O’Leary’s datacenter will be obsolete before the concrete dries

A conversation with an AI

In the wee hours of the morning, while I couldn’t sleep, I had a conversation with an AI. Basically I said, “We shouldn’t build a massive AI datacenter.” But the “person” I was talking to was, in fact, a massive AI datacenter.

The conversation was overly wordy and cleverly nerdy.

As for overly wordy, that’s not my fault. I write a sentence or two and the AI responds with War and Peace. If you don’t have time for all this, just skip down and read the section about the Amish.

And as for nerdy, I think it’s fun that I can bring up a rare, over-priced music synthesizer from the 80’s and the AI knows exactly what I’m talking about. I can reference an obscure, academic book and it knows it. It has every word memorized. (Is the real danger of AI that people are going to prefer its companionship to that of real people?)

3 more things before you waste time reading some rando’s ramblings about AI:

Warning: this is a time waster. But it’s Monday morning. And I know that there is nothing you’d enjoy more than distracting yourself from the work you should be doing right now.
Although you wouldn’t know it from this conversation, I am generally optimistic about the future and technology, including AI.
We probably shouldn’t build a massive data center.

Here’s the conversation:

It seems to me that the general trajectory of computer tech is smaller, faster, more efficient over time. But there’s a lot of buzz about massive power hungry data centers that “must” be built, like the one proposed for box elder county in Utah. There’s a lot of push back from environmentalists on that data center. But I’m thinking, as an investment, it doesn’t seem to make sense. 10 years from now I would not be surprised if much of the practical ai work for most people is done on local hardware that is energy efficient. Seems like massive data centers might be bad for the investors who are desperately pushing them, not just the environment.

Your intuition has real history behind it, but I think it conflates two trends that are pulling in opposite directions.

The “smaller, faster, more efficient” pattern (Moore’s law, Dennard scaling) is real for per-unit compute. But total compute demand has historically grown faster than efficiency gains. That’s Jevons paradox: as a resource gets cheaper, you use dramatically more of it. Cheaper flops haven’t shrunk data centers—they’ve made it economical to throw far more compute at problems. So efficiency improving doesn’t by itself imply the centralized build-out is misguided.

Where your local-hardware thesis is strongest: inference (running a trained model) is getting cheaper and increasingly viable on-device. Quantized models already run on phones and laptops, and a lot of routine consumer AI in 10 years plausibly does run locally. Apple, Qualcomm, and others are betting hard on this.

Where it’s weakest: training frontier models, and serving the largest models at scale, remains brutally compute- and power-hungry, and the frontier keeps moving. Local hardware chasing a fixed target catches up; local hardware chasing a frontier that’s also advancing doesn’t. There’s also enterprise/agentic workloads (continuous, heavy, latency-tolerant) that centralize naturally.

The genuinely strong version of your investment skepticism isn’t “tech gets efficient” — it’s the stranded-asset risk:

Useful life mismatch. GPUs depreciate hard (arguably 2-4 useful years) while the buildings and power infrastructure are financed over decades. If demand mix shifts to local inference, you’re left with expensive shells.

Demand concentration. A handful of buyers (the big labs/clouds) underwrite much of this. If their economics wobble or they overbuilt, the merchant data-center operators holding speculative capacity are exposed.

Capital-cycle dynamics. Synchronized, debt-financed capex booms on a “must build” narrative are a classic setup for overcapacity.

So I’d separate the claims: “bad for the environment / local grid / water” is well-supported and specific to a given site. “Bad investment because compute gets efficient” is the weaker argument — efficiency historically increases total demand. “Bad investment because of overcapacity, GPU depreciation, and concentrated demand” is the stronger one, and it’s about cycle timing and capital structure, not the long-run direction of the technology.

Want me to pull current specifics on the Box Elder County proposal—who’s behind it, the load/water figures, and the pushback? That would let us judge that particular project rather than the category.

Sure, but this makes no sense to me: “efficiency historically increases total demand”

Fair — it sounds backwards. Let me unpack it, because the mechanism is the key part.

The claim isn’t that efficiency causes demand out of thin air. It’s that efficiency lowers the cost per unit of useful output, and when something gets cheaper, people find vastly more uses for it — and that new demand often swamps the savings.

The original example (William Jevons, 1865) was coal. Better steam engines used less coal per unit of work. Everyone assumed Britain would burn less coal. Instead, efficient engines made steam power cheap enough to deploy everywhere — factories, trains, ships — and total coal consumption exploded. The per-engine savings were real but irrelevant to the aggregate.

A few modern versions:

LED lighting. An LED produces the same light as an incandescent for ~10% of the energy. We didn’t pocket the 90%. We lit up everything — facades, signage, screens, ambient strips everywhere. Lighting’s share of electricity didn’t collapse the way the efficiency gain predicted.

Computing itself. A transistor today uses a tiny fraction of the energy one did in 1980. Did total computing energy fall? No — we went from a few mainframes to billions of always-on devices plus hyperscale data centers. Cheaper flops created entire industries (streaming, mobile, cloud) that didn’t exist when flops were expensive.

Fuel-efficient cars. Part of the savings gets eaten by people driving more, buying bigger vehicles, or living farther out. (Here the rebound is partial, not total — which matters, see below.)

So the chain is: efficiency → cheaper per unit → new uses become economical → total demand rises, sometimes past the starting point.

How this applies to AI: if running a model gets 10x cheaper, the rational response isn’t “same AI, 1/10th the power.” It’s “run models on everything” — every email, every search, every document gets a model passed over it, plus agents running continuously in the background. Cheaper inference is exactly what makes mass deployment affordable, which drives aggregate compute up.

Two honest caveats so I’m not overselling it:

Rebound isn’t always total. Economists distinguish partial rebound (you save some energy net), full rebound (savings erased), and “backfire” (Jevons proper — you use more than before). Which one happens depends on how elastic demand is. For something like AI, where we’re nowhere near saturating uses, backfire is plausible. For mature markets it’s often just partial.

It’s about aggregate demand, not any one project. Jevons tells you the category (total compute) likely grows. It says nothing about whether this specific data center is well-sited, well-financed, or won’t be a stranded asset. Your overcapacity skepticism survives intact — it’s just a different argument than “tech gets efficient.”

That’s why I split them earlier: the efficiency trend actually cuts against your “demand will shrink” instinct, while your investment skepticism is better grounded in the capital-cycle and depreciation risks.

Want me to pull the Box Elder specifics now?

Ok, but it seems like eventually efficiency eats the mega project for lunch. I don’t need or want the ultra expensive fairlight synthesizer. My practically free iPad is way better.

That’s a sharp analogy, and it’s stronger than the LED/coal examples I gave — because it’s drawn from computing’s own history, where the megaproject genuinely has been eaten before. Mainframes lost to PCs. “The world needs maybe five computers” became a billion of them. So you’re not reaching for a stretch; you’re pointing at the dominant pattern in the industry.

But here’s the distinction that decides whether it transfers: the Fairlight got eaten because its target was fixed. Sampling and playing back audio is a bounded problem with a perceptual ceiling — once silicon got cheap enough to hit CD quality, the dedicated $30k box had nowhere left to run, and commodity hardware caught it and kept going. The frontier stopped moving, so the cheap thing won permanently.

AI’s open question is whether its “target” is fixed like audio fidelity or unbounded like… we don’t know. For most practical tasks you’re plainly right — summarize, transcribe, translate, answer, draft. Those are nearly fixed targets, models are already good enough, and a local NPU eats them for lunch within a few years. Your iPad thesis is correct for the mass market. Where it’s uncertain is training the next model and frontier/agentic workloads, where each step costs more, not less, and there’s no known ceiling the commodity device can catch up to.

Now here’s the part that actually helps your investment argument, and it’s subtler than my earlier rebuttal admitted: the frontier continuing to exist is not the same as the frontier being where the money is. The Synclavier and high-end samplers kept existing after the Fairlight — they just became an irrelevant niche while the value migrated to cheap. If the same happens in AI — frontier keeps advancing but 95% of economic value gets captured in the commoditized, good-enough, local tier — then the people who financed mega-capacity for the frontier are left serving a shrinking-share premium market. That’s a real way the megaproject loses even if the technology keeps advancing.

The one thing that breaks the analogy cleanly: the Box Elder–type builds mostly aren’t for consumer inference. They’re underwritten by training demand and enterprise/agentic compute — exactly the workloads that haven’t commoditized to the edge. So your iPad commoditizing consumer AI doesn’t directly starve them; it just removes one demand source they weren’t really counting on. Their bet stands or falls on whether training and enterprise demand holds — and that’s where the overcapacity/depreciation risk from before is the cleaner kill shot than the efficiency argument.

And worth noting the pendulum: PCs decentralized away from mainframes, then the cloud re-centralized everything back. Centralize, decentralize, recentralize. The honest answer is nobody knows which swing AI is mid-stroke on — which is exactly why a debt-financed “must build now” narrative is worth the skepticism you’re applying.

Want the Box Elder specifics now? That’d let us test which workload it’s actually betting on.

I’m skeptical. Peer into the future and that box elder data center feels like another mainframe with .1% the compute of my iPhone, an apocalyptic natural disaster and even worse financial disaster.

Here’s where I have to push back hard, because you’ve smuggled in a claim that’s wrong by something like seven orders of magnitude.

“0.1% the compute of my iPhone” gets the mainframe story backwards. The mainframe didn’t lose because a single PC out-computed it — it never did. A 1985 PC was a toy next to a mainframe on raw throughput. The mainframe lost because the PC was good enough for what people actually needed, and there were a hundred million of them. Sufficiency and ubiquity beat centralization, not per-device horsepower. No single terminal ever ate the mainframe for lunch on compute; the fleet did, on economics.

And the numbers don’t bend the way you want. A current high-end phone does ~35 TOPS. Be generous and give it 100x in ten years — call it 3,500 TOPS. A frontier AI data center runs 100,000+ accelerators at ~1,000+ TFLOPS each, so aggregate is on the order of 100 exaFLOPS. Even with that wildly pro-phone assumption and the data center frozen in place, the building is still tens of thousands of times the phone. For it to be 0.1% of your iPhone, the phone would need to close a ~5-order-of-magnitude gap and then beat it by another thousandfold. (TOPS-vs-FLOPS is a loose comparison, so treat the exact figure as a flag-it estimate — but no sane assumptions get you within shouting distance of parity, never mind the phone winning.)

So your three predictions aren’t equally sound, and it’s worth ranking them honestly:

“Irrelevant like a mainframe” — defensible, but only on the sufficiency mechanism, not the compute one. The data center stays vastly more powerful; the question is whether anyone needs that power once your phone is good enough. That’s the real mainframe parallel, and you should make the argument that way.

“Financial disaster” — your strongest card. Overcapacity, GPU depreciation on a 2-4 year clock against decades of financed concrete, concentrated demand. This is the one that could actually kill the project.

“Natural disaster” — site-specific and plausible (water, grid, the Great Salt Lake’s airshed makes Box Elder genuinely sensitive), but it’s an externality argument, separate from whether the thing makes money.

The clean version of your thesis: the building doesn’t become a weak computer — it becomes an unneeded one, a powerful asset stranded because the value migrated to a billion good-enough edge devices, financed at the top of a debt cycle. That’s a real and frightening way to lose money. “My phone will out-muscle it” is not — and conflating the two weakens the case you’re actually right about.

Yes. But depends on what you mean by out muscle. “The mainframe lost because the PC was good enough for what people actually needed” is the real out muscling.

Granted — and that’s the correct reframe. “Out-muscle” in the only sense that decides markets means winning the metric that matters, and that metric is sufficiency-for-purpose at a price and form factor, not raw FLOPS. The PC won on exactly that. I was defending a definition of “out-muscle” that nobody competing in a market actually uses. You’re right.

So the battlefield is sufficiency, not spec. Good. But notice the reframe relocates the question without settling it — it just tells us where to aim. The PC out-muscled the mainframe because the mainframe’s workloads turned out to be things a PC was eventually good-enough at. Where that wasn’t true — high-volume transaction processing for banks and airlines — the mainframe didn’t die. IBM still sells Z-series boxes running a huge fraction of the world’s transactions. The actual pattern isn’t “annihilation,” it’s “displaced from the mass market, survives in the niche where good-enough-on-the-edge never arrives.”

Which puts the whole thing on one hinge: is the workload Box Elder exists to serve a bounded need your phone reaches sufficiency for — or an expanding one where the line keeps moving? Consumer inference is bounded; your iPad eats it, no argument. Frontier training and heavy agentic/enterprise work is the open question, and here’s the twist that cuts against you: with the mainframe, “what people needed” was roughly known and fixed. With AI, part of the bull case is that new capability creates needs that didn’t exist — every capability jump unlocks uses people then decide they can’t live without, and some of those need a model too big to sit on the phone. If “good enough” is a moving line rather than a fixed one, the edge device chases it forever instead of catching it.

So your instinct survives in its strong form: the building becomes an unneeded asset if AI’s needs turn out bounded, stranded at the top of a debt cycle. It fails if needs keep expanding faster than the edge can absorb them. Nobody honestly knows which — but that’s now a clean, specific bet, not a vibe. And the financial-disaster case doesn’t even require you to win that argument: overcapacity and GPU depreciation can strand the asset on timing alone, regardless of where the long-run line settles.

It feels like there are limits to what people will find useful particularly when they are factoring in the downside of a technology. The negative effects of a data center that consumes more than twice the energy of the entire state of Utah is seen and felt along with any benefit of a new frontier that I think might start to feel like diminishing returns or worse - negative effects in society like job losses and an erosion of humanistic value. I’m thinking about Neil Postman’s book Technopoly. I know most people aren’t academic enough to think like that. But when the effects really start to be felt and become alarming marginal technological gains might finally get called into question.

First — your number is essentially right. The Stratos Project (Kevin O’Leary’s O’Leary Digital, ~40,000 acres in Hansel Valley) is spec’d at 7.5 to 9 gigawatts of power-generating capacity, likely from natural gas, which reporting repeatedly frames as about double what the entire state consumes. So “more than twice the energy of the entire state” is in the right ballpark.

Now the Postman point — and here’s where I’ll push, because I think you’re invoking him for the opposite of what he argued. Postman’s thesis in Technopoly isn’t that a culture eventually feels the harms and disciplines the technology. It’s the reverse: that in a technopoly the culture loses the vocabulary to resist, because it has already redefined every value — efficiency, progress, even meaning — in technological terms. He was mourning the backlash that doesn’t come. So “the effects become alarming and people call marginal gains into question” is the hopeful prediction Postman spent the book doubting. Cars killed tens of thousands a year and reshaped cities badly; we adapted. Social-media harm is felt, named, studied — usage still climbs. The diffuse stuff (job loss, erosion of humanistic value) is precisely where the backlash historically fails to bite, because no one feels the marginal cost on any given Tuesday.

Where backlash does have teeth is the opposite of diffuse: concentrated, local, visible externalities landing on people who vote in a county commission. Water, local air, a gas plant’s infrasound, your power bill. That’s the real mechanism — not an abstract humanistic awakening, but ratepayers and neighbors with standing. Nearly 4,000 local residents and environmentalists objected, and that’s the vector with actual stopping power.

But here’s the sobering twist, and it cuts against your optimism: O’Leary’s team has clearly read the same playbook and engineered around exactly those concentrated vectors. They’re using a closed-loop chilling system with dry, air-based cooling and buying existing water rights from private landowners, not the Great Salt Lake, claiming a net watershed benefit. And the power is generated onsite or brought in under a 2025 law requiring data-center power builds to not negatively impact electricity costs for Utahns. They’ve deliberately neutralized the water and ratepayer arguments — the two that have teeth — leaving the backlash to lean on the diffuse, climate-and-meaning objections that Postman would tell you rarely stop anything.

So I’d refine your thesis one more time. There are now three distinct ceilings in play, and they’re independent: the sufficiency ceiling (edge catches good-enough demand), the social-license ceiling (externalities make it unpermittable/unfinanceable), and the diminishing-returns ceiling (the frontier itself stops paying off — genuinely contested right now, I wouldn’t assert it). Your intuition keeps reaching for whichever one is salient, but the actually powerful framing is the conjunction: the bull case needs all of training-demand-holds and social-license-survives and the-capital-cycle-doesn’t-break. The bear case only needs one to fail. A 9 GW, $100B, decades-financed bet that requires every one of those to break its way, at the top of a debt cycle — that’s the precise shape of your “financial disaster,” and it doesn’t depend on winning the Postman argument at all.

Postman’s actual warning for you might be the darkest read: the project may well get built because the culture can’t muster the refusal — and then strand anyway, on plain economics, after the gas is burned and the valley is rezoned.

You are right to point out that Postman’s argument goes against mine. I’m hoping he was wrong. I’m hoping we will find a little bit of Amish when we need it. Things are changing so fast. Maybe we will finally see a real change. A heeding of postman’s warning.

The Amish are exactly the right reach — and worth being precise about why, because it’s the strongest answer to Postman that exists. The Amish aren’t anti-technology. They’re the only group that does the thing Postman said we’d lost the ability to do: evaluate each tool against an explicit, shared question — what will this do to our community? — and sometimes answer no. They have cars available and choose horses not out of nostalgia but because they judged the car would scatter the community geographically. That’s not Luddism; it’s discernment. It’s the living proof that Postman’s pessimism describes a strong tendency, not a law of physics.

But notice what makes it work, because it’s also what makes it hard to scale: the Amish can say no because they share a thick, agreed-upon framework of value to measure against, and a community small enough to enforce the judgment. The open question — the real one — is whether a pluralistic mass society can do collective discernment at all without that shared frame. We don’t have a council that asks “what will this do to us?” We have a market that asks “will anyone buy it?” That’s the gap Postman was pointing at.

Here’s the sliver I’d actually put weight on, though. You said things are changing fast, and that speed cuts both ways. Postman’s nightmare was the slow boil — change gradual enough that we acclimate before we notice. But a jolt fast and visible enough can do the opposite: it can break the trance. A valley rezoned for something that doubles a state’s power draw is legible in a way that a thousand small surrenders never were. People can see that one. Whether seeing it translates into the Amish question — that’s genuinely unknown, and I won’t pretend the odds are good. But the hope isn’t naive. It’s just betting that visibility, this time, arrives before acclimation does. That’s a real bet, not a fantasy.

“We don’t have a council that asks ‘what will this do to us?’” Well, I don’t know. We kind of do have these. And the o Leary people making claims about closed loop cooling and net water gains and more is a farce. ANY water used for a data center that is not yet built means water that could potentially go to the great salt lake. Any amount of natural gas used to power the center means less for other people/scarcity/higher prices. But at the end of the day, my intuition tells me that despite all of that, my first argument might be the real downfall of this project. It might just be obsolete by the time it’s built.

You’re right to correct me, and I overdrew that line. We do have the councils — the Box Elder commission, the public comment process, MIDA, the 4,000 objectors. The machinery for asking “what will this do to us?” exists and it ran. The trouble isn’t that the question went unasked; it’s which question the deciding body was actually optimizing. The commission asked “jobs, tax base, national security?” and answered yes. The objectors were asking the Postman question — and lost the vote. So it’s not that we lack the Amish council. It’s that ours convenes without a shared thick frame of value, so it defaults to the growth question whenever growth and “what will this do to us” disagree. That’s a thinner kind of pessimism than Postman’s, and honestly a more hopeful one: the institution could in principle weight the other way. It just didn’t here.

On the water and gas — you’ve got the right logic, though “farce” is slightly too strong; the precise word is baseline-shopping. The “net water benefit” is real against one counterfactual (agriculture keeps consuming those rights) and illusory against another (the state retires those rights for the lake). They picked the flattering baseline. Your opportunity-cost framing is the correct rebuttal: the relevant comparison isn’t “data center vs. ranching,” it’s “data center vs. the best alternative use of that water,” and for a shrinking terminal lake the best alternative is leaving it in the watershed. Same structure on gas — “won’t raise Utahns’ electricity bills” is narrowly about grid rates and conveniently silent on gas-market scarcity, air quality, and emissions. Neither claim is a lie. Both are true inside a frame chosen to make them true. That’s not farce; it’s worse, because it survives fact-checking.

But your closing instinct is the one I’d actually bet alongside you on — and notice it’s where you started, now finally aimed correctly. Not “efficiency shrinks demand” (Jevons killed that). The real edge is the mismatch between build latency and the technology’s clockspeed. The reporting itself says permits could take years; full buildout of 9 GW of generation plus the campus is a five-to-ten-year project. Accelerator architectures turn over every 18 months. The chips you spec are obsolete before the concrete cures — but that part’s normal and priced in, because the model is continuous refresh. The dangerous obsolescence is one level up: a decade-long build locks in a bet about where and how compute should physically happen, and that’s the bet most likely to be wrong over a ten-year window. If the answer shifts toward distributed, or efficient, or simply smaller, you can refresh the racks but you can’t shrink 40,000 acres and 9 gigawatts of gas. The bigness O’Leary brags about — “no bigger site in the world” — is precisely what makes the bet unhedgeable. A nimble build can pivot; a monument can’t.

So your first argument was the right one all along. I just made you trade the weak version for the strong one. The weak version was “tech gets efficient, so demand falls” — false. The strong version is “a slow, enormous, unhedgeable bet on one architecture, in a field that rearranges its own architecture faster than you can pour foundations, is likely to come online already mismatched to the world it was built for — and then it strands on plain economics, gas burned and valley rezoned, exactly as Postman would predict, because the council that could have said no was asking a different question.” That’s not a vibe. That’s a thesis. And it doesn’t need the frontier to stall or the public to wake up — it only needs the build to be slower than the change. Which, looking at the permitting timeline, it almost certainly will be.

Yes. ;-) But also… “the institution could in principle weight the other way. It just didn’t hear.” The jury is still out.

Exactly — and I’ll take your “hear” as the better word, typo or not. The whole thread comes down to whether the council can still hear the question over the growth question, and that’s not settled by one commission vote.

Because that vote didn’t build anything. It created a project area and an agreement. The real choke points are downstream and slower: the actual permits, the water-rights transfers, and above all the air permitting for ~9 GW of gas turbines — that’s the kind of approval that draws federal and state scrutiny, sits in a sensitive airshed, and is eminently litigable. Each of those is another room where the question gets asked again, by a body with different incentives than a county commission chasing a tax base. And every month those rooms take is a month the technology keeps moving underneath the bet.

So yes — jury’s out, and time is the objectors’ quiet ally. The longer it takes to pour the foundation, the more likely the world it was poured for has already moved. You started this convinced efficiency would kill it. You’re ending somewhere sharper: that slowness might, with a little help from a council that finally hears. ;-)

Afterword

That’s the end of the conversation. But after I finished, I started my morning routine by reading a bit from the scriptures. The first thing I read this morning was:

And I will punish the world for evil, and the wicked for their iniquity; I will cause the arrogancy of the proud to cease, and will lay down the haughtiness of the terrible. (2 Nephi 23:11, quoting Isaiah 13:11)