The article identifies a real and increasingly important risk, but the headline turns several different trillion-dollar figures into one ominous “problem.”
There is no known $1 trillion loss. The actual concern is that Big Tech is committing extraordinary amounts of capital to AI infrastructure while investors still cannot determine – with confidence – whether the resulting revenue, margins, and asset lives will justify the spending.
Yahoo restricted access to the full article during my review, so I verified its indexed thesis against company filings, earnings calls, Barclays’ forecast, Goldman Sachs’ infrastructure model, and reporting on depreciation disclosure.
What the article is really arguing
The AI investment cycle creates three separate problems:
- Cash is leaving immediately. Alphabet, Amazon, Microsoft, and Meta have collectively spent more than $1.1 trillion on capital expenditures since early 2023 and are projected to spend about $745 billion during 2026 alone. Financial Times
- The earnings expense appears gradually. Servers, GPUs, buildings, and power systems are capitalized on the balance sheet rather than expensed immediately. Their cost reaches the income statement through depreciation over several years.
- Investors cannot see the economics clearly. Companies provide limited detail about which assets are depreciating, their utilization, their remaining economic value, and the returns generated by AI-specific infrastructure. New expense-disaggregation rules should improve depreciation disclosure around 2028, but they still will not reveal true project-level AI returns. The Wall Street Journal
That accounting fog is probably the article’s most legitimate complaint.
Where the “$1 trillion” comes from
There are multiple trillion-dollar numbers being mixed together:
| Figure | What it actually represents |
|---|---|
| $1.1 trillion | Cumulative capital spending by four major technology companies since 2023 |
| $745 billion | Their estimated combined 2026 capital spending |
| $1 trillion annually | Barclays’ forecast for AI capital spending to peak around 2028 |
| $1.6 trillion annually | Goldman Sachs’ broader baseline estimate for 2031 |
| $7.6 trillion | Goldman’s cumulative 2026 – 2031 infrastructure estimate |
Barclays’ $1 trillion forecast assumes continued demand from AI labs, agentic systems and increasingly capable models. Barclays itself presents the estimate as controversial because physical constraints – including power, permits, construction and labor – could prevent spending from reaching that level. LinkedIn+1
Goldman’s model is even larger, projecting annual AI infrastructure investment rising from approximately $765 billion in 2026 to $1.6 trillion in 2031. But Goldman explicitly says this is a supply-side model based partly on projected Nvidia accelerator sales – not an independent forecast proving that end-user demand will support the resulting capacity. Goldman Sachs
Forecasted spending is not forecasted profitability. That distinction is critical.
Why the concern is legitimate
Free cash flow is already being squeezed
Alphabet raised its 2026 capex guidance to $195 – 205 billion and warned that depreciation, energy expenses and infrastructure investment would pressure both earnings and free cash flow. Alphabet
Meta spent $31.1 billion during the second quarter alone. Despite producing $31.9 billion in operating cash flow, it reported just $784 million in free cash flow. Its full-year capex guidance is now $130 – 145 billion. AtMeta
Microsoft remains in better shape: it spent $41 billion during its latest quarter but still produced $19.6 billion in free cash flow. Nevertheless, roughly two-thirds of that capex went into relatively short-lived CPUs and GPUs. Microsoft
The companies are not going bankrupt. The risk is that they become far more capital-intensive businesses than the valuations assigned to historically asset-light software platforms assume.
Useful-life assumptions materially change reported earnings
The economic life of a GPU is uncertain. It may physically function for six years but become commercially inferior much sooner because newer chips deliver radically better performance per watt and per dollar.
Goldman estimates that AI chips are commonly assigned useful lives of four to six years and identifies silicon life as the most influential variable in the entire infrastructure-spending model. Goldman Sachs
Extending depreciation periods raises current reported earnings. It does not create cash, improve utilization, or make an obsolete GPU more valuable.
That does not mean the companies are committing accounting fraud. It means earnings are unusually sensitive to estimates that cannot yet be validated.
Every company cannot simultaneously earn monopoly returns
Each hyperscaler is behaving rationally in isolation: insufficient capacity could cause it to lose the AI platform race.
Collectively, however, they can overbuild. If several companies create interchangeable compute capacity, prices fall, utilization declines and customers gain bargaining power. The infrastructure may be technologically useful while still producing mediocre shareholder returns.
This is the classic flaw in the bullish argument:
AI demand can grow enormously while AI infrastructure returns still disappoint.
Railroads, fiber-optic networks and airlines all changed the world. Many of the investors financing their expansion still lost money.
What the article underplays
Current demand is not imaginary.
Alphabet says demand continues to exceed available capacity and is accelerating infrastructure deliveries. Alphabet
Microsoft reported:
- $678 billion in remaining commercial performance obligations;
- $59.3 billion in quarterly cloud revenue;
- 27% cloud revenue growth;
- nearly 90% of annual cloud revenue coming from customers outside frontier-model companies. Microsoft
Nvidia’s most recent quarter produced $81.6 billion in revenue, up 85% year over year, with data-center revenue of $75.2 billion, up 92%. NVIDIA Investor Relations
So this is not presently a collection of empty data centers serving no customers. The bearish case depends on future capacity growing faster than monetizable demand, not on current demand disappearing.
What it means for Nvidia
For Nvidia, the spending boom is overwhelmingly beneficial today.
Nvidia records hardware revenue when it sells systems. Its customers bear the later problems of depreciation, financing, utilization and cloud pricing. Consequently, the accounting problem is more immediately damaging to hyperscaler cash flow and margins than to Nvidia’s earnings.
But the sequence can eventually reverse:
- Hyperscalers overbuild.
- Available capacity catches demand.
- Cloud pricing weakens.
- Customers delay the next GPU replacement cycle.
- Nvidia’s growth slows precisely when the market has capitalized several more years of high spending.
That is why a capex peak in 2028 is not automatically bullish for Nvidia. A peak means spending remains strong until then – but it also means the rate of growth eventually rolls over. Stocks react to the change in expectations before the actual spending peak.
Relevance to your concentrated Nvidia position
The article is not a persuasive sell signal. It does not demonstrate that AI demand has peaked, that Nvidia orders are being cancelled, or that existing infrastructure is broadly underutilized.
But dismissing it because Nvidia’s current revenue is exceptional would be equally careless.
Given your stated concentration, the dangerous assumption would be:
“Barclays expects spending through 2028, therefore Nvidia is protected through 2028.”
Forecasts are not purchase contracts. They can reverse within two quarters if utilization, financing conditions, power constraints or customer returns deteriorate. A highly concentrated portfolio turns an ordinary forecasting error into a potentially permanent personal-capital loss.
The indicators that matter are:
- hyperscaler capex growth relative to cloud and AI revenue growth;
- free cash flow after capex and lease payments;
- GPU utilization and cloud-compute pricing;
- depreciation growth and changes in asset-life assumptions;
- growth in custom accelerators such as TPUs and Trainium;
- Nvidia data-center growth, gross margin and customer concentration;
- evidence that inference demand is replacing training as the next large compute market.
Bottom line
The article’s accounting warning is stronger than its investment conclusion.
The genuine problem is not spending $1 trillion. It is spending $1 trillion without enough disclosure to determine:
- what portion is productive;
- how quickly equipment becomes economically obsolete;
- what returns each additional dollar generates;
- whether every hyperscaler’s demand forecast can be correct simultaneously.
The present evidence still supports a powerful AI infrastructure expansion. It does not yet prove attractive long-term returns for every company financing it. For Nvidia, the boom remains operationally bullish, while the eventual capex deceleration is one of the largest medium-term risks embedded in the stock.
The ruthless conclusion: the headline is exaggerated, but the underlying risk is not. Your vulnerability comes less from this article being correct than from having too little room for it to be correct.
Alert me if Big Tech cuts AI capex guidance