The five largest hyperscalers, Amazon, Alphabet, Microsoft, Meta, and Oracle, are targeting roughly $725 billion in capital expenditure for 2026, with approximately 75% directed at AI-specific infrastructure: GPU clusters, data center construction, power systems, and cooling. Goldman Sachs projects total AI capital expenditures across all players at around $765 billion in 2026, representing about 2.4% of US GDP. That ratio exceeds levels seen in past innovation cycles, including the dot-com bubble.

The question dominating investor calls, analyst reports, and policy discussions: can enterprise AI adoption generate returns that justify this spending, or is the industry building infrastructure for demand that may not materialize at the necessary scale?

The scale of spending

Multiple sources triangulate on similar numbers:

Moody’s warning (July 2026). The rating agency flagged that combined capital expenditures of the six largest AI investors are set to reach approximately $785 billion in 2026 and are likely to approach $1 trillion by 2027. The warning highlighted that ROI timelines remain uncertain.

Goldman Sachs analysis. AI capex at $765 billion in 2026 would represent 2.4% of US GDP, above the peak of the dot-com infrastructure buildout.

CNBC projections. Analysts forecast hyperscale capex to exceed $1 trillion in 2027, with some projections suggesting AI infrastructure spending could reach $3–4 trillion annually by the end of the decade.

Bank for International Settlements. A BIS systemic risk study found that the AI boom has outgrown every previous tech bubble in history by capital commitment at this stage of the cycle.

The bull case

Proponents argue the spending is rational:

Demand is real. Enterprise AI adoption is accelerating. Every major cloud provider reports strong demand for AI services, with waitlists for GPU capacity.

Infrastructure takes time. Data centers have 2–5 year construction timelines. Building now anticipates demand that will materialize over the next several years.

Winner-take-most dynamics. AI infrastructure has network effects and scale economies. The company that reaches scale first captures disproportionate value.

New use cases. Agentic AI, AI coding assistants, and AI-native applications are in early stages. Demand could grow faster than current forecasts.

The bear case

Critics point to several structural concerns:

Revenue does not match capex. AI-related revenue at the hyperscalers, while growing, remains a fraction of the infrastructure investment. The gap between spending and revenue is widening, not closing.

Depreciation mismatch. Depreciation schedules set in 2023 may understate real asset depletion by an estimated $176 billion between 2026 and 2028, according to one analysis. GPU generations turn over faster than depreciation assumes.

Off-balance-sheet exposure. Over $662 billion in lease commitments sit off balance sheets, complicating the picture of how much capital is truly at risk.

Enterprise ROI uncertainty. While pilots are widespread, production AI deployments that generate measurable ROI remain the exception. MIT Sloan’s 2026 AI trends analysis noted that the gap between AI spending and AI value extraction remains a core tension.

Concentration risk. The US stock market has become synonymous with the AI boom. Nine of the ten most valuable US companies are tech companies betting heavily on AI. A correction in AI valuations would have systemic effects.

Historical parallels

The parallels to past bubbles are contested:

The railroad comparison. Fortune ran an essay by Pulitzer Prize winner Liaquat Ahamed comparing the current AI buildout to the 1873 railroad crash. The argument: massive infrastructure spending, if it gets ahead of demand, leads to busts even when the underlying technology is valuable.

The dot-com comparison. The infrastructure buildout of 1998–2000 exceeded near-term demand, leading to a crash, but the fiber and data centers laid during that period enabled the next generation of internet services.

The counterargument. Unlike dot-com, the AI buildout is funded primarily by large, profitable companies with strong balance sheets, not by speculative startups burning venture capital.

What the companies say

The hyperscalers maintain confidence in their spending:

  • Microsoft CEO Satya Nadella has described AI infrastructure as a “generational opportunity.”
  • Meta’s Mark Zuckerberg has defended AI capex as essential to competitive positioning.
  • Amazon and Google have both emphasized the long-term nature of data center investments.

However, investor sentiment is shifting. Multiple analysts noted in July 2026 that share prices of major AI infrastructure beneficiaries were under pressure despite strong reported demand.

Why it matters for builders

The macro uncertainty has practical implications:

Pricing. If the AI infrastructure buildout proves excessive, overcapacity could lead to lower inference and training costs. If demand exceeds supply, costs could remain elevated or rise.

Provider stability. The financial health of your AI providers matters. If one of the major players faces pressure to cut back, service quality and availability could suffer.

Build vs. buy timing. In a potential bubble scenario, buying services now may be preferable to building infrastructure that could depreciate rapidly as the market shakes out.

Planning horizons. Multi-year AI strategies should account for the possibility of significant market dislocation. Commitments that assume current pricing and provider landscape continuing unchanged carry risk.

Sources

Further reading