Introduction: Driving the Next Industrial Wave — The Ghost of Webvan and the Prophecy of Premature Infrastructure
In 1999, at the feverish height of the dot-com era, an ambitious startup named Webvan promised to revolutionize the grocery industry by delivering food straight to consumers’ doors. The company was not a fly-by-night operation staffed by dreamers; it was backed by the most sophisticated venture capital firms of its generation, championed by celebrated executives, and armed with what was, at the time, one of the most aggressive infrastructure programs ever attempted by a private consumer company. Webvan raised hundreds of millions of dollars and immediately converted that capital into physical assets: massive, highly automated fulfillment warehouses packed with advanced robotics, conveyor systems, and refrigerated logistics corridors. It signed a one-billion-dollar contract with Bechtel to build twenty-six state-of-the-art distribution centers across the United States, each one designed to serve a metropolitan region that had not yet demonstrated any durable appetite for online grocery shopping.[1]
The problem was never the vision. The problem was the clock. Consumer habits had not yet caught up to online delivery. Broadband penetration was thin, smartphones did not exist, digital payments were clumsy, and the logistics software required to route thousands of perishable deliveries per day was still primitive. Webvan burned through its capital at an astonishing rate, expanded into new cities before proving unit economics in its first, and declared bankruptcy in 2001, liquidating warehouses that had cost tens of millions of dollars each for pennies on the dollar.[2] It became, for a generation of investors, the canonical cautionary tale about building infrastructure ahead of demand.
And yet, the more interesting half of the Webvan story is the half that is usually left out. While the company itself failed, its core thesis was right — it was simply too early. Decades later, as digital literacy exploded, smartphones became ubiquitous, cloud computing collapsed the cost of logistics software, and pandemic-era behavior permanently normalized home delivery, services like Amazon Fresh and Instacart stepped directly into the operational and infrastructure templates that Webvan pioneered. Online grocery delivery became a multi-billion-dollar staple of modern commerce, executed by companies that inherited a market whose earliest, most expensive lessons had been paid for by somebody else.[3] The physical warehouses were gone, but the intellectual infrastructure — the demand models, the automation designs, the painful knowledge of what breaks at scale — endured and compounded. Overbuilding did not kill the idea. Overbuilding subsidized the idea’s eventual triumph.
This paper argues that today’s technology giants are racing through an identical cycle at a scale that dwarfs anything the dot-com era ever attempted. Hyperscalers and frontier AI laboratories are pouring hundreds of billions of dollars per year into graphics processing units (GPUs), specialized data centers, power infrastructure, and cooling systems. According to first-quarter 2026 earnings compiled by the Financial Times, Google, Amazon, Microsoft, and Meta alone plan to spend roughly $725 billion on capital expenditures in 2026 — up 77 percent from the record $410 billion they deployed in 2025.[4] OpenAI’s Stargate platform, a joint venture with Oracle and SoftBank announced at the White House, has by itself disclosed nearly 7 gigawatts of planned capacity and more than $400 billion in committed investment over three years across sites in Texas, New Mexico, Ohio, and the Midwest, on its way to a full $500 billion, 10-gigawatt commitment.[5] They are betting everything on artificial intelligence, and history suggests that a capital surge of this velocity will inevitably overshoot immediate demand.
The early tremors of that overshoot are already visible in the second quarter of 2026, and they are visible precisely where economic theory says they should first appear: in the behavior of the builders themselves. Meta — a company with no legacy cloud business and no external customers for its silicon — is now reportedly standing up a cloud infrastructure unit, internally dubbed Meta Compute, to sell excess AI computing capacity to outsiders, and Reuters and CNBC reported that Mark Zuckerberg told shareholders in May 2026 that entering cloud computing was “definitely on the table” if the company found itself overbuilt.[6] Even more strikingly, Elon Musk’s Colossus data center complex — originally constructed by xAI to train its own Grok models — began renting out its overcapacity to rivals in 2026: Anthropic signed a deal in May to lease the entire output of Colossus 1 for $1.25 billion per month, and Google followed in early June at $920 million per month, both agreements landing in the weeks before and in direct anticipation of SpaceX’s initial public offering on June 12, 2026.[7] When the most aggressive builders in the industry begin monetizing idle capacity by renting it to their fiercest competitors, the market is telling us something profound about the relationship between supply and demand.
“These models are being hyped up, and we’re investing more than we should.”
— Daron Acemoglu, Institute Professor, MIT; 2024 Nobel Laureate in Economics [8]
When the overshoot fully arrives, the resulting overcapacity will reshape the global economy. It will create a massive, institutional secondary market for compute power; it will redefine the financial playbook for technological infrastructure; and — like the fiber-optic glut of 2001 and the ghost warehouses of Webvan — it will quietly lay the foundation for the next generation of adoption at a marginal cost approaching zero.
Why This Framework Is Named the “AI Infrastructure Boom”
A note on naming is warranted before the analysis begins, because the choice of the phrase “AI Infrastructure Boom” as the organizing framework of this paper is deliberate, and each of its three words carries analytical weight.
First, “AI” rather than “technology” or “cloud.” The capital cycle described in this paper is not a general technology upgrade cycle. It is monomaniacally concentrated on a single workload class — the training and inference of large artificial intelligence models — and on a single dominant hardware vendor whose accelerators absorb the majority of incremental spending. That concentration is precisely what makes the cycle fragile: when capital chases one workload, one chip architecture, and one demand thesis, any deflationary shock to that thesis propagates through the entire asset base simultaneously, rather than being absorbed by a diversified portfolio of uses.
Second, “Infrastructure” rather than “investment” or “spending.” What is being purchased is not software, marketing, or talent — categories that can be dialed down within a quarter. It is land, concrete, steel, substations, transformers, twenty-year power purchase agreements, and silicon bolted into racks. Infrastructure is illiquid, site-specific, slow to build, and slow to unwind. Historically, it is exactly this asset class — railroads in the 1870s, electrification in the 1920s, telecom fiber in the 1990s — that produces the deepest overcapacity cycles, because supply decisions made under scarcity arrive in the market years later, all at once, under conditions of abundance. Naming the framework around infrastructure signals that the analysis belongs to the long tradition of capital-cycle economics, not to a short-run market-timing debate.
Third, “Boom” rather than “bubble.” This is the most important distinction, and it defines the intellectual posture of the entire paper. A bubble is a claim about prices being wrong; a boom is a claim about capital arriving faster than demand can absorb it. Booms can occur around technologies that are entirely real and ultimately transformative — indeed, they usually do. The railroads were real, electricity was real, the internet was real, and artificial intelligence is real. This paper does not argue that AI will fail; it argues that the capital deployed to serve AI is temporarily and predictably outrunning monetizable demand, that the gap will resolve through overcapacity rather than collapse, and that the overcapacity itself will become the cheap foundation of the next adoption wave — exactly as Webvan’s failure became Instacart’s inheritance. “Boom” preserves the double meaning the evidence demands: an era of extraordinary construction, and the sound the excess makes when it lands.
Thesis and Roadmap
The argument of this paper proceeds in six movements. Section 1 documents the scale and structure of the capital expenditure overdrive, using corporate earnings reports through the first quarter of 2026 to show that the hyperscalers have entered a spending regime with no precedent in corporate history — one that is already consuming their free cash flow and re-leveraging their balance sheets. Section 2 dissects the four deflationary demand shockwaves — algorithmic efficiency, regulatory friction, enterprise monetization failure, and the nuclear-powered energy surge — that are converging on the boom from four different directions and that will, in combination, cause supply to overshoot demand. Section 3 evaluates the financial anatomy of stranded compute: margin compression, asset write-downs, the increasingly contested accounting of GPU depreciation, and the credit-market pressure now flagged by the International Monetary Fund. Section 4 maps the emerging secondary market for compute — already visible in Meta’s cloud pivot and SpaceX’s Colossus leases — and explains why compute will commoditize into a spot-traded resource resembling oil or electricity. Section 5 translates the analysis into strategic de-risking playbooks for financial officers, infrastructure operators, and enterprise buyers. Section 6 distills the argument into seven strategic pillars, and the Conclusion returns to Webvan to explain why the coming overcapacity, painful as it will be for the overextended, is best understood not as a catastrophe but as a predictable economic rite of passage — the mechanism by which the marginal cost of intelligence is driven toward zero.
Section 1: Tracking Capital Expenditure Overdrive
Every capital cycle begins with a rational decision made under scarcity. In 2023 and 2024, the scarce resource was compute: frontier AI laboratories could not buy enough accelerators, cloud providers turned away enterprise customers, and waiting lists for Nvidia’s flagship GPUs stretched beyond a year. Under those conditions, the strategic logic of the hyperscalers was airtight — being short of compute was existentially more dangerous than overspending on it, because a rival with more compute could train a better model, capture the developer ecosystem, and convert a temporary hardware advantage into a durable platform monopoly. The problem with airtight logic is that every player in the industry reasoned identically, at the same time, with effectively unlimited balance sheets. The result, visible in full by the first-quarter 2026 earnings season, is the largest concentrated private capital deployment in economic history.
1.1 Aggressive Capital Outlays: From $200 Billion Annually to $725 Billion and Beyond
The trajectory of hyperscaler capital expenditure has broken free of every historical anchor. In 2024, the four largest cloud and platform companies — Amazon, Microsoft, Alphabet, and Meta — spent roughly $228 billion combined, itself a record that alarmed analysts at the time. In 2025 that figure climbed to approximately $410 billion. For 2026, according to guidance issued during the fourth-quarter 2025 and first-quarter 2026 earnings calls and compiled by the Financial Times, the four companies plan combined capital expenditures of roughly $725 billion — a 77 percent year-over-year increase on top of a year that had itself nearly doubled the prior one.[4] Adding Oracle’s roughly $50 billion program brings the five-company total toward the $700 to $900 billion range that credit-research firm CreditSights projects for 2026.[16]
The company-level guidance, disclosed in earnings calls between late January and April 2026, is worth recording precisely, because these figures constitute the primary evidence base for everything that follows.[9] Amazon guided to approximately $200 billion in 2026 capital expenditures, up from $125 billion in 2025, and disclosed $44.2 billion of spending in the first quarter of 2026 alone as AWS revenue grew 28 percent. Alphabet guided to a range of $175 to $185 billion, up from $91 billion — a near-doubling in a single year for the world’s fourth-largest company, later supplemented by an $85 billion equity raise to help finance the buildout. Microsoft, whose chief financial officer Amy Hood attributed roughly $25 billion of incremental spending to surging memory-chip and component prices, set calendar-year 2026 capital expenditure at approximately $190 billion, far above the $152 billion consensus, while telling investors the company would remain capacity-constrained through at least the end of 2026.[4] Meta initially guided to $115 to $135 billion, then raised the range toward $125 to $145 billion in April 2026, citing higher memory prices and additional data-center construction — roughly double its $72 billion of 2025 spending, for a company that, unlike its three peers, has no external cloud business through which to monetize the hardware.[10]
Table 1. Hyperscaler capital expenditure, actual and guided, 2024–2026 (US$ billions)
| Company | 2024 (actual) | 2025 (actual) | 2026 (guidance) | Primary monetization channel |
| Amazon | ~78 | ~125 | ~200 | AWS cloud, Trainium custom silicon, Bedrock |
| Microsoft | ~56 | ~90 | ~190 [4] | Azure cloud, Copilot, OpenAI partnership |
| Alphabet (Google) | ~52 | ~91 | 175–185 | Google Cloud, Gemini, TPUs, ads |
| Meta | ~39 | ~72 | 125–145 (raised April 2026) [10] | Advertising only; cloud business in planning |
| Oracle | ~7 | ~21 | ~50 | OCI, Stargate lease revenue from OpenAI |
| Big Four total | ~228 | ~410 | **~725 [4]** | — |
Sources: company earnings releases and guidance, Q4 2025–Q1 2026; Financial Times compilation; Futurum Group; Data Center Frontier.[4], [9], [10]
Nor does the guidance appear to represent a peak. Following the first-quarter 2026 earnings season, Goldman Sachs raised its cumulative capital expenditure forecast for the four largest hyperscalers to $5.3 trillion for fiscal years 2025 through 2030 — up from $4.5 trillion before the quarter began — and modeled a baseline aggregate of $7.6 trillion between 2026 and 2031 across compute, data centers, and power.[12] Analysts at Evercore and Bank of America project that combined hyperscaler capital expenditure will clear $1 trillion annually in 2027.[14] To put these figures in historical perspective: the entire Apollo program cost roughly $280 billion in today’s dollars, spread across thirteen years. The hyperscalers now deploy the equivalent of an Apollo program approximately every twenty weeks.
The bulls, it must be said, are not without evidence. Google Cloud revenue jumped 63 percent year over year to $20 billion in the first quarter of 2026; AWS grew 28 percent; Azure grew roughly 40 percent.[4] Backlogs are enormous, and enterprise commitments continue to build. When the Financial Times surveyed the earnings season, Jefferies analyst Brent Thill dismissed the skeptics in unusually blunt terms:
“The bear thesis is garbage.”
— Brent Thill, Analyst, Jefferies, to the Financial Times, April 2026 [4]
This paper takes the bull evidence seriously — cloud revenue growth is real, and Section 6 will argue that the infrastructure ultimately compounds in value. But revenue growth and capital discipline are different questions. The issue is not whether AI generates revenue; it is whether revenue can arrive fast enough, at high enough margins, to service a capital base that is doubling annually. That is a question of arithmetic, and the arithmetic is examined in Sections 2 and 3.
1.2 Concentrated Hardware Bets: The Nvidia Chokepoint and the Memory Squeeze
A defining structural feature of this cycle — and a key difference from the diversified telecom buildout of the 1990s — is the extreme concentration of spending on a single vendor’s silicon. A massive share of incremental hyperscaler capital flows directly to Nvidia GPUs and the systems built around them, supplemented by custom accelerators (Google’s TPUs, Amazon’s Trainium, Meta’s MTIA, Microsoft’s Maia) that remain, for now, minority workloads. This concentration produces three second-order fragilities.
First, it synchronizes the industry’s depreciation clock. When every builder buys the same chip generation in the same window, every builder’s fleet ages simultaneously, and every builder faces the same obsolescence cliff when the next architecture ships. Nvidia has compressed its release cadence to an annual rhythm — Hopper in 2022, Blackwell in 2024, Blackwell Ultra in 2025, Vera Rubin arriving in 2026, and Rubin Ultra targeted for 2027 — with each generation delivering step-function gains in performance per watt and per dollar.[28] The faster the cadence, the shorter the economic life of the installed base, a problem examined in depth in Section 3.
Second, concentration transmits input-cost inflation instantly across the entire industry. The clearest 2026 example is memory: surging prices for high-bandwidth memory and DRAM pushed spending forecasts higher at both Microsoft — whose CFO attributed $25 billion of incremental 2026 capex to component costs — and Meta, which cited memory prices in raising its guidance range.[4] Industry analysis suggests memory alone will consume roughly 30 percent of hyperscaler data-center spending in 2026. When one supply chain feeds one workload for five buyers, there is no diversification benefit anywhere in the system; every shock is systemic.
Third, and most subtly, concentration creates circularity in the demand signal itself. Nvidia has committed up to $100 billion of investment into OpenAI, whose Stargate program then purchases Nvidia systems; Oracle finances data centers whose anchor tenant is OpenAI, whose payments then justify Oracle’s capex; SoftBank funds Stargate while selling Nvidia stock to do so. Critics have questioned this circular financing, in which suppliers invest in their own customers’ buildouts, because it makes it genuinely difficult to determine how much demand is organic and how much is vendor-financed.[19] MIT’s Daron Acemoglu warned that the danger of such interlocking arrangements is that they risk collapsing like a house of cards — a characterization the industry disputes but cannot yet refute with independent, arm’s-length demand data.[8]
1.3 Accelerated Depreciation Risks: Twenty-Year Buildings, Three-Year Chips
AI infrastructure is a marriage of two asset classes with radically mismatched lifespans. The physical facility — the shell, the land, the substations, the cooling plant — is a twenty-to-thirty-year asset, financeable like real estate and genuinely durable. The silicon inside it is not. Frontier-training GPUs face economic obsolescence within roughly two to three years as each new Nvidia generation renders its predecessor uncompetitive for cutting-edge work; Meta’s own published study of Llama 3 training documented meaningful annualized hardware failure rates under sustained full-power operation, underscoring that these chips are consumables run hot, not durable equipment.[28] Yet the hyperscalers currently depreciate GPU servers over four to six years — Microsoft extended its schedule from four years to six, and Meta adopted five and a half — an accounting posture that flatters current earnings and that Section 3 will examine as one of the central financial controversies of the boom.[29]
The mismatch is not merely an accounting curiosity; it changes the economics of the entire buildout. If the shell is a twenty-year asset but its contents must be replaced every three, then what looks on the balance sheet like growth capex is substantially disguised maintenance capex — a permanent, recurring obligation rather than a one-time investment. Even Microsoft’s chief executive has acknowledged engineering the company’s buildout around this reality, describing a deliberately staggered, generation-hopping deployment approach precisely to avoid concentrating exposure to any single chip generation:
“I didn’t want to go get stuck with four or five years of depreciation on one generation.”
— Satya Nadella, CEO, Microsoft [15]
When the industry’s largest and most sophisticated buyer publicly designs its capital program to dodge its own depreciation schedules, investors are entitled to ask what those schedules are really telling them.
1.4 Capacity Utilization Pressures: Building Ahead of Contracted Revenue
The final structural risk is sequencing. Mass construction is occurring before long-term enterprise software revenue is locked in. Unlike the utility industry — where a power plant is typically financed only after decades-long offtake agreements are signed — the AI buildout is largely speculative construction underwritten by internal demand forecasts. The revenue side of the ledger is growing, but it is growing from a small base relative to the capital: pure-play AI vendors led by OpenAI and Anthropic are posting rapid growth, yet their combined revenues remain a fraction of the infrastructure investment being deployed on their behalf.[10] J.P. Morgan estimates that the industry must generate approximately $650 billion in new annual revenue merely to achieve a 10 percent return on the infrastructure currently being built.[19]
The strain is now visible in the purest measure of corporate health: cash. Reaching the 2026 spending targets implies a dramatic contraction in free cash flow across the group. Alphabet’s first-quarter 2026 free cash flow fell 47 percent year over year to $10.1 billion; Amazon’s trailing free cash flow collapsed roughly 95 percent, from approximately $38 billion to $1.2 billion, with projections that it turns negative during 2026.[14] According to Epoch AI’s analysis of SEC filings, aggregate operating cash flow across the five largest builders is growing at roughly 23 percent annually while aggregate capital expenditure grows at roughly 70 percent — two curves that cross around the third quarter of 2026, after which the group as a whole spends more building infrastructure than its operations generate. Oracle has already crossed, with capital spending exceeding operating cash flow by roughly $24 billion.[14] As Longbow Asset Management’s chief executive put the matter with disarming simplicity:
“If you’re going to pour all this money into AI, it’s going to reduce your free cash flow.”
— Jake Dollarhide, CEO, Longbow Asset Management [13]
Companies that can no longer self-fund turn to debt and equity markets, and they have: the financing shift is documented in Section 3.4. For now, the essential point of Section 1 is structural. The industry has committed to a capital program that (a) is concentrated on one vendor’s rapidly obsolescing silicon, (b) marries thirty-year facilities to three-year chips, (c) runs ahead of contracted revenue, and (d) has begun consuming the free cash flow that historically made these companies the safest credits in the world. That is the posture of the boom at its apex. The next section examines the four forces converging to knock demand out from under it.
Section 2: Dissecting the Four Deflationary Demand Shockwaves
If Section 1 described a supply curve shifting outward at historic velocity, this section describes the demand side of the ledger — and the four distinct, overlapping forces that threaten to erode the projected demand for raw compute power faster than the buildout can absorb. Each force operates on a different timescale and through a different mechanism, which is precisely what makes them dangerous in combination: hedging against one provides no protection against the others. The first shockwave is technological, the second is institutional, the third is economic, and the fourth — counterintuitively — is the resolution of the industry’s most celebrated constraint. Together, they form the analytical core of the overcapacity thesis.
Table 2. The four deflationary demand shockwaves: mechanism, timescale, and evidence
| Shockwave | Mechanism | Timescale | Leading evidence (2024–2026) |
| 2.1 Efficiency gains | Algorithmic progress reduces GPUs needed per unit of output | Continuous; compounding | Inference cost of GPT-3.5-class output fell ~280× in two years; token prices falling at a median of ~50× per year across capability milestones [17], [18] |
| 2.2 Regulatory bottlenecks | Compliance friction slows enterprise deployment in regulated sectors | 2–5 years | EU AI Act phase-in; copyright litigation; state-level AI statutes; data-residency mandates |
| 2.3 Monetization friction | Pilots fail to scale to profitable production; spending pullback | 1–3 years | 95% of enterprise GenAI pilots show no measurable P&L impact (MIT); only 39% of firms report any EBIT impact (McKinsey) [18] |
| 2.4 Energy surge | Nuclear/SMR deals dissolve the power bottleneck; stalled supply floods in | 3–8 years | >13 hyperscaler nuclear deals, ~9.8 GW committed; 127 GW of firm data-center capacity contracted vs. ~108 GW of projected 2030 demand [23], [27] |
Source: author’s synthesis of sources cited in Sections 2.1–2.4.
2.1 Foundational Model Efficiency Gains: The Deflation Machine Inside the Boom
The first and most relentless shockwave comes from inside the laboratories themselves. Engineers are rapidly learning to do more with less, and the pace of that learning has no precedent in the history of industrial technology. Algorithmic breakthroughs — quantization, which compresses model weights to lower numerical precision with minimal quality loss; distillation, which trains small models to replicate the behavior of frontier ones; speculative decoding, mixture-of-experts routing, sparse attention, and improved serving infrastructure — allow smaller and cheaper systems to match performance that once required frontier-scale hardware. Each of these techniques individually delivers a two-to-four-fold efficiency gain; because they compound multiplicatively, their combined effect has been staggering.
The measured numbers deserve emphasis, because they are among the most important economic statistics of the decade. Stanford University’s AI Index reports that the inference cost of achieving GPT-3.5-level performance fell more than 280-fold between November 2022 and October 2024 — hardware costs declining roughly 30 percent annually and energy efficiency improving about 40 percent per year, with algorithms doing the rest.[17] Epoch AI’s broader analysis finds that large-language-model inference prices have fallen between 9-fold and 900-fold per year depending on the capability milestone measured, with a median decline of roughly 50-fold annually — accelerating to a median of roughly 200-fold per year since January 2024.[18] The 2025 release of DeepSeek’s V3 model, whose final training run consumed only about $5.6 million of GPU compute, triggered an industry-wide price war that demonstrated how quickly training economics, and not just inference economics, can deflate.[18]
The implication for infrastructure planning is direct and uncomfortable. A data center is a bet on the future price of the service it produces, and the service that AI data centers produce — tokens of machine intelligence — is deflating faster than any commodity in recorded economic history. If enterprise workload demand grows 10-fold while the compute required per workload falls 50-fold, aggregate GPU demand for that workload declines by four-fifths even as usage explodes. This drastically reduces the number of GPUs needed to run identical enterprise workloads year over year, and it means the buildout must be justified not by today’s model architectures but by a perpetual faith that new, more compute-hungry frontier workloads will always arrive in time to refill the racks that efficiency empties.
The industry’s counterargument is Jevons’ paradox: cheaper intelligence induces more consumption of intelligence, just as cheaper lighting induced more illumination. There is real evidence for this — inference now accounts for roughly two-thirds of all AI compute demand, up from one-third in 2023, as consumer and agentic applications scale.[50] But Jevons’ paradox is an empirical claim about elasticity, not a law of nature, and it has a threshold condition: induced demand must grow faster than efficiency deflates cost, forever, across a capital base that doubles annually. Section 2.3 examines whether enterprise demand is in fact behaving that way. The evidence to date suggests it is not.
2.2 Regulatory and Governance Bottlenecks: The Institutional Drag Coefficient
The second shockwave operates through institutions rather than algorithms, and it is chronically underweighted in infrastructure forecasts precisely because it is unglamorous. Compliance friction slows enterprise deployment, and enterprise deployment is where the demand projections that justify the buildout ultimately have to be realized. The friction arrives through at least four channels.
First, comprehensive AI statutes are phasing in across major markets. The European Union’s AI Act imposes conformity assessments, documentation, transparency, and human-oversight requirements on high-risk systems, with obligations phasing in through 2026 and 2027; a growing patchwork of American state laws governs automated decision-making, algorithmic discrimination, and disclosure; and sectoral regulators in finance, healthcare, and insurance have issued guidance that effectively requires model-risk-management regimes before deployment. None of this prohibits AI — but all of it converts deployment from an engineering sprint into a multi-quarter legal program, and capital-intensive infrastructure earns nothing while its customers’ lawyers deliberate.
Second, strict data privacy and data-residency laws prevent the frictionless pooling of enterprise data that many AI business cases quietly assume. Cross-border transfer restrictions, sovereign-cloud requirements, and health- and financial-privacy statutes force fragmented, jurisdiction-specific deployments that raise unit costs and shrink addressable workloads. Third, copyright litigation over training data continues to generate legal uncertainty that makes risk-averse general counsels reluctant to embed generative systems in revenue-critical or customer-facing processes — several landmark settlements have established that training-data provenance carries a price, and that price is still being discovered. Fourth, safety mandates and internal AI governance boards — now standard in large enterprises — add review cycles that stretch procurement timelines from weeks to quarters.
The macroeconomic institutions have begun to model this drag explicitly. The International Monetary Fund’s 2026 scenario-planning exercise on the global economic and financial implications of artificial intelligence frames the technology as a macro-critical transition whose outcomes depend on diffusion speed, institutional readiness, financial stability, and global coordination — a formal acknowledgment that adoption, not capability, is the binding variable.[20] For infrastructure economics the translation is simple: as adoption stalls in heavily regulated sectors — which happen to include the highest-value enterprise workloads in banking, insurance, healthcare, and government — raw compute demand plateaus in exactly the segments the demand forecasts assumed would grow fastest.
2.3 Enterprise Monetization Friction: The Pilot Purgatory Problem
The third shockwave is the most immediately measurable, because it is already in the survey data. Corporate proof-of-concepts are struggling to scale into profitable production, and the gap between AI enthusiasm and AI-attributable earnings has become one of the defining puzzles of the 2024–2026 period.
The headline statistic came from MIT’s Project NANDA, whose study of more than 300 enterprise deployments found that 95 percent of enterprise generative-AI pilots failed to deliver measurable profit-and-loss impact — only 5 percent of projects created measurable financial value within the study window.[18] The figure requires careful handling: it measured rapid P&L impact within roughly six months, concentrated in sales and marketing pilots, and its most actionable finding — that vendor-led deployments succeeded at roughly twice the rate of internal builds — was largely lost in the coverage.[21] But even the more charitable large-sample surveys tell a sobering story: McKinsey’s State of AI research finds that only 39 percent of organizations report any enterprise-level EBIT impact attributable to AI, and only about 6 percent qualify as high performers attributing more than 5 percent of EBIT to the technology.[18] Meanwhile, 79 percent of enterprises report AI cost overruns in the trailing twelve months.[18]
Beneath the surveys sits a unit-economics problem. Many businesses discover that the fully loaded cost of running large language models — token costs, retrieval infrastructure, evaluation, human review, integration, and governance — outweighs the marginal productivity gains for their specific workflows, particularly where error tolerance is low. The result is a predictable corporate behavior pattern: enthusiastic pilot, ambiguous measurement, budget scrutiny, and a quiet pullback in software spend that never appears in a press release. Even leading infrastructure executives concede the pricing problem: Palo Alto Networks’ chief executive argued publicly in 2026 that AI pricing needs to fall on the order of 90 percent as token consumption in agentic workloads skyrockets — an admission that current enterprise economics do not close at current prices.[6]
The most institutionally weighty statement of skepticism remains that of Goldman Sachs’ head of global equity research, whose report title — “Gen AI: Too Much Spend, Too Little Benefit?” — became shorthand for the entire bear case, and who framed the cost problem at its root:
“AI technology is exceptionally expensive.”
— Jim Covello, Head of Global Equity Research, Goldman Sachs [22]
Covello’s fuller argument is that to earn back a trillion-dollar buildout, the technology must solve complex, high-value problems rather than merely accelerating existing tasks, and that efficiency gains that are available to every competitor are competed away rather than captured as margin.[22] One need not accept his most pessimistic conclusions — this paper does not — to accept the narrower point that matters for infrastructure: enterprise monetization is arriving later, lumpier, and lower-margin than the capital plans assumed. It is worth noting the honest breadth of expert disagreement here: Stanford’s Erik Brynjolfsson, among the most rigorous empirical optimists, finds large and growing labor-market effects from AI and argues the productivity gains are real but require organizational redesign to surface — a position he defends with nearly four years of payroll data even while conceding how little the profession can yet see:
“We are flying blind into one of the most consequential periods in world history.”
— Erik Brynjolfsson, Professor and Director, Stanford Digital Economy Lab [45]
Between Acemoglu’s minimalism — his published estimate puts AI’s cumulative U.S. productivity gain at roughly 0.7 percent over a decade — and Brynjolfsson’s measured optimism lies the entire range of defensible forecasts, and the capital plans of Section 1 are consistent only with the most aggressive end of that range.[51] When $725 billion of annual spending requires the optimistic tail of the academic distribution to be correct, the spending, not the academy, carries the burden of proof.
2.4 The Nuclear and SMR Energy Surge: How Solving the Bottleneck Creates the Glut
The fourth shockwave is the strangest, because it arrives disguised as good news. Throughout 2024 and 2025, the consensus constraint on the AI buildout was not capital, chips, or even demand — it was electricity. Roughly 40 percent of announced AI data-center projects faced construction delays attributable to power infrastructure rather than hardware; the U.S. grid-interconnection queue swelled beyond 2,600 gigawatts of waiting projects with multi-year delays; and transformer shortages became boardroom vocabulary.[24] Power scarcity functioned, in effect, as an involuntary discipline mechanism: it rationed how much capacity could physically come online, regardless of how much capital wanted to build.
The industry’s response has been the most consequential corporate energy program in half a century: the rapid integration of nuclear energy and Small Modular Reactors (SMRs) into data-center supply chains. As of mid-2026, every major hyperscaler has signed at least one nuclear deal, with a tracked total of 13 announced projects committing roughly 9.8 gigawatts of nuclear capacity to AI infrastructure.[23] Microsoft anchored the movement with a 20-year, $16 billion power purchase agreement for the restart of Three Mile Island Unit 1 — renamed the Crane Clean Energy Center — an 835-megawatt facility roughly 80 percent staffed in early 2026 and expected to deliver the first nuclear electrons dedicated to AI in 2027.[23] Amazon invested $700 million in SMR developer X-energy, committed to purchase up to 5 gigawatts of its output by 2039, and is deploying more than $20 billion into a nuclear-adjacent campus at the Susquehanna plant; X-energy itself raised $1.02 billion in an April 2026 Nasdaq IPO that priced well above its range — a market signal that investors now treat SMRs as financeable infrastructure rather than science projects.[25] Google contracted with Kairos Power for a fleet of 500 megawatts of advanced reactors; Meta leads the field with commitments of up to 6.6 gigawatts across TerraPower, Oklo, Vistra, and Constellation; and Oracle has disclosed designs for a gigawatt-scale campus powered by three SMRs behind the meter, with Stargate’s later phases explicitly evaluating SMR siting.[23], [24]
Now consider what happens when this program succeeds. Energy constraints do not vanish gradually and smoothly; they vanish in cliffs, as restarts, uprates, gas peakers, grid upgrades, and eventually SMR fleets come online in discrete blocks between 2027 and the early 2030s. Every data-center project currently stalled in the interconnection queue — the majority of announced capacity, by the IMF’s estimate roughly 60 percent of planned facilities have not yet broken ground — is a coiled spring of supply waiting for electrons.[32] When the power arrives, a massive wave of previously stalled capacity hits the market nearly simultaneously. The forward numbers already show the shape of the problem: CreditSights analysts estimate U.S. data-center power demand grows to roughly 108 gigawatts by 2030, yet by the end of the first quarter of 2026, utilities had already contracted approximately 127 gigawatts of new data-center capacity deemed firm or committed — which, added to roughly 45 gigawatts of existing capacity, implies potential 2030 supply of about 172 gigawatts against 108 gigawatts of demand: a possible overbuild on the order of 64 gigawatts.[27] At construction costs that CBRE Investment Management places at $10 to $12 billion per gigawatt before land, servers, and chips, a 64-gigawatt overshoot represents roughly $640 billion of excess buildout before a single GPU invoice is paid.[27]
This is the deep irony at the heart of the boom: the industry’s most impressive achievement — conjuring a nuclear renaissance to feed its data centers — is also the mechanism that converts power scarcity, the only force currently pacing the buildout, into power abundance, the force that will unleash it. Scarcity disciplines capital cycles; abundance completes them. When the four shockwaves are assembled — efficiency deflating compute-per-workload, regulation slowing workload adoption, monetization friction capping workload budgets, and energy abundance releasing the supply dam — the conclusion is difficult to escape: sometime in the 2027–2029 window, the AI infrastructure complex crosses from undersupply into structural overcapacity. Section 3 examines what that crossing does to the financial statements of the companies that built it.
Section 3: Evaluating the Financial Implications of Stranded Compute
When supply outpaces demand in an infrastructure industry, the financial ledger changes character with brutal speed. Assets that were strategic weapons under scarcity become fixed-cost liabilities under abundance; accounting assumptions that looked conservative under growth look aggressive under saturation; and debt that was cheap against appreciating collateral reprices against depreciating collateral. This section traces that transformation through four channels — margin compression, asset write-downs, trapped capital, and credit pressure — and gives particular attention to the controversy that moved from accounting footnotes to front pages during the winter of 2025–2026: the useful life of a GPU.
3.1 Severe Margin Compression: The Arithmetic of Idle Silicon
A data center is among the purest fixed-cost businesses ever devised. Once built, its costs — depreciation, power-capacity charges, staffing, cooling, security, debt service — are almost entirely invariant to how much work it performs. Its revenue, by contrast, is entirely variable, priced per GPU-hour or per token. This operating leverage is magnificent on the way up: every incremental workload above breakeven utilization falls through to margin nearly intact. It is merciless on the way down: unused data centers generate massive fixed overhead without generating operational revenue, and every point of utilization lost subtracts revenue while subtracting almost no cost.
The industry has already produced a natural experiment demonstrating what sub-scale utilization looks like. Before Anthropic leased the entirety of SpaceX’s Colossus 1 facility, the effective utilization rate of that data center — a complex housing more than 220,000 Nvidia GPUs on roughly 300 megawatts of power — was approximately 11 percent, after xAI moved its training workloads to the newer Colossus 2 and discovered that Colossus 1’s mixed H100, H200, and GB200 architecture made efficient large-scale training difficult.[37] An 11-percent-utilized, $30-billion asset is not a strategic reserve; it is a bonfire of fixed costs. The lease to Anthropic converted that bonfire into revenue, which is precisely why Section 4 treats the transaction as the founding event of the secondary compute market. But the episode’s more important lesson is prospective: if one of the most sophisticated builders in the industry could strand a frontier-scale facility within roughly two years of constructing it — through nothing more exotic than an architecture transition — then utilization risk is not a tail scenario. It is an operating characteristic of the asset class.
For hyperscalers with cloud businesses, margin compression will arrive first as price competition. In an overcapacity regime, every operator faces the same incentive: any price above marginal cost (essentially electricity plus a sliver of operations) is better than idleness, so spot prices for compute grind down toward marginal cost exactly as they do in shipping, airlines, semiconductors, and every other high-fixed-cost industry that has ever overbuilt. Wall Street has begun to price this: hyperscaler equities recently traded at their lowest forward valuations since the launch of ChatGPT, at a discount to the S&P 500 — an extraordinary statement about the market’s confidence in the marginal return on the next capex dollar.[36]
3.2 Asset Write-Downs and the $176 Billion Depreciation Question
The second channel is the balance sheet itself, and it runs through the most consequential accounting debate of the AI era. Under both U.S. GAAP and IFRS, management estimates the useful life over which hardware is depreciated, and between 2020 and 2024, the major hyperscalers steadily extended server useful lives from roughly three years to five or six — Microsoft moved from four years to six, and Meta adopted five and a half — even as Nvidia compressed its architecture cadence to an annual release rhythm.[29], [47] Lengthening the schedule mechanically flatters earnings: capital expenditure of $660 billion depreciated over three years produces roughly $220 billion of annual expense, while the same spending depreciated over six years produces roughly $110 billion — a $110 billion swing in reported operating income that reflects an accounting assumption, not a dollar of new revenue.[28]
The debate detonated into public view in November 2025, when Michael Burry — the investor of “The Big Short” — disclosed short positions against Nvidia and Palantir and accused the hyperscalers of systematically flattering earnings through extended depreciation schedules:
“Understating depreciation by extending useful life of assets artificially boosts earnings.”
— Michael Burry, Scion Asset Management, posting on X, November 2025 [28]
Burry’s arithmetic holds that because Nvidia’s product cycle now runs at two to three years, hardware depreciated over five to six years will understate depreciation by approximately $176 billion across 2026–2028, overstating reported operating income at the most exposed companies — he singles out Oracle and Meta — by more than 20 percent.[43] He has further flagged the increasingly baroque financing structures around GPU purchases, warning that layered special-purpose vehicles and lease arrangements may obscure the true scale of hardware commitments sitting off balance sheets — a concern echoed by independent analyses that count more than $662 billion of lease commitments outside hyperscaler balance sheets.[19], [47]
The industry’s defense is the “compute cascade”: a GPU that ages out of frontier training at year two or three descends through a waterfall of progressively less demanding work — large-scale inference, fine-tuning, smaller-model serving, batch processing — and genuinely produces revenue for five or six years, just not at the cutting edge. Nvidia itself has publicly argued that customers observe four-to-six-year economic lives, and the schedules are audited and repeatedly upheld.[42] The honest verdict, supported by the most careful independent reviews, sits between the poles: the fraud characterization is legally unsubstantiated, because the standards deliberately grant management judgment; but the optimism characterization is well founded, because the cascade thesis depends on inference demand growing fast enough, forever, to absorb every generation of descending silicon — precisely the assumption that Section 2’s efficiency shockwave attacks.[29], [44] The most telling datapoint is the divergence of early 2025, when Amazon shortened the useful life of a subset of its servers, explicitly citing the accelerating pace of technology, in the same quarter that Meta extended its estimate — two of the world’s most sophisticated infrastructure operators, holding identical hardware, reaching opposite conclusions. Useful life, in other words, is functioning as an earnings-management lever, not an engineering fact.[29]
Goldman Sachs’ own sensitivity analysis quantifies what is at stake if the optimists are wrong: shortening assumed GPU life from five years to three pushes cumulative industry depreciation from roughly $3 trillion to approximately $4 trillion between 2026 and 2031 — a $1 trillion earnings swing hanging on a single estimate.[14] In an overcapacity scenario, the adjustment does not arrive politely through revised schedules; it arrives as impairment. Companies must aggressively write down the value of underutilized, rapidly aging hardware once utilization data makes the cascade thesis untenable to auditors — and unlike a schedule change, which spreads pain over years, an impairment lands in a single quarter.
Table 3. What one assumption is worth: GPU useful-life scenarios on ~$660B of annual AI hardware capex
| Assumption | Annual depreciation expense | Effect on reported operating income | Analytical interpretation |
| 6-year useful life (current posture at several hyperscalers) | ~$110B | Highest reported profits | Defensible only if the compute cascade absorbs every descending chip generation [28] |
| 5-year useful life | ~$132B | −$22B vs. 6-year | Amazon’s direction of travel after its 2025 shortening [29] |
| 4-year useful life | ~$165B | −$55B vs. 6-year | Midpoint between product cadence and cascade optimism |
| 3-year useful life (aligned to Nvidia’s cadence) | ~$220B | −$110B vs. 6-year; ~$176B cumulative 2026–2028 across the industry | Burry’s economic-life thesis; Goldman’s stress case adds ~$1T of depreciation 2026–2031 [14], [43] |
Sources: Forbes Business Council analysis of Futurum capex data; Goldman Sachs Global Institute sensitivity analysis; Scion Asset Management estimates.[14], [28], [43]
3.3 R&D Capital Traps: The Opportunity Cost of Idle Silicon
The third channel is subtler than the first two but may matter more for the long-run competitive landscape: capital locked in idle infrastructure cannot be redeployed into the next generation of algorithms. Every dollar of stranded compute is a dollar unavailable for research, talent, product development, or the efficiency innovations that Section 2.1 showed are the industry’s true engine of progress. The great strategic irony of the boom is that the deflationary force undermining the buildout — algorithmic efficiency — is itself the highest-return investment available, and it is being crowded out on the margin by hardware accumulation. DeepSeek’s demonstration that a competitive frontier model could be trained for a few million dollars of marginal compute was, among other things, a proof that the binding constraint on progress had shifted from silicon to ideas; a balance sheet loaded with depreciating silicon is optimized for the constraint that just expired.[18]
There is a portfolio-theory framing of the same point. The hyperscalers have historically been valued as capital-light compounding machines whose R&D produced software with near-infinite gross margins. The AI buildout converts them, at the margin, into capital-heavy industrial utilities whose returns depend on utilization and power costs. If overcapacity strikes, the trapped capital does double damage: it earns sub-hurdle returns directly, and it drags the corporate return on invested capital toward utility levels, inviting the valuation compression that the market began applying in 2026.[36]
3.4 Credit Rating Pressure: When the Safest Balance Sheets Start Borrowing
The final channel runs through the debt markets, and it is here that the boom’s financial architecture has changed most dramatically since 2024. For two decades, the hyperscalers were the closest thing corporate America offered to sovereign credits: fortress balance sheets, oceanic free cash flow, minimal net debt. The capital program of Section 1 has ended that era. With internal cash generation no longer sufficient — the group’s operating-cash-flow and capex curves crossing in 2026 — the builders have turned outward: AI hyperscalers including Amazon, Alphabet, Meta, Microsoft, and Oracle issued $159 billion of corporate bonds in the first five months of 2026 alone, a sum exceeding their combined borrowing over the previous five years, according to Dealogic; Nvidia issued $25 billion of bonds in its first offering since 2021; Alphabet announced an $85 billion equity raise; and Meta disclosed that new cloud contracts and infrastructure purchase agreements added $107 billion to its contractual commitments in a single quarter.[32], [49] Beyond the bond market, the buildout increasingly taps private credit, joint ventures, and special-purpose vehicles whose obligations are harder to see and harder to aggregate.
The International Monetary Fund has now formally flagged the structure. In its 2026 financial-stability commentary, the Fund argued that the principal systemic risk is not an equity-market correction in AI stocks but the mountain of debt accumulating behind the buildout — leverage extended against assets of uncertain useful life, for facilities of which roughly 60 percent have not yet broken ground, in an industry already showing construction delays.[32] The credit mechanics of an overcapacity scenario are well rehearsed from every prior infrastructure cycle: skyrocketing infrastructure debt paired with stalling cash flows spooks debt markets; spreads widen first for the levered pure-plays (the neocloud operators and data-center developers), then for the suppliers financed by their receivables, and finally the rating agencies re-examine the hyperscalers themselves — Moody’s has already flagged counterparty concentration and rising leverage at Oracle,[48] whose capital spending now exceeds its operating cash flow.[14] None of this requires a default to matter. Rising capital costs alone are sufficient to force the capex discipline that managements have so far resisted, and in every prior cycle, that forced discipline — announced as “optimization” — has been the market’s first official confirmation that the boom has crested.
It is important to state the counter-case fairly, because serious investors hold it. KKR’s infrastructure team argues that today’s data-center cycle differs fundamentally from late-1990s fiber: capacity is underwritten by long-term contracts with the world’s strongest counterparties, current absorption data show no vacancy buildup, accelerator refresh cycles soak up temporary excess, and power scarcity itself limits how far overbuilding can physically run — in their framing, temporary overbuilds behave like rolling upgrades rather than stranded assets.[33] Each point has force, and the last one is the crux: the KKR thesis is, at bottom, a bet that the power bottleneck endures. Section 2.4 documented a $50-billion-plus corporate nuclear program engineered precisely to break that bottleneck. If the industry succeeds in its own stated energy strategy, the strongest structural argument against overcapacity is the one the industry is spending most aggressively to dismantle. The financial implications traced in this section — compressed margins, impaired assets, trapped capital, repriced credit — are therefore not a doomsday forecast. They are the standard, well-documented second act of every infrastructure boom in history, arriving on schedule. The third act, examined next, is where the story becomes constructive: the birth of a market that turns stranded capacity into democratized capability.
Section 4: Navigating the Emerging Secondary Compute Market
Every commodity market in history was born the same way: producers built capacity for their own use, discovered they had too much, and began selling the surplus — first bilaterally, then through brokers, and finally on standardized exchanges where the commodity traded at a spot price divorced from any particular producer. Oil followed this arc; so did electricity, shipping capacity, and telecom bandwidth. The thesis of this section is that compute began the same journey in 2026, that the founding transactions are already on the public record, and that an infrastructure overshoot will complete the transformation — giving rise to a highly liquid, institutional secondary market for compute that fundamentally changes how machine intelligence is bought, sold, and consumed.
4.0 The Founding Transactions: Colossus for Rent and Meta Compute
The secondary market did not arrive as a theory; it arrived as a series of SEC filings in the weeks surrounding SpaceX’s June 12, 2026 initial public offering. Following its February 2026 merger with xAI — a transaction valuing the combined entity at $1.25 trillion — SpaceX inherited the Colossus data-center complex in Memphis, Tennessee, built to train Grok and running at approximately 11 percent effective utilization after training migrated to the newer Colossus 2.[37], [38] Rather than let more than $30 billion of GPU infrastructure idle, SpaceX did something no frontier AI builder had done before: it rented its overcapacity to its competitors, at scale, in public.
The deals came in rapid succession. In May 2026, Anthropic agreed to pay $1.25 billion per month through 2029 for exclusive access to the entire output of Colossus 1 — more than 220,000 Nvidia GPUs on roughly 300 megawatts — a single customer relationship worth approximately $15 billion per year at full rate.[37] On June 5, a second filing disclosed that Google would pay $920 million per month from October 2026 through June 2029 for approximately 110,000 GPUs, with capacity ramping through September; a Google Cloud spokesperson characterized the arrangement as short-term “bridge capacity” to meet surging demand for its Gemini Enterprise platform.[38] A third agreement gave AI coding startup Cursor access to xAI capacity as part of a $10 billion collaboration (with SpaceX securing an option to acquire Cursor outright), and on June 22, open-source lab Reflection AI became the fourth tenant, agreeing to pay $150 million per month from July 1, 2026 for immediate access to Nvidia GB300-class hardware.[37], [39] Combined, the Anthropic and Google leases alone put roughly $26 billion of annualized compute revenue onto SpaceX’s books in the weeks before its IPO priced at a valuation of approximately $1.75 trillion — the largest in history — with SpaceX’s own S-1 stating plainly that its strategy provides “substantial flexibility in how we allocate and monetize capacity.”[7], [41] Market observers grasped immediately what the structure meant:
“It’s a dramatic pivot of xAI from artificial intelligence provider to server farm.”
— Cory Johnson, Chief Market Strategist, Epistrophy Capital Research [40]
Weeks later, the largest pure overbuilder confirmed the pattern. On July 1, 2026, Bloomberg reported — and CNBC confirmed — that Meta is developing a cloud infrastructure business to sell excess AI computing capacity and hosted model access to outside customers, organized under an internal initiative dubbed Meta Compute and led by infrastructure chief Santosh Janardhan, Superintelligence Labs’ Daniel Gross, and president Dina Powell McCormick.[34] The move followed months of signaling: Zuckerberg told shareholders in May that companies approached Meta nearly every week asking to buy compute at a premium or to stand up an API service, that Meta had declined only because it expected to consume the capacity internally — and that if the company ever concluded it had overbuilt, selling the excess externally remained an option it would exercise.[6] Bernstein analysts noted the scale hiding behind the hedge:
“This scale easily rivals cloud provider footprints.”
— Madison Rezaei, Analyst, Bernstein, estimating Meta’s ~20 GW global footprint with ~14 GW more coming [35]
Meta’s stock jumped nearly 9 percent on the report — its sharpest rally in five months — while the shares of CoreWeave and Nebius, the new cloud specialists whose entire business is selling GPU capacity, fell 14 and 17 percent respectively, and chip stocks slid on the inference that excess capacity exists at all.[35], [36] The market reaction is itself the thesis of this paper in miniature: when a builder monetizes surplus, its own investors cheer the revenue while the rest of the supply chain reprices the demand. Wall Street simultaneously began preparing for the margin consequences, since selling raw compute carries structurally lower margins than Meta’s advertising business — the first institutional acknowledgment that hyperscaler economics converge toward utility economics as the secondary market grows.[30]
Table 4. The founding transactions of the secondary compute market, 2026
| Seller | Buyer | Terms | Assets covered | Strategic significance |
| SpaceX (xAI) | Anthropic | $1.25B/month through 2029 (~$15B/yr); 90-day cancellation after 2026 | Entire Colossus 1 output: 220,000+ GPUs, ~300 MW, Memphis, TN [37] | First frontier lab renting a rival’s surplus at scale; facility was ~11% utilized before the lease |
| SpaceX (xAI) | $920M/month, Oct 2026–Jun 2029 (~$30B total) | ~110,000 Nvidia GPUs plus CPUs and memory [38] | Hyperscaler buying “bridge capacity” from a non-cloud provider for Gemini Enterprise demand | |
| SpaceX (xAI) | Cursor | $10B collaboration; SpaceX option to acquire for $60B | xAI data-center capacity for AI coding workloads [39] | Compute as acquisition currency |
| SpaceX (xAI) | Reflection AI | $150M/month from Jul 1, 2026 (~$6.3B through 2029) | Nvidia GB300-class capacity [39] | Open-source lab gaining frontier silicon without owning any |
| Meta (planned) | External developers & enterprises | Meta Compute unit; raw capacity and hosted-model API under consideration | ~20 GW existing footprint; ~14 GW in development [34], [35] | Largest pure overbuilder formalizing surplus monetization; neocloud stocks fell 11–17% on the news |
Sources: SpaceX SEC filings as reported by TechCrunch and CNBC; Bloomberg/CNBC reporting on Meta Compute; Bernstein estimates.[6], [7], [34], [35], [37], [38], [39]
4.1 The Architecture of the Coming Market: From Bilateral Leases to Spot Exchange
The 2026 transactions are the bilateral, over-the-counter infancy of the market — enormous, bespoke, relationship-driven deals. As overcapacity broadens, the market will acquire the standard anatomy of every maturing commodity system, and the value chain will organize into four layers:
[Hyperscaler Excess Supply] → [Wholesale Compute Brokers] → [Discount Marketplaces (e.g., Together AI / Lambda / San Francisco Compute-style exchanges)] → [Agnostic Compute Layer: open-source labs, academics, startups]
Layer one: excess supply. The hyperscalers, Stargate-class ventures, and sovereign AI projects sit at the top as structural net producers. Their overbuild — planned in gigawatts, financed in decades — becomes the raw feedstock of the market, exactly as Meta’s 20-gigawatt footprint and SpaceX’s Colossus complex already have.[35], [37]
Layer two: wholesale compute arbitrage. Third-party aggregators will buy idle hyperscaler capacity at steep discounts and repackage it. The early cancellation clauses in the SpaceX deals — 90-day termination rights on multi-billion-dollar leases — reveal that even today’s flagship contracts are structured like short-tenor wholesale capacity agreements rather than utility offtakes, which is precisely the contractual raw material from which brokered markets are built.[31] As surplus deepens, expect specialist trading houses (and eventually the commodity desks of investment banks) to intermediate between builders who need utilization and buyers who need flexibility, profiting from the spread exactly as power marketers did after electricity deregulation.
Layer three: discount marketplaces and exchanges. Platforms in the mold of Together AI and Lambda — which already resell aggregated GPU capacity — will evolve into standardized marketplaces quoting prices per GPU-hour for defined hardware classes, with reservation tiers, spot tiers, and interruptible tiers. Compute will commoditize toward real-time spot pricing resembling oil or electricity markets: benchmark contracts (an “H100-hour” or “GB300-hour” as the Brent crude of intelligence), transparent forward curves, and eventually financial derivatives allowing operators to hedge utilization risk and buyers to hedge token costs. The economic precondition for all of this is homogeneity plus surplus — and the overbuild supplies both.
Layer four: the agnostic compute layer. At the bottom of the waterfall, open-source developers, academic institutions, and bootstrapped startups gain cheap access to elite silicon that scarcity pricing had reserved for the richest laboratories. The Reflection AI lease is the prototype: a $25-billion open-source lab renting frontier GB300 capacity by the month, without owning a single facility, at the exact moment when governments and enterprises are reassessing their dependence on closed-model providers.[39] This is democratized frontier access, and its consequences compound: cheaper compute lowers the cost of training open models, which pressures closed-model pricing, which accelerates the efficiency shockwave of Section 2.1, which deepens the surplus that made the compute cheap. The flywheel of overcapacity, once spinning, feeds itself.
4.2 Cloud Provider Decoupling and the End of Ecosystem Lock-In
The final structural consequence operates on the demand side. For fifteen years, enterprise computing strategy has been organized around ecosystem allegiance — a company was “an AWS shop” or “an Azure shop,” accepting proprietary services and egress fees as the price of integration. A liquid secondary market dissolves the economic logic of that allegiance. When equivalent GPU-hours are quotable from a broker at a 40 percent discount to on-demand hyperscaler rates, enterprises will migrate away from rigid cloud ecosystems toward flexible, multi-cloud compute brokers, treating the hyperscalers as interchangeable capacity providers rather than platforms.[11] Google’s own behavior in 2026 is the leading indicator: the world’s third-largest cloud provider chose to rent 110,000 GPUs from a rocket company rather than wait for its own buildout, demonstrating that even hyperscalers now treat compute as a fungible commodity to be sourced wherever it is cheapest and fastest.[38] Once the sellers of lock-in become buyers on the open market, lock-in’s days are numbered.
Two honest caveats complete the picture. First, compute is not yet as fungible as oil: interconnect topology, software stacks, data gravity, and security certifications create real switching costs, and the commoditization described here will proceed workload by workload — inference and fine-tuning first, frontier training last. Second, a deep secondary market is a double-edged instrument for the builders: it rescues their utilization in a glut, but it also caps their pricing power forever, because every customer negotiation now happens against a visible spot price. That is exactly what happened to telecom carriers after the fiber glut created bandwidth markets, and it is why Section 5 argues that the winners of the next phase will be those who prepare for commodity economics before commodity economics are imposed upon them.
Section 5: Executing Strategic De-Risking Playbooks
Analysis without prescription is commentary. This section translates the preceding argument into operational playbooks for the three constituencies most exposed to the boom’s turn: the financial officers and boards of the builders themselves; the operators and investors of data-center assets; and the enterprise buyers whose contracts will ultimately determine who survives the transition from scarcity to abundance. The unifying principle across all five playbooks is the conversion of fixed commitments into optionality — because in a capital cycle, optionality is the only asset whose value rises when everything else falls.
5.1 Implement Variable Capital Expenditure Architectures
The first playbook attacks the root exposure: ownership itself. The traditional hyperscaler model — buy the land, build the shell, own the silicon — maximizes margin under scarcity and maximizes loss under glut. The de-risked alternative shifts the capital structure from owned to contracted: flexible leasing arrangements for hardware; build-to-suit developments where a real-estate partner owns the shell; joint ventures with sovereign wealth funds and infrastructure investors that share construction cost in exchange for preferred returns; and financing structures whose tenors match the true economic life of the assets they fund — three-to-four-year facilities for silicon, twenty-year facilities for shells, never the reverse. The industry has already begun this migration under the pressure of arithmetic: Meta and Oracle’s turn to private credit and joint-venture structures, Alphabet’s equity raise, and the $159 billion bond surge of early 2026 all represent attempts to push the buildout’s risk onto balance sheets better shaped to hold it.[8], [32] The discipline the playbook adds is honesty about what is being transferred: risk-sharing structures reduce equity exposure only if their obligations are genuinely contingent — take-or-pay commitments dressed as leases transfer nothing, and the more than $662 billion of off-balance-sheet lease commitments already accumulated suggests the industry is currently using these structures to obscure exposure rather than reduce it.[19] The chief financial officer’s test is simple: in a 30-percent-utilization scenario, which payments stop? Whatever survives that test is variable; everything else is debt by another name.
5.2 Establish Algorithmic Agility Protections
The second playbook operates at the level of physical design. Because the deflationary shockwaves of Section 2 attack specific workloads rather than computing in general, the defensible facility is the one that can change workloads faster than demand can change beneath it. Concretely: design data-center cooling, power routing, and chassis layouts to be chip-agnostic, so that halls provisioned for one accelerator generation can accept the next — or a different vendor’s silicon, or CPUs for traditional cloud work — without structural rework; standardize on liquid-cooling and power-density envelopes that bracket several future hardware generations; and reserve genuine electrical headroom rather than value-engineering it away. The strategic model is Microsoft’s publicly articulated policy of staggering deployments across chip generations to ride each hardware wave briefly, which is agility applied in time, complemented by agility applied in space: facilities engineered so that specialized AI silicon can be swapped toward inference, traditional cloud, rendering, or scientific and biotech processing as relative prices move.[15] In an overcapacity regime, the spread between a convertible facility and a single-purpose one is the spread between a repriced asset and a stranded one — the fiber glut’s enduring lesson, where conduit and rights-of-way retained value long after the specific electronics lit within them were scrap.
5.3 Secure Pre-Construction Offtake Agreements
The third playbook borrows the energy sector’s oldest discipline. No rational developer builds a liquefied-natural-gas terminal on a demand forecast; construction begins only after long-term, fixed-price offtake contracts with creditworthy counterparties de-risk the revenue line. The compute industry is drifting toward this model under duress — Oracle’s Stargate economics rest on OpenAI’s committed consumption, and the SpaceX-Anthropic lease is, in substance, an offtake agreement — but the practice remains the exception rather than the underwriting standard.[5], [48] The playbook makes it the standard: secure long-term, fixed-price compute purchase contracts with enterprise clients before breaking ground on new facilities, and let the pace of contracted demand, not the pace of competitor announcements, set the pace of construction. The 2026 deal record also teaches the necessary refinement: tenor and termination rights are where these contracts succeed or fail. The Colossus leases carry 90-day cancellation windows after 2026 — headline revenue of $26 billion per year secured by one quarter of notice — which is offtake in form but merchant exposure in substance.[7], [31] Builders should trade price for tenor (discounts for five-to-seven-year committed terms), tier their capacity stack across committed, reserved, and spot layers, and treat the ratio of contracted-to-speculative megawatts as a board-level risk metric, disclosed with the same prominence as backlog.
5.4 Align Depreciation Policy with Engineering Reality — Before the Market Does It for You
The fourth playbook, absent from most strategy discussions because it lives in the accounting department, may determine which companies keep the market’s trust through the transition. Section 3.2 established that useful-life assumptions currently function as an earnings lever, that the industry’s own leaders design around the schedules they publish, and that a $1 trillion depreciation swing hangs on the estimate.[14], [15] The de-risking move is preemptive candor: shorten schedules toward observed economic life while earnings are strong enough to absorb it, as Amazon began to do in 2025; disclose fleet-level utilization and redeployment data so that the compute-cascade defense rests on evidence rather than assertion; and publish the sensitivity of operating income to useful-life assumptions rather than leaving analysts to reverse-engineer it.[29] The alternative — waiting until utilization data forces auditors’ hands — converts a manageable multi-year expense into a single-quarter impairment shock delivered at the moment of maximum market skepticism. In every prior capital cycle, the companies that wrote down early were punished briefly and trusted afterward; the companies that wrote down last were remembered as the cycle’s cautionary tales.
5.5 For Enterprise Buyers: Build a Portfolio, Not an Allegiance
The final playbook belongs to the demand side, which has historically been treated as a passive beneficiary of gluts but which can materially improve its position by preparing for one. Enterprise technology officers should manage compute the way corporate treasurers manage currency exposure: a laddered portfolio of commitments across providers, tenors, and price structures. Concretely: keep baseline, predictable inference on committed contracts negotiated against the visible spot market; keep burst and experimental workloads on spot and interruptible tiers, whose prices will fall fastest as the secondary market deepens; maintain architectural portability — containerized serving stacks, open model formats, abstraction layers above any single provider’s proprietary services — so that the migration threat in every negotiation is credible; and resist the vendor-financed bundling by which builders will try to convert their overcapacity into customers’ long-term lock-in at pre-glut prices. The buyers who enter 2028 with portable workloads and short-dated commitments will harvest the overcapacity; the buyers who signed decade-long proprietary bundles in 2026 will subsidize it.
Section 6: What Have We Learned? The Strategic Pillars
The argument of this paper compresses into seven pillars — five inherited from the framework’s original outline and refined by the evidence, and two added because the events of 2026 demanded them.
Pillar 1: Overbuilt Infrastructure Drives Later Innovation. Just as the dot-com era’s automated-logistics experiments eventually paved the way for Instacart and Amazon Fresh, and the fiber glut of 2001 became the free bandwidth on which YouTube and cloud computing were built, an early AI compute glut will act as the cheap foundation for the next generation of massive AI adoption.[3] The capital is lost to its original investors long before it is lost to civilization; historically, it is never lost to civilization at all.
Pillar 2: Efficiency Deflates Demand Faster Than Growth Scales. Algorithmic optimization is a structurally deflationary force — token prices falling at a median of roughly 50-fold per year — that regularly outruns physical infrastructure planning cycles measured in half-decades.[18] Any capacity model that does not haircut demand for compounding efficiency is not a forecast; it is a hope.
Pillar 3: Energy Abundance Accelerates Gluts. Solving the power bottleneck through nuclear restarts, gas bridges, and SMR fleets will release the coiled spring of stalled capacity nearly simultaneously — the contracted-capacity data already imply a potential 64-gigawatt, $640-billion overshoot by 2030 — hastening the very overcapacity the energy program was meant to outrun.[23], [27] Scarcity is the only discipline a capital boom respects, and the industry is spending $50 billion to abolish its own discipline.
Pillar 4: Compute Will Liquidly Commoditize. The founding transactions are complete: Colossus rents to Anthropic and Google by the month, Meta prepares to sell its surplus, and open-source labs lease frontier silicon they will never own.[34], [37], [39] The future belongs to the platforms that aggregate, broker, and trade spot-priced compute — the market-makers of intelligence — rather than to those who merely own the depreciating physical assets.
Pillar 5: Flexibility Outperforms Scale. The winners of the next decade will not be the companies with the largest capital budgets but those with the most adaptable, modular supply chains — chip-agnostic facilities, staggered deployments, contracted revenue, and honest depreciation. Scale won the scarcity era; optionality wins the abundance era.
Pillar 6: Contracts Are the New Balance Sheet. In the emerging regime, the decisive corporate asset is no longer the gigawatt owned but the offtake signed: SpaceX’s $26 billion of annualized leases transformed its IPO narrative in three weeks, while Meta’s $107 billion of quarterly contractual commitments now moves its valuation more than its data centers do.[34], [49] Analysts, lenders, and boards should evaluate AI infrastructure companies primarily through their contract books — tenor, counterparty quality, termination rights — because in a commoditizing market, the contract is where the economics live and the hardware is merely where they are performed.
Pillar 7: Depreciation Honesty Is a Competitive Weapon. The $176 billion accounting question will be answered, voluntarily or involuntarily, within the decade; the only management choice is whether the answer arrives as a controlled adjustment or as an impairment cascade at the moment of maximum vulnerability.[43] The companies that align their books with their engineering first will command the credibility — and the cost of capital — that the rest of the industry forfeits.
Conclusion: Maximizing the Upside of Excess Capacity
This paper opened with a bankruptcy and must close with an inheritance, because the two are the same story told at different distances. Webvan built the physical and intellectual infrastructure of online grocery delivery a decade before demand could pay for it; the company died, and the idea it overpaid for became one of the defining conveniences of modern life, delivered by successors who inherited its lessons at no charge. The framework developed across the six sections of this paper — and the reason it bears the name “AI Infrastructure Boom” rather than “AI bubble” or “AI winter” — is that the same double-exposure photograph now describes artificial intelligence. The name was chosen, as the Introduction argued, because each word does analytical work: “AI” marks the dangerous concentration of the capital cycle on a single workload and a single silicon supply chain; “Infrastructure” locates the analysis in the long economics of illiquid, slow-to-unwind physical capital, where every great overcapacity cycle in history has lived; and “Boom” insists on the crucial distinction between a technology being overbuilt and a technology being wrong. The railroads were overbuilt and right. The fiber networks were overbuilt and right. Webvan was overbuilt and right. The evidence assembled here — $725 billion of single-year capital guidance against enterprise adoption still trapped in pilot purgatory, token prices deflating fifty-fold annually against six-year depreciation schedules, and a nuclear program engineered to dissolve the boom’s last constraint — indicates that artificial intelligence is now traveling the same arc, at a scale that makes every predecessor look like a rehearsal.[4], [18]
The impending overcapacity of artificial intelligence infrastructure is therefore not a fatal diagnosis for the technology, but a predictable economic rite of passage. The transition from scarcity to extreme abundance will undoubtedly trigger painful financial corrections: margin compression as spot compute prices grind toward marginal cost; balance-sheet write-downs as depreciation schedules are forced into alignment with engineering reality; credit repricing as the $159-billion-per-half-year borrowing machine meets stalling cash flows; and market consolidation among hyperscalers, neoclouds, and developers who overextended their capital budgets in the scarcity years.[32] The first institutions through the wall — the Colossus leases, Meta Compute, the offtake-hungry financing structures of 2026 — are already showing the shape of the adjustment, and Sections 3 through 5 have tried to give its participants an honest map and a usable playbook.
However, this capital overshoot will ultimately de-risk the entire digital economy, and that claim is not consolation but arithmetic. By driving the marginal cost of intelligence toward zero, the wave of overbuilding will democratize high-performance computing — putting frontier silicon within monthly-lease reach of open-source laboratories, universities, and bootstrapped startups that scarcity pricing had excluded; it will dismantle the high walls of proprietary control, as liquid secondary markets convert ecosystem lock-in into commodity procurement; and it will lay down the cheap, abundant foundational infrastructure — the shells, the substations, the nuclear fleet, the trained workforce, the hard-won operational knowledge — required to power the next century of global economic growth. The overcapacity phase, in other words, is not the failure mode of the AI Infrastructure Boom. It is the delivery mechanism of its promise: the moment the boom stops enriching its builders and starts endowing everyone else. Webvan’s investors lost everything; the world got grocery delivery. The pattern is older than any of the companies in this paper, and it is the most reliable pattern in the economics of technology: the infrastructure of the future is always built by people who arrive too early, at prices that only make sense to those who come after.
Footnotes and Endnotes:
[1] HowStuffWorks / Wikipedia contributors, “Webvan: history, Bechtel contract, and automated fulfillment buildout.” https://en.wikipedia.org/wiki/Webvan
[2] SFGate staff, “Webvan’s collapse and 2001 bankruptcy.” https://www.sfgate.com
[3] Wired / StreetFins, “How Amazon Fresh and Instacart realized the online-grocery thesis Webvan pioneered.” https://www.wired.com
[4] Luke James, Tom’s Hardware (citing the Financial Times and Jefferies’ Brent Thill), “Google, Microsoft, Meta, and Amazon capex spending to hit $725 billion in 2026, up 77% from last year.” https://www.tomshardware.com/tech-industry/big-tech/big-techs-ai-spending-plans-reach-725-billion
[5] OpenAI, Oracle, and SoftBank, “OpenAI, Oracle, and SoftBank expand Stargate with five new AI data center sites (nearly 7 GW, $400B+).” https://openai.com/index/five-new-stargate-sites/
[6] CNBC (Jonathan Vanian et al.), “Meta stock pops on cloud push to sell excess AI compute power capacity; Zuckerberg: ‘definitely on the table’.” https://www.cnbc.com/2026/07/01/meta-stock-cloud-ai-compute.html
[7] Sean O’Kane, TechCrunch, “Google will pay SpaceX $920M per month for compute; Anthropic pays $1.25B/month for Colossus 1; SpaceX IPO June 12, 2026.” https://techcrunch.com/2026/06/05/google-will-pay-spacex-920m-per-month-for-compute/
[8] Bobby Allyn, NPR (quoting Prof. Daron Acemoglu, MIT), “Here’s why concerns about an AI bubble are bigger than ever.” https://www.npr.org/2025/11/23/nx-s1-5615410/ai-bubble-nvidia-openai-revenue-bust-data-centers
[9] Data Center Richness / Rich Miller, “Hyperscalers plan $630 billion in 2026 CapEx: company-by-company earnings breakdown.” https://datacenterrichness.substack.com/p/hyperscalers-plan-630-billion-in
[10] Futurum Group Research, “AI Capex 2026: The $690B infrastructure sprint.” https://futurumgroup.com/insights/ai-capex-2026-the-690b-infrastructure-sprint/
[11] Columbia Business School, “Research on cloud decoupling and multi-cloud enterprise migration.” https://business.columbia.edu
[12] Yahoo Finance (citing Goldman Sachs Research), “Meta, Microsoft, Amazon, and Alphabet are about to spend a shocking amount of money to dominate the AI era ($5.3T FY2025–2030).” https://finance.yahoo.com/sectors/technology/article/meta-microsoft-amazon-and-alphabet-are-about-to-spend-a-shocking-amount-of-money-to-dominate-the-ai-era-115359575.html
[13] CNBC (quoting Jake Dollarhide, Longbow Asset Management), “Tech AI spending approaches $700 billion in 2026, cash taking big hit.” https://www.cnbc.com/2026/02/06/google-microsoft-meta-amazon-ai-cash.html
[14] Tech Times (citing Epoch AI, Goldman Sachs, Evercore, Bank of America), “Big Tech AI spending tops $725 billion: free cash flow hits zero this summer.” https://www.techtimes.com/articles/319306/20260629/big-tech-ai-spending-tops-725-billion-free-cash-flow-hits-zero-this-summer.htm
[15] Stanley Laman Group, Signals & Noise (quoting Satya Nadella, Microsoft), “GPUs: how long do they really last? Useful lives, the computing cascade, and earnings quality.” https://www.stanleylaman.com/signals-and-noise/gpus-how-long-do-they-really-last
[16] Jason Kirsch, Forbes (citing CreditSights), “AI spending is surging faster than revenue and markets are repricing.” https://www.forbes.com/sites/jasonkirsch/2026/06/02/the-ai-capex-to-revenue-gap-is-widening—and-markets-are-starting-to-notice/
[17] Stanford University Human-Centered AI Institute (HAI), “The AI Index Report (2025–2026 editions): 280-fold inference cost decline; adoption and productivity data.” https://hai.stanford.edu/ai-index/2026-ai-index-report
[18] 200 OK Solutions (compiling Epoch AI, MIT Project NANDA, McKinsey State of AI, DoiT/Sapio, PwC), “Enterprise AI adoption statistics you need to know in 2026.” https://www.200oksolutions.com/blog/enterprise-ai-adoption-statistics-you-need-to-know-in-2026/
[19] Anomaly Investments (citing J.P. Morgan estimates), “This obviously is an AI bubble. The math says so.” https://anomalyinvestments.substack.com/p/this-obviously-is-an-ai-bubble-the
[20] K. Barhoumi et al., International Monetary Fund, “Global economic and financial implications of artificial intelligence: lessons from a scenario planning exercise (2026).” https://arxiv.org/html/2606.01575v1
[21] B. Sykes, “The state of AI adoption in the enterprise (Q1 2026 review): reframing the ‘95% failure’ finding.” https://bsykes.substack.com/p/the-state-of-ai-adoption-in-the-enterprise
[22] Jim Covello et al., Goldman Sachs Global Investment Research, “Gen AI: too much spend, too little benefit? (Top of Mind No. 129).” https://www.goldmansachs.com/insights/top-of-mind/gen-ai-too-much-spend-too-little-benefit
[23] SMR Intel, “Every nuclear-powered data center deal: Google, Amazon, Meta & Microsoft (2026) — 13 deals, 9.8 GW committed.” https://smrintel.com/nuclear-data-center-deals/
[24] Informed Clearly, “AI data centers hit grid wall: Big Tech pivots to nuclear in 2026 (interconnection queue >2,600 GW).” https://informedclearly.com/en/ai/53909/ai-data-centers-nuclear-power-2026
[25] Informed Clearly, “AI-nuclear pivot: tech giants become energy companies in 2026 (X-energy $1.02B IPO).” https://informedclearly.com/en/energy/54001/ai-nuclear-tech-giants-energy-2026
[26] Introl Research, “Nuclear power for AI: inside the data center energy deals (10 GW+ contracted).” https://introl.com/blog/nuclear-power-ai-data-centers-microsoft-google-amazon-2025
[27] Douglas in Vegas (citing Grant’s Interest Rate Observer, CreditSights’ Andy DeVries, CBRE IM’s Tania Tsoneva), “The data center bubble: 127 GW contracted vs. 108 GW of projected 2030 demand.” https://www.douglasinvegas.com/blog/2026/6/3/the-data-center-bubble
[28] Brian Anderson, Forbes Business Council (quoting Michael Burry), “The hidden variable in the AI rally: a depreciation reality check.” https://www.forbes.com/councils/forbesbusinesscouncil/2026/04/17/the-hidden-variable-in-the-ai-rally-a-depreciation-reality-check/
[29] Dave Friedman, “The $176 billion accounting question at the heart of the AI boom.” https://davefriedman.substack.com/p/the-176-billion-accounting-question
[30] CNBC, “Meta’s push into cloud computing means Wall Street has to prepare for lower margins.” https://www.cnbc.com/2026/07/02/metas-push-into-cloud-excites-wall-street-despite-lower-margins.html
[31] Yahoo Finance, “SpaceX, Google compute deal raises eyebrows ahead of IPO (90-day cancellation windows; broker commentary).” https://finance.yahoo.com/markets/stocks/article/spacex-google-compute-deal-raises-eyebrows-ahead-of-ipo-120522033.html
[32] Moneywise (citing the International Monetary Fund and Dealogic), “The IMF says the real threat is the mountain of debt behind the AI buildout — $159B in bonds in 5 months; 60% of planned data centers not yet broken ground.” https://moneywise.com/news/economy/imf-ai-debt-leverage-data-centers-2026
[33] KKR Global Institute / KKR Infrastructure, “Beyond the bubble: why AI infrastructure will compound long after the hype.” https://www.kkr.com/insights/ai-infrastructure
[34] Rebecca Bellan, TechCrunch (confirming Bloomberg), “Meta, like SpaceX, looks to turn excess AI compute into cash (Meta Compute; Janardhan, Gross, Powell McCormick).” https://techcrunch.com/2026/07/01/meta-like-spacex-looks-to-turn-excess-ai-compute-into-cash/
[35] Axios (quoting Madison Rezaei, Bernstein), “Meta stock soars as Mark Zuckerberg explores cloud business (~20 GW footprint, ~14 GW pipeline).” https://www.axios.com/2026/07/01/meta-cloud-mark-zuckerberg
[36] Sherwood News, “Meta surges on report it’s starting a cloud business to sell excess AI compute; neoclouds slammed; hyperscalers at lowest forward valuations since ChatGPT.” https://sherwood.news/markets/meta-surges-report-entering-into-cloud-business-excess-compute/
[37] INDmoney Research, “SpaceX-xAI deals with Google & Anthropic: Colossus 1 (~220,000 GPUs, ~300 MW) was ~11% utilized before the Anthropic lease.” https://www.indmoney.com/blog/us-stocks/spacex-xai-compute-deals-google-anthropic-ipo-valuation
[38] Lora Kolodny / MacKenzie Sigalos, CNBC (quoting Google Cloud spokesperson), “Google to pay SpaceX $920 million a month for compute capacity at xAI data centers (‘bridge capacity’ for Gemini Enterprise).” https://www.cnbc.com/2026/06/05/google-to-pay-spacex-920-million-a-month-for-xai-compute-capacity.html
[39] CNBC, “SpaceX signs compute deal with open-source AI startup Reflection AI ($150M/month from July 1, 2026).” https://www.cnbc.com/2026/06/22/spacex-ai-colossus-data-center-reflection.html
[40] Yahoo Finance (quoting Cory Johnson, Epistrophy Capital Research), “SpaceX, Google compute deal raises eyebrows ahead of IPO.” https://finance.yahoo.com/markets/stocks/article/spacex-google-compute-deal-raises-eyebrows-ahead-of-ipo-120522033.html
[41] Euronews Business (quoting the SpaceX IPO filing), “Google rents SpaceX/AI supercomputers for $920M a month, ahead of IPO (‘substantial flexibility in how we allocate and monetise capacity’).” https://www.euronews.com/business/2026/06/06/google-rents-spacexai-supercomputers-for-920m-a-month-ahead-of-ipo
[42] Olga Usvyatsky, Deep Quarry / National Law Review, “Useful lives of GPUs: key accounting considerations in the Burry–Nvidia depreciation debate.” https://natlawreview.com/article/deep-quarry-useful-lives-gpus-key-considerations
[43] Bernardo, Level-Headed Investing (summarizing Michael Burry’s estimates), “Are AI chip ‘useful lives’ creating useless earnings? ($176B understated depreciation, 2026–2028).” https://www.levelheadedinvesting.com/p/are-ai-chips-useful-lives-creating-useless-earnings
[44] Interesting Engineering++, “Why Michael Burry is wrong about AI depreciation (but still might be right): the computing-cascade defense.” https://interestingengineering.substack.com/p/why-michael-burry-is-wrong-about
[45] Fortune (quoting Prof. Erik Brynjolfsson, Stanford Digital Economy Lab), “’It’s not going away’: the Stanford economist who called the AI entry-level jobs crisis early has the receipts.” https://fortune.com/2026/06/27/what-is-ai-impact-entry-level-jobs-stanford-adp-canaries-brynjolfsson-richardson/
[46] Sam Altman, OpenAI / SoftBank Group press release, “OpenAI, Oracle, and SoftBank expand Stargate with five new AI data center sites (‘AI can only fulfill its promise if we build the compute to power it’).” https://group.softbank/en/news/press/20250924
[47] Crypto Briefing editorial team, “Michael Burry revives bear case for AI chips amid GPU depreciation concerns (off-balance-sheet financing structures).” https://cryptobriefing.com/michael-burry-bear-case-ai-chips-gpu/
[48] IntuitionLabs Research, “Oracle & OpenAI’s $300B deal: AI infrastructure analysis (Stargate status through April 2026; Moody’s counterparty warning).” https://intuitionlabs.ai/articles/oracle-openai-300b-deal-analysis
[49] Yahoo Finance / Barchart, “Mark Zuckerberg doubles down on raw computing power to challenge AWS and Microsoft ($107B of new quarterly contractual commitments).” https://finance.yahoo.com/technology/articles/mark-zuckerberg-doubles-down-raw-183002829.html
[50] GPUnex Research, “AI inference economics: the 1,000× cost collapse reshaping GPUs (inference now ~two-thirds of AI compute).” https://www.gpunex.com/blog/ai-inference-economics-2026/
[51] Dean Baker, Center for Economic and Policy Research (CEPR), “The AI Bubble Monitor (Acemoglu’s ~0.7% ten-year cumulative productivity estimate).” https://cepr.net/publications/ai-bubble-monitor/
