The United States is in the middle of the largest single-sector infrastructure build in modern economic history. The numbers are staggering in isolation; taken together, they describe something with no clear precedent in how private capital has organized itself around a technology thesis.
The four largest hyperscalers, Amazon, Google, Microsoft, and Meta, were on track to spend over $350 billion on capital expenditure in 2025, with the broader universe of tech players pushing total AI-related infrastructure spend toward $500 billion for the year.¹ The top five hyperscalers are collectively committed to $660 to $690 billion in 2026, with approximately $450 billion of that specifically to AI infrastructure.² Over 23 gigawatts of data center capacity were under construction globally as of late 2025, with roughly three-quarters located in the United States.³ Global data center capacity is on a trajectory to roughly double by 2030.
In the first half of 2025, AI-related capex contributed approximately one full percentage point to US GDP growth, which Barclays estimates represents roughly half of total GDP growth during that period.⁴
This is not a normal technology investment cycle. This is a structural reorientation of private capital around a thesis that has not yet been proven out. The thesis presented here is straightforward: the promise of future revenue is driving the capex push, which is raising construction and spending across the broader economy, and the concern is that a bubble is forming that will burst when actual value is not realized. Given the sophistication of current models on existing technology, the need for extensive additional computing doesn’t obviously follow, making the whole setup worth examining carefully.
The dominant theory of value underlying the build is that AI is a general-purpose technology, analogous to electrification or the internet, that will reshape every sector of the economy and generate revenue over decades. Under this thesis, the returns on infrastructure laid today will accrue over a long horizon, and the risk of under-building permanently outweighs the risk of over-building.
KKR has explicitly highlighted the internet boom in its investment research: telecom consumption grew faster than aggregate consumption before, during, and after the dot-com bust, even as many individual operators went bankrupt. The bulls argue the same will hold for AI compute. Even in a shake-out, they say, aggregate demand will justify the infrastructure.⁵
This creates a forward-looking discounted value framework in which enormous present spending is rationalized by probabilistic future cash flows that have not yet materialized and, importantly, cannot yet be measured.
Perhaps the most important structural driver is that no single hyperscaler can rationally reduce spending unilaterally, regardless of what any individual firm believes about the risk of overbuild.
The logic is straightforward. If Microsoft slows its data center build while Google maintains pace, Microsoft’s Azure capacity will lag, enterprise customers seeking AI-native cloud services will shift, and Microsoft will permanently cede market share in the most important computing transition in a generation. The cost of being late is existential in a winner-take-most market.
This is a textbook prisoner’s dilemma operating at the scale of global capital markets. Every individual hyperscaler has private incentives to maintain spending even if the collective outcome, massive overbuild, is visible to all participants. Google co-founder Larry Page was quoted as saying, “I’m willing to go bankrupt rather than lose this race.”⁶ That is not hyperbole. It is the rational game-theoretic equilibrium.
The pressure is self-reinforcing: each quarter of announced capex expansion triggers equity analyst upgrades and investor signaling that, in turn, informs competitors’ own board decisions. Cutting capex now would be read by capital markets as a signal of retreat, compress multiples, and structurally disadvantage the firm.
Hyperscalers are increasingly issuing debt to bridge the gap between capex commitments and internal cash generation. Big tech companies issued $100 billion of bonds in 2026, and investors began demanding record protections via Credit Default Swaps, effectively insurance against default, reflecting rising market skepticism about sustainability.⁷
According to CreditSights, aggregate capex for the top five hyperscalers, after buybacks and dividends, now exceeds projected cash flows, requiring external funding. These firms maintain strong balance sheets, with liabilities-to-assets around 48%, near 2015 lows, but the trajectory is meaningful. Historically, cash-rich businesses are becoming leveraged infrastructure builders.⁸
Hyperscalers routinely describe their markets as supply-constrained rather than demand-constrained, a framing that removes the standard market discipline that would otherwise slow investment. Microsoft disclosed an $80 billion backlog of Azure orders it cannot fulfill due to power constraints.⁹ This is used internally and in investor communications to justify continued aggressive spend.
The problem with this framing is that supply-constrained demand is not the same as demand-constrained demand. The existence of unfilled orders does not tell you whether, once supply is met, customers will maintain those commitments at scale, or whether a portion of that demand is itself speculative, stemming from enterprises bulking up on commitments they believe they may need.
AI-related services generated between $25 billion and $100 billion in revenue in 2025, depending on whether AI revenue or general cloud revenue is counted, compared with hyperscaler capex of $400 billion or more for the same year.¹⁰ Even using the more generous figure, the revenue-to-capex ratio runs roughly one dollar of revenue for every four dollars of infrastructure being built.
Only about 25% of AI initiatives have delivered their expected return on investment to date, and fewer than 20% have been scaled across entire enterprises.¹¹ An MIT study found that 95% of generative AI pilot programs fail to achieve business value, and only 5% of enterprises report significant EBIT impact from AI investment despite widespread adoption.¹²
OpenAI ended 2025 with approximately $20 billion in annual recurring revenue, impressive growth by any measure, and a threefold increase from the prior year, but still a fraction of the infrastructure investment being deployed on its behalf across the ecosystem.¹³
Goldman Sachs projects total hyperscaler capex from 2025 through 2027 will reach $1.15 trillion, more than double the $477 billion spent from 2022 through 2024.¹⁴ The implied revenue needed to justify that at conventional tech multiples has not been modeled publicly in any form that squares with current enterprise adoption rates.
The investment thesis relies on a single untested premise: that today’s massive infrastructure outlays will translate into durable, asymmetric revenue growth. That premise has not yet been tested. As one technology strategist put it, “The biggest untested assumption in the 2026 AI narrative is that today’s valuations are justified by fundamentals that have yet to materialize.”¹⁵
AI adoption is inherently gradual. It requires reskilling workforces, restructuring workflows, navigating regulatory environments, managing data governance and security concerns, and building trust in system reliability. None of these constraints yields to more computing spending. The result is a temporal mismatch: infrastructure depreciation clocks start running the moment hardware is installed. Enterprise revenue realization happens on a human organizational timeline, not a silicon one.
The foundational empirical work by Kaplan et al. (2020) at OpenAI and the Chinchilla paper (Hoffmann et al., 2022) at DeepMind established that large language model performance scales as a power law with model size, training data, and compute.¹⁶ This relationship gave labs a reliable North Star: more of everything reliably produced better models, with predictable loss curves.
The Chinchilla result specifically showed that the optimal training ratio is roughly 20 tokens per parameter; that prior large models had been undertrained relative to their size, not undersized. More data in proportion to model size produced more efficient gains than scaling model size alone.
By 2025, researchers studying advanced reasoning systems found that adding more computational steps no longer delivered proportionate improvements.¹⁷ The power-law relationship is hitting natural ceiling effects. As a model approaches the theoretical limits of what can be predicted from next-token prediction on the available corpus of human text, each additional unit of compute buys exponentially less performance improvement. One arxiv paper formalizes the dynamic: without ongoing efficiency gains in hardware and algorithms, “reaching R(t) = 0.68 could demand 3,000x the current GPU capacity.”¹⁸
A fundamental constraint on further training-scale gains is data exhaustion. EpochAI estimates that roughly 510 trillion tokens of data exist on the indexed web, with the largest known training datasets totaling around 18 trillion tokens. There is headroom on paper, but most remaining data is low-quality or repetitive. Critically, a significant and growing portion of new internet text is itself LLM-generated, meaning training on it creates recursive degradation.¹⁹
The power-law relationship for data is not linear. To get the first unit of performance improvement requires 1 unit of data; the next requires 10; the next, 100. The low-hanging fruit of high-quality, high-diversity human-generated text has been largely consumed.
The most structurally important signal of 2025 was DeepSeek’s R1 model, which demonstrated frontier-level reasoning capability at a fraction of the training cost of competing models. DeepSeek was trained on approximately 2,048 NVIDIA H800 GPUs and matched the performance of OpenAI’s o1 reasoning model, which was developed with resources estimated at roughly 10 times that scale.²⁰
DeepSeek achieved this through several architectural innovations: a mixture-of-experts design that activates only 37 billion of its 671 billion parameters during any given forward pass; FP8 precision training instead of the more compute-intensive FP16 format; multi-head latent attention that reduces memory bandwidth requirements and accelerates inference; and speculative decoding that predicts two tokens simultaneously rather than one.²¹
This was not a one-off. Inference costs dropped from $20 per million tokens to $0.07 per million tokens between 2022 and 2025.²² Mixture-of-experts architectures allow large models to activate only a fraction of their parameters per token, drastically reducing effective compute per inference step. The efficiency curve runs directly counter to the justification for a centralized hyperscale buildout.
The standard rebuttal to the efficiency argument is Jevons Paradox: when a technology becomes more efficient, its use tends to increase rather than decrease, because falling unit costs expand the total addressable market. Applied to AI: cheaper inference means more enterprises can afford it, more applications become economically viable, and aggregate compute demand rises even as per-task compute falls.
This argument has merit. Inference is projected to grow from 33% of AI compute in 2023 to 65% or more by 2029.²³ Reasoning models like DeepSeek R1 consume 150 times more compute per inference than traditional models for complex tasks.²⁴ Total energy demand for AI in the US is projected to increase by 130% by 2030, even accounting for efficiency gains, according to the International Energy Agency.²⁵
However, Jevons Paradox has a critical limitation here: it assumes demand is elastic at the expanded price point. The history of enterprise software says this is far from guaranteed. Enterprise adoption follows organizational readiness cycles that are largely insensitive to reductions in computing prices. A hospital system, a law firm, or a manufacturing conglomerate does not adopt AI 10 times faster because inference costs fell 10 times. It adopts AI when its governance, workflows, regulatory environment, and talent are ready; regardless of unit economics. The inference market may expand, but almost certainly not fast enough to absorb $5 trillion in infrastructure investment on the timeline being built.
The current data center boom has become a macro GDP support mechanism. Barclays estimates that AI infrastructure contributes approximately 0.8-1.0 percentage points to US GDP growth in an environment where total growth runs at around 1.6%.⁴ Deutsche Bank notes that private business investment outside AI-related categories has been essentially flat since 2019, and traditional commercial construction is in decline.⁴
This creates structural feedback risk. A slowdown in AI capex is not merely a tech-sector correction. It is a direct GDP contraction event, with secondary wealth effects (JPMorgan calculates that rising AI equity prices added $180 billion in consumer spending through the wealth effect⁴), potential labor market deterioration, and contagion into broader investment sentiment. Barclays estimates that a 20 to 30% correction in stock prices could cut GDP growth by 1 to 1.5% over the subsequent year. BCA Research’s chief global strategist is direct: “If you take a fragile labor market and you kick it with a capex bust, you’re probably going to get a recession out of it.”⁴
The hyperscalers have, perhaps inadvertently, made the US economy systematically dependent on a capex program justified by returns that have not yet materialized.
Research firm Omdia has modeled a bubble scenario for data center investment, which it assigns a 5% probability, but which illuminates the structural mechanics. In this scenario, capex accelerates through 2027 to approximately $1.4 trillion, then collapses as productivity gains from AI fail to materialize quickly enough. Post-burst capex drops to roughly $1 trillion in 2028; still above 2025 levels in absolute terms, but a 28% drawdown from the peak.²⁶
Even in this scenario, the floor is high. The underlying cloud infrastructure business is real and growing. But the delta between the speculative investment and the rational floor is where value destruction occurs, and that delta is now measured in trillions.
The shift from cash-funded to debt-funded capex introduces a refinancing risk timeline. At some point, capital markets lose the capacity to absorb further AI-infrastructure debt at acceptable rates. Bond investors are already demanding record CDS protection.⁷ If monetization metrics don’t materially improve, the cost of new debt issuance rises, compressing free cash flow further, which accelerates the feedback loop toward capex reduction.
The comparison to the 2000 telecom bust is real but incomplete. That collapse was driven primarily by debt-funded spending at companies without profitable underlying businesses. This buildout is being funded by the most cash-generative companies in history, companies with massive non-AI revenue streams that allow them to sustain speculative infrastructure investment longer than any prior actor could.
This does not eliminate the risk. It extends the timeline and raises the eventual floor for correction. The core question, whether AI-driven productivity gains generate sufficient enterprise value to justify the infrastructure, remains unanswered, and the longer it goes unanswered while capital keeps flowing, the larger the gap that eventually must be closed.
The following metrics provide the clearest early warning signals of stress developing in the system.
Enterprise AI monetization rate. Is the gap between AI-attributable revenue and infrastructure capex narrowing or widening? The current ratio of roughly 1:4 needs to approach 1:1 within three to five years to avoid stranded asset dynamics at scale.
Hyperscaler free cash flow. Combined free cash flow is already shrinking as capex consumes operating income. A sustained negative FCF quarter from multiple hyperscalers simultaneously would be a significant signal.
CDS spreads on hyperscaler debt. Investors are already demanding record protection. Material spread widening indicates capital markets pricing in meaningful default risk; the bond market tends to see the turn before equities.
Inference utilization rates. Are the GPU clusters being built actually running at high utilization, or sitting partially idle? Under-utilization is the physical manifestation of the demand hypothesis being wrong, and it will eventually show up in earnings calls if operators are honest about it.
Enterprise contract behavior. Watch for large enterprise buyers reducing AI cloud commitments rather than expanding them. The supply-constrained narrative falls apart if committed demand softens before supply is met.
Architecture efficiency acceleration. If models continue to improve in capability per FLOP at the pace DeepSeek demonstrated, the case for hyperscale centralized compute weakens materially in favor of edge and distributed architectures, meaning the infrastructure being built today is not the infrastructure that ultimately gets used.
The entire structure of the current cycle rests on whether general-purpose AI automation generates sufficient enterprise productivity gains to create new pricing power, new revenue lines, and new market categories at a pace that justifies the infrastructure. The evidence to date, 5% of enterprises reporting significant EBIT impact, 25% ROI delivery rates, 95% pilot failure rates, suggests the organizational and integration challenges are more profound than the technical community anticipated.
That may change. Early AI adoption metrics in software development, legal document review, and some medical applications are genuinely interesting, but “interesting” is not a substitute for the revenue trajectory needed to close a 1:4 capex-to-revenue ratio across a $5 trillion investment program. The bet is enormous. The evidence base for it is thin at the moment.
KKR Global Macro & Asset Allocation, “Beyond the Bubble: Why AI Infrastructure Will Compound Long after the Hype,” February 2026. https://www.kkr.com/insights/ai-infrastructure
Futurum Group, “AI Capex 2026: The $690B Infrastructure Sprint,” February 2026. https://futurumgroup.com/insights/ai-capex-2026-the-690b-infrastructure-sprint/
BloombergNEF, “AI Data Center Build Advances at Full Speed: Five Things to Know,” March 2026. https://about.bnef.com/insights/commodities/ai-data-center-build-advances-at-full-speed-five-things-to-know/
Data Centre Magazine, “What Risks the US Economy Faces if AI Data Centre Boom Slows,” November 2025. Covers Barclays, Deutsche Bank, JPMorgan, and BCA Research estimates on GDP exposure and correction risk. https://datacentremagazine.com/news/how-the-us-economy-faces-risk-if-ai-data-centre-boom-slows
KKR Global Macro & Asset Allocation, “Beyond the Bubble: Why AI Infrastructure Will Compound Long after the Hype,” February 2026. (Internet boom parallel and telecom consumption analysis.) https://www.kkr.com/insights/ai-infrastructure
IEEE ComSoc Technology Blog, “Hyperscaler Capex > $600bn in 2026,” December 2025. (Larry Page quote and debt issuance data.) https://techblog.comsoc.org/2025/12/22/hyperscaler-capex-600-bn-in-2026-a-36-increase-over-2025-while-global-spending-on-cloud-infrastructure-services-skyrockets/
IEEE ComSoc Technology Blog, “Hyperscaler Capex > $600bn in 2026,” December 2025. (Bond issuance and CDS protection data.) https://techblog.comsoc.org/2025/12/22/hyperscaler-capex-600-bn-in-2026-a-36-increase-over-2025-while-global-spending-on-cloud-infrastructure-services-skyrockets/
CreditSights / MUFG Americas, “AI Chart Weekly: Financing the AI Supercycle,” December 2025. (Liabilities-to-assets ratio and leverage analysis.) https://www.mufgamericas.com/sites/default/files/document/2025-12/AI_Chart_Weekly_12_19_Financing_the_AI_Supercycle.pdf
Futurum Group, “AI Capex 2026: The $690B Infrastructure Sprint,” February 2026. (Microsoft $80B unfulfilled Azure backlog.) https://futurumgroup.com/insights/ai-capex-2026-the-690b-infrastructure-sprint/
Invezz / TradingView, “Looking Ahead to 2026: Why Hyperscalers Can’t Slow Spending Without Losing the AI War,” December 2025. (AI revenue vs. capex gap; $25B AI-related services revenue estimate.) https://www.tradingview.com/news/invezz:751717ae0094b:0-looking-ahead-to-2026-why-hyperscalers-can-t-slow-spending-without-losing-the-ai-war/
Invezz / TradingView, “Looking Ahead to 2026: Why Hyperscalers Can’t Slow Spending Without Losing the AI War,” December 2025. (25% of AI initiatives delivering expected ROI; fewer than 20% scaled enterprise-wide.) https://www.tradingview.com/news/invezz:751717ae0094b:0-looking-ahead-to-2026-why-hyperscalers-can-t-slow-spending-without-losing-the-ai-war/
Cresset Capital, “Market Update 12/17/25: 2026 Outlook: Is AI a Bubble?” December 2025. (MIT study on generative AI pilot failure rates; 5% of enterprises reporting significant EBIT impact.) https://cressetcapital.com/articles/market-update/market-update-12-17-25-2026-outlook-is-ai-a-bubble/
Futurum Group, “AI Capex 2026: The $690B Infrastructure Sprint,” February 2026. (OpenAI $20B ARR, threefold year-over-year growth.) https://futurumgroup.com/insights/ai-capex-2026-the-690b-infrastructure-sprint/
Goldman Sachs, “Why AI Companies May Invest More than $500 Billion in 2026,” December 2025. ($1.15 trillion cumulative hyperscaler capex projection 2025–2027.) https://www.goldmansachs.com/insights/articles/why-ai-companies-may-invest-more-than-500-billion-in-2026
Invezz / TradingView, “Looking Ahead to 2026: Why Hyperscalers Can’t Slow Spending Without Losing the AI War,” December 2025. (Technology strategist quote on untested valuation assumptions.) https://www.tradingview.com/news/invezz:751717ae0094b:0-looking-ahead-to-2026-why-hyperscalers-can-t-slow-spending-without-losing-the-ai-war/
Kaplan, J. et al., “Scaling Laws for Neural Language Models,” OpenAI, 2020. https://arxiv.org/abs/2001.08361 — Hoffmann, J. et al., “Training Compute-Optimal Large Language Models” (Chinchilla), DeepMind, 2022. https://arxiv.org/abs/2203.15556
Sapien.io, “When Bigger Isn’t Better: The Diminishing Returns of Scaling AI Models,” November 2025. https://www.sapien.io/blog/when-bigger-isnt-better-the-diminishing-returns-of-scaling-ai-models
Lu, Chien-Ping, “The Race to Efficiency: A New Perspective on AI Scaling Laws,” arXiv, January 2025. (3,000x GPU capacity formalization of diminishing returns under static efficiency.) https://arxiv.org/pdf/2501.02156
Jon Vet, “A Brief History of LLM Scaling Laws and What to Expect in 2025,” December 2024. (EpochAI data on available token counts; LLM-generated text contamination of training corpora.) https://www.jonvet.com/blog/llm-scaling-in-2025
S&P Global Market Intelligence, “Potential Impacts of DeepSeek on Datacenters and Energy Demand,” March 2025. (DeepSeek R1 training hardware comparison to GPT-4.) https://www.spglobal.com/market-intelligence/en/news-insights/research/potential-impacts-of-deepseek-on-datacenters-and-energy-demand
Brookings Institution, “Why AI Demand for Energy Will Continue to Increase,” August 2025. (DeepSeek R1 architectural innovations: MoE parameter activation, FP8 training, MLA attention, speculative decoding.) https://www.brookings.edu/articles/why-ai-demand-for-energy-will-continue-to-increase/
Introl, “AI Inference vs. Training Infrastructure: Economics Diverging,” April 2026. (Inference cost decline from $20 to $0.07 per million tokens, 2022–2025; Stanford 2025 AI Index.) https://introl.com/blog/ai-inference-vs-training-infrastructure-economics-diverging
Introl, “AI Inference vs. Training Infrastructure: Economics Diverging,” April 2026. (Inference share of AI compute projected to grow from 33% in 2023 to 65%+ by 2029.) https://introl.com/blog/ai-inference-vs-training-infrastructure-economics-diverging
Introl, “AI Inference vs. Training Infrastructure: Economics Diverging,” April 2026. (DeepSeek R1 consuming 150x more compute per inference than traditional models for complex tasks; NVIDIA GTC demonstration data.) https://introl.com/blog/ai-inference-vs-training-infrastructure-economics-diverging
International Energy Agency, “Electricity 2025,” cited in Brookings Institution, “Why AI Demand for Energy Will Continue to Increase,” August 2025. (US data center energy demand projected to increase 130% by 2030 versus 2024 levels.) https://www.brookings.edu/articles/why-ai-demand-for-energy-will-continue-to-increase/
Data Center Dynamics, “Data Center Investment Likely to Hit $1.6 Trillion by 2030,” March 2026. (Omdia bubble scenario modeling; capex trajectory and post-burst floor analysis.) https://www.datacenterdynamics.com/en/news/data-center-investment-likely-to-hit-16-trillion-by-2030-report/
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.