What is $B in AI infrastructure spending actually buying?

Six companies spent $B on capital expenditure in 2025, up from $B the year before. Three of them have run cloud infrastructure businesses for over a decade. AWS, Google Cloud, and Azure earned $B combined last year. The other three don't report an equivalent revenue segment. Every CEO who addressed the increase pointed to the same thing. Sundar Pichai, February 2025: "" Mark Zuckerberg, January 2025: "" Satya Nadella, January 2025: "" No company discloses what AI specifically earns.

Three questions follow from those numbers:

1. How much of the spending is backed by current revenue?
2. What are they building it for?
3. What happens if the demand doesn't arrive?

The previous cloud infrastructure cycle (2015–2021) roughly doubled spending every six years. This one doubled in two.

In January 2025, DeepSeek, a Chinese AI lab, released a reasoning model that matched GPT-4-class benchmarks for roughly $ million in training compute. Nvidia lost $ billion in market cap in a single day. Satya Nadella posted on X that evening: "Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket." None of the six cut spending. All accelerated.

Amazon has led total spending since 2022, with 2026 plans still climbing

The hatched 2026 column shows what each company says it plans to spend. Those plans have shifted by ±% in a single year before.

Between September and November 2025, Microsoft signed over $B in leases with neocloud companies (firms that build and lease AI compute capacity rather than running their own cloud services): Nebius ($B), Nscale ($B), Iren ($B), plus Lambda. Meta financed $B of Louisiana data center construction through a separate shell company called Beignet Investor LLC, with Blue Owl Capital providing 80% of the funding. Together, just those two commitments add $B in infrastructure obligations that don't appear in reported capex totals.

These are special purpose vehicles: legal structures that hold assets separately from the parent company. Because the arrangements are classified as operating expenses rather than capital investment, they never appear in reported capex.

A note on accounting: US rules that took effect in 2019 require long-term leases to appear on a company's balance sheet. But the payments still count as operating costs, not capital investment, so they drop out of the infrastructure spending totals that analysts and media typically report.

1. How much of the spending is backed by current revenue?

Amazon, Alphabet, and Microsoft are the only three that report cloud revenue as a separate segment (AWS, Google Cloud, Intelligent Cloud), so a direct comparison between spending and revenue is possible. Meta, Oracle, and Nvidia don't break out cloud or AI revenue in a way that allows direct comparison.

Microsoft's Intelligent Cloud is the only segment where quarterly capex still hasn't crossed above cloud revenue. That gap fell from $B to $B in a single quarter.

"" Sundar Pichai told analysts on Alphabet's Q4 2024 earnings call (February 2025). David Cahn at Sequoia Capital frames the question differently: his "AI's $B Question" (September 2024) estimates the industry needs $B in annual AI revenue to justify the current infrastructure pace. No company discloses what AI specifically earns.

MIT economist Daron Acemoglu estimated AI would add % to total factor productivity over a decade (NBER Working Paper 32487, May 2024). Goldman Sachs economist Joseph Briggs estimated %. Goldman's own head of global equity research, Jim Covello, asked in June 2024: ""

Near-term signals are mixed. Microsoft's Azure AI business reached a $B annual run rate by January 2026, growing 175% year-on-year. Google Cloud's remaining performance obligations reached $B. But a significant portion of cloud demand comes from AI startups whose spending depends on continued investor funding rather than their own revenue. McKinsey's 2025 survey found only about 25% of businesses testing generative AI saw meaningful bottom-line impact.

2. What are they building it for?

Andy Jassy (Amazon CEO) said in September 2025 that about Mark Zuckerberg, January 2025: ""

A September 2025 New York Times investigation (Metz & Weise) mapped at least six distinct demand theses being funded simultaneously, from products with revenue measured in quarters to bets with no disclosed timeline at all.

Marc Benioff claimed 5,000 Agentforce deals on Salesforce's Q4 FY2026 earnings call (February 2026). Satya Nadella reported GitHub Copilot had crossed a $2 billion annual run rate. Andy Jassy described AWS AI services as a "multi-billion-dollar" business growing at triple-digit rates. None of the three broke out agent revenue as a separate line item. No company has reported revenue from scientific AI or artificial general intelligence in any filing. A data center built in 2025 runs for 15–20 years.

Spending Forecasts Have Been Wrong Before, in Both Directions

The six companies have collectively forecast $B in infrastructure spending for 2026.

3. What happens if the demand doesn't arrive?

The last infrastructure cycle that looked like this was fiber optic cable in the late 1990s. US telecom companies invested roughly $B between 1996 and 2002, financed largely by $T in issued debt. Their capex-to-revenue ratio peaked at %, roughly double the industry norm. By 2002, an estimated 3–5% of installed fiber was carrying traffic. The infrastructure survived; the companies that built it mostly did not.

Those companies borrowed to build. Today's AI builders spend from operating cash flow. The six companies generated $B in trailing-twelve-month operating cash flow, enough to cover most of their infrastructure spending without issuing debt. If demand disappoints, the equity holders absorb the loss through lower earnings. No debt cascade.

Under the bear scenario, both capex and revenue fall. But the gap stays wide, because the same weakness that cuts spending also cuts the revenue it was meant to generate. For Amazon and Alphabet, infrastructure spending already exceeds what their cloud businesses earn. Sustaining current valuations requires that gap to close.

The financial commitments behind those valuations convert to physical assets on a different timeline. A data center under construction, a signed power purchase agreement, a transformer in the interconnection queue: these don't reprice in a day. What's already sunk and what's still convertible is what the next article traces.

This is the first of three DD-001 articles. Next: Only 14% of grid connection requests reach operation traces how slowly capital converts to working infrastructure. Then: Who Holds the Downside follows the risk to its final bearers.

Sources & Methodology

Capital expenditure (2015–2025). SEC 10-K and 10-Q filings for Amazon, Alphabet, Microsoft, Meta, Oracle, and Nvidia, retrieved via yfinance. Annual figures are fiscal-year sums of quarterly reported capex. 2015–2021 figures from company annual reports where quarterly data was unavailable.

2026 guidance. Q4 2025 earnings calls and investor presentations. Amazon ($B), Alphabet ($B), Microsoft ($B), Meta ($B). Uncertainty band (±%) based on the largest single-year guidance revision in the 2020–2025 period (Meta 2022–2023 cut; Microsoft FY2025 raise).

Cloud segment revenue. AWS, Google Cloud, and Intelligent Cloud quarterly revenue from SEC 10-K/10-Q filings. These are the only three hyperscalers that report cloud as a separate segment.

Off-balance-sheet commitments. Microsoft neocloud leases: Nebius ($B), Nscale ($B), Iren ($B), plus Lambda — reported in company announcements and New York Times, Dec 15, 2025 (Weise & Tan). Meta Beignet Investor LLC: $B Louisiana data center financing via Blue Owl Capital.

Revenue demand theses. Classification based on New York Times, Sep 2025 (Metz & Weise), cross-referenced with company earnings calls and product announcements.

Telecom historical baseline. Andrew Odlyzko, "Internet traffic growth: Sources and implications" (2003), University of Minnesota. Cumulative telecom capex ~$B, debt issued ~$T, capex-to-revenue peak ~%. Fiber utilization estimates (3–5%) from industry analysts and trade press, 2002; no authoritative government survey measured this directly.

Executive quotes. Sundar Pichai: Alphabet Q4 2024 earnings call, February 4, 2025. Mark Zuckerberg: Meta Q4 2024 earnings call, January 29, 2025. Andy Jassy: New York Times (Metz & Weise), September 16, 2025. Satya Nadella: post on X, January 27, 2025 ("Jevons paradox strikes again"), confirmed by CNBC (Novet, January 27, 2025); Microsoft Q2 FY2025 earnings call, January 29, 2025. Jensen Huang: Nvidia GTC 2024 keynote, March 18, 2024. All quotes reported verbatim by multiple outlets; verification against official transcripts recommended.

Skeptic and analyst sources. Jim Covello, Goldman Sachs: "Gen AI: Too Much Spend, Too Little Return?" (June 25, 2024), paywalled equity research — quote widely reported by Bloomberg and FT. Daron Acemoglu: "The Simple Macroeconomics of AI," NBER Working Paper 32487 (May 2024), freely available abstract. Joseph Briggs, Goldman Sachs: productivity estimate referenced in comparison to Acemoglu.

Azure AI revenue. Microsoft Q2 FY2026 earnings call (January 29, 2026): $B annual run rate for Azure AI. "Annual run rate" = last quarter's revenue × 4; it is a snapshot, not a contractual commitment. Google Cloud RPO: Alphabet Q4 2025 10-K filing.

DeepSeek training cost. DeepSeek-V3 technical report (arXiv:2412.19437, December 2024) discloses 2,788,000 H800 GPU hours for the final pre-training run; at approximately $2/GPU-hour market rate, this implies ~$5.6M in compute. Covers final training run only — excludes prior experiments and hardware acquisition. GPT-4 training cost: Sam Altman public statement, April 2023 (reported by Business Insider and Fortune). Inference pricing: DeepSeek official API (platform.deepseek.com) vs OpenAI API (openai.com/api/pricing).

Nvidia market cap loss. $ billion on January 27, 2025 — the largest single-day market cap loss in US stock market history. Sources: Reuters, Yahoo Finance, and CNBC, all reporting January 27, 2025.

Scenario projections. TZD Labs analysis. Cloud revenue range ($–$B) based on trailing growth rates applied to 2025 actuals. Bull case assumes acceleration; base case assumes current trajectory; bear case applies the ±% guidance band to both capex and revenue.

All data from public, freely accessible sources. No proprietary data or paywalled databases. Analysis: TZD Labs, February 2026.