Yes, the US AI Bubble Is Real. Here's Why.

Position taken. The CrossVol desk does not equivocate. The US AI infrastructure complex is in a bubble by any honest metric — capex to revenue, debt maturity to depreciation, optionality to fundamentals. The full framework, with dated catalysts and trade construction, is in The China AI Disruption Thesis (Amazon Kindle, May 2026).

Almost every sell-side desk has converged on the same narrative: hyperscaler capex grows 25–30% CAGR through 2028, US power demand doubles by 2027, memory stocks enter a supercycle extending through 2028. The data tables across Goldman, Morgan Stanley, JPMorgan are nearly identical. The retail crowd has internalized it. NVDA, Vertiv, the datacenter REITs, the power complex — all are priced for the consensus.

The consensus is wrong, and it is wrong on quantifiable grounds. Below, the five arguments. None of them are speculative. Each one resolves on a specific date or against a specific number. If you want the construction — the trades, the sizing, the calendar — read the book. If you want to know why we're calling a bubble, keep reading.

1. Hyperscaler Capex Is Outrunning Hyperscaler Revenue

In 2023, Amazon, Microsoft, Google and Meta spent a combined $147 billion in capital expenditures. In 2024 it was $230 billion. In 2025 it was $310 billion. The 2026 consensus run-rate is $390 billion. Over the same three-year window, cloud and AI-related revenue at these four companies grew from roughly $230 billion to roughly $340 billion — a 47% expansion against a 165% expansion in capex.

The narrative explanation is "front-loaded investment that will pay back over a decade." The quantifiable problem is duration mismatch. The depreciation schedule on the GPUs being deployed today is three to six years. The corporate bonds being issued to finance the spend are seven to thirty years. By 2028, on the consensus's own numbers, the four hyperscalers will have $1.4 trillion of AI infrastructure to depreciate against revenue streams that have not been written yet.

2. The Corporate Bond Wall

In 2025, the five largest US hyperscalers issued $121 billion of long-dated investment-grade debt — a 4.3× jump from the prior-decade average. Year-to-date 2026, the run-rate tracks toward $230–240 billion. The debt is long-dated (10y, 20y, 30y), the underlying revenue stream is short-cycle (enterprise contracts, ad spend, cloud renewals). Pension funds, insurance companies, sovereigns have absorbed the issuance because the spread looked attractive against treasuries.

When the revenue assumptions get revisited — and they will, because token prices are collapsing — the credit market repricing happens in days. The 2022 episode in long-dated tech IG (Meta, Salesforce) gave a preview: a 65 bp spread widening in eleven sessions. Repeat that at 2× scale, on hyperscaler paper held by every pension fund in the West. The equity drawdown is one consequence. The credit drawdown is the bigger one.

3. Token Commoditization Is Already Underway

On 22 May 2026, DeepSeek announced that its 75% discount on V4 Pro inference would become permanent. The discount had been treated as a promotional subsidy since January. The permanent print changes the calculation. DeepSeek V4 Pro inference is now $0.87 per million output tokens. The same task on GPT-4o or Claude Sonnet runs $25–30 per million output tokens. The intelligence delivered is, by published benchmarks, within 5–10% on most enterprise workloads.

This is the precise pattern of every commoditization cycle in technology history. The frontier player invests heavily, the fast follower delivers 90% of the capability at 5% of the cost, the gross margin on the frontier collapses. In storage it was the 2007 Chinese DRAM entry. In solar it was the 2011 Chinese polysilicon push. In LED it was 2014. The frontier didn't disappear in any of those — but the equity rerated 50–70% before the bottom.

4. The US Power Grid Bottleneck Is Math, Not Narrative

The PJM 2026/2027 capacity auction cleared at $329.17 per MW-day in May 2025 — an 11.4× increase over the prior auction. Of the announced US data center projects through 2030, roughly 50% have been delayed or cancelled because of grid interconnection queues that now run four to seven years in the densest markets. The capex narrative assumes the data centers get built. Half of them won't get built on schedule, and the half that do are 25–40% more expensive than 2023 underwriting assumed.

This isn't a regulatory risk. It is electrical engineering. Transformers take eighteen months to manufacture. High-voltage transmission corridors take five years to permit. The bottleneck is real, the numbers are public, and the equity prices do not reflect them.

5. The Dot-Com Analog Is Tighter Than Everyone Wants It To Be

Comparison to 1999–2000 invites lazy thinking — every market downturn gets called "the new dot-com." Most are nothing of the kind. This one is. The shared pattern is not "internet bad." The shared pattern is capex front-loading against unproven unit economics, financed by long-dated debt, sold to retail and pension allocators on a single dominant narrative.

In 1999, the telecom buildout assumed exponential bandwidth demand. Real bandwidth demand grew but at one-third the projected rate. Cisco, Lucent, JDS Uniphase — the picks-and-shovels names — fell 80–95% between March 2000 and October 2002. The internet did get built. The companies that financed it got destroyed. Two decades later, the same pattern reasserts itself on a different layer of the stack. The names are different. The mechanics are not.

The Window: Q4 2026 → Q1 2027

This is not a forecast that the unwind starts next week. Bubbles end on catalysts, not on valuation alone. Two dated catalysts are visible:

• 10 November 2026 — the US–China tariff truce expires. Whatever the renegotiated terms, the supply chain assumptions baked into hyperscaler unit economics will be revisited.
• 27 November 2026 — the suspension of Chinese export controls on gallium, germanium and antimony expires. These are the rare-earth inputs without which 30%+ of US AI-infrastructure equipment cannot be manufactured at current price points.

Between these two dates and the Q1 2027 earnings cycle, the consensus narrative gets stress-tested against reality. The CrossVol desk expects pure AI-infrastructure equities to draw down 25–40% across that window. The companion outperformance is in open-source AI architectures, edge inference platforms, critical mineral miners outside China, and Chinese AI platforms with monetization paths.

What This Article Doesn't Cover

This piece states the case. It does not lay out the trade. Constructing a position with appropriate sizing, duration, vega exposure and hedge legs requires the full framework — not bullet points. The China AI Disruption Thesis covers it in detail across fourteen chapters: the five vectors, the four-front geopolitical chessboard (Iran, Greenland, Venezuela, Cuba), the calendar of catalysts, and a chapter dedicated to trade construction (dispersion, variance, single-name shorts, credit hedges).

The four-lens market-reading framework that produced the call — Gamma, Vega, Risk Reversal, Term Structure & Physical — is in the companion volume Beyond Gamma Exposure, also on Amazon Kindle.

---