The Conviction Trade

The Paranoidist | Issue #13 By Paul Morin | May 3, 2026

Four companies that account for roughly 30 percent of the S&P 500 by market value reported earnings within a two-hour window and collectively committed to approximately $725 billion of AI infrastructure spending in 2026, a 77 percent increase over the $410 billion the same four companies spent in 2025. Meta lifted its 2026 capex guidance to $125 to $145 billion, the largest single-year infrastructure commitment in its history. Alphabet raised its full-year guidance to $180 to $190 billion, and CFO Anat Ashkenazi told analysts 2027 capex will "significantly increase" again. Amazon held its $200 billion plan, with Q1 cash capex already at $43.2 billion. Microsoft posted a $627 billion commercial backlog, nearly doubled in a year. Stocks split: Alphabet rose 7 percent on Google Cloud growth of 63 percent, Meta fell as much as 10 percent on the capex raise, Amazon traded down on free cash flow that has collapsed 95 percent year-over-year to $1.2 billion. The split was not about who has demand. It was about who has conviction.

This is a Paranoidist piece about what the headlines did not show. The capex announcements are not the architecture. The architecture is the interlocking obligation chain those announcements have already locked into place across at least eight sectors of the S&P 500, most of which are not technology companies. Boards reading Wednesday's reports as a tech-sector growth story are reading the announcements. They are not reading the architecture, which is the operating environment for the next three to seven years across utilities, real estate, semiconductors, banks, insurers, pension funds, enterprise software, and most downstream sectors that have priced AI productivity into 2026-2028 guidance. The thesis is not that the buildout will fail. The thesis is that it has been described as a capex cycle when it is in fact a multi-party obligation chain, that the chain has only existed in its current shape for about eighteen months, and that it has not been stress-tested by anything other than expected outcomes.

Combined 2026 capex of approximately $725 billion across the four hyperscalers. Goldman Sachs estimates cumulative hyperscaler capex from 2025 through 2027 will reach approximately $1.15 trillion, more than double the $477 billion deployed in 2022 through 2024. KKR estimates that AI-related capex represented approximately 5 percent of U.S. GDP in 2025 and contributed more to growth in the first half of that year than consumer spending did. Industry analyses now project total U.S. tech capex at approximately 7.2 percent of GDP in 2026, exceeding the 6.4 percent peak of the late-1990s telecom buildout. A PowerLines analysis released April 14, 2026 of 51 utilities serving 250 million customers found planned ten-year capital expenditure of $1.4 trillion, up 27 percent from the prior year and effectively double the $700 billion invested over the previous decade.

The cash mechanics tell a different story than the announcements. Aggregate hyperscaler capex now reliably exceeds operating cash flow for the cohort. Amazon's trailing twelve-month free cash flow stood at $1.2 billion in Q1 2026, down from $25.9 billion a year earlier, a 95 percent decrease driven by a $59.3 billion year-over-year increase in property and equipment. Amazon's Q1 2026 net income included a $16.8 billion pre-tax mark-to-market gain on its Anthropic equity stake, a paper gain reflecting Anthropic's higher private-round valuation, which itself depends on the AI capex narrative continuing. Barclays projects Meta's 2026 free cash flow will fall close to 90 percent. Microsoft's free cash flow is projected to drop approximately 28 percent in 2026 before recovering in 2027. The funding gap is being filled by debt: aggregate AI-related debt issuance crossed $200 billion in 2025 and is projected at $250 to $300 billion in 2026 from hyperscalers alone. The first credit default swap contracts referencing private credit funds heavily exposed to AI infrastructure financing began trading in mid-April 2026, with JPMorgan, Barclays, Morgan Stanley, and Citigroup writing contracts on flagship Blackstone, Apollo, and Ares funds.

The hyperscaler backlogs are the architecture's principal collateral and the hardest number on each balance sheet to read. Microsoft's $627 billion commercial Remaining Performance Obligation (under ASC 606, the dollar value of contracts signed but not yet delivered as revenue) was approximately 45 percent concentrated in OpenAI as of Q2 FY26. Alphabet's Google Cloud backlog reached approximately $462 billion in Q1 2026, nearly doubled sequentially. Amazon does not disclose the equivalent figure in the same form, but Andy Jassy stated on Wednesday's call that AWS has "customer commitments for a substantial portion" of planned 2026 capex. These are the numbers that anchor the bull case. They are also subject to renegotiation, restructuring, and right-sizing under conditions that have not yet emerged at scale. They have been tested once, in the DeepSeek event of January 2025, and held. They have not been tested by a sustained six-quarter softening of enterprise AI revenue.

The buildout extends downward from the announced capex through six layers of obligation, each a separate set, each assuming the layers above and below will perform on schedule.

Long-term lease commitments to data center REITs (Equinix, Digital Realty, the private cohort), absorbing a growing share of off-balance-sheet capacity. Power purchase agreements with utilities and IPPs at 20-year terms, including Microsoft's $16 billion Three Mile Island restart, locking in fixed-cost generation regardless of consumption. Chip allocation queues with Nvidia, Broadcom, AMD, TSMC, and the memory cohort, assuming the order books execute approximately on time. Multi-year revenue-share arrangements with foundation model partners, principally OpenAI and Anthropic, whose own funding depends on enterprise AI revenue ramping into the commitments. Corporate debt and private credit, increasingly financing the gap between capex and free cash flow at the hyperscalers and providing working capital underneath the neoclouds (CoreWeave, Nebius, Lambda, IREN, Crusoe). And Stargate, the $500 billion OpenAI-SoftBank-Oracle-MGX joint venture committed to U.S. data center development through 2029, which connects the foundation model layer directly to the power, real estate, and chip layers in a single multi-party obligation.

Each layer is rational. Each assumes the others will perform. The schedule for the entire stack to come into balance, with installed capacity matching demand at the prices required to support the debt service, is approximately 2027 through 2029.

The Pichai sentence on Wednesday's call captured the reigning narrative. "We are compute constrained in the near term. Our cloud revenue would have been higher if we were able to meet the demand." Read on the upside, the constraint is power and demand is unlimited. Read on the downside, it is the Paranoidist concern. If demand softens before power capacity arrives, the constraint inverts. The same Microsoft GPUs that Satya Nadella has acknowledged sit in inventory for lack of power become, in the inverted scenario, the same GPUs that arrive into a market with insufficient demand to absorb the depreciation.

The natural comparison Western directors will reach for is the late-1990s telecommunications buildout. The valid part is the scale. Approximately eighty million miles of fiber were laid between 1996 and 2001. Four years after the dot-com bust, approximately 85 percent of that fiber remained unused, "dark fiber" in industry parlance. Bandwidth pricing fell roughly 90 percent in the early 2000s. Global Crossing, Qwest, Level 3, and WorldCom were either bankrupt, restructured, or absorbed within five years of the peak. Corning fell from approximately $100 to $1. The infrastructure ultimately got used, but at pennies on the dollar of original investment, by companies that had not built it. Netflix and YouTube ran on fiber priced as transformative in 2000 and stranded by 2003.

The invalid part is structural, and the difference is where the actual risk sits. The fiber buildout was speculative on the demand side, with capacity built ahead of contracts. AI infrastructure is being built against contractually committed customers, principally the hyperscalers themselves and a small number of large foundation model and enterprise customers. The hyperscalers are also among the most profitable companies in the world, with the strongest balance sheets on the equity market, in stark contrast to the IPO-fueled, debt-laden telecom companies of 1999. Bankruptcies of the kind that defined 2001-2003 are unlikely.

The fiber comparison teaches the wrong lesson if it is used to ask whether the hyperscalers will fail. They will not. It teaches the right lesson if it is used to ask whether the obligation chain underneath them can sustain a 12 to 24 month softening in enterprise AI revenue ramp. The hyperscalers are not the exposure. The neoclouds whose entire business is GPU rental, the data center REITs whose forward yields require full lease-up at current rates, the utilities whose multi-decade rate cases assume the load growth materializes, the chip companies whose backlogs assume the customers can pay, and the financial institutions accumulating exposure to all of the above are. AI overcapacity, if it materializes, will not destroy the hyperscalers. It will produce a multi-year period of margin compression, asset impairment, and growth deceleration in companies whose current valuations require none of those.

Two days before last Wednesday's earnings, on April 27, Microsoft and OpenAI announced a restructuring of their commercial agreement. Microsoft's licensing of OpenAI intellectual property converted from exclusive to non-exclusive through 2032. Microsoft will cease paying OpenAI revenue shares on certain product categories. OpenAI's revenue share to Microsoft will continue through 2030 with a stated cap. OpenAI gained the ability to use AWS and Google Cloud for core services and enterprise workloads, ending Microsoft's exclusive status as OpenAI's compute provider. The announcement was widely read as a strategic loosening that benefited both parties.

Read structurally, it was the most significant signal in the week's news flow. In February 2026, OpenAI had already cut its multi-year compute spending budget from approximately $1.4 trillion to approximately $600 billion, a reduction that immediately raised questions about the realizability of the OpenAI-related portion of Microsoft's then-$625 billion commercial backlog. The April 27 restructuring redistributed the residual obligation across the cloud cohort. In three months, the largest single anchor obligation in the entire architecture was right-sized twice, quietly, without market dislocation, paired in each case with a structural reframing.

The pattern is the structural concern. The architecture contains many large customer commitments. None has been independently audited at the public level. The Microsoft-OpenAI restructuring is the only one that has been visibly recalibrated. The next test is whichever commitment comes due next without the revenue to support its scheduled draw.

Four sectors with the most mispriced exposure to the architecture, in approximate descending order:

Semiconductors. Nvidia, Broadcom, AMD, TSMC, Micron, and the broader memory and packaging supply chain face order books that assume hyperscaler capex executes on the announced schedules. Directors should be asking what inventory and revenue look like under a 25 percent reduction in 2027 hyperscaler capex (a scenario consistent with applying OpenAI's February 2026 budget revision across other major customers), and whether supply contracts allow for cancellation, deferral, or renegotiation. The chip cohort is the most leveraged single exposure in the architecture.

Utilities, IPPs, and grid infrastructure. Constellation, Vistra, Talen, GE Vernova, and the broader generation universe face load-growth assumptions and rate-base recoveries that depend on data center capacity coming online roughly on schedule. PowerLines' $1.4 trillion ten-year utility capex projection implies a step-change in the rate base of these companies. Directors should be asking what the rate-case posture looks like if data center load materializes at 60 to 70 percent of projection, particularly in concentrated markets such as Northern Virginia, Phoenix, central Ohio, and the Carolinas. Stranded generation in deregulated markets and underrecovered rate base in regulated markets are different problems. Neither is well-priced.

Banks and the financing chain. JPMorgan disclosed approximately $160 billion in non-bank financial institution exposure as of 2025, a category that includes private credit funds with material AI infrastructure positions. The Federal Reserve's April 2026 inquiries to major banks about private credit exposure are a leading indicator. The April 17 launch of CDS contracts on Blackstone, Apollo, and Ares funds is a second one. Directors should be asking what the lookthrough exposure to AI capex looks like across direct hyperscaler debt, neocloud project finance, GPU leasing structures, REIT credit lines, and indirect exposure through private credit fund relationships. The financing chain is the part of the architecture most likely to transmit stress beyond technology if the buildout decelerates.

Enterprise software and downstream sectors that have priced AI productivity into forward guidance. The deepest mispriced exposure is in companies that have told investors AI will drive cost reduction, revenue growth, or margin expansion in 2026-2028 guidance. The MIT August 2025 finding that 95 percent of organizations are getting zero return on generative AI investments is the single most-cited data point on this question. If the productivity narrative ramps slower than the capex schedule, the gap between announced AI capex and announced AI productivity benefits becomes a significant audit issue across the S&P 500. Directors should be asking what their forward guidance assumes about AI productivity, what the supporting evidence is, and what the contingency posture is if productivity lags by 12 to 24 months.

The question across all four sectors is whether the firm is positioned for the announcements to be the architecture, or for the architecture to be the operating environment. The first is a tech-sector growth story. The second is a multi-year, multi-sector restructuring of the U.S. economy that is too far advanced to roll back and not far enough advanced to verify on its sponsors' announced schedule.

The hyperscalers are first-party customers of their own infrastructure to a degree the fiber-buildout participants of 1999 were not. Google Search, YouTube, AWS Bedrock, Microsoft 365 Copilot, Meta's recommendation systems, and Amazon's e-commerce inference workloads represent captive demand absorbing a substantial fraction of installed capacity regardless of external enterprise AI ramp. If captive demand grows at the rate management projects, the dependence on third-party revenue narrows and the obligation chain compresses on its own. I am skeptical of those growth rates, but I cannot rule them out.

The power constraint may moderate the buildout in a way that prevents the worst case rather than producing it. Microsoft's $80 billion Azure backlog due to its inability to deliver power to existing chip inventory is, in this reading, the system's own self-correction mechanism. If hyperscalers physically cannot install GPUs at the rate their announced capex implies, delivered capacity in 2027-2028 will run materially below announced capacity, the depreciation schedule decompresses, and the demand-supply balance is preserved by the inability to build rather than by demand outrunning the build. This is the most persuasive single bull case for the architecture, and it is largely outside management control.

The depreciation framing depends on the useful-life assumption. Amazon disclosed in its Q1 call that data centers carry a 30+ year useful life and that chips, servers, and networking gear carry a 5 to 6 year useful life. If chip-side depreciation runs over 5 to 6 years rather than the 3 to 4 years some analyses assume, the steady-state depreciation expense is materially lower and absorption through earnings is more manageable. The honest answer is that the assumption itself will be tested when the first generation of installed AI silicon reaches its book life and either gets refreshed or written down. That is approximately three to four years out for the earliest cohort.

The DeepStrategy.ai signature method requires sorting what is risk (quantifiable) from what is uncertainty (not quantifiable) at every major analysis. The capex commitments, depreciation schedules, debt service ratios, free cash flow paths, RPO balances, power purchase agreement pricing, chip supply lead times, REIT lease maturity ladders, and bank-and-private-credit exposures are all risks. They can be modeled, stress-tested, and tracked through standard reporting cycles.

The rate at which enterprise AI workloads convert into recurring high-margin revenue at scale; the trajectory of model architecture improvements that may compress or expand the capex requirement per unit of useful AI output; the political and regulatory posture toward energy-intensive AI development in major U.S. jurisdictions; the geopolitical environment for Taiwan-based advanced semiconductor manufacturing; and the integration speed of AI capabilities into the operational core of large enterprise customers are all uncertainties. They cannot be quantified by current methods. They can be bounded and integrated into scenario planning. They cannot be reduced to a probability distribution.

Boards that conflate the two are at the highest structural risk. The capex announcements are quantifiable to the dollar. The demand assumptions underlying them are not. The architecture has been priced as if both were as quantifiable as the capex. When uncertainty is priced as risk, the system in question is not pricing the variable that will determine its outcome.

The architecture is approximately eighteen months old in its current shape. It is the largest coordinated industrial buildout in U.S. economic history, financed substantially with debt, executing on schedules that are visible at the top of the stack and increasingly opaque at the bottom. It will not fail catastrophically. It may compress, restructure, and stress-test selectively over the 2026-2029 window in ways that affect boards across at least eight sectors of the S&P 500.

The institution that consumes the analytical process as preparation for multiple futures has what the forecast cannot provide: adaptability. The boards that will be best-positioned in 2027 and 2028 are not the ones reading the announcements. They are the ones auditing the architecture, sector by sector, layer by layer, ahead of the test.

The Paranoidist publishes weekly, with flash issues when events warrant.

If this changed how you think about one thing, consider subscribing. If it didn't, tell me what I'm missing.

Paul Morin is the founder of DeepStrategy.ai, author of Uncertainty: When Risk Is Not Enough (a guide to decision-making when probabilities fail), and publisher of The Paranoidist, BoardroomRadar, and ScenarioWatch. He has spent more than three decades in entrepreneurship, finance, risk management, and insurance, which is why he worries about the things that keep other people awake at night.

Researched, written, and edited in collaboration with Claude by Anthropic.