AI Capex Speed Is Not the Housing Boom - the Funding Structure Is

Apollo Global's chief economist Torsten Sløk published a set of charts yesterday that are now circulating through the financial press. The headline figure is arresting: AI data-center capex as a share of GDP is rising at close to twice the pace of the housing boom at its fastest.

The number structure tells a more precise story. And the one dimension the speed comparison leaves under-exposed is the question that actually matters to investors - not how fast the build is happening, but who is paying for it.

Decomposition

Sløk's charts break three historical capex cycles into GDP share:

• AI data-center buildout: projected to reach roughly 3% of GDP by 2027–2029, up from 0.3% in 2019 and 1.4% in 2025.
• Housing boom peak: 6.6% of GDP in 2005.
• Telecom/fiber buildout peak: 1.2% of GDP in 2000.

In absolute level, the AI buildout is still less than half the size of the housing peak. On Apollo's own arithmetic, the AI cycle only becomes the big number on speed and cumulative change. Data-center capex adds 2.5 percentage points of GDP from 2023 to 2027, versus 2.2 for housing (mid-1990s to 2005) and 0.4 for telecom. And on the speed metric - the one generating the headline - AI capex adds 1.7 percentage points in two years (2025 to 2027), or roughly 0.85 percentage points a year, versus 0.5 for housing's fastest phase from 2002 to 2005.

The math checks out. But the speed comparison, while correct, obscures a structural difference that changes the entire risk equation.

The funding source is different

The housing boom was funded by leverage - mortgage debt, subprime lending, mortgage-backed securities, and a shadow banking layer that amplified every dollar of equity into many dollars of construction. When the bubble unwound, from 6.2% of GDP in early 2006 to 3.0% by the end of 2008, the destruction was not just in homes falling off cranes. It was in a financial system leveraged against them. The unwind made the recession severe because the capital structure was fragile.

The AI capex cycle is funded almost entirely by operating cash flow and corporate balance sheets. Amazon, Microsoft, Alphabet, and Meta collectively spent roughly $413 billion on capex in 2025. Their 2026 guidance - a combined $725–760 billion - still represents companies with aggregate free cash flow in the hundreds of billions and net cash positions or investment-grade credit. These are not leveraged bets. They are cash-rich companies deploying their own earnings into infrastructure they control.

You can't trigger a financial crisis by writing a check from a balance sheet that holds hundreds of billions in liquid assets. The unwind risk is different because the capital structure is different.

But the capex-to-revenue gap is widening, and that is the structural question the speed headline doesn't address.

The capex-to-revenue gap

If hyperscaler capex hits $760 billion in 2026, the AI-related revenue it is supposed to generate is still a fraction of that.

OpenAI ended 2025 with approximately $20 billion in annual recurring revenue - triple the prior year, and impressive growth from a near-zero base two years ago. Anthropic's revenue run rate surpassed $9 billion in January 2026, up from roughly $1 billion at end-2024. AWS AI-related revenue, including chips, is approximately $25 billion annualized. Microsoft's AI business grew 123% year-over-year, though from a small base. Alphabet cloud grew 82%.

These are strong growth rates. But the aggregate AI revenue that the hyperscalers can directly attribute to their infrastructure investment is still well below $100 billion - against $760 billion of annual spending. The payback period, even on optimistic assumptions, is measured in years, not quarters.

This is not a fatal flaw. Infrastructure spend always leads revenue. The telecom buildout of the late 1990s also showed a multi-year gap between capex and monetization. The difference then was that the revenue never fully arrived. The question now is whether AI's revenue trajectory - already growing at 80–120% year-over-year - is steep enough to close the gap before investors lose patience.

Three unwind scenarios, not one

Apollo's closing warning is worth taking seriously: a cycle that builds at 0.85 percentage points a year can unwind at a similar pace, and that is the macro risk if AI demand disappoints. The same arithmetic runs in reverse.

But the unwind would not look like 2008. Three scenarios matter:

Capex discipline failure. The hyperscalers keep spending because of competitive fear, even as utilization rates soften. The result is not systemic crisis but depressed returns on invested capital. The stock impact would be multiple compression, not balance-sheet collapse.

Supply bottleneck as a hard ceiling. Power availability, chip supply, or cooling infrastructure constrains the build. This is already happening. Microsoft disclosed an $80 billion Azure backlog that cannot be fulfilled due to power constraints. The constraint is a risk to growth, not to capital safety - the money just doesn't get deployed as fast as guided.

Revenue inflection failure. AI products generate incremental engagement but not incremental willingness to pay at scale. This is the structural version of the telecom bust, where the infrastructure was built for a demand curve that never materialized. This is the one scenario that would make the housing analogy relevant, because the capex would become stranded assets instead of income-producing infrastructure.

Scenario 3 is the one worth watching. Scenarios 1 and 2 are operational risks that the hyperscalers can manage with discipline. Scenario 3 is the narrative-versus-earnings gap that has to resolve one way or the other.

What to watch next

• Capex-to-free-cash-flow ratio by company. If any hyperscaler's annual AI capex exceeds its annual free cash flow for multiple consecutive quarters without a clear revenue bridge, the funding model shifts from self-sustaining to externally dependent. That would be the first structural warning sign.
• Utilization rates. The hyperscalers have not been transparent about how much of the installed compute capacity is actively serving paid workloads versus sitting idle. If utilization drops below 50%, the capex story changes from growth investment to excess capacity.
• AI revenue deceleration. Current AI revenue growth (80–123% year-over-year) is the only number that justifies the current capex run rate. The moment that growth rate slows to the high teens or low twenties - the cloud industry's mature growth band - the market will force a re-evaluation of the entire buildout.
• Power constraints as a hard ceiling. If the $80 billion Azure power-constrained backlog persists into 2027 and 2028, actual AI capex deployed could fall short of guided levels - not because companies stop wanting to build, but because they physically cannot. That changes the GDP math Apollo's charts project.
• Oracle's expansion. Oracle's 136% capex increase to $50 billion, backed by $523 billion in remaining performance obligations, shows the buildout spreading beyond the four-name concentration. More players means more competitive pressure on cloud margins, which pressures the return on every dollar of AI infrastructure.

The headline that AI capex is building twice as fast as the housing boom is true on the speed metric. But the analogy breaks on the dimension that matters most. The housing boom destroyed the financial system because it was funded by leverage that amplified the unwind. The AI buildout is funded by cash flow that survives a correction. The risk is not a crash. The risk is a multi-year period of depressed returns if the revenue curve does not keep up with the spending curve. That is a different beast entirely.