Up to $725 Billion in 2026 AI Capex Meets Fast-Obsolescing Hardware

Amazon, Microsoft, Alphabet and Meta planned combined capital expenditure of up to $725 billion for 2026, according to Business Insider's reporting this month. Most of that money buys data centers, and data centers take years to build. The hardware inside them can lose commercial value faster than the concrete dries.

That timing gap is the core of the current AI infrastructure debate. Yahoo Finance separately cited a $730 billion bill and asked who is footing it. We use the lower $725 billion figure, because it is the more conservative estimate and the one tied to the Business Insider report.

Four buyers, one year, one sum. When demand for chips, power gear, networking and construction concentrates in so few hands, the supply chain organizes itself around their plans, and suppliers and lenders feel any shift first.

Why do data centers outlast their contents?

A data center is a long-lived container with short-lived contents. Tomlinson told Business Insider that projects planned two years ago are still being built while AI technology keeps evolving. A shell and a power connection hold value for decades; a rack of accelerators does not.

The stack also ages unevenly. Land and grid connections, cooling and electrical systems, and chips and servers each depreciate at different speeds and carry different financing, and the obsolescence risk sits almost entirely in the top layer, the accelerators themselves.

Business Insider does not supply project-level detail, specific hardware models, replacement schedules, project locations or investor returns. It also does not give the full names or titles of Tomlinson or Pfirsching, the two people quoted. Their comments should be read as industry views, not official company positions.

What did the named voices actually say?

Tomlinson framed the return equation precisely: "For infrastructure investors, returns depend on whether growth in paid usage compensates for falling prices and covers electricity, financing and equipment replacement."

Pfirsching named three compounding vulnerabilities: "The greatest vulnerability is borrowing heavily to buy hardware, relying on a handful of customers and assuming today's rental prices will last."

Each leg deserves separate stress. Debt works when the asset outlasts the loan; it fails when the market reprices the collateral, GPUs assumed to hold value for six years, in half that time. A handful of anchor customers can fill a hall quickly and then vanish or renegotiate. And rental prices reflect scarcity, which $725 billion of planned supply is designed to end.

What did the October share drop signal?

CoreWeave and Nebius, both young GPU rental companies, saw their shares fall in October as markets priced in higher interest rates. Business Insider does not tie the drop to any operational failure. CoreWeave reported Q3 revenue of $3.66 billion, up 168 percent, alongside a backlog of $132.5 billion. Strong growth and fragile financing can occupy the same balance sheet.

Is the depreciation clock too long?

Princeton technology-policy researcher Mihir Kshirsagar, quoted by Delta News, said: "The combination of wear and tear along with technological obsolescence makes the six-year assumption hard to sustain." D.A. Davidson analyst Gil Luria connected the loop to capital costs: "If they look like they're a lot less profitable because equipment must be replaced more frequently, it will become more expensive for them to raise the capital."

Investor Paul Kedrosky, writing in July 2025, put AI data center capex at roughly 20 percent of peak railroad spending as a share of GDP, and rising. But he noted these are not railroads; rails lasted generations, while a GPU cluster may not last a loan term. He called AI data centers "short-lived, asset-intensive facilities riding declining-cost technology curves, requiring frequent hardware replacement to preserve margins."

Model release cadence sets the commercial clock. Trackers report nine named launches in roughly five weeks in September and October 2026, including Claude Opus 5.5, GPT-6 Luna and Gemini 4 Argon, with one tracker counting 199 releases in October alone. We have no efficiency figures for these releases, so no specific cost-per-token drop can be claimed. The direction, however, is clear.

What should buyers and investors watch?

The late-1990s fiber buildout offers one comparison: infrastructure endured while first owners did not. Fiber aged slowly; accelerators do not.

For buyers of compute, the practical steps follow from the risks:

• Avoid long fixed-price contracts that assume today's rates hold.
• Keep workloads portable between providers.
• Check the financial health of smaller GPU cloud vendors.
• Re-benchmark every few months against newer, more efficient models.

None of this says AI demand will fade. It says profits will be uneven, and that owners of long-lived assets and owners of short-lived ones are holding very different bets on the same boom.