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.

Why the AI Revolution Is Breaking the Link Between Output and Labor Income

Artificial intelligence is transforming the modern workplace at a breathtaking pace. Generative AI tools are drafting legal briefs, diagnosing medical images, writing software code, and managing supply chains with superhuman efficiency. Yet a landmark report from the International Labour Organization, released on June 15, 2026, reveals a troubling disconnect: while global labor productivity has accelerated to a 3.2% annual clip, real median wages in advanced economies have risen a mere 0.8% (ILO World Employment and Social Outlook, June 2026). The AI boom, it appears, is delivering a productivity miracle that primarily rewards capital owners and the highest‑skilled technologists, leaving the typical worker behind.

The Labour Share in Freefall

The ILO’s most alarming finding is the labor share decline. The labor income share—the slice of national income that goes to workers in the form of wages, salaries, and benefits—has fallen to a historic low of 51% globally, down from 54% in 2004. The decline is sharpest in the United States and Northern Europe, where AI adoption is most advanced. In the US, the labor share has dropped to 56.5%, a level not seen since the Gilded Age. The ILO attributes 40% of this decline since 2020 to technological displacement, with AI being the primary driver.

The mechanism is subtle but powerful. AI automates cognitive routine tasks, not just physical ones. When a financial analyst’s report that once took five days can be produced by an AI in five minutes, the marginal value of that analyst’s time plummets. The analyst may keep her job, but her bargaining power for raises evaporates. Meanwhile, the firm’s profits surge because output per worker rises dramatically. The ILO found that in the top 500 AI‑adopting firms globally, operating margins expanded by an average of 4.8 percentage points between 2022 and 2026, but the wage‑to‑revenue ratio contracted by 2.3 points (McKinsey Global Institute, “The State of AI in 2026”).

Technology Unemployment 2.0

The term “technological unemployment” has moved from academic journals to mainstream policy debates. The ILO estimates that while AI will create 50 million net new jobs by 2030, it will displace or fundamentally transform 400 million roles. The occupations most exposed are those that involve information processing, pattern recognition, and language generation: paralegals, accountants, call‑center agents, radiologists, and software developers themselves. In a striking case, a major global bank announced in April 2026 that it had reduced its compliance department headcount by 35% while simultaneously cutting error rates, replacing human reviewers with a combination of natural‑language processing and robotic process automation (Financial Times).

What makes this wave different from previous automation cycles is the speed and the educational threshold. Historically, automation hit blue‑collar manufacturing; this time, it is hitting white‑collar, university‑educated professionals. A paper from the National Bureau of Economic Research circulated in May 2026 shows that for the first time, workers with a bachelor’s degree are seeing a negative return to experience in AI‑exposed roles; their earnings trajectory is flattening relative to peers in less automatable trades such as plumbing or elderly care (NBER Working Paper 31050).

The Gig Economy Entrenchment

AI is also accelerating the fissuring of the traditional employment relationship. Platforms that match freelancers with tasks, from graphic design to legal research, are increasingly using AI to manage work allocation, evaluate performance, and even set piece‑rate prices. The ILO found that 38% of the global workforce is now engaged in some form of non‑standard employment, up from 34% in 2019. While this provides flexibility, it strips away the training, benefits, and career progression that traditional employment offered. Workers in these arrangements have seen their real incomes stagnate or fall, as algorithmic management squeezes task‑by‑task compensation.

Policy Responses: From AI Taxes to Universal Basic Capital

Governments and international bodies are scrambling to rewrite the social contract. The European Parliament’s Committee on Employment is debating an AI training levy that would require firms deploying automation to contribute 1% of payroll to a reskilling fund. The idea, inspired by Singapore’s SkillsFuture credit, has drawn support from trade unions and even some tech leaders. Sam Altman’s concept of a “universal basic capital”—an ownership stake in the AI‑driven economy distributed to all citizens—has moved from concept to pilot in Finland and Kenya, where blockchain‑based digital trusts allocate shares in a portfolio of AI‑intensive public companies to citizens (World Economic Forum, “AI Governance in Practice”).

The OECD has issued new guidelines urging members to strengthen collective bargaining rights in the digital economy and to enforce antitrust laws that prevent algorithmic wage‑fixing (OECD Employment Outlook 2026). In the United States, the Federal Trade Commission has opened investigations into several large HR‑tech platforms over allegations that their “optimal wage” algorithms constitute illegal coordination among employers.

What Workers and Employers Can Do

For individuals, the advice is increasingly nuanced. The ILO recommends “AI literacy” not as a coding skill but as the ability to supervise, critique, and collaborate with AI outputs. Skills in emotional intelligence, complex negotiation, and ethical judgment are commanding a premium. Employers, on the other hand, are facing a talent paradox: they need workers who can manage AI, but if they hollow out the middle tier of employees, they lose the pipeline for future managers. Firms that invest in robust apprenticeship programs and internal mobility, such as Bosch and Siemens, are finding that they can deploy AI without triggering the toxic wage compression that hurts morale and long‑term innovation (Harvard Business Review, “The Smart Way to Automate”).

The AI productivity boom is real, but the ILO’s message is stark: without deliberate policy intervention, the link between rising output and rising living standards will remain broken. The labor share decline is not an iron law of technology; it is a consequence of institutional choices. Whether nations choose to tax, redistribute, or upskill will determine whether the 2020s are remembered as the decade of shared prosperity or of deepening divide.

China’s domestic chip ecosystem is accelerating even as US export controls tighten. With tungsten up 557% and Nvidia’s China share halving, we map the permanent splitting of the global semiconductor supply chain.The global semiconductor supply chain is bifurcating. This statement was contested in 2023, hedged in 2024, and is now — as of 2026 — treated as a structural baseline by supply chain strategists, chipmakers, and government planners on both sides of the Pacific. The question has shifted from whether the split will happen to how deep and permanent it will become.

The evidence is visible in multiple datasets simultaneously. Nvidia, which once commanded over 90% of the Chinese AI chip market, had seen that share decline to approximately 50% by early 2026 — not because US export controls had successfully denied China access to capable chips, but because the combination of tariffs, “buy local” mandates, and regulatory uncertainty had accelerated Chinese enterprises’ migration to domestic alternatives. Meanwhile, China’s semiconductor output surged 87% year-on-year in May 2026, underscoring that domestic production capacity was advancing at a pace that few had forecast five years ago.

The Tungsten Shock: A Materials Leverage Beijing Chose to Use

In February 2026, China added tungsten to its export control list as trade tensions with the United States escalated. The consequence was rapid and severe. Tungsten prices rose 557% in just over a year — outperforming gains in gold, copper, and oil by a wide margin. Chinese exports of restricted tungsten products fell approximately 40% in 2025. The strategic logic was precise: China controls roughly 79% of global tungsten mine production, and tungsten’s exceptionally high melting point and density make it an essential input for chipmaking — both in chips themselves and in multiple fabrication processes at advanced nodes.

The move demonstrated that materials leverage extends far beyond rare earths. For semiconductor supply chains already under AI-driven demand stress, the tungsten shock added a new category of critical bottleneck that western efforts to build alternative supply chains cannot resolve in the near term.

Nvidia’s Paradox: Export Controls and the H200 Restart

The Nvidia-China relationship in 2026 illustrates the inherent contradiction of export controls applied to commercially motivated technology companies. After a roughly ten-month freeze on advanced chip exports to China — during which Nvidia absorbed a $5.5 billion charge tied to stranded inventory — a December arrangement allowed H200 sales to approved Chinese customers, with the US government taking a 25% cut of revenues. The arrangement normalised commerce while creating a fiscal mechanism for the US government.

Chinese tech firms collectively placed orders for more than two million H200 units for 2026 delivery — a volume that simultaneously demonstrates unmet demand and the limits of export control effectiveness. Where legal channels are closed, demand finds other pathways: a DOJ indictment unsealed in 2026 detailed a scheme involving approximately $2.5 billion in Supermicro servers containing restricted Nvidia GPUs being smuggled to Chinese buyers.

China’s Domestic Progress: Real but Incomplete

China’s semiconductor self-sufficiency ambitions are advancing, but the trajectory is uneven across subsectors. SMIC and Hua Hong have made genuine progress at mature nodes. Equipment vendors Naura and AMEC are gaining market share globally. The country’s AI chip domestic alternatives — while not yet matching Nvidia’s leading-edge capability — are advancing at an accelerating pace under the pressure of necessity.

The critical constraint remains high-bandwidth memory. CXMT, China’s domestic HBM producer, is targeting viable HBM3 yields in 2026 and HBM3E by 2027. If those milestones are achieved on schedule, Nvidia’s current China advantage — which exists precisely because China’s domestic HBM production remains constrained — will narrow materially. The competitive window is real but finite.

The Strategic Implication: Permanent Bifurcation as Business Baseline

For supply chain strategists, the most consequential shift is not any individual export control or price spike — it is the recognition that the global semiconductor supply chain’s bifurcation is permanent. Semiconductor leaders navigating this environment most effectively are treating the US-China bifurcation as a structural feature of the landscape, not a temporary disruption awaiting resolution.

This means conducting detailed audits of supplier dependencies, stress-testing revenue models against scenarios where China access is restricted or structurally changed, and tracking China’s domestic chip progress as a competitive variable rather than a geopolitical curiosity. Revenue projections that assume stable China market access now carry geopolitical risk that most financial models have not historically priced.

The age of a single, integrated global semiconductor supply chain is over. The question is how many chains will replace it, and at what cost.