The AI Data Center Boom: Who Pays If the Investment Doesn't Pay Off?

01 / The story

What companies are actually spending

The numbers are enormous, but they mean different things. Microsoft reported $41 billion in quarterly capital expenditure on its July 29 earnings call, including $35.8 billion in cash paid for property and equipment and $5.6 billion in finance leases; roughly two-thirds of its capex was short-lived assets, primarily CPUs and GPUs. Its revised calendar-2026 expectation was approximately $175 billion—guidance, not money already spent. [1]

Alphabet reported Google Cloud revenue of $24.8 billion, up 82%, on July 22, and Reuters reported management raised 2026 capex guidance to $195–205 billion. Meta's July 29 results showed quarterly capex including finance-lease principal of $31.08 billion, with revenue up 28% but operating income down 8%. Amazon's July 30 results showed trailing-twelve-month operating cash flow of $161.4 billion alongside negative $7.6 billion in free cash flow—implying roughly $169 billion in net cash property and equipment purchases over that period, across the whole company, not AWS alone. [2, 3, 4, 5]

Oracle's September 10 results showed cloud infrastructure revenue up 121% to $7.4 billion and remaining performance obligations—contracted future revenue—of $664 billion. Backlog is not cash collected, and not proof of its ultimate margin. [6]

Then there are the partnership announcements. The September 2025 Nvidia–OpenAI announcement was a letter of intent for at least 10 gigawatts of Nvidia systems, with Nvidia intending to invest up to $100 billion progressively as capacity deploys. The Stargate update described nearly 7 GW of planned capacity and over $400 billion of investment over three years. These are conditional intentions and overlapping plans—not cash transferred, and not additive totals. [8, 9]

02 / The story

Follow the contract, not the logo

When a data center fails to earn its keep, the loss lands wherever the contracts put it. Each structure has a different first victim.

If a company funds infrastructure with its own cash, its shareholders take the hit first, through lower returns, depreciation and possible write-downs. With corporate bonds or loans, equity still loses first, but creditors are exposed to default or restructuring. With leased buildings and equipment, the lessee owes contracted rent while the owner or lender carries the residual-value risk. [1, 13]

Newer structures spread risk further. Project financing and private credit can place losses with project equity, lenders and guarantors. Long-term cloud commitments can push unused-capacity costs onto customers. And when a supplier invests in its own customer—as Nvidia's intended investment in OpenAI would—the investor can lose equity at the same time as it loses future product demand. That linkage is real, but it is not automatic proof of fake revenue. [8, 13]

03 / The story

The ratepayer question is real

One group rarely mentioned in AI announcements is electricity customers. When a utility builds grid capacity for a data center that never fully arrives, who pays for the stranded equipment?

Ohio has already answered once. On July 9, 2025, the Public Utilities Commission of Ohio ordered AEP Ohio to create a data-center-specific tariff, explicitly identifying cost-shifting risks from underused infrastructure and adding safeguards for other customers. AEP says the tariff took effect on July 23, 2025. This establishes that stranded-grid-cost concerns are real regulatory issues—not that households inevitably pay every loss. Outcomes depend on each jurisdiction's rules. [12, 14]

04 / The story

Electricity, cooling and the depreciation problem

The International Energy Agency's base case projects worldwide data-center electricity demand of around 945 TWh in 2030—more than double the level when the projection was made—with AI-optimized facilities' demand more than quadrupling. That is a forecast covering all data centers, not an AI-only measurement. [10]

Cooling's share of facility electricity ranges from about 7% at efficient hyperscalers to over 30% at less-efficient enterprise facilities, according to the IEA. To make that concrete: a facility drawing a constant 1 GW uses 8.76 TWh per year; at an assumed $0.07 per kWh, energy alone costs about $613 million per year. That is an illustrative calculation, not a market quote, and it excludes demand charges, equipment, water, financing and taxes. [11]

Depreciation is the quieter risk. Faster chips can undermine an older cluster's price-performance before the hardware physically fails. Microsoft's own disclosure—that about two-thirds of its capex is short-lived compute—shows how much of this spending races against obsolescence. But 'every GPU is worthless after two years' is not supported by the evidence; residual value depends on workload, energy costs and rental prices. [1]

05 / The story

The bubble debate, honestly stated

The evidence supporting concern: capital outlays can outrun internally generated cash, as Amazon's negative trailing free cash flow shows. Reported capex can shift with accounting choices, as Microsoft's lease-classification explanation shows. And supplier-invests-in-customer arrangements create observable feedback loops that deserve scrutiny. [1, 5, 8]

The evidence against 'it's all hype': Microsoft, Alphabet, Amazon and Oracle report substantial current cloud revenue and rapid growth, and Meta has a large existing revenue base. These are operating businesses, not pre-revenue ventures. [2, 4, 5, 6]

What remains unsettled is the part that matters most: whether the incremental AI dollar earns its cost of capital. Broad cloud and advertising revenue does not prove that. Falling utilization and rental prices, impairments, cancellations and refinancing stress would strengthen the bubble case; sustained externally funded demand and positive cash returns after replacement costs would weaken it. Watch those signals, not the headline totals.