The AI Capital Trap

The technology may be transformative. The investment can still be disastrous.

That distinction is the foundation of this argument. The bear case for AI does not require models to stop improving, people to stop using them, or the technology to become irrelevant. It requires something much more ordinary: revenue arriving more slowly than depreciation, financing costs, power commitments, and the next hardware cycle.

AI is being financed like durable infrastructure while its core assets age like consumer electronics and its output is increasingly priced like a commodity.

The central claim: Usage can compound while returns on the underlying compute compress.

This is a bear case, not a prediction of certainty. Reported figures are sourced. Financial-model assumptions are exposed and adjustable. The strongest counterarguments are included because a thesis that cannot survive them is not useful.

The outlay

Amazon, Alphabet, Microsoft, and Meta together point to approximately $665 billion of company-wide capital expenditure across their current 2026 reporting years.

Company	2026 reference figure	Basis
Amazon	$200B	Company-wide guidance
Alphabet	$185B	Guidance midpoint
Microsoft	$145B	Reported FY2026 quarterly total, rounded
Meta	$135B	Guidance midpoint
Combined	$665B	Reference total

These totals are not pure AI spending. The companies do not publish a clean, comparable “AI capex” line item. But each identifies AI or technical infrastructure as a central driver, and the comparison is useful precisely because it shows the scale of the capital cycle without pretending that disclosure is more precise than it is.

The total is not automatically irrational. These are profitable companies funding infrastructure that supports cloud computing, advertising, logistics, and other businesses in addition to generative AI. Scale alone does not prove a bubble.

The risk comes from the clocks attached to that scale.

Three clocks are running against the buildout

1. The price clock

Intelligence is getting cheaper faster than demand can become proprietary.

Model competition reduces token prices. Open-weight models narrow performance gaps. Inference optimizations reduce the compute required per task. Better hardware performs more work per watt and per dollar. All of this is excellent for adoption.

It is less obviously good for the owner of a data center underwritten on scarcity pricing.

Commodity industries frequently experience rising demand alongside disappointing returns. When every supplier expands capacity in response to the same signal, the eventual abundance transfers pricing power to customers. AI compute has an additional source of pressure: the product itself becomes more efficient while physical capacity is still being constructed.

2. The hardware clock

The highest-cost assets sit in the fastest-moving layer of the stack.

Microsoft reported that roughly two-thirds of recent quarterly capital expenditure went to CPUs and GPUs—assets it described as short-lived. A new accelerator does not need to make an older chip physically useless. It only needs to deliver materially better performance per dollar or per watt.

That creates a difference between accounting life and economic life. A server may continue operating for years while becoming uncompetitive for the workloads that justified its purchase.

The cash leaves first. Depreciation reaches the income statement later. By the time the full expense is visible, a replacement cycle may already be underway.

3. The power clock

Compute is scaling into the slowest parts of the physical economy.

The International Energy Agency projects global data-center electricity consumption to roughly double by 2030. Transmission, substations, transformers, generation, land, and interconnection queues do not improve at software speed.

Those constraints can make existing capacity more valuable in the short run. They can also turn growth assumptions into long-dated contractual obligations that are difficult to unwind when economics change.

The trap is a loop, not a single bad bet

1. Scarcity: Demand outruns available compute.
2. Overbuild: Capital floods into chips, power, and data centers.
3. Abundance: Physical capacity and model efficiency compound together.
4. Compression: Prices fall, customers gain leverage, and margins narrow.
5. Impairment: Revenue misses the capital-recovery schedule.

Every decision inside the loop can be rational. No executive wants to be the company that lacks compute if AI becomes the next dominant platform. No cloud provider wants its largest customers to leave because capacity is unavailable. No investor wants to miss the apparent infrastructure layer of a generational transition.

The collective outcome can still be destructive. When every participant responds to the same scarcity signal, the industry builds against demand that assumes competitors will not build as aggressively.

The interactive model below is intentionally simple. It asks how much annual revenue a single infrastructure cohort must generate to cover annualized capital recovery and a cost-of-capital charge. It excludes power, labor, taxes, and future replacement capex, so it should be read as a lower-bound stress test rather than a complete valuation model.

The honest countercase

There are at least four serious reasons this thesis could be wrong.

The buyers are not fragile

The hyperscalers are among the most profitable companies ever built. They can absorb years of weak direct returns while smaller competitors cannot. Infrastructure that looks uneconomic in isolation may protect a much larger franchise.

Demand is already constrained

Microsoft and Alphabet report demand that exceeds available supply, strong cloud growth, and expanding contracted backlogs. The buildout may be catching up to paying demand rather than manufacturing speculative capacity.

Commitments may transfer the risk

Amazon says a substantial portion of expected AWS capital expenditure is supported by customer commitments. Long-term contracts could make much of the investment more predictable than public revenue disclosures imply.

AI can monetize indirectly

Better advertising, retention, productivity, and product quality can create returns without appearing as a standalone “AI revenue” line. A narrow comparison between AI subscriptions and infrastructure spending may miss the most valuable benefits.

These points do not invalidate the bear case. They define the evidence required to evaluate it honestly.

What would change the conclusion

• Reported, recurring AI-linked revenue approaches the run rate required to cover the assets behind it.
• Inference prices and gross margins stabilize after new capacity reaches the market.
• Capital expenditure slows before depreciation and financing charges peak.
• Utilization remains high without discounts, credits, or circular financing arrangements.
• Older accelerator fleets retain economically valuable workloads after newer generations arrive.

Primary sources

1. Amazon 2026 capital expenditure guidance — Amazon / SEC
2. Alphabet Q1 2026 earnings call — Alphabet Investor Relations
3. Microsoft FY2026 Q4 earnings — Microsoft Investor Relations
4. Meta Q1 2026 results and outlook — Meta / SEC
5. Energy and AI: data-center demand — International Energy Agency
6. Microsoft FY2026 Q3 capital expenditure mix — Microsoft Investor Relations
7. Amazon 2025 shareholder letter — Amazon Investor Relations

This essay is analysis, not investment advice. Forecasts are uncertain, and the scenario language above is explicitly hypothetical.