Marcus Bennett here, and let me tell you what I see when I look at the shimmering edifice of the AI trade: the same vertigo I felt watching condo flippers in 2006 explain that “housing never goes down.” Except this time the collateral isn’t stucco in Phoenix — it’s rectangular warehouses full of GPUs depreciating at roughly 20 to 30 percent a year, powered by electrical grids held together with duct tape and interconnection queues measured in half-decades.
The latest excitement comes courtesy of the White House itself. Kevin Hassett, National Economic Council director and a man whose optimism has never once been accused of being premature, took to the airwaves to reassure the masses. The occasion: an election that swept in lawmakers openly hostile to unfettered AI development, plus a string of local ballot measures halting data center construction in their communities. And Hassett’s response? A verbal shrug. He “respects the localities,” mind you, “but America is a big place,” and besides, centers already under construction probably aren’t affected. Plenty of room elsewhere, he says.
Plenty of room. Truly, the econ-speak equivalent of “the fundamentals are strong.”
Let’s unpack this, because buried in that breezy dismissal are several assumptions that deserve the same scrutiny we’d apply to, say, a CFO claiming “non-recurring” losses for the fifth consecutive year.
First, the fiscal architecture. Data centers are not popping up in “plenty of room” by accident. They cluster where power is cheap, fiber exists, and — crucially — where local governments have offered tax abatements so generous they’d make a 1970s stadium deal blush. The localities voting to halt construction aren’t NIMBY cranks; they’re communities doing the arithmetic Hassett apparently hasn’t. A hyperscale data center employs perhaps 50 to 100 permanent workers while consuming electricity on the scale of a mid-sized city and straining water supplies that, in places like Northern Virginia and central Ohio, were already contested. Residents have looked at this trade — subsidies in, service jobs modest, utility rates potentially socialized — and concluded the return on investment accrues to Mountain View and Redmond, not to them. How delightfully un-American of them: voting against capital.
Hassett’s “America is a big place” line deserves historical framing, because we’ve heard it before. When a federal official tells you that localized resistance doesn’t matter because there’s plenty of land elsewhere, remember that the transcontinental railroads were also built on “plenty of room” — room generously provided by land grants, federal guarantees, and a financing structure that collapsed spectacularly in 1873. The lesson wasn’t that the railroads were useless; they transformed the continent. The lesson was that overbuilt infrastructure financed on rosy projections produces panics, and the panics don’t care how transformative the technology ultimately was. Railroads were real. The panic of 1873 was also real. Both things, and this is the part the AI bulls find cognitively inconvenient, can be true simultaneously.
Second, the “centers under construction aren’t affected” argument. This is the reasoning of a man describing the last car of a train that has already left the station and calling it proof of momentum. Yes, the ~$300-plus billion in committed hyperscaler capex for next year is largely spoken for — Microsoft, Google, Meta, and Amazon aren’t canceling poured foundations. But capital expenditure is a forward-looking decision, and the marginal dollar is what prices assets. If communities are now voting to block expansion, if state legislatures are drafting moratoria, if lawmakers swept into office on anti-tech sentiment start eyeing everything from permitting to power pricing, then the growth trajectory that justified trillion-dollar market capitalizations develops a very visible crack. Markets don’t price current cash flows; they price the narrative about cash flows in year five and year ten. Hassett’s reassurance addresses year one. Touching.
Third — and here we arrive at the genuine comedy — the market turmoil he was presumably sent out to soothe. AI-exposed equities had been wobbling precisely because a few sober voices noted that the industry’s revenue projections require roughly 30 percent annual growth in AI-related spending for years, a growth rate that historically precedes exactly nothing good. Cisco in 2000 sold actual routers to customers with actual budgets, and its stock still lost 80 percent. Today’s analog involves circular financing arrangements — chipmakers investing in the cloud providers who buy their chips, who lend to the AI startups that buy compute from the cloud providers — a structure that would have drawn a knowing chuckle from anyone who remembers vendor financing in telecom, circa 1999. When Lucent and Nortel were lending customers money to buy Lucent and Nortel equipment, it looked like demand. It was, in fact, a receivables problem wearing a demand costume.
Now, to be fair to the optimists — a burden I shoulder with great reluctance — the bears aren’t obviously right either. Unlike pets.com, the current AI infrastructure buildout is being funded overwhelmingly from the cash flows of four of the most profitable enterprises in human history, not from credulous retail money. Microsoft and company can absorb writedowns the way a casino absorbs a bad weekend. And the electricity demand is real; hyperscalers signing twenty-year power purchase agreements aren’t hallucinating the need for compute. The buildout isn’t fictional. The question — the only question, really — is whether the returns on roughly a trillion dollars of cumulative spend will arrive before the depreciation schedules and investor patience run out. Depreciation is the great honest accounting of the technology sector: those GPUs on the balance sheet are worth progressively less every year, and if the software monetization doesn’t keep pace, the writedowns will be biblical in both scale and narrative drama.
Which brings us back to the election results and why Hassett’s nonchalance should worry you more than comfort you. When a White House economic official makes the rounds to dismiss a story, the standard read is that the internal polling on that story is bad. “Plenty of room elsewhere” is not an economic argument; it’s a press release with a macro adjective attached. And the political economy here is genuinely adverse in a way Wall Street, eternally surprised by politics, has not priced: data centers have become the rare issue uniting rural conservatives annoyed by transmission lines and progressive urbanists annoyed by everything. When a buildout’s opponents span the ideological spectrum and show up to town halls with their utility bills in hand, that’s not a fringe. That’s a constituency. Constituencies, unlike GPU clusters, do not depreciate — they compound.
Consider also what the ballot measures actually signal for the cost stack. If the easy sites are legislated away, the marginal data center migrates to harder terrain: longer transmission builds, costlier power, more expensive land, more litigation. The marginal cost curve of AI infrastructure was already steepening on its own — transformer shortages, grid interconnection queues stretching five to seven years, electrical equipment lead times that would embarrass a Soviet planning ministry. Layer political friction on top and you get a subtle but corrosive effect: returns that were projected at, say, mid-teens IRR drift lower, and at some threshold the marginal project doesn’t get the green light. Hyperscalers will claim they’re being disciplined rather than deterred. Analysts will repeat the claim. The gap between the two will be filled by history.
So where does this leave the sophisticated observer? Not in the doomsday bunker, certainly. The legitimate uses of AI — in code, in drug discovery, in the unglamorous plumbing of enterprise software — are real and will generate real profits for someone. But “someone” is doing heavy lifting in that sentence. The history of transformative technologies from canals to fiber optics is unambiguous: the infrastructure gets built, the technology changes the world, the initial financiers get destroyed, and the returns accrue to late-arriving, sober operators using the now-cheap infrastructure. The internet didn’t lie about being important. Global Crossing still went bankrupt.
Hassett wants you to believe that a few local ballots are noise against the grand sweep of American technological destiny. Maybe. But note what he did not address: the circular financing, the depreciation math, the power constraint, or the fact that “America is a big place” is doing no work in an argument where the binding constraint is electrons, interconnection capacity, and now voters — none of which relocate to the empty space he’s so keen on. Officials who answer structural questions with geographic ones are not reassuring you. They are stalling for time, and markets, whatever else one says about them, are rarely fooled for long by a change of subject.
The railroads got built, the country got wired, and yet 1873 and 2001 still happened. Keep that pairing in your head as the AI trade trades sideways-to-violently. The future may indeed be intelligent. Whether the current owners of that future’s equity will profit from it is an entirely different question — and no amount of “plenty of room” answers it.
