For about a year, the recession versus soft-landing debate has been quietly reorganizing itself around a single line item: AI capital expenditure.
Hyperscalers are spending hundreds of billions on data centers, power, chips, cooling, fiber, and land. Goldman has floated estimates approaching $800 billion in annual AI infrastructure spend by the end of 20261. Whether or not the number lands there, the trajectory is the same…
a buildout that has gone from rounding error to economic load-bearer in the span of about thirty months.
Naturally, the historical analogies have started showing up. Railroads. Telegraph. Fiber. Electricity. And just as naturally, the analogies are getting flattened into a single number. Hyperscaler capex is running at roughly 2.2% of GDP this year and projected to climb past that by 20272.
Cue the debate: is 3% the line? Is 2.5%? Is there a line at all?
The threshold framing was always borrowed from the wrong shelf. It treats the historical record as a series of identically-shaped events that can be detected by a single magnitude trigger.
The pattern is sharper than that, and the lesson is more uncomfortable.
The 180-year pattern
Start in the 1830s. The United States built canals, the first generation of railroads, and a wildcat banking system on top of debt issued largely to British creditors. The whole architecture was a bet on future productivity that had not yet shown up. When tightening hit and capital reversed, the bet broke. The Panic of 1837 followed, state governments defaulted, and the financial system spent the better part of a decade unwinding.
The template then repeated with variations. The post–Civil War railroad expansion ended in the Panic of 1873 and structural overcapacity that persisted through the 1880s, with another leg down in 1893. Urban rail, utilities, and the early electrical grid built through the 1890s and 1900s gave way to the Panic of 1907. Autos, roads, electrification, radio, and consumer credit drove the 1920s and were among the layered causes of the 1929 collapse, though not the sole cause; the gold standard, banking structure, and policy failure did most of the killing.
Telecom and fiber built the late 1990s and broke in 2001, taking the equity markets and a generation of telecom balance sheets with them. Housing and the suburban credit infrastructure that surrounded it broke in 2007–2009. Shale, pipelines, and LNG ran hot through the 2010s before a sector-specific bust collided with the 2020 COVID shock to wipe out a meaningful slice of independent producers.
The pattern is real. The pattern is also not as clean as the headline version suggests, which is exactly why it’s worth thinking about properly.
The bust isn’t caused by the buildout. The buildout is what makes an otherwise-routine credit event systemic.
The real mechanism
If you try to make the strong claim, “infrastructure booms cause recessions”, you lose the argument on first contact with the data. Recessions have many fathers. Railroads in 1873 had policy company. Housing in 2008 had a global shadow banking system attached. Shale in 2020 had a pandemic.
The defensible version of the thesis is more interesting. Infrastructure capex, when it becomes large enough to materially carry GDP, creates an asymmetric vulnerability in the macroeconomy. Any disruption to the funding loop — a rate move, a capability disappointment, a financing tightening, a narrative shift — removes a major growth contributor while leaving the overcapacity and the debt overhang exactly where they were the day before.
The buildout doesn’t have to cause the credit event. It only has to amplify one. A routine financial wobble in an economy where the marginal dollar of growth is coming from speculative infrastructure becomes a systemic event. The same wobble in an economy where infrastructure capex is incidental gets absorbed.
This is why the historical recessions cluster around late-cycle buildouts. Not because the buildouts were uniquely toxic, but because they made the economy fragile to perfectly ordinary shocks.
Four conditions, not one threshold
If 3% of GDP is the wrong anchor, what’s the right one? Track four conditions together.
- Capital intensity. Is the buildout large enough to move the macro numbers?
- Debt intensity. Is it being financed by leverage that requires future cash flow to service?
- Revenue delay. Is the productive payoff arriving on a slower timeline than the financing demands?
- Narrative overreach. Is the underwriting case dependent on a story about transformation that hasn’t been validated yet?
Any one of these is normal. Two is a sector story. Three starts to look like a cycle. Four is when you should be watching the financing structure of the marginal player, because that’s where the unwind will start.
The railroads of the 1870s had all four. The telecom buildout of the late 1990s had all four. Housing in 2005–2007 had all four. Shale in 2014–2019 had all four. AI data centers in 2026, and this is the honest part, have most of three and a contested fourth.
Debt intensity is the contested one.
At the headline level, the hyperscalers are funding capex out of operating cash flow… they look unlevered, and they are. But the marginal data center builder isn't a hyperscaler. It's a neocloud provider, a private-equity-backed operator, a joint venture stitched together with creative off-balance-sheet structures. The system's leverage is hidden in the seams: vendor financing from chip suppliers, private credit funds with rapidly growing exposure to data center special purpose vehicles, operating leases and power purchase agreements that don't appear as debt on the prime players' books.
Whether you count the fourth condition as present or absent depends entirely on where you stand in the capital stack. And the unwind, when it comes, always starts at the marginal player… not the prime one.
The productive bust
Here is the part of the story that most “bubble” pieces get wrong.
Railroads paid off. Electricity paid off. Telephone paid off. Fiber paid off, eventually… the late-90s overbuild is exactly what made the 2000s consumer internet economically possible. Even shale paid off, in the sense that American oil production is structurally higher today than anyone in 2005 would have forecast.
The bust wasn’t about the infrastructure being useless. It was about a timing mismatch between financing structure and revenue arrival. The original capital stack couldn’t survive the gap. The capacity got built. The builders got wiped out. New owners picked up the assets at cents on the dollar. And the productivity dividend accrued to the second wave.
Railroads paid off over forty years. GPUs depreciate in five.
This is the inheritance paradox of every infrastructure cycle. The boom is simultaneously the mechanism by which productivity gets installed and the mechanism by which the installers get destroyed. If you’re trying to short the AI buildout, you have to be right about the financing structure, not the technology. They’re different bets.
The other side of that trade is more interesting.
The historical pattern says the smart positioning isn’t shorting the cycle… it’s being ready to buy the substrate when it gets reassigned. Railroad bondholders lost in the 1870s; the consolidated networks that emerged from the wreckage dominated American industry for the next half-century. Telecom equity got wiped out in 2001; whoever owned the fiber after the bankruptcy auctions ran the consumer internet. Shale producers went under in 2020; the majors bought their acreage and pumped it for free cash flow.
The asset doesn’t go away. It changes hands.
For AI, the question worth asking isn’t whether to short. It’s who is positioned to be the second-wave buyer.
- The hyperscalers themselves, probably… they have the balance sheets to consolidate distressed competitors.
- Sovereign wealth funds with multi-decade horizons.
- Private credit firms with existing exposure who would rather take ownership of the data center than write off the loan.
- Strategic buyers from adjacent industries who couldn’t justify building from scratch but can justify buying for cents on the dollar.
None of them are short the bubble. All of them are reading the depreciation schedule and the financing covenants and waiting for the first defaults at the margin.
Why shale is the AI analogue, not railroads
Most AI-bubble pieces reach for the railroad and fiber comparisons because they’re flattering… long-arc productivity stories with eventual vindication. The closer analogue is shale.
Shale shares the profile that actually matters. Capital-intensive. Debt-financed. Fast-depreciating. Narrative-heavy. Structurally dependent on a continuous funding loop. Shale producers drilled to service debt that was issued to drill. Break the loop and the math collapses.
AI data centers share most of this. The depreciation schedule is the part nobody wants to discuss. A rail line built in 1875 was still moving freight in 1925. A fiber line built in 1999 is still lit in 2026. A top-tier GPU bought in 2024 will be operationally obsolete by 2029, possibly sooner. This compresses the window in which the buildout has to generate the cash flows that justify it. There is no forty-year amortization curve to lean on. The revenue has to show up almost immediately, or the asset is impaired before it’s paid for.
The circular financing problem is also present, in a softer form. AI labs buy compute from hyperscalers who buy chips from a vendor whose largest customers are the labs. This is not the same as the shale debt loop, but it is closer to it than to a railroad bond.
Three twists that break the analogy
The shale comparison is the best historical analogue, but it isn’t perfect. AI has three unique features that don’t appear in any prior infrastructure cycle, and each one bends the unwind in a different direction.
First, this is the first infrastructure cycle whose explicit purpose is labor substitution. Railroads needed conductors, engineers, station agents, and track crews. Electrification created millions of jobs in line work, manufacturing, and appliance distribution. Fiber and the consumer internet built around it spawned entire labor categories… developers, content creators, gig workers, support staff.
Every prior buildout, even when it disrupted existing industries, also generated new ones. AI is different.
The buildout is the substrate for systems whose value proposition is replacing labor inputs. If it succeeds on its own terms, GDP can grow even as employment compresses. The productivity dividend accrues to capital. This changes the political economy of both outcomes — the bull case is dystopian for most participants, and the bust case lands on workers who never participated in the upside in the first place.
AI is the first infrastructure cycle whose explicit purpose is labor substitution. The historical pattern says these buildouts pay off over decades. It doesn’t say who they pay off for.
Second, software efficiency erodes the asset base. A rail line built in 1875 was still moving the same freight in 1925. A fiber line lit in 1999 is still carrying packets today. The infrastructure was static; the application layer changed, but the underlying asset kept delivering the same unit of service. AI breaks this. Model efficiency keeps improving by orders of magnitude… quantization, distillation, mixture-of-experts architectures, sparse attention, smaller models matching the capability of larger ones from twelve months prior.
Per-token inference costs are falling roughly tenfold a year.
The same physical data center generates less revenue per unit of utility every year, not because the hardware degrades but because the software running on it gets smarter. The bull case assumes demand grows fast enough to absorb the efficiency gains. The bear case is that software eats the hardware’s runway before the capex pays for itself. Railroads got more valuable as the country grew. AI data centers get less valuable as the models get better.
Third, energy is now the binding constraint, not capital. Capital is patient. Power isn’t. Grid interconnection queues are running at multiple years in most regions. Transformer lead times stretch past 18 months. Combined-cycle gas turbines have multi-year backlogs. Nuclear is a decade-plus proposition. Local opposition to new transmission is increasing, not declining.
The hyperscalers can write the checks; they can’t actually plug in the data centers fast enough. This caps the buildout independent of any financial dynamic, and it distorts the economics of the survivors… whoever secures the power first wins, and the others are stuck waiting on grid capacity that may not exist on their timeline. Railroads could lay track. Fiber could pull cable. AI data centers depend on a physical system that operates on the timeline of public utility planning, which is decades slower than the financing cycle.
Stacked together, these three twists produce an uncomfortable picture.
The buildout can’t grow as fast as planned, because energy gates it. Each unit of capex generates less revenue over time, because efficiency deflates it. And even if the whole thing works exactly as the bull case predicts, the productivity flows to capital and not to labor… which means both the bull case and the bear case produce political destabilization, just on different timelines.
What this means for now
None of this predicts a recession. The NBER’s most recent dating still has the U.S. expansion intact, and there are perfectly plausible paths in which AI revenue scales fast enough to validate the current capital stack. The bull case is not stupid.
But the asymmetry is what matters.
If the revenue arrives, the buildout is brilliant and the productivity dividend is enormous. If it doesn’t arrive on the depreciation schedule, the unwinding will be sharp… because the buildout is now load-bearing for headline growth, and because the financing chains running through it are not visible to most observers until they break.
The historical record doesn’t tell us what happens next. It tells us where to watch. Look at the marginal financier of the marginal data center. Look at the depreciation assumptions in the hyperscaler 10-Ks. Look at the implicit revenue ramp baked into the chip vendor’s forward guidance. Look at whether the AI labs’ top-line projections require a category of customer that does not yet exist.
The 2.2% question isn't the right question. The four conditions are. And even the four conditions don't capture the full picture, because this cycle has features the others didn't.
If the past 180 years are any guide, the bust, if it comes, won't kill the technology. It will reassign it to a second set of owners. The novel question, the one history doesn't answer, is who gets to participate in the economy that gets built on top of what the first set financed.
That's the trade this time. And it's not quite the same trade as before.