The four largest hyperscalers—Microsoft, Alphabet, Meta, and Amazon—closed the second quarter of 2025 with aggregate capital expenditure above $60 billion. The year-over-year growth for that single ledger line is roughly twice the annual pace recorded during the peak of the 2000s housing boom. Ledger doesn't lie, but the ledger is incomplete.
Crypto Briefing published the comparison under the headline “AI capex boom is growing twice as fast as the housing boom did.” The piece is an alert, not an analysis: four information points, no unique dataset, no sourcing, no author byline. Its function is to join an established chorus—Goldman Sachs's “AI: too much spent, too little benefit” note, Sequoia's $600 billion annual AI revenue gap argument—that artificial intelligence infrastructure spending has reached bubble territory. As a data analyst who tracks capital flows professionally, I treat the chorus as a hypothesis, not a verdict. The verdict requires a ledger.
From an audit perspective, the visible condition is an unexplained variance: a capital expenditure line growing at a rate that historical depreciation schedules do not support, with an assumption that future revenue will reconcile the difference. My job is to determine whether that assumption is sound. Based on my audit experience—DeFi bridge reconciliations in 2021, the Terra collapse verification in 2022, and spot Bitcoin ETF flow mapping in 2024—the pattern is always the same at the start. The flow is real. The question is whether the receiving side will pay enough to keep the system solvent.
Context: What the capex line actually captures
Capital expenditure at hyperscaler scale is not a single purchase. It covers GPU clusters and custom ASICs; data center land and construction; network gear; cooling systems; and increasingly, multi-year power-purchase agreements. The buyers are few: Microsoft, Alphabet, Meta, Amazon, and now Apple, with OpenAI's infrastructure partners entering the ledger through special-purpose vehicles. The selling side is even more concentrated: NVIDIA for compute, with dedicated suppliers for memory, power, and cooling.
The accounting matters more than the headline. Hyperscalers depreciate computing hardware over four to five years. Data centers are depreciated over 20 to 30 years, but they must generate utilization to support the principal. Power contracts run 10 to 20 years. The engine behind the “twice as fast” growth narrative is that these commitments are made in advance: NVIDIA's order book extends 12 to 18 months; data centers take two to three years to construct; power delivery requires electrical grid upgrades that outlive a typical technology budget cycle. Tracing the source. The order book, expressed as a multi-quarter backlog, is the origin of the rigidity. The spending is committed before the demand is verified.
The housing comparison captures the growth rate but not the composition. In the 2000s, U.S. residential investment grew at roughly 15 to 20 percent annually at its nominal peak. The hyperscaler capex line is growing at 40 to 60 percent over the same comparable periods. The base, the leverage, and the end-buyer are different. Comparing growth rates without reconciling those variables is a journalistic device, not an audit.
The public audit trail for this market is the SEC EDGAR database, not a blockchain explorer. The XBRL-tagged elements for capital expenditure, purchase obligations, and depreciation are machine-readable and updated quarterly. I scan those tags the same way I scan Etherscan for transfer anomalies: looking for the variance between the reported flow and the disclosed commitment. The filing is the ledger. The ledger has no sentiment column.
Core: The locked collateral structure
What I find when I audit the current build-out is a structure resembling a DeFi protocol approaching a liquidity limit: a large volume of value locked in illiquid positions, a depreciation schedule that functions as ongoing debt service, and revenue that must arrive on schedule to cover the maintenance. Follow the outflows. The first critical line is depreciation.
The model I use for protocol audits treats liabilities as claims on productive assets, adjusted for a withdrawal or usage curve. When a DeFi bridge reports total value locked, that TVL number is not coverage of its outstanding wrapped assets. The collateral must be liquid, or the audit fails. For hyperscalers, the GPU clusters and data centers are the collateral, and the depreciation schedule is the debt. A $10 billion GPU cluster, depreciated straight-line over five years, requires $2 billion per year in accounting recovery just to keep book value flat. That recovery only occurs if the compute is utilized. If inference demand misses, the depreciation charge does not pause. The ledger doesn't negotiate.
In 2021, I spent 400 hours verifying transaction hashes for cross-chain bridge liquidity during my master's research. I identified a $2.5 million discrepancy caused by off-chain oracle manipulation that inflated a bridge's reported reserves. The bridge's own dashboard matched its smart contracts; the mismatch sat between the smart contracts and the off-chain asset layers. I submitted a 50-page technical audit, citing specific block numbers and gas fees, to the project repositories. The engineers confirmed the issue. That experience became my template: verify the asset side, then check whether the liability side matches. The same template applies to hyperscaler capex.
When I apply that template today, the reported asset is “future AI capacity.” The liability is the multi-year commitment on the order book: NVIDIA's backlog, construction in progress, energy obligations. The mismatch is not hidden. It sits in the quarterly earnings statement in three visible forms: depreciation growth, operating margin movement, and purchase commitments in the footnotes. In the December 2024 and March 2025 filings, those three lines moved in the same direction: depreciation climbing at a rate that cloud AI revenue growth barely covers. That gap is the risk.
The concentration compounds the risk. For fiscal 2024, Microsoft, Alphabet, Meta, and Amazon reported combined capital expenditure of roughly $180 billion. The pace in the second half of 2024 pointed to an annualized run rate above $200 billion for calendar 2025. NVIDIA's data center revenue—the upstream confirmation line—grew sequentially from roughly $18 billion per quarter in early fiscal 2024 to more than $25 billion per quarter by calendar Q4 2024. The order book, expressed as a multi-quarter backlog, kept growing.
Receipt-less purchase orders are the core problem. In a normal equipment cycle, order book growth reflects demand. In a four-entity market, order book growth can also reflect defense. Each hyperscaler must secure the same limited supply or lose the model-quality race. That kind of bidding is a prisoner's dilemma. It inflates orders without a proportional increase in verified end-user demand. The behavior mirrors what I documented in the 2024 ETF flow work: institutions acting in concert created a flow pattern that looked like momentum but was actually a compliance timing artifact. Only reconciliation with the actual holders revealed the trend.
The crypto hardware cycle offers a direct precedent for the repricing mechanics. In 2021–2022, mining companies bought GPUs and ASICs at peak prices to lock in hashrate. The hash price did not hold; the hardware market re-priced before the firms admitted impairment. Used GPU prices fell 40 to 60 percent within months. The claim “mining is a productive asset” did not change the depreciation math. The same mechanism applies today to AI-focused small-scale operators—start-ups that bought GPU servers at 2023–2024 prices and face a four-year amortization on assets whose resale value is set on a secondary market. If the top hyperscalers slow purchases, that secondary market reprices first, and the impairment announcements follow one quarter later.
Energy is the hidden second ledger. Data center build-outs are driving new power purchase agreements with utilities and, in several documented cases, accelerating nuclear and natural gas capacity planning. This is not optional capex; it is a precondition for the data center to operate at planned utilization. Energy costs enter the income statement as an operating expense, not a capital expense. The effect is that the P&L absorbs the risk that the balance sheet deferred. When utilization drops, power contracts do not pause. The costs continue, and margin pressure appears before the headline numbers turn down. The equivalent in a protocol audit is a smart contract that keeps charging fees to an empty vault.
The demand half of the audit
The counterweight is that AI services are generating real revenue today. Microsoft reported Azure AI services with triple-digit annualized growth in fiscal 2024 and continued strong growth into fiscal 2025. Google Cloud disclosed AI-related backlog in the double-digit billions. AWS described AI run-rate revenue in the billions per quarter. The revenue exists. The problem is the ratio. At the industry level, AI capex is growing 40 to 60 percent per year, while AI service revenue, growing faster from a much smaller base, still covers only a fraction of the new depreciation and operating cost. If revenue growth slows, that ratio worsens every quarter.
The metric to watch is the income-to-capex ratio: trailing twelve-month AI-related operating income divided by quarterly capex run rate. In the housing era, the analogue was household debt service relative to income. For hyperscalers, the ratio is currently below one and falling. A ratio below one is not insolvency. But it is a measured deficit that only revenue acceleration can close.
Compliance note: none of these numbers require special access. They are in XBRL-tagged filings, quarterly investor decks, and the footnotes of the largest public companies on earth. The data pipeline is open. The interpretation requires method. That is why I publish detection logic rather than conclusions.
The training-versus-inference distinction is critical. Training clusters are lumpy, high-risk investments tied to a specific model release. Inference capacity is recurring infrastructure tied to actual product usage. The current spending wave was ordered during the training race of 2023–2024 and is being delivered during the inference build-out of 2025–2026. If no application emerges that requires inference at mass scale, the delivered capacity sits underutilized. Utilization is the closest thing to a chain of custody for the capex claim. If inference utilization drops to single digits, the perceived value collapses regardless of balance-sheet accounting.
Contrarian: The housing analogy is structurally flawed
The warning “twice as fast as housing” carries emotional weight, but the comparison fails on three verifiable dimensions.
Leverage. Housing booms are debt phenomena: new home purchases backed by mortgages and banking credit. When prices fell, household insolvencies and bank failures defined the crisis. AI capex is funded almost entirely from the operating cash flows of four highly profitable technology firms. The burden is a depreciation charge, not a debt repayment. If growth stops, the consequence is an asset markdown and lower forward earnings, not a default cascade across the financial system. The risk concentrates in equity valuations—particularly NVIDIA's multiple—rather than the banking sector balance sheet.
Demand verification latency. A housing boom is verified by family occupancy over decades. AI demand is verified every 90 days in the cloud revenue line. The short information window means the market can correct early, before misallocation reaches systemic scale. The mechanism that keeps a system honest is its block time. In crypto, the block time is seconds. For hyperscaler capex, the block time is one quarter. Housing bubbles featured none of that transparency. Housing data was opaque and financing fragmented. AI capex has none of that opacity. The repricing can begin while construction continues.
Source bias. The editorial source of the “twice as fast” frame is Crypto Briefing, an outlet whose covered sector—digital assets—competes directly with AI infrastructure for the same risk capital. The AI capex boom is, in measurable terms, a rival narrative to crypto adoption. An AI bubble narrative serves the readership's interests. That does not make the warning wrong. It makes the source a data point in the bias assessment, not a neutral oracle. I apply the same scrutiny when AI-native media discuss cryptocurrency. Verify the numbers independently.
There is also a recovery asymmetry. Housing is tied to location; a vacant house in the wrong city cannot be moved. Compute infrastructure is re-deployable. A data center built for inference can serve scientific computing, rendering, or Bitcoin mining if power economics work. The GPU asset may lose 30 to 50 percent of its value, but it does not go to zero. The impairment is real, but it is not a systemic extinction event. What is lost is the accounting expectation that utilization would match the promise. That is a writedown, not a collapse.
Takeaway: The next 90 days
Audit complete. The third-quarter earnings guidance from Microsoft, Alphabet, Meta, and Amazon will set the short-term direction. The signal is not the absolute capex number; it is the first downward revision in forward capex guidance. If that revision appears, the oversupply repricing has begun. If all four firms raise guidance again, the build-out continues, and the housing comparison remains commentary rather than evidence. If the discrepancy between revenue and capex persists for four consecutive quarters, no narrative will shield the correction.
The second signal is NVIDIA's data center revenue growth. A sequential slowdown in recognized revenue—not orders, but recognized revenue—means the contracted backlog is converting to delivered assets, and the market must begin asking about utilization.
The third signal sits in the GPU secondary market. Used ASIC prices collapsed before mining companies admitted impairment in 2022. The same applies to inference capacity. If GPU cloud spot pricing falls below replacement cost, the ledger changes before the narrative does.
The ledger for AI capex runs on the same principles as any market: committed capital, utilization, and realized revenue. The flows are centralized, but they are auditable. The next quarterly filing is the next block. Follow the outflows.