Why AI Capex Cycles Are Different From Previous Technology Booms

Here is the framing we use: every capital cycle has a financing identity and a depreciation clock, and together they determine the topology of the bust — who fails, through what channel, how fast the excess clears, and who inherits the assets. A boom financed by capital markets ends when sentiment ends, because the funding is reflexive. A boom financed by operating cash flow ends only when the operators choose to stop, because the funding is internal. A boom in durable assets leaves a capacity overhang that suppresses returns for years. A boom in perishable assets self-corrects, because the excess literally depreciates away. The dot-com telecom buildout was capital-markets-financed durable capacity. The AI buildout, in its current form, is operating-cash-flow-financed perishable capacity. Same mania energy, opposite mechanics.

The financing identity: who writes the checks

The great infrastructure booms of market history were overwhelmingly financed by external capital flowing to challengers. The railway manias of the nineteenth century were funded by public share subscriptions to companies with prospectuses and no traffic. The telecom buildout of the late 1990s was funded by high-yield debt and equity issuance to competitive carriers whose business plans assumed demand that had not yet materialized. The structure matters because it is reflexive: the capacity gets built only as long as capital markets stay open, and capital markets stay open only as long as the narrative holds. When sentiment turned in 2000, the funding stopped mid-buildout, the leveraged challengers defaulted, and the defaults cascaded into the equipment vendors who had been booking those customers' purchases as revenue — often purchases the vendors themselves had financed. The bust was a credit event first and a demand event second.

The AI buildout is financed differently at its core. The dominant share of the spending comes from a small number of incumbent platforms funding capacity out of the operating cash flow of some of the most profitable business models ever constructed — search, cloud software, digital advertising, e-commerce logistics. This is not a moral improvement; it is a structural one, and it cuts both ways. On one side, the cycle is robust to sentiment in a way the fiber era never was: an equity de-rating does not stop the checks, because the checks were never conditional on the equity market. On the other side, the cycle is extraordinarily concentrated: an entire supply chain's revenue outlook is a derivative of the capital-allocation decisions of a handful of buying committees. In the fiber era, demand for equipment was fragile but diversified across dozens of funded challengers. Here it is durable but narrow. The relevant nightmare is not a wave of customer bankruptcies; it is a small number of rational CFOs revising a growth rate in the same quarter.

The depreciation clock: how the asset ages

The second inversion is the asset itself. Fiber in the ground is close to immortal, and — this is the crucial, underappreciated detail — it historically became more valuable while sitting idle, because the transmission equipment at the endpoints kept improving and multiplied the carrying capacity of the same strand. That is why the fiber glut suppressed bandwidth pricing for the better part of a decade: the excess did not decay, it compounded. The owners were wiped out, but the assets survived to be bought out of bankruptcy for a fraction of construction cost, and that cheap durable capacity became the substrate on which the next generation of internet businesses was built. The bust subsidized the future.

Accelerated computing ages in the opposite direction. A GPU fleet is competing against the next generation's performance per watt, and in a workload where power and space are the binding constraints, an older accelerator is not merely slower — it occupies scarce megawatts that a newer part would use several times more productively. The economic life of the asset is set by competitive obsolescence, not physical wear. This has two consequences that the fiber analogy misses entirely. First, an overcapacity episode in compute is partially self-correcting: excess capacity depreciates toward irrelevance on a clock measured in a few years, not decades, so a glut cannot suppress pricing for a decade the way dark fiber did. Second, there will be no equivalent of the bankruptcy-fiber subsidy: a future builder cannot buy mothballed frontier compute for cents on the dollar, because by the time it is mothballed it is no longer frontier. Whatever the next era is built on, it will not be built on the cheap leftovers of this one.

The illustration: what the telecom bust actually punished

The strongest case against this framework

The serious objection is that this is "this time is different" wearing an accounting costume. Overbuilding is overbuilding: if the revenue to justify the capacity does not arrive, then cash-flow financing merely means the losses are absorbed by incumbent shareholders through squandered retained earnings rather than by creditors through default — a distinction of channel, not of outcome. A dollar of value destroyed is destroyed. And the skeptic can add that the tidy structure described above is already eroding at the edges: financing structures around the buildout have grown more elaborate, and arrangements in which suppliers invest in customers whose purchases become the supplier's revenue carry an unmistakable echo of the vendor-financing era.

We accept most of this, and the framework absorbs it rather than denying it. The claim is not that internally financed booms end well; it is that they end differently — through the income statement rather than through the credit system — and that this changes what an investor should monitor and how fast the damage propagates. More importantly, the objection supplies the framework's own regime-change test. A boom's financing identity is observable, and it can migrate. As long as the marginal dollar of AI capex is funded from operating cash flow, the cycle retains its shock-absorbing character. If the marginal dollar migrates toward debt, special-purpose vehicles, vendor financing, and capacity contracts from counterparties who need capital markets to honor them, the cycle is acquiring the reflexive fragility of its predecessors — one structure at a time. That migration, not any demand statistic, is the variable this framework says to watch. Sterling's Embedded Intelligence treats it as the cycle's master indicator: not how much is being spent, but whose balance sheet the spending is really resting on.

What this changes for a portfolio

Three consequences follow. First, concentration replaces credit as the thing to underwrite. A supplier's revenue durability is a function of a handful of budget decisions, so position sizing in the supply chain should reflect customer-concentration risk that never appears as a ratio on the balance sheet. Second, the equity market, not the credit market, carries the adjustment: because there is little leverage at the core, a capex deceleration converts almost entirely into equity de-rating along the chain, which argues for paying attention to what growth rate each supply-chain multiple already assumes. Third, the self-correcting depreciation clock means an eventual downturn is more likely to be sharp and comparatively short than long and grinding — the fiber decade is the wrong template for the aftermath, which changes how much terminal pessimism a de-rated supplier deserves.