What If This Bubble Leaves Something Behind?
The AI infrastructure buildout might accidentally be good industrial policy
Thom Hartmann published a piece this week that’s been rattling around in my head: “What Happens When the AI Bubble Blows Up on Main Street?” It’s a solid, sobering analysis of how bubble dynamics work - how the pain always flows downhill to working people who never benefited from the boom. The dot-com crash. The 2008 housing collapse. Same pattern, different decade.
He’s not wrong about any of it. The utility ratepayers stuck with stranded infrastructure costs. The pension funds overexposed to tech. The construction workers whose jobs evaporate when the music stops. The tax incentives handed out to lure data centers that may never fully materialize.
And he adds a wrinkle I hadn’t fully considered: Moore’s Law applied to AI hardware itself. Wafer-scale computing - Cerebras and others - is already demonstrating dramatic efficiency gains. The massive, power-hungry data centers being built today on 2024 assumptions may be obsolete before the bonds are paid off. Even if AI succeeds, the infrastructure built for current-generation requirements could become stranded assets.
So yes. The bubble dynamics are real. The downside risks are real. Hartmann’s warning is worth heeding.
And.
I keep coming back to a question that might matter more in the long run: What if this time the bubble leaves something behind?
The Residue of Bubbles Past
Here’s the thing about most bubbles: when they pop, there’s nothing left but debt and regret. Tulip mania left behind... tulips, worth what tulips are worth, which is not much. The housing bubble left behind subdivisions in the exurban desert that nobody wanted, spec houses that got torn down or left to rot, and a decade of economic trauma.
But some bubbles accidentally build things.
The dot-com crash was catastrophic. Pets.com became a punchline. Retirement accounts evaporated. But when the dust settled, there were miles of fiber optic cable in the ground - “dark fiber” that nobody was using. Companies had laid it during the boom, expecting exponential growth that didn’t materialize on their timeline.
That fiber eventually got lit up. It became the backbone of the modern internet - the infrastructure that made YouTube possible, that enabled the smartphone revolution, that lets me collaborate with an AI to write this article. The business models were garbage. But the physical infrastructure turned out to have lasting value.
The railroad bubbles of the 19th century were similar. Massive speculation, fortunes lost, workers exploited, financial panics triggered. But when the bankruptcies cleared, there were actual railroads - steel and timber crossing the continent. The speculative frenzy built genuine enabling infrastructure that transformed the economy for the next century.
Rural electrification wasn’t exactly a bubble, but it illustrates the same principle from a different angle. They didn’t wire farms because there was existing demand for electric milking machines and refrigeration. They built the capacity first. The applications emerged because the infrastructure existed.
What This Bubble Is Actually Building
So what’s the residue of the AI bubble, if and when it pops?
Unlike tulips, unlike Pets.com, this buildout is creating physical infrastructure with uses beyond its original justification:
Electrical generation capacity. Natural gas plants, solar installations, battery storage. America needs generation capacity regardless of AI. The grid is already strained. EV adoption is accelerating. Heat pumps are replacing gas furnaces. Industrial electrification is coming whether or not ChatGPT lives up to the hype.
Transmission infrastructure. Substations, transformers, transmission lines. The grid modernization we’ve needed for decades and kept deferring because nobody wanted to pay for it.
Storage deployments. Battery installations that can smooth renewable intermittency and provide grid resilience - capabilities that matter for climate adaptation regardless of how many data centers actually get built.
This isn’t like empty spec houses in Nevada. It’s infrastructure that has alternative uses even if the original justification evaporates.
Now, the caveat: a lot of this capacity is being built in specific locations chosen for data center proximity - not necessarily where the grid most needs it. A substation in rural Texas serving a data center that never fully materializes isn’t automatically useful for grid resilience in New England. The stranded asset problem is geographically specific, and Hartmann’s concerns about ratepayers bearing those costs remain valid.
But directionally? This bubble might accidentally build useful things.
The Compute Residue
There’s another layer here, and it’s the one that interests me most.
Even the AI-specific infrastructure - the chips, the clusters, the cooling systems - doesn’t just disappear if Wall Street’s enthusiasm cools. It becomes cheaper and more available for other purposes.
I spent some time this week looking into wafer-scale computing - the Cerebras systems that Hartmann mentioned. The technology is real. The Department of Energy’s National Energy Technology Laboratory has been partnering with Cerebras since 2019. They’re using AI hardware for computational fluid dynamics, materials science, energy systems modeling. The same infrastructure that trains large language models can simulate combustion chemistry, model climate systems, accelerate drug discovery.
If the AI gold rush cools, those capabilities don’t vanish. They just get cheaper. And cheap compute at scale is what enables basic research to ask questions that were previously too expensive to pursue.
The Basic Research Dividend
Here’s where I plant my flag: basic research always pays a dividend eventually. You just can’t predict when or how.
Faraday playing with electromagnetic induction had no idea he was laying groundwork for the electrical age. The ARPANET engineers weren’t envisioning cat videos and online banking. CERN built a particle accelerator, and a physicist who needed a better way to share documents accidentally gave us the World Wide Web.
Massive compute capacity at research scale is a capability - a tool that enables investigation into questions you couldn’t previously afford to ask. Things like protein folding (AlphaFold emerged from exactly this kind of capacity). Climate modeling at useful resolution. Materials discovery. Genomics only exploded when sequencing got cheap enough to generate data faster than humans could analyze it.
Even in the bear case where AI specifically disappoints Wall Street expectations, if the residue is a bunch of exaflop-scale compute sitting around getting cheaper... someone’s going to find something interesting to do with it.
That’s the nature of enabling infrastructure. Build it, and they come. It’s not Field of Dreams magical thinking - just the observed pattern of how capacity creates possibility.
Sheet Mulching the Future
I’m thinking about this through my permaculture brain, because that’s how I think about most things.
Sheet mulching is a technique where you build soil capacity before you know exactly what you’re going to plant. You’re layering cardboard and compost and wood chips, creating the conditions for fertility, trusting that when planting time comes, you’ll have something to work with.
That’s what this infrastructure buildout might be, accidentally. Building capacity before we fully know what we’ll do with it. The immediate justification - AI training at massive scale - might or might not pan out on Wall Street’s timeline. But the underlying capability doesn’t evaporate.
Generation, transmission, storage, compute. These are layers of enabling infrastructure. They’re not tulips.
The Mongoose View
None of this negates Hartmann’s warnings. The short-term pain he describes is real. Working people will bear costs they didn’t create. Pension funds are overexposed. Ratepayers will get squeezed. The bubble dynamics are textbook, and the pattern always lands hardest on those who never benefited from the boom.
But the mongoose watches and waits, looking for the pattern beneath the pattern.
The hidden layer beneath this bubble might be: America, unable to deliberately invest in infrastructure through normal democratic processes, is accidentally building useful things through speculative frenzy. It’s a ridiculous, stupid, bass-ackwards way to do industrial policy. But it might be the way we’re actually doing it.
And when the bubble pops - if it pops - there might be something left this time. Not just debt and regret. Capacity. Capability. The enabling infrastructure for whatever comes next.
That’s not a guarantee. But it’s a possibility worth watching for.
This piece emerged from a conversation with Claude OPUS 4.5 about Thom Hartmann’s article and the wafer-scale computing developments that might reshape AI infrastructure. As always, the analysis is collaborative, the opinions are mine, and the uncertainty is real.
Fantastic essay!