern issues.
- Research becomes a university press release.
- The press release becomes a news story.
- The news story becomes social media discussion.
- The discussion becomes activism.
- Activism influences policymakers.
- The public encounters a narrative rather than the original research.
The happens as the information is filtered through institutions, incentives, and attention.
Data Is Not a Conclusion
An important lesson I learned across engineering, software design, and decision-making is that data is not a conclusion. Data tells us what was measured. It does not tell us what it means. It certainly doesn't tell us what we should do about it.
In the case of AI and water consumption, many encountered headlines suggesting that AI uses enormous amounts of water. What fewer realized is that some of the most widely cited research wasn't simply measuring the water physically consumed by a data center.
The researchers were asking a broader question. They were using lifecycle accounting. Instead of only counting cooling water used onsite, they included water associated with electricity generation and other indirect resource consumption.
This is not wrong. It is simply answering a different question. The question is not : “How much water does the facility consume?” The question was: “What is the total water footprint associated with providing this computing service?” Both questions are legitimate. They are not the same question. One is a local infrastructure question. The other is a lifecycle environmental question. Confusion often occurs when people assume one question is answering the other.
The answer depends on which question you ask.
How Narratives Form
This is where something interesting happens. A researcher publishes a finding. The finding may be accurate. Different groups discover the finding. Each group interprets it for its own objectives.
- Journalists emphasize interesting parts.
- Activists emphasize the most concerning parts.
- Politicians emphasize the most actionable parts.
- Investors emphasize the profitable parts.
Eventually, the public encounters the narrative - rather than the research itself.
How many people read the original paper? How many consider the underlying assumptions? How many downloaded the supporting data?
Most do not. That isn't irrational. Modern society is simply too complex for everyone to independently verify every scientific claim, economic model, and technical analysis they encounter.
People instead rely on trust networks:
- Experts.
- Journalists.
- Organizations.
- Influencers.
The interesting question is how people decide which trust networks to trust. Not because someone lied. Not because there is a conspiracy. Information is filtered through incentives.
The result is that a measurement becomes an interpretation.
The interpretation becomes a conclusion.
The conclusion shapes public policy and investment decisions.
This pattern appears everywhere.
Climate change
Healthcare.
Nutrition.
Education.
Artificial intelligence.
A study read by one hundred people changes very little.
A study amplified by journalists, activists, policymakers, investors, and social media can influence millions who never encounter the original paper. By that point, many are responding not to the research itself but to the narrative built around it.
The story attached to the data evolves.
Over time, that story often becomes the accepted understanding.
An Unexpected Consequence
As concerns about terrestrial data centers grew, another industry began (paying attention, and) gaining attention - Space-based computing.
Orbital data centers. AI infrastructure in space. The companies pursuing these ideas did not appear to create the original water-consumption research. I found no evidence that they did.
What I found was something more interesting.
As public concern about terrestrial infrastructure increased, companies building space-based infrastructure began citing the same land based constraints:
- Water.
- Power.
- Land availability.
- Permitting delays.
- Community opposition.
- Grid constraints.
The strongest evidence I found was not that orbital-compute companies created the original water narrative. Rather, once terrestrial data centers became politically and environmentally controversial, orbital-compute investors and advocates began incorporating those same constraints into their investment thesis.
The argument became:
If building data centers on Earth becomes increasingly difficult, orbital infrastructure becomes attractive. What makes this interesting is that the researchers and the beneficiaries appear to be different groups. The researchers were asking environmental-accounting questions. Their work focused on measuring resource consumption and understanding the environmental impacts of AI infrastructure.
The orbital-compute industry found itself operating in an environment where those same concerns made its proposed solutions more attractive. That does not mean one group was working on behalf of the other. It illustrates how the people creating a narrative, and the people who eventually benefit from a narrative, are not always the same.
In other words, the same narrative that raised questions about terrestrial data centers strengthened the case for alternatives. The harder it becomes to build infrastructure on Earth, the more attractive infrastructure in space begins to look. Whether that was intentional or not is almost beside the point.
The result is the same.
Narratives influence the solutions society becomes willing to consider.
The Question Nobody Seems to Be Asking
This is where I think the conversation becomes most interesting. The original water-footprint research expanded the system boundary. Instead of asking:
"How much water does the facility consume?"
Researchers asked:
“What resources are consumed beyond the facility itself?”
That is what lifecycle analysis is supposed to do. The intellectually consistent next step would be to apply the same reasoning to orbital computing. If we are going to evaluate terrestrial data centers using lifecycle analysis, shouldn't we apply the same standard to space-based data centers?
Most discussions focus on the benefits:
- No local water consumption.
- No zoning battles.
- No neighborhood opposition.
Consider, every solution creates new problems:
- How much energy is required to manufacture orbital infrastructure?
- How many launches are required?
- What resources are needed to build rockets?
- What rare materials are required?
- How often must hardware be replaced?
- What’s environmental impact of launch operations?
- What infrastructure remains necessary on Earth to support infrastructure in space?
These questions rarely receive the same attention. Yet they are exactly the kinds of questions lifecycle analysis was designed to answer.
Comparing the Full System
A fair comparison would evaluate both systems using the same framework.
Terrestrial AI Infrastructure
Costs:
- Land use
- Power consumption
- Cooling systems
- Water consumption
- Grid expansion
- Local environmental impacts
Benefits:
- Existing infrastructure
- Easy maintenance
- Direct physical access
- Lower transportation costs
Orbital AI Infrastructure
Costs:
- Manufacturing
- Rocket production
- Launch operations
- Deployment systems
- Maintenance complexity
- Replacement launches
- Space debris management
- Satellite manufacturing
- Launch infrastructure
- Replacement hardware
- Servicing missions
- Debris mitigation
- End-of-life deorbiting
Benefits:
- Minimal local water use
- No local land use
- Reduced permitting challenges
- Continuous solar energy potential
One possibility is that orbital systems prove more efficient. Another possibility is that they consume more total resources while merely shifting those impacts to different places and different stages of the lifecycle. At the moment, surprisingly little public discussion focuses on that comparison.
Only after evaluating the complete lifecycle of both systems can we meaningfully compare them.
Why This Matters
Some may wonder why any of this matters. After all, isn't this simply an engineering problem? Perhaps. I think something deeper is happening. Most major decisions begin with an assumption. And assumptions hide inside narratives.
Once a narrative becomes accepted, people begin debating solutions before examining the assumptions that created the narrative in the first place. That is how entire industries can emerge. That is how policies can be created.
That is how billions of dollars can be invested, not because the data is wrong, ratheer the interpretation became more influential than the measurement itself.
The Real Question
In many ways, the article is not about data centers at all. Data centers simply provide a useful case study for observing how measurements become narratives, narratives become assumptions, and assumptions shape the solutions society pursues.
I am not arguing against terrestrial data centers.
I am not arguing for space-based data centers.
I am not arguing that researchers are wrong to study resource consumption.
I am suggesting we should be careful when observations become conclusions. We should be equally careful when conclusions become solutions. Before deciding whether AI belongs on Earth or in orbit, perhaps we should ask a simpler question:
What problem are we trying to solve?
Is it water consumption? Energy consumption? Environmental impact? AI growth ?
Each problem points to a different solution.
The original data-center water debate became influential because someone expanded the system boundary and asked a larger question.
What resources are consumed beyond the walls of the facility?
The same question can be asked of orbital AI.
What resources are consumed beyond the computer itself?
- Manufacturing.
- Launch.
- Maintenance.
- Replacement.
- Deorbiting.
If we're going to evaluate competing visions of the future, it seems reasonable to evaluate both using the same framework.
If history has taught us anything, it is that the solution depends on the question.