The current capital expenditure on AI, as a percentage of GDP and nonresidential fixed investment, is larger and happening at a much faster pace than historical projects like railroads, electrification, or the fiber optic build-out.
External capital providers are financing data centers by analogizing them to multi-tenant buildings. They evaluate investments based on expected cash flow and comparable cap rates from commercial real estate (around 6-7%), not on speculative tech market sizes.
Unlike a building with upfront CapEx, data centers are like non-regulated utilities. They require wholesale replacement of hardware (GPUs) every 4-7 years, creating ongoing capital needs that dilute investor returns and break the simple real estate investment model.
AI data centers produce "tokens," a commodity whose price falls 70-80% annually. Investors provide capital based on fixed-return expectations (like real estate cap rates), but the underlying revenue-generating asset is rapidly deflating, creating a fundamental economic mismatch.
The mean time between failure for GPUs depends heavily on their use. Chips used for intensive model training have much higher failure rates than those used for inference. Investors often conflate the two, underestimating the high churn and replacement costs for training hardware.
When AI companies like OpenAI and Anthropic compete with customers, it's a defensive strategy driven by the commoditization and price collapse of their core product (tokens). They are desperately searching for higher-margin revenue streams as their fundamental business model erodes.
If a frontier model company truly had technology to "eat the economy," its most profitable move would be to apply it directly, not sell access via tokens. The very act of selling the service is proof that the technology isn't as all-powerful as claimed.
Unlike past bubbles (e.g., railroads), today's investors explicitly justify AI overspending by citing the eventual positive outcomes of prior bubbles. This creates a self-reinforcing, reflexive loop that encourages even greater excess than was seen in previous historical cycles.
The argument that AI investment is a "call option on AGI" is not the justification used in partner meetings. It’s a marketing narrative used to persuade limited partners (LPs) and shut down debate. Internally, these are cold, calculated project finance decisions.
Sovereign funds managing hundreds of billions filter opportunities by "check size." Giant AI data center projects are uniquely attractive because they are one of the few asset classes that can absorb $50-100 billion checks, creating a perverse incentive to fund them regardless of underlying economics.
Much of the perceived improvement in AI over the last 18 months comes from better "harnesses"—software layers that orchestrate and manage models. This suggests massive spending on new training runs is becoming less critical, threatening the business model of hyperscalers.
The ultimate fate of AI is to become a background utility, similar to the power grid. Consumers will no more know who provides their AI "tokens" than they know which hydroelectric dam powers their laptop. This implies a future of low, utility-like returns, not high-margin tech profits.