Power, Ping, and Pipe: The Hidden Industrial Base of AI

This text examines the physical and industrial foundation of artificial intelligence, challenging the popular perception of AI as a weightless, digital entity. It reveals how every interaction with a foundation model activates a global supply chain involving massive electricity consumption, specialized semiconductor fabrication, and intensive freshwater usage for cooling. The author argues that AI has transitioned from a software tool into a critical public utility, yet its governance remains insufficient because its environmental and social costs are largely invisible to users. To address these scaling challenges, the source proposes radical architectural shifts, such as relocating data centers to subsea environments or Earth’s orbit to leverage natural cooling and stable energy sources. Ultimately, the text serves as a call for rigorous accountability and a holistic understanding of the material infrastructure that sustains modern machine intelligence.

Transcript

SPEAKER_01 0:00

Every time uh a leader approves an enterprise AI integration, they basically believe they're purchasing this weightless, frictionless software.

SPEAKER_00 0:08

Trevor Burrus Right, exactly. They sign a vendor agreement, and uh they think they've just plugged into this ethereal, invisible service.

SPEAKER_01 0:15

Aaron Powell Yeah, but in reality, what they're actually doing is plugging their entire organization into a massive, heavily resource-intensive industrial supply chain. I mean, we're talking about millions of tons of concrete, uh ultra-high voltage power grids, and quite literally millions of gallons of water.

SPEAKER_00 0:34

Aaron Powell It is a profound miscalculation. And for the executives listening right now, especially if you sit in the role of chief information and resilience officer, you know, a CIR treating artificial intelligence like some sort of atmospheric cloud is no longer just uh a misunderstanding of the technology. No, not at all. It is a critical systemic vulnerability for your enterprise.

SPEAKER_01 0:53

Aaron Powell Absolutely. So welcome to this deep dive. Today we are unpacking a really pivotal advisory paper. This is by Dr. Stephen Wilson, and it's titled Power, Ping, and Pipe: The Hidden Infrastructure Behind Artificial Intelligence.

SPEAKER_00 1:06

It's a fascinating read.

SPEAKER_01 1:08

It really is. And our mission today is to trace the actual physical footprint of AI. So we're going to go from the absolute bedrock laws of thermodynamics all the way to uh global water dependencies. And finally, we're going to explore some genuinely radical future solutions like sub-sea and even orbital computing.

SPEAKER_00 1:28

Aaron Powell, which sounds like science fiction, but it's very real.

SPEAKER_01 1:32

Yeah, completely. And the goal here is to help you rigorously account for the true cost and the systemic risks of the AI systems you are adopting right now.

SPEAKER_00 1:41

Aaron Powell And to do that, I mean, we really have to start by completely dismantling this marketing abstraction that we all casually use, right? The cloud.

SPEAKER_01 1:49

Yeah, the cloud. It sounds so soft and harmless.

SPEAKER_00 1:51

Exactly. The paper actually leans on the philosopher Alfred North Whitehead, who had this brilliant observation. He essentially said that civilization advances by extending the number of important operations we can perform without thinking about them. Aaron Powell Right.

SPEAKER_01 2:05

So like when you flip a white switch, you don't want to think about the voltage regulation at the local substation, right? Yeah. That complexity is basically hidden by design. Trevor Burrus, Jr.

SPEAKER_00 2:12

Precisely. That is the hallmark of modern civilizational progress. And AI has really just become the newest member of this invisible infrastructure. I mean, when you ask generative model a question, the interface is completely minimalist.

SPEAKER_01 2:28

Trevor Burrus, Jr. It's this little text box.

SPEAKER_00 2:29

Trevor Burrus, Jr. Right. And it feels instantaneous. But beneath that blinking cursor is an incredibly intricate, heavy industrial lattice, you know, of electrical generation, fiber optic backbones, and just raw material extraction. Trevor Burrus, Jr.

SPEAKER_01 2:44

And I think what makes this so difficult for enterprise leaders to grasp is that historically, enterprise IT was a highly predictable domain. Things like email servers, web hosting, transactional databases, systems engineers actually call those bounded problems. Trevor Burrus, Jr.

SPEAKER_00 2:58

Right. They have a ceiling.

SPEAKER_01 2:59

Trevor Burrus, Jr. Exactly. You could easily forecast the growth curve in the server space you'd need. But AI is fundamentally different. It isn't bounded by typical IT growth metrics. It is bounded by thermodynamics.

SPEAKER_00 3:10

And that is the crucial shift every tech leader needs to understand. Every prior technological epoch had a physical limit, right? Like the Industrial Revolution was bounded by coal. The early information age was bounded by silicon manufacturing. While AI is bounded by the second law for thermodynamics, computation isn't magic. It is a physical process.

SPEAKER_01 3:30

It's a flip and switches.

SPEAKER_00 3:31

Exactly. Every single arithmetic operation, every single token generated by a large language model requires electrical energy to flip a transistor.

SPEAKER_01 3:40

And because no processor is perfectly efficient, a portion of that energy unavoidably turns into waste heat.

SPEAKER_00 3:46

Yes, exactly. It's a physical byproduct. You cannot virtualize it away, and you cannot um, you know, optimize it out of existence with just better software.

SPEAKER_01 3:55

Aaron Powell Let's talk about the sheer density of that heat because the hardware specs in Dr. Wilson's paper are just wild. A traditional enterprise server used, what, a few hundred watts of power?

SPEAKER_00 4:04

Yeah, around there.

SPEAKER_01 4:05

But today, a single modern AI accelerator chip can exceed 700 watts all on its own. So when you pack those into a data center, we've gone from a standard rack running maybe 10 to 20 kilowatts up to racks demanding 50, 100, or even several hundred kilowatts.

SPEAKER_00 4:20

It's a staggering jump.

SPEAKER_01 4:22

It really is. To put that in perspective, trying to cool a modern AI data center with traditional air conditioning is uh it's like trying to cool an industrial blast furnace with a desk fan.

SPEAKER_00 4:33

It's physically impossible. I mean, air simply lacks the thermal density to absorb and move that much heat away from the processors fast enough. At those extreme power densities, the silicon would literally melt.

SPEAKER_01 4:45

Wow.

SPEAKER_00 4:45

Yeah, the facility will experience catastrophic thermal failure unless you fundamentally change the cooling medium.

SPEAKER_01 4:51

Aaron Powell Okay, let me play devil's advocate for a second, though. Right. Because if I'm looking at tech history, components just become more efficient over time, right? Yeah. Like my smartphone has way more computing power than a desktop from the 90s and it doesn't burn my hand. Won't Moore's Law just naturally solve this? Won't engineers just design cooler AI chips?

SPEAKER_00 5:09

Aaron Powell It's a very fair assumption, but it conflates software bloat with thermodynamic structural limits. We aren't talking about, you know, poorly written code making a processor run hot. We are talking about the unavoidable physical cost of executing trillions of floating point operations per second. Yes, individual chips become marginally more efficient per calculation, but the scale of the AI models we are training is growing exponentially faster than any of those efficiency gains. The heat isn't an engineering bug, it is a structural reality.

SPEAKER_01 5:43

So if air cooling totally fails against that structural reality, the tech industry is essentially backed into a corner here. Yeah. They have to return to a 19th century industrial standard just to keep these machines running.

SPEAKER_00 5:54

They have to use water.

SPEAKER_01 5:55

Which takes us right into the core vulnerability defining AI's growth right now. Heat necessitates cooling, and that shifts our focus from electrical power directly to global water supplies.

SPEAKER_00 6:05

Right. Because water is the dominant thermal medium for a very specific physics reason. It has an exceptionally high specific heat capacity.

SPEAKER_01 6:13

Aaron Powell Meaning it can hold a lot of heat.

SPEAKER_00 6:15

Exactly. It can absorb substantially greater quantities of thermal energy per unit mass than air can before it actually increases in temperature. It's been the foundation of heavy industrial process cooling for over a century in chemical plants, in nuclear reactors, and now in hyperscale AI data centers.

SPEAKER_01 6:34

But Dr. Wilson is very particular about the terminology we use when tracking this. And for good reason, there's a massive difference between water withdrawal and water consumption. And if I'm an executive reading, say a sustainability report, I need to know the difference.

SPEAKER_00 6:49

Oh, it is arguably the most important distinction in infrastructure reporting today. Withdrawal is when you temporarily take water from a source like, say, a river or a municipal reservoir, you use it to absorb heat, and then you return a substantial portion of it to the watershed, usually just slightly warmer.

SPEAKER_01 7:06

Right. So the water still exists in the local environment.

SPEAKER_00 7:08

Exactly. Consumption, however, means that water is permanently removed from the local hydrological cycle. And in AI data centers, this most commonly happens through evaporative cooling towers.

SPEAKER_01 7:18

Let's break down the mechanics of that for a second. Why does cooling a server require destroying the water, so to speak?

SPEAKER_00 7:26

Aaron Powell It's due to the latent heat of vaporization. When you turn liquid water into a vapor, it requires a massive amount of energy to break those molecular bonds. By spraying warm water through an airstream in a cooling tower, a small portion of it evaporates. And that phase change pulls a tremendous amount of heat out of the remaining water, chilling it so it can be sent back to cool the servers. But that evaporated water, it drifts away into the atmosphere. It is gone from the local watershed.

SPEAKER_01 7:56

Aaron Ross Powell Okay, so the evaporation is the feature, not a bug, but it comes at the steep cost of local water supply. Right. And the paper dives into a taxonomy by Lai and colleagues that splits this footprint into two distinct categories on-site and off-site water. Trevor Burrus, Jr.

SPEAKER_00 8:12

Which is a vital framework for a CRO to grasp.

SPEAKER_01 8:14

On-site water is exactly what we just described, right? The direct operational water consumed at the data center itself to run those evaporative cooling towers. It's visible, it's local, and it's what most people think of.

SPEAKER_00 8:25

Right. And off-site. Off-site or indirect water is the water consumed upstream by the power plants that are generating the massive amounts of electricity the grid is sending to the data center in the first place.

SPEAKER_01 8:36

Aaron Powell Because generating electricity boils water.

SPEAKER_00 8:38

Yes. Thermoelectric power generation, whether it's coal, natural gas, or nuclear, is hugely water intensive. They use water to create steam to turn turbines, and then they need even more water to cool that steam back down.

SPEAKER_01 8:52

Aaron Powell Knowing that upstream complexity really makes you reconsider those viral metrics going around. Like you've probably seen the headline that one AI prompt equals one bottle of water. The paper argues that's a massive analytical flattening of what's really going on under the hood.

SPEAKER_00 9:06

Aaron Powell It is a completely reductive metric. The true water cost of a prompt is not some fixed physical constant. It fluctuates wildly based on the geographic location of the facility, the seasonal ambient temperature, the specific cooling architecture they're using, and uh the regional electrical generation mix.

SPEAKER_01 9:24

Aaron Powell So if the local grid is heavily solar or wind, the off-site water cost drops, but if it relies on a water-cooled nuclear plant, it spikes.

SPEAKER_00 9:33

Exactly. It also depends on the facility's power usage effectiveness, or PUE, and water usage effectiveness, WUE.

SPEAKER_01 9:42

I see those acronyms constantly in vendor pitches.

SPEAKER_00 9:44

Yeah.

SPEAKER_01 9:44

Conceptually, how should a leader read those? It's kind of like a golf score, right? The lower the number, the better.

SPEAKER_00 9:49

That's a great way to look at it, yeah. PUE is simply the ratio of the total amount of energy used by the data center facility divided by the energy delivered specifically to the computing equipment.

SPEAKER_01 9:58

Got it.

SPEAKER_00 9:59

A POE of 1.0 would mean 100% of the power goes directly to the processors. But because you need power for lights, you know, cooling systems, and pumps, it's always higher. And WUE measures liters of water consumed per kilowatt hour of IT energy.

SPEAKER_01 10:15

Aaron Powell Okay, but playing the skeptic here again, if the exact water cost changes based on whether I'm querying an AI model in the dead of winter in Norway versus like the middle of summer in Arizona, isn't it basically impossible for a CIRO to accurately measure and report their company's physical AI footprint?

SPEAKER_00 10:32

Aaron Powell It is incredibly complex, yes, but it is entirely possible, and more importantly, it is no longer optional. What that variability means is you just cannot rely on viral marketing metrics or generalized sustainability averages provided by a vendor.

SPEAKER_01 10:48

Right, you need specific.

SPEAKER_00 10:50

You have to demand rigorous localized infrastructural accounting. You need to know exactly which geographical zone your compute is occurring in and what the specific resource mix of that facility is.

SPEAKER_01 11:01

Because the water used to cool the servers and even the off-site water at the power plant, that's still only half the story. To understand AI's true footprint, we have to look at the global supply chain before the hardware is even plugged into the wall.

SPEAKER_00 11:14

Yes. Dr. Wilson refers to this as embodied water, and the scale is staggering. I mean, think about the semiconductor fabrication facilities, the fabs making the GPUs. Manufacturing these chips requires millions of gallons of ultra-pure water daily.

SPEAKER_01 11:29

And just to clarify the physics here, when we say ultra-pure water or UPW, we are talking about water purified down to parts per trillion. It doesn't even act like normal water anymore.

SPEAKER_00 11:39

It really doesn't. It's essentially an industrial solvent. At the nanoscale of modern transistors, and we are talking about gates just a few nanometers wide, a single microscopic speck of calcium, or just a solitary bacterium, would completely ruin the circuit path.

SPEAKER_01 11:54

That's insane.

SPEAKER_00 11:55

It is. UPW is used in wafer cleaning, focalithography, and etching. And without millions of gallons of this highly processed water, you physically cannot manufacture the silicon that makes AI possible.

SPEAKER_01 12:07

So this massive water footprint is fully realized and locked in before the GPU ever executes a single line of code. The paper heavily references Kate Crawford's book, Atlas of AI, which drives this point home beautifully. She argues that AI is fundamentally an extractive industry mining copper, lithium, silicon that is just wearing the aesthetics of software.

SPEAKER_00 12:29

The aesthetics of software, it's a perfect phrase because it really is a planetary scale industrial supply chain. And the sheer scale of the construction happening right now to support this is difficult to overstate. Hyperscale campuses are no longer measured in megawatts, they are being planned in gigawatts.

SPEAKER_01 12:43

A gigawatt? I mean, that is a unit of measure previously reserved for entire national power grids or large-scale nuclear reactors, not just a single campus of server racks.

SPEAKER_00 12:53

And that energy demand is actively reshaping the terrestrial power grid. The paper details how these AI facilities are so incredibly power-hungry that they are actually delaying the retirement of coal-fired power plants across the country.

SPEAKER_01 13:07

They're even slated to restart decommissioned nuclear plants like Three Mile Island, strictly under long-term power purchase agreements, to feed AI compute demand.

SPEAKER_00 13:17

Because these data centers are essentially acting as municipal-scale public utilities. They require dedicated substations, multi-year transmission queue positions, and massive expansions of local municipal water treatment capacity just to handle the cooling tower blowdown.

SPEAKER_01 13:32

But wait, if I look at local news, municipalities often bend over backwards to attract these data centers. They offer massive tax abatements, sometimes stretching for decades. From a city planner's perspective, don't these facilities bring high-paying tech jobs, infrastructure investments, and prestige? Isn't it a win-win?

SPEAKER_00 13:48

It can absolutely bring short-term commercial gains and, you know, an initial influx of construction capital. But Dr. Wilson warns about an asymmetry here that mirrors the old company-town dynamic of the steel or coal eras.

SPEAKER_01 14:01

Oh, interesting.

SPEAKER_00 14:02

A municipality might get construction jobs and a temporary tax boost, but they are inheriting a decades-long infrastructure and resource debt.

SPEAKER_01 14:10

Right. What happens if the local water basin is depleted from a multi-year drought?

SPEAKER_00 14:14

Exactly. Or what if the AI company achieves a breakthrough in chip architecture in five years, abandons the facility, and moves to a cheaper grid elsewhere?

SPEAKER_01 14:24

They just back up and leave.

SPEAKER_00 14:25

And the local weightpayers and the community are left holding the bag for the expanded utility substations and the diminished resource rights. The corporate strategic horizon is often much, much shorter than the infrastructural obligation they leave behind.

SPEAKER_01 14:38

Which naturally forces a really radical question. If citing these massive resource-hungry AI factories on land is causing such immense friction with human populations competing for our drinking water and straining our local power grids, why are we building them on land at all?

SPEAKER_00 14:53

That is the ultimate first principles design challenge the paper poses. We cite data centers near human cities mostly due to path dependence. In the early internet era, we needed web servers geographically close to users to reduce latency to make sure an e-commerce page loaded instantly.

SPEAKER_01 15:11

But AI training isn't like loading a web page.

SPEAKER_00 15:14

No. AI training workloads are batch-oriented. They crunch massive data sets continuously for weeks or months at a time. They are largely latency insensitive.

SPEAKER_01 15:24

So they don't need to be fast in that way.

SPEAKER_00 15:26

Right. They don't need to live next door to a major population center.

SPEAKER_01 15:28

Aaron Powell, which opens up some wild engineering possibilities. To go back to an earlier point, if you treat the Earth as a legacy operating system, instead of constantly patch it with more cooling towers on land, you can just move the workload to an entirely new OS. Let's talk about sub-sea computing, basically using the ocean as a natural thermal reservoir.

SPEAKER_00 15:47

It's a highly elegant solution to the thermodynamic problem we discussed earlier. Just below shallow depths, seawater temperature is cold and incredibly stable.

SPEAKER_01 15:56

And there's a lot of it.

SPEAKER_00 15:57

Right.

SPEAKER_01 15:58

More importantly, the thermal mass of the ocean is effectively infinite compared to anything we can engineer on land. The paper points to Microsoft's Project Matic, which proved that putting sealed data centers underwater actually reduced hardware failure rates.

SPEAKER_00 16:13

That sounds completely counterintuitive. Why would putting a computer in the ocean make it break less?

SPEAKER_01 16:18

Because you remove oxygen, which causes corrosion, you remove moisture variation, and honestly, you remove human technicians bumping into the racks.

SPEAKER_00 16:27

But Dr. Wilson proposes taking it even further with an open frame architecture, right?

SPEAKER_01 16:31

Yes. Instead of putting a traditional air-cooled data center inside a sealed pressure vessel, you engineer the compute modules to leverage direct liquid thermal exchange with the seawater itself. And you co-locate this architecture directly with offshore wind farms and submarine fiber optic cables. So it's all out there. Exactly. You have your power generation, your infinite cooling sink, and your data connectivity all happening miles out in the ocean. It totally bypasses the terrestrial resource competition. You aren't fighting a municipality for their drinking water. But of course, the ocean is a pretty hostile environment. You introduce severe engineering challenges around saltwater corrosion, biofouling, like literal barnacles growing on your heat exchangers and remote maintenance.

SPEAKER_00 17:14

And if the ocean isn't extreme enough for you, the paper goes one step further into extraterrestrial or orbital computing.

SPEAKER_01 17:22

Space data centers.

SPEAKER_00 17:23

Space is arguably the ultimate environment for this specific type of high-density, latency, insensitive workload. I mean, you have continuous, unattenuated solar power. There are no clouds, no atmospheric interference, and no night cycle if you position the orbit correctly.

SPEAKER_01 17:41

Earlier we established that air is terrible to cooling chips because it lacks density. Space is a vacuum. There's no air or water at all to carry the heat away. How do you cool a 700-watt GPU in a vacuum?

SPEAKER_00 17:51

Aaron Powell You have to transition entirely away from convective cooling, which relies on fluids like air or water absorbing heat and moving it away, and you rely entirely on radiative cooling.

SPEAKER_01 18:00

Okay, how does that work?

SPEAKER_00 18:01

You engineer massive infrared radiator surfaces. The waste heat from the chips is transferred to these panels, which then emit the thermal energy as infrared radiation directly out into the near zero Kelvin vacuum of deep space. It's exactly how thermal control systems on the International Space Station currently operate.

SPEAKER_01 18:19

Aaron Powell That is fascinating.

SPEAKER_00 18:20

The European Ascens study actually evaluated this recently, looking at orbital data centers specifically as a way to reduce the terrestrial carbon footprint of cloud infrastructure. It entirely removes terrestrial externalities. I mean, zero land use, zero freshwater competition.

SPEAKER_01 18:37

Aaron Powell But practically speaking, I have to ask the obvious operational question. This sounds brilliant for infinite cooling and, you know, solar power, but if a processor fails in low Earth orbit or at the bottom of the Atlantic, I can't just send a technician down the hall with a replacement part on a Tuesday afternoon. How does the maintenance math possibly pencil out?

SPEAKER_00 18:56

That is the binding constraint of off-world and offshore compute. You have to design for what engineers call infrequent, expensive servicing intervals. It requires highly autonomous robotic maintenance and incredibly resilient radiation-hardened components.

SPEAKER_01 19:10

Which has to be incredibly expensive.

SPEAKER_00 19:12

No, the initial capital expenditure and the launch costs are astronomical. But the paper's core argument is that the projected multi-gigawatt power requirements of future AI training clusters represent a demand profile so unfathomably large it might actually change the underlying capital calculus of space infrastructure. The terrestrial friction is becoming so high that space becomes economically viable.

SPEAKER_01 19:36

These offshore and off-world solutions are incredible feats of engineering, but honestly, they highlight just how massive the physical footprint of AI has become, which brings us right back to what leaders must do right now down here on Earth.

SPEAKER_00 19:49

If you are a CIRO or sitting anywhere in enterprise leadership, the takeaway from Dr. Wilson's brief is absolute. You must stop governing AI as a pure software deployment. It is not software, it is a critical infrastructure dependency.

SPEAKER_01 20:02

So what is the tactical move for them?

SPEAKER_00 20:04

You must mandate transparent disclosures from your vendors. Do not accept a generic green cloud badge, require audits of their embodied water from chip fabrication, and demand visibility into the upstream power consumption and off-site water footprint of the specific facilities running your models.

SPEAKER_01 20:23

That makes total sense.

SPEAKER_00 20:24

And most importantly, you must incorporate these physical industrial supply chain risks directly into your enterprise risk registers. Because if your vendor's data center loses its water rights, your AI capability goes offline.

SPEAKER_01 20:39

Because you can't manage what you don't measure. And right now the industry's just largely measuring the wrong things, or worse, hiding them behind smooth interfaces. The paper leaves us with a really provocative philosophical thought to mull over. Systems that become invisible also become highly resistant to public governance. As AI becomes as ambient and assumed as electricity, will we lose the ability to see its physical costs altogether?

SPEAKER_00 21:03

We only ever notice infrastructure when it fails. A rolling blackout makes the power grid visible. A dried up riverbed makes the reservoir visible. The question this advisory paper forces us to ask is what happens when the infrastructure of cognition runs out of water?

SPEAKER_01 21:19

It brings us right back to the cold reality we started with. Every time you authorize a new enterprise AI tool, you are not buying a weightless cloud. You are buying concrete, ultra pure water, copper, and a massive supply chain of industrial power. And the bill for that physical infrastructure is eventually going to come due. Thank you for joining us for this deep dive. Take a hard look at your vendor agreements this week, push past the marketing abstractions, and keep questioning the hidden realities of the technologies you rely on.