If you hold cloud stocks, chip stocks, or any “AI infrastructure beneficiary” — this piece answers a question deadlier than “is the demand real”: can the returns outrun the asset’s decay?

Scale first. Per company guidance, Microsoft, Amazon, Google, and Meta plan a combined ~$725 billion of capital expenditure for 2026, up nearly 80% from ~$410 billion in 2025. Only three buildouts in history compare: nineteenth-century railroads, the early electric grid, and the fiber-optic boom of the late nineties.

Each time, capital was convinced it was laying the infrastructure of the next era. Each time, it was right — the infrastructure did change the world. And in two of the three, the people who paid for it lost everything. The railway mania erased UK market value worth roughly half of GDP; the telecom bust’s fiber sold for cents on the dollar to latecomers — who used it to build someone else’s internet era.

Infrastructure success and investment success are two different things. Which will this round be? The answer hides in a variable that none of the historical precedents had.

The Secret on the Asset Side: GPUs Are Not Rails

Rail steel lasts a century. Grid transformers last forty years. Fiber still carries data twenty-five years after burial — when those bubbles burst, the assets remained, waiting for the next cycle to light them up.

A GPU is not that kind of asset. A flagship training chip’s economic life — not physical life, but “still worth running” life — is roughly three to five years: each chip generation cuts unit compute cost off a cliff, and software stacks keep making old silicon uneconomical for new workloads.

That creates a key gauge: the depreciation–replacement scissors. Cloud providers typically depreciate servers over five to six years, while frontier workloads actually turn over closer to every three. Book depreciation slower than economic decay means part of today’s reported profit is borrowed from the future — the write-down simply hasn’t happened yet. The wider the scissors open, the harder they eventually snap shut, in the form of a massive impairment charge in some future quarter.

Plug asset life back into the math: railroads could repay over fifty years; a GPU must earn itself back in four or five — there is no “wait for the next cycle.” $700B+ a year of spend against a four-year payback window demands AI revenue materialize at a speed with no historical precedent.

Who Carries the Debt Decides the Shape of the Burst

Suppose part of the return never materializes. Who eats the loss? Four layers:

Big Tech balance sheets — funded from their own cash flow; they can absorb write-downs (the biggest difference from the debt-built telecom bubble); neocloud debt — borrowing to buy GPUs, living on contracts (CoreWeave alone has $4.2 billion of principal due within 2026, per SEC filings) — the most fragile link in the system; lab funding dependence — compute commitments far beyond their cash flows, with the funding window as oxygen; sovereign buyers — price-insensitive, the cycle’s extender.

Stack the four layers and the shape emerges: a remarkably solid core with highly fragile edges. The likely burst isn’t a 2000-style collapse, but cascading defaults at the edge plus one-off impairments at the core — the infrastructure stays, the wealth changes hands. And history’s rule for the handover: assets flow from those who borrowed to those who hold cash.

Three sentences to close. One: distinguish “built with their own money” from “built with borrowed money” — when the first falls, it’s an opportunity; when the second falls, it’s a trap. Two: watch the scissors — whoever’s depreciation policy strays furthest from chip cadence has the most water in their reported earnings. Three: in the panic, core assets will be mispriced downward — the most reliable buying moment of this cycle.

The “AI infrastructure” in your portfolio: whose money built it, and over how many years is it depreciated? Comments open.

Data sources (verified, June 2026): Big Four 2026 capex ~$725B, +77% YoY (company guidance); CoreWeave $4.2B principal due in 2026, $8.5B/$3.1B facilities (SEC 8-K); server depreciation 5–6 years (company policies; to be anchored to 10-K text before publication).

— From Chapter 2 of a book in progress, working title The Deflation Sandwich

#AI #AIIndustry #CapexCycle