AI Capex Bubble 2026: The Hidden $662B Debt Nobody Reports

Every earnings season now brings a fresh wave of headlines about hyperscaler AI capital expenditure hitting a new record. The “big four” — Amazon, Microsoft, Alphabet, and Meta — are on track to spend roughly $725 billion combined in 2026, a 77% jump from the $410 billion deployed in 2025 (UnboxFuture). That number gets reported constantly. What almost nobody is reporting with the same prominence is a separate figure that may matter more: roughly $662 billion in data center lease commitments that hyperscalers have already signed but not yet begun — obligations that currently sit entirely off balance sheet.

Why the Off-Balance-Sheet Number Changes the Whole Picture

Under GAAP accounting rules governing when a lease “commences,” these signed-but-not-started commitments don’t appear in the capital expenditure figures analysts and investors typically scrutinize when assessing hyperscaler financial health. According to reporting citing Moody’s early-2026 analysis, this shadow liability is larger than the combined on-balance-sheet debt of the same companies (Anomaly Investments).

That detail matters enormously for one specific argument AI infrastructure bulls have relied on: the claim that this buildout is being conservatively self-funded from operating cash flow rather than risky leverage. Once the full picture of committed-but-unrecognized obligations is accounted for, that defense becomes much harder to sustain.

The Debt Is Already Showing Up, Not Just Theoretical

This isn’t a purely hypothetical concern about future liabilities. Big tech companies have already issued more than $100 billion of bonds in 2026 specifically to help fund AI capital expenditure, and investors have responded by demanding record levels of protection against potential defaults through credit default swaps — essentially insurance policies against bond default (IEEE ComSoc).

Individual company examples illustrate the shift toward leverage: Oracle issued an $18 billion bond specifically tied to its data center expansion; CoreWeave secured a $2.6 billion loan alongside a $1.75 billion bond package; and OpenAI and Oracle reportedly entered into a $100 billion vendor financing arrangement (Anomaly Investments). At Amazon specifically, capital expenditure over the trailing twelve months has reached $151 billion — a figure that now exceeds the company’s entire operating cash flow, pushing free cash flow into negative territory.

The Depreciation Assumption Almost No Coverage Questions

Here’s an angle genuinely underexplored across most financial media: the depreciation schedules hyperscalers use for AI hardware assume a five-to-six-year useful life. But given how rapidly GPU generations are turning over and how intensively AI workloads are pushing hardware utilization, critics argue the real economic life of this equipment is closer to two to three years. That gap between assumed and actual depreciation is estimated to understate true asset depletion by roughly $176 billion between 2026 and 2028 alone — a figure that grows as accelerating token consumption pushes hardware utilization beyond the assumptions built into current depreciation schedules (Anomaly Investments).

Layered on top of that is the energy cost curve: running the current roughly 30-gigawatt installed base of AI infrastructure costs approximately $27 billion annually today, but that figure is projected to climb to between $45 and $90 billion per year as capacity scales toward 2029 — and crucially, these are first charges against revenue, not optional or deferrable costs.

The Revenue Gap: Who’s Actually Paying for All This?

The most commonly cited justification for the capex surge is that the pure-play AI vendors — OpenAI, Anthropic, and others — represent a massive and rapidly growing revenue opportunity. The reality is more nuanced. OpenAI’s roughly $20 billion annualized revenue run rate, while genuinely impressive for a company with barely any consumer products three years ago, represents only about 3% of projected 2026 hyperscaler capex. Anthropic’s roughly $9 billion run rate, despite showing 9x year-over-year growth, occupies a similarly small share. The entire cohort of pure-play AI vendors combined — including Cohere, Mistral, Perplexity, and others — likely accounts for less than $35 billion in projected combined 2026 revenue against a hyperscaler capex figure exceeding $700 billion (Futurum Group).

That gap is the crux of the bubble debate: hyperscalers are betting the infrastructure will ultimately serve enterprise adoption and their own AI services broadly, not just third-party AI vendor revenue — but that bet requires enterprise AI monetization to arrive at a scale that, as of mid-2026, remains largely unproven outside of code generation and basic customer service automation.

The Skeptic’s Case, From Inside Goldman Sachs Itself

The most prominent voice of institutional skepticism doesn’t come from an outside critic — it comes from within Goldman Sachs itself. Jim Covello, the bank’s Head of Global Equity Research, has consistently argued the economics of the generative AI transition are fundamentally flawed, stating in mid-2026 that the industry has moved “further away” from justifying the scale of capital expenditure compared to two years prior (UnboxFuture). Covello has specifically flagged circular capital flows between cloud providers and AI startups — where hyperscalers invest in AI companies that then spend that same capital purchasing compute from those same hyperscalers — as a red flag reminiscent of vendor financing patterns seen in the dot-com era.

The valuation comparison to that era is explicit and increasingly common among strategists: US technology and AI equities carry EV/EBITDA multiples near 25x, close to historical extremes and above the telecom valuations that preceded the 2000 dot-com peak. More specifically, capex is currently expanding roughly 46 percentage points faster than revenue growth — a gap that exceeds the 32-point divergence observed during the 2001 telecom excess cycle (Allianz Research). Separately, Bank of America strategists have pointed out that AI stock concentration has reached levels matching prior bubble peaks, with the “AI Big 10” (Nvidia, Microsoft, Alphabet, Amazon, Meta, Apple, Tesla, Broadcom, Micron, and AMD) now making up 41% of the S&P 500 — comparable to the concentration of tech and telecom stocks during the actual dot-com bubble (Yahoo Finance).

The Bull Case Isn’t Naive Either

It would be inaccurate to frame this purely as informed skeptics versus blind enthusiasm. Goldman Sachs’ own broader research (distinct from Covello’s individual view) models roughly $7.6 trillion in cumulative AI capital expenditure between 2026 and 2031, built on the expectation that token consumption will increase 24-fold by 2030, driven largely by enterprise AI agents becoming embedded in production workflows rather than remaining experimental (Sesame Disk / Goldman commentary). Microsoft has disclosed an $80 billion backlog of Azure orders it currently cannot fulfill due to power constraints — genuine evidence that demand, at least for existing capacity, is outpacing even the current aggressive build-out pace (Futurum Group).

Leverage levels also remain more conservative than headlines suggest in absolute terms: the top five US capex providers reported a combined $385 billion in debt at the end of 2025, with leverage ratios still roughly 20% below the “high spender” cohort from the 2000 dot-com peak, according to Allianz Research analysis — meaning rising debt levels are a trend worth monitoring closely, not yet an acute crisis.

What Happens If the Bubble Skeptics Are Right

Historical infrastructure cycles offer a specific and somewhat counterintuitive lesson: the investors who fund the initial frenzied build-out phase rarely capture the long-term rewards. If the AI capex cycle follows the pattern of the 1998-2001 fiber optic buildout, hyperscalers may eventually be forced to write down the value of data centers and GPUs purchased at today’s prices and utilization assumptions. But that collapse in computing costs, paradoxically, could pave the way for a new generation of leaner, genuinely profitable software companies to build on top of the resulting cheap, overbuilt infrastructure — much as fiber-optic overbuild eventually enabled the 2000s streaming and cloud computing boom, even after the original telecom investors were wiped out.

What This Means for Investors and Businesses

For equity investors, the practical signal to watch isn’t the headline capex number — it’s the widening gap between capex growth and revenue growth, and whether that gap begins narrowing through 2027 as enterprise adoption either accelerates or disappoints. For businesses evaluating AI vendor relationships, the circular-financing pattern flagged by Covello is worth diligence: understanding whether an AI vendor’s revenue depends partly on capital originally supplied by the same hyperscaler providing its compute is a legitimate red flag for assessing that vendor’s underlying financial independence. For fixed-income investors, the rising credit default swap pricing on hyperscaler-linked debt is itself a market signal worth tracking as an early indicator of shifting sentiment, independent of equity price action.

The Bottom Line

The AI infrastructure buildout genuinely is the largest corporate capital expenditure cycle in recorded history, and it’s happening for real, defensible reasons tied to a genuine technology shift. But the debate over whether it constitutes a bubble isn’t really about whether AI technology is useful — it’s about whether the timing of returns can keep pace with public equity markets’ patience, and whether the $662 billion in off-balance-sheet lease commitments, aggressive depreciation assumptions, and circular vendor financing arrangements represent manageable financial engineering or the early architecture of a genuinely serious correction. Both cases have real evidence behind them. What’s clear is that the headline capex figure everyone quotes is no longer the most important number in this story.