Introduction: The AI Capex Surge — Why This Moment Is Unlike Any Other

There are moments in economic history when the sheer scale of capital commitment by private firms begins to rival that of sovereign infrastructure programs. The railroad booms of the nineteenth century were one such moment. The fiber-optic explosion of the late 1990s was another. We are living through a third, and it is faster, more concentrated, and more debt-financed than its predecessors.

The numbers are staggering in their own right, but they become genuinely sobering when placed against historical context. According to first-quarter 2026 earnings compiled by the Financial Times, Google, Amazon, Microsoft, and Meta collectively plan to spend $725 billion on capital expenditure in 2026, up 77% from last year’s record $410 billion.[1] To provide a sense of proportion: in all of 2024, the combined free cash flow of these four companies was approximately $200 billion, which itself was down from $237 billion in 2024. That is to say, the capital these companies intend to deploy in a single year now substantially exceeds the cash they once generated in the same period.[2]

What makes this cycle categorically different from those of the past is the financial strength of the builders. When Global Crossing and WorldCom laid fiber-optic cable across the globe in 1999 and 2000, they were doing so with highly leveraged balance sheets that had no margin for error once demand disappointed. Today’s hyperscalers, by contrast, enter this investment cycle with net-cash positions, investment-grade ratings anchored at the top of the credit spectrum, and cash flows that, even diminished by capex, represent a buffer most industrial corporations can barely imagine. UBS has raised its 2026 US investment-grade issuance forecast to $1.8 trillion from $1.725 trillion, with technology supply estimates rising to $360 billion from $300 billion, citing a sharp rise in hyperscaler capital expenditure.[3]

Yet credit markets are not entirely sanguine. The transition from entirely self-funded capital expenditure to systematic investment-grade bond issuance represents a structural shift with lasting consequences. Duration in the IG index is lengthening as hyperscalers issue thirty-year and even one-hundred-year paper. Sector weightings are shifting. And as Oracle’s experience in late 2025 demonstrated, when a debt-funded AI strategy is perceived as concentrated or insufficiently diversified, the market can reprice credit risk with brutal speed — even for an issuer carrying investment-grade ratings.

This paper names the overall phenomenon the Capex Surge — not simply to describe the scale of spending, but to capture the surge’s self-reinforcing, competitive logic. These companies cannot stop spending because their rivals will not stop spending. The infrastructure being built today will determine who controls the compute capacity that runs the AI economy for the next decade. And the debt being issued to finance that infrastructure is now a permanent fixture of the investment-grade credit landscape.

Why Capex Surge? Because the word surge is more precise than boom or bubble. A surge has directionality, momentum, and a point of potential exhaustion. The surge is not yet over. But the question of when and how it moderates — and what happens to the credit markets it has reshaped when it does — is among the most consequential analytical questions facing fixed-income investors today.

Section 1: Foundational Definitions — Setting the Analytical Vocabulary

1.1  Capital Expenditure (Capex): What It Means and Why It Matters

Capital expenditure, commonly abbreviated as capex, refers to the funds a company deploys to acquire, upgrade, or maintain physical or intangible assets that are expected to generate economic benefits over multiple future periods. Unlike operating expenditure, which is consumed in the period in which it is incurred and flows directly through the income statement, capex is capitalized on the balance sheet as a long-term asset and then depreciated or amortized over the asset’s useful life.

For the purposes of this paper, capex encompasses the full spectrum of infrastructure investment that hyperscalers are undertaking: the acquisition of land and buildings for data center campuses; the procurement of semiconductor chips, primarily NVIDIA H100 and H200 GPUs and successor architectures; the build-out of networking, switching, and storage infrastructure; the installation of cooling and power delivery systems; and increasingly, long-term leasing commitments to third-party data center operators, which function economically as capex even when classified differently on the balance sheet.

The critical financial relationship at the heart of this paper is the one between capex and free cash flow (FCF). Free cash flow is computed as operating cash flow minus capital expenditure. When capex rises dramatically, FCF compresses, even if revenue and operating income are growing robustly. It is this compression — and its duration — that drives the decision to access external debt markets rather than fund investment entirely from internal sources. The capex-to-FCF relationship is thus the mechanism through which AI infrastructure investment becomes a credit market event.

1.2  AI Infrastructure Buildout: The Physical Substrate of Intelligence

When this paper refers to the AI infrastructure buildout, it is describing the construction and operation of the physical and digital systems required to train, host, and serve large-scale artificial intelligence models. This infrastructure comprises several interdependent layers, each with its own cost structure and capital intensity profile.

At the base layer are the data center campuses themselves — large-format facilities that may occupy hundreds of acres and consume hundreds of megawatts of power. Above that is the compute layer, dominated by graphics processing units (GPUs) designed for parallel computation at the scale required by transformer-based AI models. Connecting these GPU clusters is a fabric of high-speed networking, including InfiniBand and Ethernet switching infrastructure from vendors such as NVIDIA, Arista, and Broadcom. Above the compute and networking layers sit the storage systems that house training datasets and model weights, and above that the software orchestration infrastructure that schedules workloads.

The AI infrastructure buildout is not a one-time construction project. It is an ongoing program because the models being trained are growing rapidly in size and complexity, because the inference workloads required to serve AI applications at scale are enormous and growing, and because the underlying chip architecture is evolving quickly enough that equipment deployed today may need to be replaced or supplemented within three to five years. This creates what some analysts have called a perpetual capex trap: the necessity of continuous re-investment simply to remain competitive.

1.3  Investment-Grade Credit: The Market That Finances the Surge

Investment-grade credit refers to debt obligations — primarily bonds — issued by entities whose creditworthiness has been assessed by one or more of the major rating agencies (S&P Global Ratings, Moody’s Investors Service, or Fitch Ratings) as meeting a minimum threshold of quality. Specifically, investment-grade designations correspond to ratings of BBB- or above on the S&P and Fitch scales, or Baa3 or above on the Moody’s scale.

The investment-grade bond market in the United States is one of the deepest and most liquid capital markets in the world, representing the primary channel through which large corporations raise long-term debt capital. According to S&P Global Ratings, technology sector issuance reached 16.7% of global nonfinancial bond supply in 2025 alone, reflecting the degree to which hyperscaler debt issuance has come to dominate the market’s flow.[4]

For investors, investment-grade bonds represent the core of fixed-income portfolios seeking predictable income with limited default risk. The duration, sector composition, and credit spread dynamics of the IG index directly affect the return profiles of pension funds, insurance companies, sovereign wealth funds, and the broad community of institutional investors who hold these instruments. When the composition of the IG index shifts — as it is now doing due to massive tech sector issuance — the implications cascade through every portfolio benchmarked to that index.

Section 2: The Scale of the AI Build-Out — From Earnings Rooms to the $1 Trillion Threshold

2.1  Q1 2026 Earnings Reports and Projections Through 2030

The earnings season of late April 2026 brought a set of disclosures that, taken together, constitute the clearest picture yet of how seriously the world’s largest technology companies are committed to AI infrastructure investment. The numbers were not merely large — they were large and rising, in some cases materially above what analysts had penciled in only weeks before.

Alphabet reported Q1 2026 revenue of $109.9 billion, beating consensus by nearly $3 billion, with Google Cloud growing 63% year-over-year to $20.02 billion — well above analyst estimates of $18.05 billion. Capital expenditure for the quarter reached $35.67 billion, more than doubling year-over-year, while Google Cloud backlog jumped to over $460 billion. Sundar Pichai told investors: [5]

“Our AI investments and full stack approach are lighting up every part of the business.” [5]

— Sundar Pichai, CEO, Alphabet — Q1 2026 Earnings Call

Amazon committed approximately $200 billion in capital expenditure for 2026, the largest absolute commitment of any single company. Free cash flow for the trailing twelve months compressed to $1.2 billion — a 95% decline year-over-year — as AI infrastructure spending accelerated, even as AWS grew 28% and Amazon’s chip business hit a $20 billion revenue run rate. Second quarter revenue guidance was set at $194 to $199 billion, above the $189.2 billion consensus.[6]

Meta raised its full-year 2026 capex guidance to $125 to $145 billion from the prior $115 to $135 billion range, citing higher component costs and expanded data center capacity. The contentious number triggered a roughly 6% after-hours decline in the stock, reflecting investor unease about the pace of spending relative to visible near-term returns. Meta CFO Susan Li stated on the earnings call that the ‘highest order priority is investing our resources to position ourselves as a leader in AI.’[7]

Microsoft reported fiscal Q3 capital expenditure of $30.88 billion, up 84% year-over-year, with AI revenue surpassing a $37 billion annual run rate. CFO Amy Hood attributed $25 billion of the company’s $190 billion calendar-year 2026 capex projection to rising memory chip and component costs, and told investors the company expected to remain capacity-constrained through at least 2026 as it works to bring GPU, CPU, and storage infrastructure online. Barclays estimates Microsoft’s free cash flow will slide by 28% in 2026 before recovering in 2027.[8]

Beyond these four, Oracle committed to $50 billion in capital expenditure for fiscal 2026, a figure that, as discussed in Section 4, produced considerable market anxiety when it was revealed in late 2025 given Oracle’s more concentrated customer base and heavier reliance on debt financing. The Stargate project — a consortium of OpenAI, SoftBank, and Oracle targeting $500 billion in aggregate AI infrastructure investment — adds a further dimension to the scale of the buildout.

The aggregate picture: combined 2026 capex from the five largest US cloud and AI infrastructure providers — Microsoft, Alphabet, Amazon, Meta, and Oracle — is projected at $660 billion to $690 billion, nearly doubling 2025 levels. With Q1 upward revisions factored in, UBS estimates total hyperscaler capex could reach $770 billion, approximately 23% above the firm’s prior estimate.[9]

For the period from 2026 to 2030, Goldman Sachs projects total hyperscaler capex of $1.15 trillion from 2025 through 2027 alone, more than double the $477 billion spent from 2022 through 2024. BlackRock’s 2026 Investing Outlook projects global AI infrastructure spending will exceed $2.2 trillion by 2028, characterizing the investment cycle as simultaneously a capital expenditure boom, a labor substitute, and a margin expander.[10]

On the private side, OpenAI and Anthropic are posting rapid revenue growth, though their combined revenues remain a fraction of the infrastructure investment being deployed on their behalf. OpenAI raised $110 billion in the largest-ever private technology funding round, with NVIDIA contributing $30 billion. The Stanford HAI 2026 AI Index Report notes that leading frontier companies are reaching meaningful revenue scale in a short period of time, but compute spend has increased significantly year-over-year.[11]

2.2  Capital Intensity Drivers: Chips, Campuses, and the Constraint of Power

Three forces drive the extraordinary capital intensity of the AI build-out, and they are interdependent: the cost and scarcity of specialized semiconductor chips; the construction and operation of the physical data center campuses that house those chips; and the energy infrastructure required to power and cool the entire system. Understanding each driver separately — and their interactions — is essential to evaluating the durability and sustainability of the build-out.

On chips: the NVIDIA H100 and H200 GPU clusters that constitute the workhorses of AI training cost between $25,000 and $40,000 per unit, and training a frontier model requires clusters of tens of thousands of such chips operating simultaneously for weeks or months. The successor Blackwell architecture carries similar or higher price points at deployment. Microsoft’s CFO explicitly called out rising memory chip and component costs as a driver of its $25 billion upward revision to capital expenditure guidance, illustrating how supplier pricing power directly feeds capex inflation. NVIDIA’s position as the dominant supplier of AI-grade GPUs has given it pricing leverage that keeps infrastructure costs elevated even as production scales.

On campuses: data centers built for AI workloads are categorically different from the cloud data centers of the previous generation. AI facilities require dramatically higher power density — a rack of H100 GPUs may draw 60 to 80 kilowatts, versus 5 to 10 kilowatts for a conventional server rack — which demands entirely new cooling architectures, including direct liquid cooling and immersion cooling systems. A single hyperscale AI campus can cost between $1 billion and $5 billion to build, with the largest campuses occupying over a million square feet and requiring dedicated power substations. A single AI task can use up to 1,000 times more electricity than a traditional web search, creating highly concentrated, large-scale power demands that regional electricity grids were not built to handle.[12]

On power: energy has become the binding constraint of the entire build-out. According to the International Energy Agency’s April 2026 report, electricity demand from data centers soared by 17% in 2025, and AI-focused data centers climbed even faster, well outpacing growth in global electricity demand of 3%. The capital expenditure of five large technology companies surged to more than $400 billion in 2025 and is set to increase by a further 75% in 2026.[13]

The World Economic Forum, in a May 2026 analysis, identified grid connectivity as the strategic bottleneck of the AI era: connecting a new facility to the power grid can take four to ten years, while AI data centers are typically planned and built within two to three. The misalignment increasingly determines which projects advance and which stall.[14]

The Brookings Institution estimates that energy consumption from data centers could approach 1,050 TWh by 2026, which would make data centers the fifth-largest energy consumer in the world, between Japan and Russia. US utilities have responded by committing $1.4 trillion in capital spending plans, with more than 30 of 51 surveyed utilities citing data centers as a top growth driver.[15]

2.3  Historical Parallels and Differences: Learning from Fiber-Optic Ghosts

Every analyst covering the AI capex surge has reached, at some point, for the comparison to the late 1990s telecom build-out. The parallel is instructive, but so are the differences, and conflating the two frameworks produces analytical error in both directions.

The similarities are genuine and should not be dismissed. In the late 1990s, telecoms such as Global Crossing and AT&T spent over $500 billion laying fiber-optic cable in anticipation of rapid Internet adoption. Their projections proved overoptimistic, leaving the industry to suffer for years amid a glut of capacity and collapsing prices.[16]

The scale analogy is also stark. Data center capital expenditures as a share of GDP have already surpassed peak telecom spending during the dot-com bubble. Paul Kedrosky calculates that AI data center spending is already at 20% of peak railroad spending as a percentage of GDP from the nineteenth century, and it is still accelerating.[17]

But the differences are at least as important as the similarities, and they matter enormously for credit analysis. First, the issuers are categorically different. Global Crossing, WorldCom, and most of the telecom-era infrastructure builders were highly leveraged, economically thin enterprises with no operating cash flow capable of absorbing a demand disappointment. Today’s hyperscalers are the opposite: they are the most profitable corporations in human history, generating hundreds of billions in annual revenue from established, diversified businesses. Alphabet entered 2026 with a Google Cloud backlog of $460 billion. Microsoft’s AI revenue exceeded $37 billion annually before most of its capex investment had come online. These are not speculation; they are cash-generating machines investing in the next generation of cash generation.

Second, the assets being built are different in kind. Fiber-optic cable, once laid, is largely inert — useful only for transmitting data, and economically worthless if demand does not materialize. GPU clusters and data centers, by contrast, can be redeployed across an enormous range of computing workloads. They are not single-purpose assets. Even in a scenario where AI demand disappoints relative to the most optimistic projections, the infrastructure could be redirected to traditional cloud computing, high-performance computing, or specialized scientific workloads. The Janus Henderson research team, in a November 2025 analysis, identifies eight structural differences between the AI wave and the dot-com era, including the macroeconomic environment, the financial strength of the issuers, and the strategic nature of the geopolitical competition driving investment.[18]

Third, the competitive and geopolitical dynamic is different. The telecom build-out was driven by market speculation about consumer internet adoption. The AI build-out is partly driven by that, but also by explicit national competition — between the United States and China, between American hyperscalers themselves, and between hyperscalers and well-funded AI startups. Larry Page was quoted saying he was willing to go bankrupt rather than lose the AI race. That is not the language of rational capital allocation; it is the language of strategic necessity, and it changes the behavioral calculus entirely.

The honest conclusion is that the historical parallel counsels caution without predicting catastrophe. The risk of overcapacity and demand disappointment is real. The financial strength of today’s builders provides a buffer that their 1990s predecessors lacked. But no buffer is infinite, and the duration of the investment cycle is still unknown.

2.4  The $1 Trillion Club: Free Cash Flow Compression and the Debt Imperative

The threshold of $1 trillion in combined AI-related infrastructure spending — across hyperscalers plus the broader ecosystem of utilities, real estate investment trusts, and private cloud operators — was effectively crossed in 2025 and 2026 taken together. Morgan Stanley estimates that large technology companies committed more than $1 trillion of spending in the 2025 to 2026 period. The investment bank notes that strategic financing will be a critical enabler, and that growth may be reliant on the robust balance sheets of mega-cap hyperscalers, who can tap their own cash flows to finance about half of their spending.[19]

The free cash flow dynamics are telling. In 2025, the four largest US internet companies generated a combined $200 billion in free cash flow, down from $237 billion in 2024. CoBank, in an April 2026 analysis, frames the question through the lens of return on invested capital: if ROIC is rising alongside increased capex, the investments are creating shareholder value. That is, broadly, what the data shows for the leading hyperscalers — but the lag between investment and return is growing.[20]

The consequence of free cash flow compression is the migration from internal financing to external capital markets. In prior years, hyperscalers paid for their infrastructure almost entirely from the operating cash flow their core businesses generated. Beginning in late 2025, that model broke down. The scale of investment outpaced the capacity of cash flow to fund it, and the hyperscalers began issuing bonds — in size, at long duration, and with increasing frequency. The transformation of America’s most cash-rich companies into systematic bond market issuers is one of the defining credit market events of the current decade.

Section 3: Mechanics of the Investment-Grade Corporate Bond Market — How the Capex Surge Reaches Fixed-Income Portfolios

3.1  Supply Surge: The Great Wall of Technology Issuance

The transition from self-funded capex to external bond issuance began in earnest in the second half of 2025. According to LPL Financial, big tech companies issued nearly $250 billion in debt in 2025, the most on record. Alphabet, for instance, raised $20 billion in the US investment-grade market, including issuing a 100-year sterling bond to fund AI capital expenditure.[21]

In 2025 alone, hyperscalers issued $121 billion in bonds, more than four times the five-year average of $28 billion. The pace accelerated into 2026: year-to-date US investment-grade issuance reached $296 billion by early 2026, 31% higher than a year earlier, with technology issuance more than doubling. Alphabet’s June 2026 announcement of an $80 billion equity capital raise — including a $10 billion investment from Berkshire Hathaway — represents the most dramatic single financing event of the cycle, signaling that even equity markets are being tapped to fund infrastructure.[22]

The supply dynamics have directly reshaped the technology sector’s footprint in the IG index. S&P Global Ratings’ Liquidity Outlook 2026 identified technology issuance as a primary driver of credit market liquidity dynamics, noting that tech represented 16.7% of global nonfinancial bond supply in 2025. S&P analyst Nicolas Charnay stated: [23]

“We identified six key issues that could affect credit market liquidity over the coming year. In addition to leverage at nonbank financial institutions, risks related to AI valuations, and events at the Fed, we will also be watching for clarification of U.S. bank capital regulations, which could gradually impact systemwide liquidity.” [23]

— Nicolas Charnay, S&P Global Ratings — Liquidity Outlook 2026, February 17, 2026

The market’s absorption of this supply has, to date, been remarkably orderly. Hyperscaler bond offerings are typically heavily oversubscribed, reflecting the enormous institutional demand for the combination of credit quality (ratings anchored at AA and AAA for the top issuers), relative yield (tech spreads remain tight by historical standards), and duration (institutional liability-matching demand is structural). The oversubscription dynamic reflects a genuine alignment between what hyperscalers want to issue — long-dated paper to match the life of their infrastructure investments — and what the largest buyers of fixed income want to own.

3.2  Duration Shock: When the Index Changes Shape

Duration is a measure of the sensitivity of a bond’s price to changes in interest rates, roughly corresponding to the weighted average time to receive the bond’s cash flows. Longer-dated bonds carry more duration risk — their prices fall more for a given rise in interest rates — and the composition of the IG index in terms of aggregate duration has material implications for how benchmark portfolios behave.

When hyperscalers issue thirty-year and fifty-year notes — and in Alphabet’s case, one-hundred-year sterling bonds — they add incremental duration to the IG index. This is not a trivial effect: as tech sector issuance has grown from under 10% to nearly 17% of the IG index, the average duration of the index has lengthened meaningfully. Institutional investors who benchmark against the IG index must either accept this additional duration or actively manage against it by tilting their portfolios toward shorter-dated instruments.

The duration effect is compounded by the pricing dynamics of long-dated tech paper. Because hyperscaler bonds are perceived as extremely high-quality credits, they trade at tight spreads — often only modestly above US Treasury bonds of equivalent maturity. This means that when the Treasury curve steepens or rates rise, long-dated hyperscaler bonds absorb significant price declines, amplifying the mark-to-market volatility of IG portfolios.

The duration shock is, in essence, a structural transformation of the benchmark. It is not a temporary distortion that will reverse when rates normalize. The infrastructure being financed by these bonds will remain in place for years, and the bonds themselves will remain in the index until maturity. Fixed-income portfolio managers must build duration risk management frameworks that account for this permanent shift.

3.3  Market Absorption: Oversubscription, Spread Dynamics, and the Limits of Demand

The investment-grade credit market has, for the most part, absorbed the technology supply surge with equanimity. Hyperscaler bond deals are routinely three to five times oversubscribed, implying that demand significantly exceeds supply even at the record issuance volumes of 2025 and 2026. Credit spreads for the top hyperscalers — Alphabet, Microsoft, Amazon — remain historically tight, reflecting the market’s confidence in their financial strength and the resilience of their underlying cash flows.

This dynamic reflects a structural feature of the IG investor base: it is dominated by liability-matching institutions — life insurers, pension funds, sovereign wealth funds — that have a structural need to own long-duration, high-quality assets. The supply of such assets has historically been insufficient to meet demand, which is why spreads are tight and oversubscription is the norm. Hyperscaler bonds fill this gap precisely. The Janus Henderson analysis noted that in September 2025, each of the hyperscalers moved to raise capital externally for the first time, and that broader markets granted ‘permission’ for these players to continue their capex spend because of continued momentum in core business metrics.[24]

However, the market’s tolerance is not unconditional. The Oracle experience in late 2025 and early 2026 illustrates the conditions under which IG credit markets can reprice quickly. Oracle’s five-year credit default swap spreads more than tripled in the weeks following its December 2025 earnings report, reaching levels not seen since 2009, even though Moody’s and S&P maintained their investment-grade ratings. The market was signaling concern not about Oracle’s current financial position but about the trajectory — the speed of debt accumulation, the concentration of customer risk in a single large AI counterparty, and the uncertainty about the return profile of a $50 billion capex commitment.

The Oracle case is instructive because it demonstrates that even within the investment-grade universe, the market can bifurcate dramatically. The hyperscalers at the top — with diversified revenues, net-cash balance sheets, and demonstrated AI monetization — continue to enjoy the deepest investor demand at the tightest spreads. Those lower in the hierarchy, with more concentrated exposures and heavier leverage, face meaningfully different pricing and a more skeptical buyer base.

Section 4: Credit Risk and Rating Agency Perspectives — Navigating the Valuation Labyrinth

4.1  Defining Credit Risk and Credit Ratings in the Context of the Capex Surge

Credit risk is the probability that a borrower will fail to meet its contractual obligations — that is, that it will default on interest payments, fail to repay principal, or restructure its debt on terms less favorable than those originally agreed. In the investment-grade context, credit risk is not simply about default probability in isolation; it encompasses the full spectrum of events that can impair the value of a debt instrument, including rating downgrades, covenant violations, changes in capital structure, and market spread widening that erodes the mark-to-market value of a portfolio holding even without an actual default.

Credit ratings are shorthand summaries of an issuer’s credit risk, produced by independent rating agencies using a combination of quantitative financial analysis and qualitative judgment about business model stability, competitive position, management quality, and macroeconomic sensitivity. The three major agencies — S&P Global Ratings, Moody’s Investors Service, and Fitch Ratings — together control roughly 95% of the global credit rating market. Their ratings function as gatekeepers: pension funds, insurance companies, and sovereign wealth funds are often prohibited by mandate or regulation from owning below-investment-grade instruments, making an IG rating an economic necessity for large-scale bond issuance.

The four largest hyperscalers enter the AI capex cycle with ratings anchored at the highest levels of the investment-grade spectrum. Alphabet and Microsoft carry Aaa/AAA ratings from Moody’s and S&P, respectively, reflecting balance sheets with net-cash positions, diversified and highly profitable revenue streams, and demonstrated financial discipline even under significant investment pressure. Amazon and Meta carry somewhat lower ratings but remain solidly in the upper tiers of investment grade. Oracle, by contrast, carries ratings of Baa2 (Moody’s) and BBB (S&P) — still investment grade, but at the lower end, with less buffer against deterioration.[25]

4.2  Rating Agency Scrutiny: What Moody’s and Fitch Are Watching

The rating agencies are not ignoring the capex surge. Their public commentary and rating action frameworks make clear that they are evaluating hyperscaler expenditure programs along several dimensions, and that the tolerance for deterioration in key financial metrics is bounded — even for the highest-rated issuers.

Moody’s and Fitch have both stated that AI risks are not uniform across software and technology issuers. Differences in business models, customer dependency, and the breadth of revenue diversification provide buffers for some issuers that others lack. The LPL Financial Rate and Credit View from February 2026 summarizes this perspective: overall default exposure in CLOs remains low, with breakeven cushions sufficient to protect senior notes from widespread losses, but pressures from concentrated AI strategies are receiving increased scrutiny.[26]

The key metrics the agencies are monitoring include: leverage ratios (net debt to EBITDA), free cash flow generation and coverage of interest obligations, the proportion of capex financed by debt versus internal cash flows, the pace of revenue growth from AI-related services, and the off-balance-sheet obligations arising from long-term data center leases. Goldman Sachs noted in November 2025 that consensus hyperscaler capex estimates had risen from $467 billion at the start of the earnings season to $533 billion, and flagged that if the 75% annual growth pace were maintained, spending could push to $700 billion — a level that would test even the strongest balance sheets if revenue did not scale commensurately.[27]

The agencies have also been explicit that their tolerance for FCF compression is time-limited. If the gap between infrastructure investment and monetization persists beyond the three- to four-year horizon that most hyperscalers have articulated as their payback timeline, ratings pressure will build. The agencies are not yet in panic mode, but they are watching closely, and their commentary has grown measurably more cautious since early 2025.

4.3  Hidden Liabilities: Off-Balance-Sheet Leasing and the True Debt Burden

One of the most significant and underappreciated credit risks embedded in the AI infrastructure build-out is the system of off-balance-sheet commitments that hyperscalers have accumulated through long-term data center leasing arrangements. Rather than building and owning every facility themselves, several hyperscalers have increasingly relied on third-party data center operators — firms such as Digital Realty, Equinix, and a growing roster of hyperscale-focused real estate investment trusts — to construct and operate campuses under long-term lease contracts.

Alphabet’s 2025 Annual Report explicitly acknowledges this risk: the company is ‘entering into significant leasing arrangements with third party operators, which may increase costs and operational complexity,’ and notes that ‘large, long-duration commercial agreements could increase our liabilities and obligations in the event of nonperformance.’ In a scenario of industry slowdown or overbuilding, excess capacity that cannot be easily redeployed would generate lease obligations that continue regardless of revenue performance.[28]

The accounting treatment of these arrangements matters enormously for credit analysis. Under US GAAP and IFRS 16, many long-term leases must now be capitalized on the balance sheet as right-of-use assets and lease liabilities. But the magnitude and duration of these commitments — some running ten, twenty, or thirty years — means that even after capitalization, the true economic exposure may not be fully appreciated by investors relying on reported leverage ratios alone.

The hidden liability problem is most acute for companies at the lower end of the investment-grade spectrum, where the margin for error is smaller and the capacity to absorb unexpected lease-related charges is limited. For the top hyperscalers, the issue is manageable given their financial strength. But for smaller data center operators, co-location providers, and second-tier cloud companies that are building out AI infrastructure on the assumption that demand will grow into their capacity, lease commitments represent a genuine solvency risk if that assumption proves wrong.

4.4  Bifurcated Risks: Suppliers Win, Utilities Strain, Mid-Tier Issuers Are Exposed

The capex wave is not uniformly benign or uniformly dangerous across the credit landscape. Its credit effects are deeply bifurcated, and understanding that bifurcation is essential to positioning a fixed-income portfolio intelligently.

For hardware and networking suppliers — NVIDIA, Broadcom, Arista Networks, Marvell Technology — the capex surge is unambiguously positive. Their customers are committing to purchase hundreds of billions of dollars of equipment over multiple years, providing a level of revenue visibility that most technology companies can only dream of. Their balance sheets are strengthening, their free cash flow is growing, and their credit profiles are improving. The capex surge is a tailwind for their debt.

For power and utility companies, the picture is more complicated. On the revenue side, the growth in electricity demand from data centers is a multi-decade growth driver that justifies significant investment. On the cost side, the capital required to build out transmission and generation capacity is enormous, and it is being deployed in a regulatory environment that moves slowly. The risk is not that utilities will fail — they are regulated monopolies with captive customers — but that the rate of investment outpaces their ability to recover costs through regulated tariffs, pressuring cash flow and leverage metrics in the medium term.

For mid-tier technology issuers — companies that are building AI infrastructure but lack the diversified revenue base of the hyperscalers — the credit environment is genuinely challenging. Oracle is the most visible example, but it is not the only one. Any company that is committing to a debt-funded capex program in the range of 30% to 50% of annual revenue, with customer concentration in a small number of AI frontier model developers, is carrying a credit profile that has limited tolerance for revenue disappointment. Barclays downgraded Oracle’s debt to underweight in November 2025, warning it could fall to BBB-, the lowest investment-grade rating before junk status. Oracle’s credit default swap spreads widened to above 125 basis points — levels not seen since the 2009 financial crisis — even while official ratings remained investment grade.[29]

Section 5: Monetization vs. Capital Intensity — The Revenue Question and the Risk of Overinvestment

5.1  The Revenue Question: The Uncomfortable Gap Between Investment and Return

At the heart of the Capex Surge framework is a question that no earnings release has fully answered: will the AI infrastructure being built today generate returns commensurate with the capital being deployed, and will those returns materialize on a timeline consistent with the debt servicing obligations that finance the investment?

The optimists have genuine evidence on their side. Google Cloud’s 63% year-over-year growth in Q1 2026 is not the performance of an infrastructure investment that is failing to monetize. Microsoft’s AI revenue run rate exceeding $37 billion annually demonstrates that enterprise adoption is occurring at scale. The $460 billion Google Cloud backlog, the reacceleration of AWS growth to 28%, and the reported 15% to 30% efficiency gains at companies like Intuit, ServiceNow, and Salesforce all suggest that AI is delivering measurable economic value to customers who are willing to pay for it.

But the optimists’ evidence is, at this stage, concentrated in the cloud and infrastructure layer. The question that remains open is whether the AI economy — the layer of applications and business models that sit above the infrastructure — will generate demand at the scale required to justify the current investment. T. Rowe Price, in a January 2026 analysis, estimates that the AI capex cycle can continue for another two to three years before facing its first true test, and identifies the greater risk as monetization lagging investment, creating volatility in sentiment.[30]

The Stanford HAI 2026 AI Index provides a measured perspective on adoption. Generative AI is now used in at least one business function at 70% of organizations, and AI agent deployment was in the single digits across nearly all business functions. The consumer surplus from AI tools was estimated at $172 billion annually by early 2026, up from $112 billion a year earlier. But the same report notes that adoption varies widely and correlates strongly with GDP per capita, suggesting that the total addressable market for AI services may be more concentrated and slower to develop in lower-income markets than the most optimistic projections assume.[11]

The critical dynamic to watch is what economists call the J-curve of infrastructure investment: returns are typically front-loaded in capital deployment and back-loaded in revenue realization. The fiber-optic networks laid in 1999 are now the backbone of the global internet economy; the companies that built them went bankrupt, but the infrastructure itself generated enormous value. The question is not whether AI infrastructure will ultimately prove valuable — it almost certainly will — but whether the specific companies financing it will capture that value, and on what timeline.

5.2  Overinvestment Risk: The Perpetual Capex Trap and the Pause Scenario

The possibility of a medium-term capex pause or inventory digestion cycle — in which demand for AI services undershoots investment assumptions, triggering a period of reduced spending and excess capacity — is the central tail risk for both the technology sector and the credit markets that finance it.

Nobel Prize-winning MIT economist Daron Acemoglu has been the most prominent academic voice of skepticism about the timeline and scale of AI’s economic impact. Writing in Fortune in August 2024, Acemoglu stated: [31]

“The problem with the AI bubble isn’t that it is bursting and bringing the market down — it’s that the hype will likely go on for a while and do much more damage in the process than experts are anticipating.” [31]

— Daron Acemoglu, Institute Professor, MIT — Fortune, August 2024

Acemoglu projects that AI will increase US GDP by only 1.1% to 1.6% over the next ten years, with a roughly 0.05% annual gain in productivity. This assessment stands in stark contrast to Goldman Sachs’ projection of a 7% GDP boost, illustrating the extraordinary range of credible expert opinion on the question. When asked about his broader framework, Acemoglu has said: ‘I’m trying not to be bearish. There are things generative AI can do, and I believe that, genuinely. However, I believe there are ways we could use generative AI better and get bigger gains, but I don’t see them as the focus area of the industry at the moment.’[32]

The Morningstar analysis of Kai Wu’s October 2025 paper ‘Surviving the AI Capex Boom’ provides a rigorous historical framework for evaluating overinvestment risk. Wu’s analysis covers the railway boom of the nineteenth century, the telecom bubble, and multiple other infrastructure cycles, and finds a consistent pattern: infrastructure booms typically result in overinvestment, excess competition, and poor stock returns for the infrastructure builders themselves, even when the underlying technology delivers enormous social value. As Wu notes: ‘One of the key lessons of past technological revolutions is that the builders of the underlying infrastructure often do not capture much of the value created. Rather, the value accrues to their customers and the rest of society.’[33]

The RBC Wealth Management analysis of February 2026 frames the overinvestment risk in terms of history: ‘transformative general-purpose technologies — such as railroads and the internet — often require periods of overinvestment and lengthy diffusion-adoption cycles before delivering durable and widespread benefits.’ The AI build-out, on this view, is not necessarily misguided — it is simply early, and the mismatch between capital deployment timing and return realization timing is the source of the credit risk.[34]

The overinvestment scenario most relevant to credit markets is not a catastrophic collapse but a more orderly pause: a scenario in which one or more hyperscalers materially reduces capex guidance in a future quarter, triggering a reassessment of the growth narrative and wider spreads across the technology sector. The Oracle CDS experience shows how quickly repricing can occur even without a fundamental credit event. A broader capex pause could produce spread widening across the entire technology sector, with particular violence in the names most dependent on the continuation of the AI investment cycle.

Section 6: Strategic Lessons and Implications for Investors — The Six Pillars of the Capex Surge Framework

The Capex Surge framework is not simply a description of what has happened; it is a structured approach to evaluating what investors, CFOs, and policymakers should do in response to the dynamics described in the preceding sections. I organize the strategic lessons and implications into six pillars, each representing a distinct analytical posture and investment approach. None of these pillars is correct for all investors or all market conditions; the art of applying the framework lies in calibrating one’s position across all six in a way that reflects one’s specific risk tolerance, liability structure, and market outlook.

Pillar 1 — Continue the Build-Out: The Case for Staying Long the Infrastructure Cycle

The first pillar is the most straightforward: the evidence, as of mid-2026, supports continued investment in and exposure to AI infrastructure. Google Cloud growing at 63% annually, Microsoft’s AI revenue exceeding $37 billion, and AWS reaccelerating to 28% growth are not signs of a cycle that is about to roll over. For investors with long time horizons and the capacity to tolerate near-term FCF compression, the hyperscalers’ bonds and equities remain compelling on a risk-adjusted basis.

The credit case for staying long the top hyperscalers is particularly robust. Their ratings are among the highest in the corporate bond universe. Their balance sheets carry net-cash positions even after the capex surge. Their revenue diversification means that a disappointing quarter in AI does not impair their ability to service debt. And their long-dated bond issuance provides duration exposure that liability-matching institutions genuinely need. For institutional fixed-income investors, underweighting Alphabet, Microsoft, and Amazon bonds on credit grounds would require a much more pessimistic view of the underlying economics than the data currently supports.

Pillar 2 — Preserve Cash and Maintain Optionality: The Case for Caution

The second pillar acknowledges the limits of the first. The pace of capex escalation — with companies repeatedly raising guidance above analyst estimates, with free cash flow declining, and with debt issuance accelerating — creates conditions in which even fundamentally strong credits can be surprised by events. A meaningful recession, a sharp rise in interest rates, or a broader loss of confidence in the AI monetization timeline could produce spread widening that impairs the mark-to-market value of IG tech portfolios even without any default.

For investors in this pillar, the prescription is to maintain above-average cash positions or short-duration allocations within the technology sector. Rather than extending duration to match the long-dated issuance of the hyperscalers, these investors position at the short end of the curve, accepting somewhat lower yield in exchange for reduced sensitivity to rate and spread movements. They also place strict limits on position sizes in any single issuer, recognizing that the correlation of AI-related credits in a risk-off scenario is likely to be higher than their historical correlations would suggest.

Pillar 3 — Wait and Observe: Identifying the Inflection Point

The third pillar is a more active form of patience: rather than holding static positions, these investors are watching for specific signals that would confirm or deny the overinvestment thesis, and are prepared to act decisively when those signals emerge. The key signals to watch are: quarter-on-quarter deceleration in Google Cloud or AWS growth rates; management commentary about capex reduction or ‘digestion’ periods; widening in Oracle’s CDS as a leading indicator of broader AI credit stress; and any guidance from NVIDIA about softening GPU demand, which would suggest that the build-out is approaching a natural pause point.

This posture is not passive; it requires active monitoring and well-defined trigger criteria. But it avoids the twin errors of over-committing to a cycle that may not have reached its peak, and of prematurely reducing exposure to a cycle that may run for several more years. The IMF’s January 2026 World Economic Outlook, which raised its global growth forecast to 3.3% partly on the strength of AI investment, but also warned that higher leverage and frequent hardware upgrades could pressure firms if financial conditions tighten, exemplifies the watch-and-see intellectual posture.

Pillar 4 — Rent Rather Than Own: The Apple Model of Infrastructure Asset-Lightness

The fourth pillar observes that not every company that benefits from the AI economy needs to build its own infrastructure. Apple is the paradigmatic example: a technology company with a market capitalization and revenue base comparable to the hyperscalers, which has elected to rent data center capacity from partners rather than build its own. This keeps Apple’s balance sheet asset-light, its free cash flow generation high, and its credit profile unencumbered by the capex cycle that is reshaping its peers.

For credit investors, the lesson is to identify the Apple-model beneficiaries within each sector: companies that are leveraging the AI build-out without bearing its capital costs. Application-layer AI companies that pay cloud API fees rather than owning GPU clusters fall into this category. So do enterprises deploying AI through Software-as-a-Service platforms. These entities are the customers of the infrastructure cycle, not its builders, and their credit profiles benefit from AI productivity gains without absorbing capex risk. Building a portfolio that combines hyperscaler credit with application-layer and enterprise credit provides exposure to the AI economy without concentration in the infrastructure layer.

Pillar 5 — Assess Latecomer Risk: Whether the Window Has Closed

The fifth pillar addresses a concern that several analysts have raised: that for companies that have not yet committed to a major AI infrastructure program, the moment of strategic entry may have passed. The argument runs as follows: the hyperscalers have accumulated GPU inventory, talent, and data assets at scale; they have established customer relationships and technical capabilities that will compound over time; and the capital cost of attempting to build a competitive infrastructure at this stage would be prohibitive for any company that has not already started.

This ‘too late’ concern is most relevant for mid-tier cloud providers, regional telecommunications companies, and enterprise software vendors that are contemplating major AI infrastructure investments in 2026 or 2027. For these companies, the credit question is acute: can they afford the debt required to build a competitive AI position, and if they do take on that debt, will the revenue upside materialize before the leverage triggers a ratings downgrade? Joshua Mahony, Chief Market Analyst at Scope Markets, summarized the structural tension: ‘Today’s mega-cap AI valuations assume that the current surge in AI spending is not a one-off infrastructure build, but the start of a highly profitable, self-reinforcing industry.’ The latecomers must assess whether they are entering a self-reinforcing cycle or buying into a cycle that is already past its economic peak for new entrants.[35]

Pillar 6 — Power as the Binding Constraint: The Energy Credit Theme

The sixth pillar addresses what is, in practical terms, the most important near-term constraint on the AI build-out: the availability of reliable, affordable electric power. As noted in Section 2.2, grid connectivity delays of four to ten years are emerging as the primary bottleneck for new data center development in many US and European markets.

For credit investors, the energy constraint creates a distinct opportunity. Utilities with strategically located grid infrastructure — particularly those with excess transmission capacity in data-center-dense regions — are likely to benefit from long-term, contracted electricity revenues that provide exactly the kind of stable, predictable cash flow that supports investment-grade credit quality. Nuclear energy operators, which offer the combination of carbon-free power and 24/7 baseload capacity that AI data centers require, are attracting particular attention. The World Economic Forum analysis from May 2026 observes that ‘access to the grid — rather than chips, capital, or algorithms — is increasingly the binding constraint’ for AI development, and that ‘strong leadership is needed to align clean energy investments, grid build-out and AI growth.’[14]

The credit theme here is not simply long power utilities. It is selective: utilities with the strongest grid connectivity, the highest proportion of contracted long-term capacity, and the most transparent regulatory relationships will be the beneficiaries. Those carrying significant construction risk on new generation capacity, or those facing unfavorable regulatory environments, may see the AI demand tailwind offset by cost inflation and capital expenditure requirements of their own.

Conclusion: The Name, the Framework, and the Investor’s Imperative

Why I Call This Framework Capex Surge

The name Capex Surge is deliberate in its vocabulary. Surge is more precise than boom because it captures the velocity and directionality of the phenomenon without presupposing its ultimate resolution. A boom implies duration and a subsequent bust; a surge implies momentum that may or may not be sustained. The AI infrastructure spend has the character of a surge: it has been building for years, it is accelerating rather than decelerating as of mid-2026, and its endpoint — in terms of both total capital deployed and the point at which monetization catches investment — remains genuinely uncertain.

The capex dimension of the name is equally precise. This is not a technology hype cycle in the traditional sense — the companies involved are not speculative ventures without revenue. They are the most profitable corporations in human history, spending prodigiously from a position of strength. The capital they are deploying will build real assets with real useful lives. The credit risk is not that these companies will fail; it is that the debt they are accumulating to fund this investment will, in certain scenarios, change the nature of their credit profiles and the composition of the fixed-income market in ways that reward careful analysis and penalize complacency.

Summary of Findings

This paper has established the following core findings. First, the scale of the AI build-out is unprecedented in absolute terms and historically unusual in its concentration: four companies are committing $700 to $725 billion in a single year, with the number rising. Second, the transition from self-funded capex to systematic bond market issuance is a structural shift with permanent implications for IG benchmark duration and sector composition. UBS has raised its 2026 US investment-grade technology supply estimate to $360 billion, and total IG issuance forecast to $1.8 trillion.[36]

Third, the leading hyperscalers — Alphabet, Amazon, Microsoft — maintain net-cash or low-leverage positions that preserve their investment-grade creditworthiness even under significant free cash flow compression. They are building from a position of genuine financial strength. Fourth, the credit environment is bifurcated: the hyperscalers’ strength coexists with meaningful vulnerability in the ecosystem of suppliers, data center operators, utilities, and mid-tier cloud companies that are absorbing the capital intensity without equivalent revenue diversity or balance sheet depth.

Fifth, the rating agencies are watching closely, and their tolerance for deterioration is bounded. Off-balance-sheet lease commitments represent a material hidden liability that investors must assess independently of reported leverage ratios. Sixth, the overinvestment risk, while not the base case, is the most important tail risk for credit portfolios: a capex pause triggered by demand disappointment could produce rapid spread widening, particularly in names most exposed to the AI build-out narrative. The IMF’s Global Financial Stability Report notes that financial stability risks remain elevated, with stretched asset valuations and growing pressure in sovereign bond markets compounding the tech sector’s credit dynamics.[37]

Broader Market Impact

The Capex Surge will leave a permanent mark on the investment-grade credit market. The technology sector’s share of the IG index has grown from under 10% to nearly 17% in a single year, and it is unlikely to retreat meaningfully as long as hyperscaler infrastructure programs continue at their current scale. The duration of the index has lengthened. The largest single-issuer positions in IG portfolios are increasingly technology names rather than the financial institutions or industrial companies that historically dominated.

These structural shifts will permanently alter the behavior of IG-benchmarked portfolios in stress scenarios. In a rising-rate environment, the longer duration of tech-heavy portfolios amplifies mark-to-market losses. In a credit risk-off scenario, the higher tech concentration amplifies spread-driven losses. Portfolio managers who have not explicitly modeled these changed dynamics are operating with an outdated framework.

Implications for Investors: The Analytical Imperative

The Capex Surge framework concludes with a call to analytical rigor. The environment described in this paper rewards investors who are willing to go beneath the surface of reported metrics, assess off-balance-sheet obligations, model FCF trajectories under multiple capex and revenue scenarios, and maintain differentiated views across the bifurcated credit landscape.

Covenant assessment matters more than it has in years. As hyperscalers issue bonds with covenants designed to preserve flexibility — including the ability to incur significant additional debt and to make large capital expenditures without triggering protective provisions — the investor relying on covenant protection as a risk management tool must scrutinize the specific terms of each issuance with care. Asset longevity analysis — assessing the depreciation profile and redeployability of AI infrastructure assets — is essential for evaluating whether capex creates durable value or merely sustains competitive parity. And credit monitoring must be continuous, not episodic: the pace of change in this environment means that quarterly earnings reviews are insufficient.

The Capex Surge is one of the defining financial phenomena of the 2020s. It is reshaping the investment-grade credit market, testing the limits of hyperscaler balance sheets, and creating a bifurcated credit environment that offers both opportunity and risk in approximately equal measure. Investors who understand its mechanics, track its evolution, and apply the Six Pillars framework with discipline will be positioned to navigate it. Those who treat it as background noise will be surprised by its consequences.

Footnotes / Endnotes

[1]  Tom’s Hardware / Financial Times. Google, Amazon, Microsoft, and Meta collectively plan to spend $725 billion on capex in 2026, up 77% from last year’s record $410 billion, according to Q1 2026 earnings compiled by the Financial Times. April 30, 2026. https://www.tomshardware.com/tech-industry/big-tech/big-techs-ai-spending-plans-reach-725-billion

[2]  CNBC. The four hyperscalers are projected to increase capital expenditures by more than 60% from historic 2025 levels; combined 2025 free cash flow was $200 billion, down from $237 billion in 2024. February 6, 2026. https://www.cnbc.com/2026/02/06/google-microsoft-meta-amazon-ai-cash.html

[3]  Yahoo Finance / UBS Research. UBS raised its 2026 US investment-grade issuance forecast to $1.8 trillion and US IG technology supply estimate to $360 billion from $300 billion, citing sharp rise in hyperscaler capex. 2026. https://finance.yahoo.com/news/ai-capex-surge-drives-higher-150004056.html

[4]  S&P Global Ratings / PRNewswire. Private Credit, Tech Issuance fuelled by AI, and Increasing Leverage Among Key Driving Factors Impacting Credit Market Liquidity in 2026. Tech issuance hit 16.7% of global nonfinancial bonds in 2025. February 17, 2026. https://press.spglobal.com/2026-02-17-Private-Credit,-Tech-Issuance-fuelled-by-AI,-and-Increasing-Leverage-Among-Key-Driving-Factors-Impacting-Credit-Market-Liquidity-in-2026-according-to-S-P-Global-Ratings

[5]  Yahoo Finance / Futurum Group. Hyperscalers Hit $700 Billion in 2026 AI Spending Plans; Sundar Pichai quote from Q1 2026 Alphabet earnings call. May 1, 2026. https://finance.yahoo.com/sectors/technology/articles/hyperscalers-hit-700-billion-2026-111243744.html

[6]  The Next Web. Q1 2026 Big Tech earnings: $650 billion in AI capex and compute constraints. Amazon FCF, revenue guidance, and AWS details from Q1 2026 earnings. April 30, 2026. https://thenextweb.com/news/alphabet-amazon-meta-q1-2026-earnings-ai-cloud

[7]  CNBC / Fortune. Meta CFO Susan Li statement on capital allocation priorities; Meta Q1 2026 earnings disclosure; shares fell ~6% after-hours on capex guidance raise. April 30, 2026. https://fortune.com/2026/04/30/big-tech-hyperscalers-will-spend-700-billion-on-ai-infrastructure-this-year-with-no-clear-end-in-sight-eye-on-ai/

[8]  Tom’s Hardware. Microsoft set calendar-year 2026 capex at $190 billion; CFO Amy Hood attributed $25 billion to rising memory chip costs; Barclays estimates 28% FCF decline in 2026. April 30, 2026. https://www.tomshardware.com/tech-industry/big-tech/big-techs-ai-spending-plans-reach-725-billion

[9]  Yahoo Finance / UBS. Amazon, Meta, and Google each raised 2026 capex guidance materially above consensus; UBS projects total hyperscaler capex at ~$770 billion, ~23% above prior estimate. 2026. https://finance.yahoo.com/news/ai-capex-surge-drives-higher-150004056.html

[10]  BlackRock. Investing in 2026: AI, War, and Income. Investment cycle projected to exceed $2.2 trillion in global infrastructure spend by 2028. March 26, 2026. https://www.blackrock.com/us/financial-professionals/insights/ai-war-and-income

[11]  Stanford HAI. 2026 AI Index Report — Economy chapter. Generative AI used in at least one business function at 70% of organizations; estimated US consumer surplus $172 billion annually by early 2026. 2026. https://hai.stanford.edu/ai-index/2026-ai-index-report/economy

[12]  Enkiai / Data Center Knowledge. AI Data Center Grid Strain: Power Halts Growth in 2026. A single AI task can use up to 1,000 times more electricity than a traditional web search. April 8, 2026. https://enkiai.com/data-center/ai-data-center-grid-strain-power-halts-growth-in-2026/

[13]  International Energy Agency (IEA). Data centre electricity use surged in 2025 — Key Questions on Energy and AI. Electricity demand from data centres soared 17% in 2025; capex of five large tech companies to increase 75% in 2026. April 16, 2026. https://www.iea.org/news/data-centre-electricity-use-surged-in-2025-even-with-tightening-bottlenecks-driving-a-scramble-for-solutions

[14]  World Economic Forum. Is power grid connectivity the strategic bottleneck for AI? Grid connectivity delay of 4–10 years vs. 2–3 year data center build timeline. May 2026. https://www.weforum.org/stories/2026/05/electricity-data-grid-connectivity-strategic-bottleneck-ai-transformation/

[15]  Brookings Institution / PowerLines. Global Energy Demands Within the AI Regulatory Landscape. Data center energy consumption approaching 1,050 TWh by 2026; US utilities committed $1.4 trillion in capital spending plans. April 21, 2026. https://www.brookings.edu/articles/global-energy-demands-within-the-ai-regulatory-landscape/

[16]  Sparkline Capital — Kai Wu. Surviving the AI Capex Boom. Historical analysis of telecom buildout: Global Crossing and AT&T spent $500 billion on fiber-optic cable; industry suffered years of glut and collapsing prices. October 2025. https://www.sparklinecapital.com/post/surviving-the-ai-capex-boom

[17]  IEEE ComSoc Technology Blog. Big tech spending on AI data centers vs the fiber optic buildout during the dot-com boom; Paul Kedrosky GDP comparison to railroad and telecom eras. September 2025. https://techblog.comsoc.org/2025/09/27/big-tech-spending-on-ai-data-centers-and-infrastructure-vs-the-fiber-optic-buildout-during-the-dot-com-boom-bust/

[18]  Janus Henderson Investors. AI versus the Dotcom Bubble: 8 reasons the AI wave is different. Analysis of macroeconomic environment, issuer financial strength, and geopolitical competition. October 16, 2025. https://www.janushenderson.com/corporate/article/ai-versus-the-dotcom-bubble-8-reasons-the-ai-wave-is-different/

[19]  Morgan Stanley. Powering AI: Energy Market Outlook 2026. Large technology companies likely to commit more than $1 trillion in the 2025–2026 period; strategic financing as a critical enabler. 2026. https://www.morganstanley.com/insights/articles/powering-ai-energy-market-outlook-2026

[20]  CoBank. AI’s Capital Supercycle Means Big Spend, Bigger Returns. ROIC analysis; U.S. hyperscalers spent ~$400 billion in 2025, likely close to $700 billion in 2026. April 8, 2026. https://www.cobank.com/knowledge-exchange/digital-infrastructure/ais-capital-supercycle-means-big-spend-bigger-returns

[21]  LPL Financial. Rate and Credit View, February 25, 2026. Big tech issued nearly $250 billion in debt in 2025, most on record; Alphabet raised $20 billion including 100-year sterling bond. February 2026. https://www.lpl.com/content/dam/edam/research/publications/rate-and-credit-view/rate-and-credit-view-february-2026.pdf

[22]  SEC EDGAR — Alphabet Form FWP. Alphabet Announces Proposed $80 Billion Equity Capital Raise to Expand AI Infrastructure; includes $10 billion investment from Berkshire Hathaway. June 1, 2026. https://www.sec.gov/Archives/edgar/data/0001652044/000119312526251733/d160205dfwp.htm

[23]  S&P Global Ratings / PRNewswire. Liquidity Outlook 2026: Six Questions, Six Answers — Nicolas Charnay quote on credit market liquidity drivers. February 17, 2026. https://press.spglobal.com/2026-02-17-Private-Credit,-Tech-Issuance-fuelled-by-AI,-and-Increasing-Leverage-Among-Key-Driving-Factors-Impacting-Credit-Market-Liquidity-in-2026-according-to-S-P-Global-Ratings

[24]  Janus Henderson Investors. Mega-issuance and the AI arms race: Big Tech’s impact on credit spreads. September 2025 shift to external capital raising; market granted ‘permission’ based on core business momentum. November 13, 2025. https://www.janushenderson.com/corporate/article/mega-issuance-and-the-ai-arms-race-big-techs-impact-on-credit-spreads/

[25]  InvestmentGrade.com / S&P, Moody’s, Fitch. Bond Ratings Chart: S&P, Moody’s and Fitch Compared. Rating scales with current Q1 2026 issuer examples. April 27, 2026. https://investmentgrade.com/bond-ratings/

[26]  LPL Financial. Rate and Credit View, February 2026. Moody’s and Fitch note AI risks are not uniform across software issuers; overall default exposure in CLOs remains low at around 0.6%. February 2026. https://www.lpl.com/content/dam/edam/research/publications/rate-and-credit-view/rate-and-credit-view-february-2026.pdf

[27]  Reuters / iTiger. LIVE MARKETS: Hyperscaler capex spending brings back memories of 1990s telecom investment boom. Goldman Sachs consensus capex estimates from $467 billion to $533 billion. November 20, 2025. https://www.itiger.com/news/2584923189

[28]  SEC EDGAR — Alphabet 2025 Annual Report (Form ARS). Alphabet acknowledges risks from significant leasing arrangements with third-party operators; large, long-duration commercial agreements as potential liabilities. Filed 2026. https://www.sec.gov/Archives/edgar/data/0001652044/000130817926000344/goog014907-ars.pdf

[29]  Tomasz Tunguz. Is Your AI Funded by Junk Bonds? Barclays downgrade of Oracle debt to underweight; Oracle CDS widened above 125 bps; Oracle carries Baa2 (Moody’s) and BBB (S&P) despite market pricing like junk. December 15, 2025. https://tomtunguz.com/is-your-ai-funded-by-junk-bonds/

[30]  T. Rowe Price. Why the AI Capex Cycle is Built to Persist. Cycle can continue 2–3 more years before first true test; greater risk is monetization lagging investment. January 2026. https://www.troweprice.com/financial-intermediary/us/en/insights/articles/2026/q1/why-the-ai-capex-cycle-is-built-to-persist.html

[31]  Daron Acemoglu, MIT — Fortune. Markets have overestimated AI-driven productivity gains. ‘The problem with the AI bubble isn’t that it is bursting… it’s that the hype will likely go on for a while and do much more damage.’ August 6, 2024. https://fortune.com/2024/08/06/mit-economist-markets-overestimated-ai-driven-productivity-gains

[32]  MIT Economics / MIT Technology Review. Daron Acemoglu: What do we know about the economics of AI? Quote: ‘I’m trying not to be bearish… I believe there are ways we could use generative AI better and get bigger gains.’ 2024–2026. https://economics.mit.edu/news/daron-acemoglu-what-do-we-know-about-economics-ai

[33]  Morningstar / Sparkline Capital — Kai Wu. Why the AI Spending Spree Could Spell Trouble for Investors. ‘The builders of the underlying infrastructure often do not capture much of the value created.’ October 30, 2025. https://www.morningstar.com/markets/why-ai-spending-spree-could-spell-trouble-investors

[34]  RBC Wealth Management. Big Tech’s AI expansion: From investment to scalable returns. General-purpose technologies often require overinvestment before delivering durable benefits. February 13, 2026. https://www.rbcwealthmanagement.com/en-us/insights/big-techs-ai-expansion-from-investment-to-scalable-returns

[35]  TradingView / Invezz — Joshua Mahony. Looking ahead to 2026: why hyperscalers can’t slow spending without losing the AI war. ‘Today’s mega-cap AI valuations assume that the current surge in AI spending is not a one-off infrastructure build.’ 2026. https://www.tradingview.com/news/invezz:751717ae0094b:0-looking-ahead-to-2026-why-hyperscalers-can-t-slow-spending-without-losing-the-ai-war/

[36]  Yahoo Finance / UBS Research. AI capex surge drives higher US investment-grade issuance forecast. UBS raised IG tech supply estimate to $360 billion; year-to-date IG issuance $296 billion, 31% above prior year. 2026. https://finance.yahoo.com/news/ai-capex-surge-drives-higher-150004056.html

[37]  IMF. Global Financial Stability Report, October 2025: Shifting Ground beneath the Calm. Financial stability risks remain elevated; stretched asset valuations and growing pressure in sovereign bond markets. October 2025. https://www.imf.org/en/publications/gfsr/issues/2025/10/14/global-financial-stability-report-october-2025