The Building Pays Back.
The Silicon Might Not.

First, the news — accurately

Blackstone's QTS sold its stakes in three Northern Virginia data centers for US$3.5 billion — less than three years after buying in. And QTS just withdrew its final appeal on the 2,100-acre "Digital Gateway," the campus planned to be the largest on earth.

But read closely before concluding anything. The buyer of that US$3.5 billion stake is Digital Realty — another data center operator — paying US$1.2 billion in cash plus US$2.3 billion in its own shares. One specialist selling to another isn't an industry fleeing; it's capital recycling. And the giant campus didn't die of bad economics — it died in court, after years of zoning battles beside a Civil War battlefield and, finally, a newspaper-notice technicality.

So the headlines don't prove data centers are a bad business. The math might. Let's do it.

The math nobody wants to do

Compute price-performance improves roughly 20–30% per year. New chip architectures now arrive every 18–24 months. Compound that: after two years, the same money buys 36–46% more computing than your installed hardware delivers — on roughly the same electricity and cooling.

Flip it around: a two-year-old AI data center is selling compute that costs its owner up to nearly twice what a brand-new competitor pays per unit — with an identical power bill.

That's not a maintenance problem or a management problem. It's structural. The product you sell gets cheaper every year; the asset you bought doesn't.

Even the giants can't agree how fast the value dies

Watch what the biggest owners of AI hardware are doing with their own accounting — they're quietly fighting over exactly this question.

Between 2020 and 2024, the hyperscalers stretched server depreciation from 3–4 years to 6 — a change worth roughly US$18 billion a year in delayed expenses. Then in 2025 the consensus cracked: Amazon shortened a set of servers to 5 years and took a US$920 million accelerated charge, citing the pace of AI development. Meta extended to 5.5 years and booked US$2.9 billion less expense. Same hardware reality, opposite bets.

Whether he's right or not, notice what the argument is about: how long a chip stays economically alive. Nobody serious says six years anymore with a straight face. The published numbers on one flagship GPU: strong profit in year two, negative returns by year four.

The answer: it depends which layer you own

Here's where my one-year-old question finally got its answer. "Data center" is two businesses wearing one name.

The landlord layer — land, building, power connection, cooling. These assets last 20–30 years, get leased to tenants on 10–15 year contracts, and don't care what chips are inside. The tenant carries the obsolescence. This layer can recover its capex — which is exactly why Digital Realty is buying at US$3.5 billion while everyone reads the same headline as a fire sale.

The silicon layer — the GPUs themselves. Economic life of two to three years. This layer is a race: recover your capex at today's compute prices before the next generation resets the price of compute for everyone. Miss the window and the math never closes. The window is the business.

The market is already doing this math

If the layer split sounds theoretical, watch what just happened to Oracle — the one giant funding its AI build-out primarily with debt. Capital spending up 162% to US$55.7 billion in fiscal 2026. Free cash flow at negative US$23.7 billion. Roughly US$130 billion of debt, with a plan announced in June to raise US$40 billion more. And behind all of it, nearly US$250 billion in long-term data center lease commitments running 15–20 years — to be filled with customer contracts on far shorter horizons.

June 2026 delivered the verdict: after the bigger raise was announced, Oracle's stock had its worst week since the 2001 dot-com bust. That is not the market saying AI demand is fake. It's the market repricing a company locked into twenty years of building-layer obligations while selling two-year silicon — on borrowed money. The revenue has to arrive inside the silicon window. The obligations run for two decades. That gap is exactly the math above, priced live.

Why this matters in Malaysia

This isn't an American story. Malaysia's data center capacity is set to more than double to over 2,000MW by end-2026 — Johor alone has 850MW built, 1,800MW under construction and 2,700MW more in the pipeline, with billions committed by global players. The government has already paused non-AI data center approvals over power and water constraints.

And understand why the silicon layer has any chance here at all: electricity. Power is the biggest operating cost of compute, and Malaysia's industrial electricity has long been among the cheapest in the region — the core reason hyperscalers picked Johor over land-scarce, expensive Singapore. Cheap power is the one variable in the depreciation math that a location can actually change: it widens the margin on every unit of compute sold, which shortens the capex-recovery window. Without it, silicon-layer payback is close to impossible anywhere.

But two cautions before anyone builds a business case on that. First, the cushion just shrank: from July 2025, data centers above 100MW sit in the ultra-high-voltage tariff class — energy costs up an estimated 10–14%, an extra US$15–20 million a year for a large site — and analysts are already pointing investors toward Vietnam and Thailand. Second, even at the cheapest tariff on earth, cheap power buys time, not escape. It stretches the two-year window; it doesn't stop the melting. If a compute business only works because electricity is cheap, that's a subsidy race, not a moat — and subsidy races end at the next tariff review.

Malaysia's other location advantage costs nothing and can't be repealed at a tariff review: geology. Peninsular Malaysia sits outside the Pacific Ring of Fire and below the typhoon belt — no earthquakes rattling racks, no typhoons taking the grid down, no volcanoes. That matters double in this business: when your payback window is two years, an asset knocked offline for weeks — or destroyed outright — by a single event never repays anything. Insurers price this. Hyperscaler site selectors price it too. Cheap power shortens the window; geological stability protects it. That combination, not power alone, is Malaysia's real pitch.

Every one of those Johor projects is making the layer choice right now, mostly invisibly. Land, shell and megawatts leased to hyperscalers on long contracts — that's the durable business. Any Malaysian player tempted to own the compute itself should first answer the two-year question: what is your rack worth the day the next generation lands, and who is contractually holding that loss? If the answer is "us," the ribbon-cutting photo is the high point of the investment.

The bottom line

Data centers aren't one business. The building pays back over decades; the silicon must pay back in about two years or it never does. Blackstone selling to Digital Realty is the landlord layer changing hands at full price. The melting happens at the chip layer — and the industry's own accounting fight tells you the giants know it.

Published by IMA AI — July 2026. We run our own AI infrastructure decisions on the same math we publish — rent the depreciating layer, own the compounding one.