AI Capex: The Next Source of Systemic Financial Risk?

In most people's mental model, capital expenditure (Capex) is the most benign line item on a corporate financial statement: buy equipment, build factories, cash goes out, assets remain, and the income statement is unaffected in the current period. Over the past two years, this logic has been amplified by the AI industry to a historically unprecedented scale — the five major hyperscalers (Alphabet, Amazon, Meta, Microsoft, Oracle) collectively invested over $1 trillion building AI infrastructure in 2025–2026, while income statements continued to look healthy on the surface. The prevailing market narrative: this is industrial upgrading, the investment of winners, not a risk.

Then, on June 28, 2026, the Bank for International Settlements (BIS) — the "central bank of central banks" — published its Annual Economic Report and formally placed this narrative on the risk register. The report listed the sustainability of AI investment alongside persistent inflation risks, intensifying financial fragilities, and deteriorating fiscal conditions as four major pressure points in the current global economy. The language was direct and rare: "a disappointment in returns could trigger a sudden withdrawal of financing, turning a capex boom into a protracted investment bust and spilling over into financial conditions."

This was not the first alarm. The Bank of England had already noted in October 2025 that the cyclically adjusted price-to-earnings ratio (CAPE) had fallen to a 25-year low, "comparable to the peak of the dot-com bubble," explicitly warning that "the risk of a sharp market correction has risen." But the BIS intervention was different — it directly named the financing structure and financial transmission mechanisms, not merely valuation concerns.

This is not an analyst debate. This is the official position of regulators.

The central question this analysis addresses: through exactly what mechanism does AI Capex evolve into systemic financial risk? To what extent is this risk real, and where has it been overstated? Answering this requires unpacking four dimensions: financial mechanics, financing structures, historical precedents, and transmission pathways.

Understanding the "Time Bomb": The Temporal Mismatch Across Three Financial Statements

Any discussion of the financial risks of AI Capex must begin with a fundamental accounting mechanism: capital expenditure (Capex) and depreciation charges are separated on the timeline — and in the case of hyperscalers, this separation has been amplified to a historically unprecedented scale.

When a company purchases servers or builds a data center, the outlay is immediately reflected in the cash flow statement under "investing activities," but does not enter the income statement in the current period. Equipment is recorded on the balance sheet under "Construction in Progress (CIP)" — and only when the project is completed and placed into service does it transfer to "Property, Plant & Equipment (PP&E)" and begin depreciating into the income statement.

This produces an accounting phenomenon that is highly misleading to external observers: the larger the Capex, the worse the cash flow looks — but the current-period income statement may show no anomaly whatsoever. Only when these assets gradually "transfer to fixed assets" and begin amortization will the income statement feel pressure — and by that point, the Capex may have been paid out two or three years prior.

Now consider these mechanics against real numbers. The five major hyperscalers (Alphabet, Amazon, Meta, Microsoft, Oracle) collectively deployed over $1 trillion in AI-related Capex in 2025–2026. Balance sheet changes already reflect the accumulation of this "bomb": Alphabet's assets not yet in service jumped from $50.6 billion to $78.6 billion in 2025, a year-on-year increase of approximately 55%; Meta's full-year capital expenditures (including finance leases) surged from $39.2 billion to $72.2 billion, an increase of 84%; Oracle's FY2026 full-year Capex reached approximately $55.6 billion, up about 162% year-on-year and more than six times the $6.9 billion of two years prior. These assets — still sitting on balance sheets, not yet depreciating — will concentrate their "transfer to fixed assets" in 2026–2028, at which point depreciation charges will grow explosively.

The Wall Street Journal, citing analyst forecasts, projects that Alphabet alone will see depreciation jump from $21.1 billion in 2025 to approximately $78 billion in 2029, with its share of revenue rising from 5% to 11%. Morgan Stanley forecasts that four major companies will accumulate over $520 billion in depreciation over three years. According to analyst estimates, the five hyperscalers will cumulatively place approximately $2 trillion in AI-related assets onto their balance sheets by 2030; assuming a 5–6 year useful life, the annual depreciation burden at peak will approach or exceed the combined current net income of these five companies — not a profit collapse, but structural margin compression before AI revenues scale substantially is now nearly certain.

One critical accounting nuance: depreciation is a non-cash expense. It affects GAAP earnings but not operating cash flow (which adds it back). Depreciation alone therefore does not trigger a liquidity crisis — its impact is on valuation. When margins compress and EPS is revised downward, stock prices reprice. That is the first transmission channel into the financial system.

But depreciation-driven valuation compression — making stock prices look ugly — is insufficient on its own to trigger systemic financial contagion. What has genuinely alarmed the BIS is that the financing model for this Capex boom has fundamentally changed. That is the critical inflection point where an industry cycle becomes a financial stability issue.

From Internal Cash Flows to Debt-Driven: A Historic Business Model Transformation

For the past decade, hyperscalers represented the world's most archetypal "asset-light, high cash return" business model: high operating margins, low Capex, abundant free cash flow. This was the fundamental reason they commanded premium technology valuation multiples.

This model underwent a fundamental transformation between 2024 and 2026.

In 2023, Capex for the five major hyperscalers consumed approximately 40% of operating cash flow — still manageable. PIMCO projects that by 2026–2027 this ratio will climb to approximately 94%, essentially reinvesting nearly all operating cash flow into AI infrastructure. Self-generated funds can no longer cover the continuing expansion, and the companies have systematically turned to bond markets for financing.

BIS research explicitly documented this transformation: hyperscalers' financing model shifted from operating cash flows to debt financing, with corporate bond issuance topping $100 billion in 2025, predominantly long-duration debt with maturities exceeding five years, locking in funding for multi-year build-outs. BofA Securities estimates that actual 2025 issuance reached approximately $121 billion — 4.3 times the annual average of approximately $28 billion from 2020 to 2024.

According to JPMorgan credit research, AI-related debt has risen to approximately $1.2 trillion, representing about 14% of its US Liquid Index (JULI) — surpassing banking to become the single largest sector in the investment-grade credit market. Companies competing in the AI race will need to issue as much as $1.5 trillion in investment-grade bonds over the next five years.

This is one of the fastest corporate bond issuance waves in modern capital market history, comparable in scale to 19th-century railroad construction financing. Unlike railroads, however, these debts are backed not by physical land usable as collateral — only GPU servers, software contracts, and expectations of future AI revenues.

"Shadow Borrowing" in Off-Balance Sheet Financing: What Regulators Actually Fear

If bond market financing represents the "visible" risk, then the "Shadow Borrowing" specifically named by BIS in its March 2026 Quarterly Review is what is genuinely causing alarm in regulatory circles.

The mechanics operate as follows:

• Hyperscalers want to build data centers, but prefer not to put all debt on their balance sheets (which would affect leverage ratios and credit ratings)
• They establish or participate in a Special Purpose Vehicle (SPV) or joint venture (JV); the SPV acquires or develops the data center assets
• The SPV raises financing through Private Credit — debt held by private credit funds, insurance companies, and similar institutions
• The hyperscaler signs long-term operating leases or offtake agreements with the SPV, committing to multi-year payments
• The SPV's debt is collateralized by these lease cash flows, typically with some form of hyperscaler guarantee

BIS characterizes these arrangements as "shadow borrowing": economically equivalent to debt, but largely existing off corporate balance sheets. Moody's estimates that the five hyperscalers have signed but not yet recognized on-balance-sheet data center leases totaling approximately $662 billion — equivalent to 113% of their combined adjusted debt.

BIS explicitly identified the systemic risk transmission pathway for these arrangements: Private credit funds → Insurance companies (holding private credit shares) → Banks (providing credit facilities to SPVs, or indirectly exposed through insurance company channels) → Retail investors (reached through BDC channels). A risk transmission chain running from hyperscalers to ordinary savers is quietly forming through this structure.

The numbers confirm this concern. Private credit lending to SaaS software companies grew from nearly $8 billion in 2015 to over $500 billion by end-2025, representing 19% of total direct loans. Business Development Companies (BDCs) account for one-fifth of all direct lending in the United States, with over 15% of their loans issued to SaaS companies. These loans are typically based on future AI revenue projections rather than hard asset collateral. When a BDC issues loans backed by an AI software company's "future recurring revenue" and then sells BDC shares to pension funds or retail investors, the fragility of the entire chain has quietly penetrated the core of the traditional financial system.

Circular Financing: A Self-Inflating Ecosystem

If "shadow borrowing" is the structural vulnerability, then the Circular Financing specifically called out by the BIS Annual Report is the most thought-provoking phenomenon of this investment boom.

The operating model: chipmakers and hyperscalers invest in AI labs or NeoClouds → AI labs commit to multi-year purchases of chips or compute from those shareholders → shareholders record AI revenue growth; AI labs record funding received and "contracted" compute orders on their books.

BIS explicitly stated: "the terms of such deals are typically poorly disclosed, with risks of the same asset being pledged multiple times."

Take Anthropic as a case study. The company has received massive strategic investments from both Amazon and Google — Amazon has cumulatively injected approximately $8 billion historically, with an additional commitment of up to $25 billion announced in April 2026; Google holds approximately 14% of Anthropic's equity, with an additional commitment of up to $40 billion announced in April 2026. Simultaneously, Anthropic has committed to purchasing over $100 billion in compute from AWS over ten years, while Google has committed to providing compute resources to Anthropic. Investor and payer roles are highly overlapping — Amazon and Google are simultaneously Anthropic's largest shareholders and its largest revenue sources. This compute spending will become Amazon's and Google's "AI Cloud revenue," while appearing on Anthropic's balance sheet as long-term lease liabilities, forming a textbook circular financial structure.

From a three-statement perspective, this structure creates an analytically thorny problem: a significant portion of hyperscalers' "AI Cloud revenue" derives from demand they themselves have "pre-installed" through equity investments. This is not an arm's-length market transaction — it is a closed loop where capital accounts circulate internally within the value chain, artificially inflating the nominal revenue scale of the entire AI industry.

Sequoia partner David Cahn's analysis quantified this gap: using a 50% gross margin and 2x total cost of ownership multiplier, the AI industry needs to generate approximately $600 billion in real end-user revenue annually to cover infrastructure capital costs. Note that $600 billion was his mid-2024 estimate; applying the same methodology to 2026's higher Nvidia data center revenue run-rate of approximately $270 billion annually, the threshold has risen to approximately $1 trillion. Meanwhile, AI revenue attributable to genuine end-user consumption in 2026 is estimated by multiple sources at approximately $50–$150 billion — the gap remains multiple-fold.

BIS Asia-Pacific representative Zhang Tao offered a cautionary observation: "If the market has any sort of correction, the interconnectedness of the financial system and interplay of vulnerabilities could mean the speed of a correction could be much faster than previous banking crisis episodes." Traditional banking crises have deposit insurance and lender-of-last-resort mechanisms as firewalls; private credit funds, hedge funds, and insurance companies — these non-bank financial intermediaries — lack equivalent institutional buffers.

Historical Precedents: Similarities and Structural Differences

Every era's bubble claims to be the exception. The BIS Annual Report draws on four historical analogies: the canal mania of the 1830s, British railway mania in the 1840s, late-1920s electrification, and the dot-com bubble of the late 1990s.

These four episodes share a common trajectory: genuine technological breakthrough → capital inflows exceeding commercial return capacity → overbuilding → demand falling short of expectations → asset impairment → economic recession. And every time, the most enthusiastic participants believed this time was different.

The parallel with the telecom fiber optic bubble (1999–2002) is most intuitive: oversupplied infrastructure (85% of fiber was never used), debt-driven expansion, depreciation hitting income statements, ultimately large-scale asset impairments.

But the AI story has several genuinely structural differences that invalidate simple analogies:

Difference 1: Demand visibility. The "future bandwidth demand" of the dot-com era was primarily analyst projection; the current AI cloud Backlog consists of genuine contracted commitments — Microsoft, Google, and Amazon all carry hundreds of billions in signed remaining performance obligations. The question of how much represents artificial demand created by circular financing remains unverifiable.

Difference 2: Player buffer thickness. WorldCom and Global Crossing were highly leveraged operators — once cash flow broke, bankruptcy followed. Alphabet's 2025 net income of approximately $132 billion means that even as depreciation increases substantially, no existential crisis is triggered. But analysts forecast Alphabet's depreciation will increase by approximately $57 billion over four years, which will materially compress margins and drive valuation repricing. Hyperscalers will not go bankrupt, but that does not mean share prices will not come under pressure.

Difference 3: AI's self-reinforcing demand logic is stronger. AI capability improvements endogenously create new use cases — something bandwidth itself could not do. ChatGPT's success created the entire GenAI application layer, which in turn requires more compute. This positive feedback loop has few complete analogues in historical precedents.

Yet PIMCO stated in its May 2026 credit analysis report: "AI is in the midst of a capex boom with genuine risks: uncertain monetization, potential overbuild, shortening asset lives, and growing reliance on debt. But for now, it is a more disciplined and far more financeable cycle than the late-1990s telecom boom." More disciplined does not mean without risk.

Transmission Pathways: How a Bubble Becomes a Systemic Event

The core distinction between an industry bubble and systemic financial risk lies in whether risk can spread outward through financial channels. The BIS 2026 Annual Report provides a complete description of these transmission pathways.

Pathway 1: Equity Markets → Wealth Effect → Consumption Contraction

US equities represent approximately 73% of the global MSCI index, while AI-related stocks have significantly increased their share of S&P 500 market capitalization. US household equity holdings relative to total wealth have increased substantially compared to a decade ago. This means a major AI-led stock market correction would directly compress consumption through the wealth effect — with far greater macroeconomic impact than the dot-com era, when household equity wealth ratios were far lower.

Pathway 2: Fixed Income Markets → Credit Tightening → Corporate Financing Cost Jump

AI-related debt already represents approximately 14% of JPMorgan's US investment-grade index (JULI), surpassing banking as the largest single sector, with concentration continuing to rise in 2026. Current investment-grade credit spreads of approximately 77 basis points are at their lowest level since 1998 — itself a risk signal, as market optimism about AI prospects is already fully priced in.

Should AI commercialization disappoint, three categories of debt would be first in line: investment-grade bonds issued by hyperscalers themselves; bonds from EPC engineering contractors (data center builders) with relatively thin balance sheets and high dependence on hyperscaler orders; and high-yield bonds from NeoClouds. Credit spread repricing would rapidly transmit to the entire corporate credit market, raising borrowing costs for all issuers.

Pathway 3: Private Credit → Non-Bank Financial System → Liquidity Freeze

This is BIS's most concerning transmission pathway, and the one most underestimated by mainstream discussion. Direct lending funds' exposure to AI and IT sectors has reached 15% of portfolios, with these loans based on future AI revenue projections rather than hard asset collateral.

When private credit comes under stress, its linkages to banks begin to surface: banks provide credit facilities to SPVs, insurance companies hold private credit shares, BDCs are absorbing growing retail investor capital. These non-bank intermediaries lack the regulatory safety net of commercial banks. Should large-scale redemption pressure emerge, an event similar to the March 2020 "Dash for Cash" liquidity freeze could materialize.

The FSB's May 2026 Private Credit Vulnerability Report identified an additional hidden risk: the private credit market's extreme opacity means regulators' data on risk scale lags significantly behind reality. This means that when problems surface, regulators may be equally unprepared.

Pathway 4: Depreciation Shock and Accounting Opacity → The "Delayed Detonator" of Valuation Repricing

FASB requires public companies to make more granular classified disclosures of expenses in footnotes beginning with fiscal year 2027, meaning investors will only be able to systematically understand how depreciation is distributed across hyperscaler expense categories beginning in 2028. The period of most concentrated depreciation pressure (2027–2029) overlaps substantially with the period of lowest information transparency. This information asymmetry will itself create "surprise effects" in market expectations for earnings downgrades, amplifying volatility during valuation corrections.

Current Market Pricing of Risk: Contradictions and Signals

Notably, when BIS issued systemic risk warnings, markets did not respond with a crash — instead displaying an internally differentiated pricing logic.

Credit markets are pricing risk ahead of equity markets. Hyperscaler bonds have already widened relative to the broader IG index, with 10–30 year yield curves steepening noticeably — bond markets demanding more duration premium as they recognize that repayment of this ultra-long-duration debt depends on long-term certainty in AI commercialization. Meanwhile, equity markets are still debating "whether these AI Capex investments are worth it," with overall valuations remaining at historical highs.

Equity markets show internal divergence. The market's judgment is no longer simply about revenue growth rates — it now probes whether Capex has a clear closed-loop monetization pathway. After Q1 2026 earnings season, Alphabet — backed by direct AI revenue from Google Cloud — saw its stock reach record highs; Meta raised its full-year Capex guidance to $125–145 billion, and its stock fell over 7% after earnings — despite its revenue growth of +33% being the fastest among major hyperscalers that quarter. The same revenue acceleration, radically different market pricing. The core of the divergence: Meta's AI Capex must monetize indirectly through advertising efficiency improvements, a path that is opaque, with free cash flow already showing compression.

Private credit markets are showing early stress signals. SaaS software company stocks fell approximately 30% cumulatively between October 2025 and February 2026; BDC stocks fell approximately 10%, with discounts to net asset value widening; BDCs with high SaaS exposure underperformed peers by approximately 5 percentage points. These signals are viewed by JPMorgan as "not yet systemic," but the BIS views them as leading indicators warranting vigilance.

The most critical contradiction: corporate bond spreads remain near their lowest levels since 1998, even as hyperscalers are issuing bonds at record pace while the AI commercialization gap remains multiple times the required threshold. This distortion in price signals may be the source of outsized amplitude when the next repricing arrives.

This Is Not 2008 — But May Be a New Form of Slowly Accumulating Systemic Risk

The 2008 financial crisis exploded through mortgage loans, CDOs, and a highly leveraged banking system. AI Capex risk transmission is closer to a chronic accumulation followed by episodic shocks — not a single-point ignition systemic collapse.

The most reasonable scenario framework:

Base Case (probability ~40–50%): Depreciation pressure releases as expected in 2027–2028; hyperscaler margins compress from ~35% to ~25–28%. Stock prices undergo P/E multiple compression-style repricing — not profit collapse, but absolute profits remain positive. AI commercialization revenue growth lags depreciation growth, but the gap gradually narrows. Credit markets see localized pressure, with SaaS exposure in private credit triggering some BDC asset impairments, but not escalating to systemic credit freeze. Macro impact limited — closer to a high-valuation correction concentrated in technology stocks.

Tail Risk Scenario (probability ~15–20%): A critical node in circular financing breaks (e.g., a major AI lab cannot honor compute purchase commitments to hyperscalers), triggering large Backlog write-downs and revenue downgrades. AI vendors (EPC contractors, NeoClouds) face serial credit rating downgrades; bond markets see rapid credit spread contagion. Private credit experiences large-scale redemptions; BDC forced asset liquidation transmits liquidity freeze to the banking system. Equity wealth effects drive meaningful consumption contraction, compounding existing inflationary pressure from the energy shock of the 2026 Strait of Hormuz closure referenced in the BIS report. The result: a mild US-European recession. Not 2008 in magnitude — but a modern replication of the 2000–2002 NASDAQ crash combined with a mild recession.

Optimistic Scenario (probability ~30–40%): Agentic AI applications commercialize at scale in 2026–2027; enterprise adoption exceeds expectations; genuine AI end-user revenue grows faster than depreciation. The Capex/FCF ratio begins to decline; bond market credit concerns fade; spreads recover. Hyperscalers complete a valuation repricing from "technology companies" to "high-growth utilities" — multiples compress moderately, but the fundamental narrative shifts to "AI infrastructure is the next internet foundation layer," a framework the market accepts.

Regulatory Response and Its Limitations

Facing BIS's explicit warnings, the actual regulatory toolkit is quite limited.

FASB's new depreciation disclosure rules are an important step toward improving information transparency — but they address information asymmetry, not overinvestment itself. Even if investors fully understand depreciation in 2028, the Capex already spent and the lease contracts already signed cannot be undone.

The regulatory gap in private credit is a greater challenge. BIS reports that banks, through providing credit facilities to SPVs, have become deeply intertwined with the private credit system in opaque ways. Yet no single regulatory authority can conduct global surveillance of risk exposure spanning the bank/non-bank boundary — and the FSB report directly acknowledges this.

Federal Reserve Bank of Chicago data shows that banks' outstanding loan balances to AI-adjacent industries average approximately 0.8% of total bank assets — manageable at face value. Including committed but undrawn credit facilities, this ratio jumps from 9% of Tier 1 capital (outstanding balances) to 25% (including committed facilities) — in a stress scenario, borrowers may draw on their lines before becoming delinquent, with actual losses far exceeding the apparent 0.8% surface exposure. And this is only the "known, on-balance-sheet" direct exposure. Indirect exposure transmitted through SPV credit facilities, insurance company holdings, and BDC retail channels — no institution currently has complete data.

BIS's prescription: regulators should prioritize off-balance-sheet disclosure requirements for AI financing structures, particularly targeting multiple-pledging and circular financing arrangements. But regulatory institutions are typically built after risk has actually accumulated. History tells us that when regulation finally catches up, it is usually after the inflection point has already arrived.

Conclusion: Clarity Within the Paradox

This discussion contains an inherent paradox: if AI commercialization delivers as promised, this $1 trillion in Capex will prove to be the greatest infrastructure investment since 19th-century railroads and electrification, and the BIS warning will become a footnote of excessive caution; if commercialization falls short of expectations, this will be an unprecedented overbuilding — a triple wave of depreciation pressure, debt pressure, and private credit stress compounding to trigger a profound global financial adjustment.

Before the verdict arrives, several things are certain:

First, the risks are real — only the timing and magnitude are uncertain. BIS is not predicting a collapse; it is identifying vulnerability nodes with systemic contagion potential. Those nodes genuinely exist.

Second, "profits still look fine" is the most dangerous judgment in the current environment. Depreciation pressure is lagging, off-balance-sheet obligations are invisible, private credit exposure is opaque — these three layers of information concealment make surface income statement health highly misleading.

Third, this will not arrive in the form of a traditional financial crisis. Hyperscalers will not go bankrupt, because they have real operating profits. The risk is more likely to manifest as years-long valuation repricing and margin compression, alongside localized liquidity stress breaking out in scattered fashion across the private credit and non-bank financial system.

Fourth, differentiation is already occurring. Companies with clear AI monetization pathways (Alphabet via Google Cloud, AWS via direct compute revenue) and those with questionable paths (Meta's AI Capex still monetizing primarily through advertising efficiency; Oracle with $638 billion RPO backing but extreme timing risk as aggressive Capex precedes revenue realization) will face divergent fates in this adjustment. Credit markets have begun pricing this differentiation; equity markets are following.

The BIS warning is not a recommendation to exit AI — it is a systemic reminder about financing structures, information transparency, and cross-market risk transmission. That is precisely what the central banking community exists to do: speak unwelcome truths during periods of exuberance.