Detailed Analysis
The central argument advanced in this piece cuts against the prevailing narrative of an AI bubble by insisting on a more precise analytical framework: the question is not whether AI is a bubble, but which specific layers of the AI ecosystem exhibit bubble dynamics and which are anchored to verifiable, existing demand. The author points to OpenAI's revenue trajectory — from roughly $2 billion annualized in 2023 to more than $20 billion by 2025 — and to Anthropic's even faster growth from a smaller base as foundational evidence that underlying consumption is real. Nvidia's fiscal 2026 data center revenue of approximately $193.7 billion serves as a further corroborating signal: capital expenditure at that scale reflects purchasing decisions made by serious institutional actors with live inference and training workloads, not speculative pilots. The hyperscalers — Google, Microsoft, Amazon, and Meta — are collectively on pace to deploy roughly $700 billion in AI infrastructure spending in the current year, a figure the piece acknowledges is "frankly absurd" but frames as a supply-chain response to demand that already exists rather than a bet on demand that might materialize.
The article draws a meaningful distinction between enterprise revenue and consumer curiosity, arguing that the former is structurally more durable. Enterprises now constitute approximately 40% of OpenAI's revenue and an even larger share at Anthropic, and the piece notes that companies are struggling to onboard customers fast enough to meet demand. This is a qualitatively different signal from consumer subscription churn. Businesses committing real budget to AI tools — for coding, compliance, research, customer operations, and sales workflows — are doing so because an internal champion has identified a quantifiable process improvement, not because a chatbot demo was entertaining. The piece acknowledges that some of this enterprise spending is FOMO-driven, but argues that even fear-driven adoption does not render the underlying utility irrational, particularly when the tools are genuinely replacing hours of slower or more expensive labor.
Where the piece concedes ground to the bears is on the mismatch between revenue and infrastructure spending. The $700 billion in annual capital expenditure by the hyperscalers is not yet matched by a clean, legible return-on-investment story at the enterprise level, and the article does not pretend otherwise. Some data centers will be financed on assumptions that fail under real-world conditions. Some suppliers will be overpaid. Some companies will overbuild the wrong capacity in the wrong geography. These are legitimate bubble dynamics — but they are bubble dynamics in the financial and logistical superstructure around AI, not in the demand signal itself. The piece uses the telling detail of company leaders complaining that developers are burning through Claude credits too fast as evidence that demand pressure and asset-price froth can coexist without one invalidating the other.
This framing connects to a broader pattern in technology cycles where the speculative and the fundamental become conflated during corrections, leading to analytical errors in both directions. The dot-com collapse is the canonical case: internet stocks were catastrophically overvalued, but internet demand was not imaginary — it simply took longer to monetize than investors priced in. The article implicitly positions AI in a similar structural moment, where a stock correction reflects investor impatience and stretched valuations rather than a verdict on the technology's utility. The growth curves at Anthropic and OpenAI — described as the fastest revenue ramps in private company history — are the clearest available evidence that this distinction has practical weight. If demand were fake, enterprise customers with procurement processes and budget cycles would not be generating those curves. The more productive analytical question, the piece suggests, is not "bubble or not" but rather which specific bets within the AI supply chain are financed on assumptions that contact with reality will not support — and which are simply serving demand that the infrastructure has not yet caught up to.
Read original article →