Investor Andrew Chen Shares AI Infrastructure GDP Estimate Chart, AI to Account for 3.63% from 2025 to 2032
Venture capitalist Andrew Chen shared a chart showing the proportion of infrastructure spending relative to GDP, captioned "epic." The chart notes that AI spending is an estimate, sourced from economist Stijn van Nieuwerburgh.
The average annual proportions are: AI 3.63% from 2025 to 2032; telecommunications and fiber optics 1.1% from 1996 to 2003; highways 1.13% from 1956 to 1973; electrification 0.5% from 1905 to 1925; railroads 2.24% from 1870 to 1890; canals 0.66% from 1836 to 1841. Brookings estimates that from 2025 to 2032, investments in U.S. data centers and AI-related infrastructure will total about $10.3 trillion, averaging about 3.6% of GDP, and this could represent the most significant expansion in U.S. history reliant on a single industry.
The same study extrapolates a cost of $8.2 billion per 200 MW of computing power, assuming some planned installations will not materialize and some will only come online after 2032. Capital expenditures for hyperscale cloud providers are projected to rise from about $97 billion in 2020 to over $400 billion in 2025, potentially exceeding $800 billion in 2026, surpassing the combined operating cash flow of five companies. Goldman Sachs also estimates that AI investment could account for about 1.9% of GDP in 2026. The author states that without this round of spending, the U.S. economy might have already fallen into recession; financing is shifting from on-balance-sheet to leasing, project financing, securitization, and special purpose vehicles, with incomplete debt and information disclosure identified as risks.
The peak of railroad investment is often referenced as the last time a single industry held a higher share of the economy. Chip depreciation occurs faster than rail tracks, and data centers require frequent upgrades to acceleration cards, with asset lifespans shorter than canals and interstate highways. The credit spreads in the bond market for major spenders have widened this year, indicating that the pricing reflects the notion that "today's spenders may not be tomorrow's winners."
In market mechanisms, buyers are model companies and cloud providers looking to lock in computing and power, while sellers include chips, electricity, real estate, and private credit. Funds are entering data centers from both operating cash flow and off-balance-sheet debt. The event is driven by the dissemination of research charts within the investment community. Beneficiaries are those able to sell chips, substations, and project debt during the expansion cycle; those under pressure are the overall economy betting on a single wave of capital expenditure, and creditors leveraged on specialized, easily outdated assets.
The area of the AI bubble in the chart is significantly larger than that of railroads, yet the timeline is the shortest, comparing proportions rather than cumulative absolute amounts.
Source: Public Information
ABAB AI Insight
Chen only replied with two letters, spreading the proportion rather than the profit margin. The 3.63% positions AI as a "heavier" national economic project than railroads, assuming the planned spending from 2025 to 2032 is treated as an already occurring annual flow. Railroads leave operational rights for a century, while data centers leave acceleration cards that need replacement every three to five years; the same percentage corresponds to completely different residual values. The $10.3 trillion is a pipeline plus assumptions, not signed checks.
Capital is migrating from the cash flow of five hyperscale companies to banks, private credit, and special purpose vehicles. When on-balance-sheet cannot cover, risks shift from tech stock earnings reports into the credit chain, with widening spreads indicating the market pricing ahead of the narrative. The author states that without this round of spending, the economy is close to recession, acknowledging that growth figures are highly sensitive to data center construction: halting construction would simultaneously undermine both tech capital expenditure and GDP.
Benchmark comparisons include 19th-century railroad bonds, late 1990s fiber optic overcapacity, and public financing amortized by fuel taxes for interstate highways. The AI expansion is in a period where private companies bear a growing share of quasi-public infrastructure, with power and land constraints being harder than model parameters.
Structural changes belong to capital concentration. Spending, debt, and growth accounting are simultaneously piled on a few cloud providers and a few chip suppliers. The mechanism is: when the proportion is large enough, the capital budget of a single industry becomes a macro stabilizer; once the stabilizer relies on debt for survival, the macro environment will inherit the depreciation speed of chips.
ABAB News · Law of Cognition
A proportion exceeding that of railroads does not mean the remaining assets will last as long as railroads.
Growth supported by data centers will disappear together when construction halts.
Off-balance-sheet financing turns the tech bubble into a credit cycle.