AI Capex Looks Unique Because the Chips Wear Out

Five U.S. cloud giants have guided to about $800 billion of 2026 capital spending, including leases. Most of that money now buys chips those companies write off in five or six years.

Financial columnist Matthew Lynn argues the AI frenzy may be “one of a kind,” unlike the railroad mania, 1920s electrification, or the dot-com fiber boom. The uniqueness is real, and it is the problem: railroad track could still earn its keep decades after the first owners were wiped out, while a server fleet has to be bought again on a five- to six-year cycle.

The Railroad Share of National Investment

The analog will not die because the scale, on paper, is in the same family. Pulitzer-winning historian Liaquat Ahamed, who has just finished a book on the 1870s railroad boom, puts planned AI infrastructure outlays at $800 billion to $1 trillion a year for the next few years, or about 2% to 3% of GDP, a band he also uses for that earlier flood of private capital. Between 1865 and 1872, he notes, the U.S. rail network doubled from 35,000 miles to 70,000 miles.

A longer NBER series on steam railroads is blunter about weight inside the investment budget. From 1870 through 1890, railroad gross capital spending averaged 15 to 20 percent of all U.S. investment. In the 1870s it was 20.1% of national capital formation. In the 1880s it was still 15.6%.

The fiber build of 1995 to 2000 is the closer modern cousin. Ahamed writes that information-technology investment then climbed from 3% of GDP to nearly 5% at the peak. That boom left glass in the ground. A lot of the companies that paid for it did not survive to collect the tolls.

TWO BOOMS, TWO ASSET LIVES

Warren Buffett, explaining Berkshire Hathaway’s Alphabet stake, put the same gap in plainer words: “That’s real money. That kind of money wasn’t even put in the railroad business.” He was talking about hundreds of billions leaving software-like cash machines and going into physical plant. He was not talking about how long a GPU remains the right GPU.

Most of the New Money Buys Chips

FactSet’s cut of Alphabet, Amazon, Meta, Microsoft, and Oracle is the cleanest single ledger for 2026. Cash capex for those five is expected to exceed $690 billion in their FY26 years, up more than 80% from the prior year, after about 70% growth in FY25. Calendar 2026 guidance near $800 billion once finance leases and customer prepayments are counted. The same five spent about $95 billion in FY20 and about $490 billion in the twelve months to May 2026. Estimates already point past $900 billion by FY28.

The mix inside those dollars is the tell. Compute, meaning processors and servers, is expected to be about $380 billion in 2026, roughly double 2025, and about 60% of capex, up from about 43% in 2022. Buildings, land, and other 10-year-plus kit are a shrinking share. A growing slice of each year’s budget is replacement, not a one-time pour of concrete.

THE 2026 CAPEX MIX

• Five-firm cash budget: More than $690 billion in FY26, the largest annual increase of this cycle.
• Including leases: Calendar guidance close to $800 billion after finance leases and prepayments.
• Short-lived share: Compute about 60% of 2026 capex, versus about 43% in 2022.
• Chip bill: About $380 billion of processors and servers in 2026, roughly double the year before.

Alphabet has already lifted 2026 capex guidance to $180 billion-$190 billion. Meta is in a $125 billion-$145 billion range. Memory prices, especially high-bandwidth memory, are part of why those numbers keep moving up. The other part is that last year’s fleet does not cancel this year’s order.

How Long an AI Server Lasts

On the books, an AI server now lives about five and a half to six years. Microsoft, Alphabet, and Oracle depreciate servers over six years. Meta extended most servers to 5.5 years in January 2025. Amazon still uses six years for most of the fleet and five for a subset. None of them books GPUs as a separate asset class, so the “server” life is the figure investors actually get.

Those extra years are not a rounding error. In the year each firm stretched its schedule, net income rose by about 4% to 4.6%. Microsoft’s FY2023 move from four years to six added $3.7 billion to operating income and $3.0 billion to net income, or $0.40 a share. Alphabet’s FY2023 change cut depreciation by $3.9 billion and added $3.0 billion to net income. Amazon’s FY2024 move from five years to six cut depreciation and amortization by $3.2 billion and added $2.5 billion to net income. Oracle’s FY2025 step from five to six trimmed operating expense by $733 million and added $573 million to net income.

Meta’s 2025 change is the latest and the largest single-year print in that set. Extending lives to 5.5 years reduced 2025 depreciation by $2.92 billion and lifted net income by $2.59 billion, or $1.00 a share.

SERVER LIVES ON THE BOOKS

Over the same recent twelve months, the five firms’ capital spending of about $482 billion ran at nearly four times their $125 billion of property-and-equipment depreciation. Stretch the assumed life and that gap stays open longer. Investor Michael Burry has argued the economic life is closer to two or three years and that depreciation through 2028 is understated by a huge margin. The filings do not show anything near his $176 billion figure. They do show five accounting teams using different clocks for similar hardware, and one of those teams already reversing course.

Amazon Cut Its Own Clock Back

Amazon is the only hyperscaler that has moved the life of some AI kit backward. After a useful-life study in the fourth quarter of 2024, it shortened a subset of servers to five years, effective January 1, 2025, from six. For assets already on the books, it said the change would cut 2025 operating income by about $0.7 billion. It also took about $920 million of accelerated depreciation and related charges in the fourth quarter of 2024 for early retirements, with about $0.6 billion more of that hitting 2025. Combined, those two AI-pace changes were a $1.3 billion 2025 operating-income hit, mostly at AWS.

These two changes above are due to an increased pace of technology development, particularly in the area of artificial intelligence and machine learning.

Amazon.com, 2024 Form 10-K

A year earlier the same company had lengthened servers from five years to six, citing hardware, software, and data-center design. The reversal is the filing that matters. Amazon is telling investors, in its own words, that AI kit ages faster than the six-year clock the rest of the group still uses.

That does not prove Burry’s two-year scrap heap. CoreWeave chief financial officer Nitin Agrawal has said the company signed a contract to rent Nvidia A100 chips, launched in 2020, into 2029, and older cards can drop from training to cheaper inference. Physical life and economic life are not the same thing. Amazon’s point is narrower and harder to wave away: when the next architecture ships, some of what you just bought is already on a shorter clock.

Debt, Leases and a Record Stock Sale

For a decade these firms funded capex from operating cash. That habit is breaking. FactSet finds incremental debt rose from 9% of capex in FY24 to 32% in the twelve months before June 2026, with aggregate total debt around $700 billion. Free cash flow in FY26 is expected to sit near zero or turn negative for everyone in the group except Alphabet and Microsoft. Gross leverage is still about 1 times EBITDA or lower for four of the five, which is why the bond market has been willing to listen.

THE FUNDING SHIFT

1. FY 2020: Combined investing cash flow for the five is about $95 billion, funded inside the firms.
2. FY 2024: Incremental debt is 9% of capex. Cash still does most of the work.
3. January 1, 2025: Amazon shortens a subset of servers from six years to five and books early retirements on AI-pace grounds.
4. June 2026: Alphabet prices an $84.75 billion equity raise, including a $10 billion private placement with Berkshire Hathaway, with $44.75 billion aimed at general corporate uses such as AI capex.
5. July 9, 2026: S&P cuts Oracle to BBB- from BBB, one notch above speculative grade, citing surging capex, negative free cash flow, and customer concentration.
6. Mid-2026: Incremental debt is 32% of capex. Off-balance-sheet lease commitments across the five are about $820 billion.

Oracle is the outlier on purpose. It is building for a concentrated set of AI-lab customers, and FactSet notes lease obligations with 15- to 19-year terms due to start between FY27 and FY29, totaling almost three times its FY27 capex guidance. A 16-year lease on a building that will hold five or six generations of servers is a duration mismatch, not a railroad. Amazon’s recent $25 billion bond showed the bid is still there. It also showed these cash machines are now regular issuers.

Meta has even floated selling spare compute. Mark Zuckerberg has said a cloud business is “definitely on the table,” which is a polite way of saying some of the capacity being built is looking for a second customer. That is a hedge, and it is also an admission that internal models are not soaking up every rack.

What Survives If the Spending Stops

Railroad bankruptcies in the 1870s still left a network. Fiber gluts in 2001 still left a backbone later companies could buy for cents on the dollar. A halt in AI capex leaves a different residue, because so much of the spend is the board, not the shell.

WHAT A BUST LEAVES THIS TIME

• The building: Data-center shells, power hookups, and cooling can last 15 to 20 years and can be re-leased or resold.
• The chips: GPUs and the servers around them are already on five- to six-year books, and Amazon has said a subset ages faster than that.
• The contracts: Multi-year chip purchases, 15- to 19-year leases, and lab offtakes do not shrink just because training demand cools.
• The circular bit: Chipmakers and clouds taking stakes in the labs that buy their product, a loop the Bank for International Settlements has flagged as a contagion path.

A BIS working paper on the AI investment race models the build as a winner-take-most contest and finds over-investment of about 50% above the efficient level in a conservative baseline, or about 1.5 times efficient spend, rising toward three times if demand is less elastic. That is not a forecast of a crash date. It is a claim that a race among a handful of firms produces more plant than a planner would want, and that debt plus circular equity makes a bust more likely once revenue disappoints.

The other side of the analog still holds in one respect. After the British railway mania and after 1873, later operators made money on assets the first owners could not. Someone will run inference on last year’s cards. The salvage value of a Blackwell rack is not the salvage value of a main line into Chicago.

Energizing the Fleet Is the Hard Part

None of this requires AI demand to be fake. Gavin Baker, managing partner at Atreides Management, has argued that spot prices for GPU rentals are at least twice contracted rates, which would mean hyperscalers are under-earning on installed compute while private buyers overpay. If he is right, cash from operations can still fund a lot of the build as contracts roll off. Credit spreads that have widened this year then look like a governor, stretching the timetable and trimming the odds of a pure overbuild, rather than a verdict that the racks are empty.

The constraint that keeps showing up instead is power. Baker’s own close is that “bringing power online and energizing all these GPUs is really hard.” Token prices can fall while token volumes rise, which is good for users and rough for anyone who paid 2025 prices for 2026 utilization. Residual value on last year’s GPUs is the sleeper line in every “this is just infrastructure” pitch, because infrastructure that dies on a six-year clock is a consumable with a building around it.

Nvidia chief executive Jensen Huang did not sound like a man winding the cycle down. On September 6, 2026, he said OpenAI’s GPT-6 Astra had been trained on more than 100,000 Grace Blackwell NVLink72 chips, and that 400,000 more GPUs are coming online next.

https://x.com/JensenHuang/status/2096700264569090384

Those 400,000 cards will land in halls that can stand for decades. They will still need to be bought again before a lot of the leases written to house them have run even a third of their term.

Disclaimer: This article is news reporting and analysis of public filings, research notes, and company comments on AI infrastructure spending. It is informational only and is not a recommendation to buy or sell any stock, bond, or other security, and it is not investment, tax, or accounting advice. Readers should consult a licensed financial adviser or registered investment professional before making decisions based on capital-expenditure figures, depreciation assumptions, or credit ratings discussed here. Figures and ratings reflect the cited company reports, FactSet analysis, and other sources as dated in the piece and can change with the next earnings cycle.