Map

AI Bubble Pale Horses

Wiki analysisai-bubblepredictionseconomics ↳ show in map Markdown
title
AI Bubble Pale Horses
type
analysis
summary
Checklist of warning signs that the AI bubble is deflating, tracking which have fired
tags
ai-bubble, predictions, economics
sources
subprime-ai-crisis, ai-too-expensive
created
2026-04-06
updated
2026-07-18

Edward Zitron first published a set of "pale horses of the AIpocalypse" in August 2024. The March 2026 subprime-ai-crisis piece updates the list. These are observable signals that the AI bubble is deflating. Tracking them over time will show whether Zitron's thesis holds.

The Signals

Already Happening (as of March 2026)

  • Rate limit tightening at Anthropic: Peak hours introduced March 26, 2026. Users on $100-200/mo plans hitting limits in minutes. Weekly limits added July 2025.
  • AI startup pricing austerity: Cursor switched to token-pass-through (June 2025). Replit moved to effort-based pricing (June 2025). Augment Code went per-message then credits (Oct 2025). Perplexity slashed deep research from 600/mo to 20 (Feb 2026).
  • CoreWeave debt issues: Trouble raising for Lancaster PA data center (Feb 2026).
  • OpenAI cutting products: Sora killed (March 2026) — burning $1-15M/day.
  • SoftBank leveraging up: $40B debt (payable in one year) to fund OpenAI round, exceeding its own 25% loan-to-asset ratio.

Not Yet Happened — Watch For

  • Big tech capex reduction: If Microsoft, Google, Amazon, or Meta cut GPU spending, NVIDIA's growth story breaks. This is the single biggest domino.
  • NVIDIA earnings miss: Markets would panic. All current growth depends on hyperscaler and data center GPU purchases.
  • Data center deal collapse: Three escalating tiers — (1) unbuilt deal cancelled, (2) mid-construction project abandoned, (3) completed center goes dark. None fully visible yet, but the 200GW announced vs 5GW building gap suggests many will never happen.
  • Stargate Abilene problems: OpenAI's flagship data center being built by Oracle. If construction stalls or Oracle can't finance it, the symbolic impact would be enormous.
  • OpenAI or Anthropic IPO delays: Both companies are heading for public listings. Any delay signals the financials are too ugly to dress up. Both lose billions annually with no profitability path.
  • Blue Owl trouble: The most aggressive AI lender. Problems with its loan book or limited partners would tighten credit across the sector.
  • ARM stock below $80: SoftBank holds a $15B margin loan against ARM shares. A drop below $80 threatens Masayoshi Son's ability to fund AI bets.
  • AI company layoffs: Would break the narrative of unstoppable growth.
  • NVIDIA customer payment issues: If data centers or hyperscalers can't pay for GPUs they've ordered, the top of the chain cracks.

How to Use This List

This is a checklist, not a prediction of timing. Zitron's thesis is that these are structural inevitabilities given the subsidy structure, not speculative risks. The question is sequence and speed.

For future reference: compare this list against actual events quarter by quarter. Which signals fired? Which ones didn't? Did the cascade happen in the predicted order?

For the view from inside the building — what mandate-driven AI adoption actually looks like at the team level — see ai-great-leap-forward.

May 2026 update: signals that have fired

ai-too-expensive is Zitron's May 2026 follow-up; several signals can now be marked off:

  • Enterprise token-budget revolt has begun — Uber, ServiceNow, Salesforce, Stripe, Zillow all blowing through annual budgets in months. New concept page: ai-token-budget-explosion.
  • Data-center cancellations hit record levels in Q1 2026 (also reported by Axios polling coverage).
  • Anthropic and OpenAI fundraising cadence is accelerating beyond even the inflated 2026 projection: Anthropic raised $75B in six months and is reportedly raising $30B more in May. OpenAI's $122B February raise is followed by reports of more capital being sought.
  • Hyperscaler capex commitments for 2026 ($700B) and 2027 ($1T) are now public, against a combined four-company revenue of $1.6T. Math from ai-too-expensive: needs $3T+ in AI revenue to break even.
  • Anthropic margin compression — gross margins 40% in 2025, 10 points below projection. OpenAI's fell from 40% to 33% over the same period. The knife-catching-compute dynamic is now visible in the numbers.

The picture has tilted further toward Zitron's thesis since the March piece, without any single signal yet producing the herd-trigger event (a public CFO budget cut).

The counterweight: a bubble can pop and still leave a legacy

Zitron's frame treats the correction as the end of the story. Caballero's July 2026 speculative-growth-ai-bubble argues it isn't necessarily. In his model a temporary, unsustainable overvaluation can install enough real capital to push the economy into a permanently higher-capital equilibrium — the technology is real, the peak prices are not, and the capital stays. This doesn't contradict any signal on this list; it accepts the overvaluation and the correction and disputes only the aftermath. The speculative-growth mechanism is fragile (the legacy survives only if the correction comes after enough capital is installed) and turns on one empirical question this checklist doesn't track: whether the installed capital — data centers and power, but especially fast-depreciating GPUs — retains value after the crash the way dot-com fiber did. Worth holding next to the pale-horses list as the "even if it pops" case.

Sub-pages
Linked from