5 Reasons AI is A Bubble (And 5 It's Not)

October 8, 2025

ai-daily-brief-podcast

Is AI a Bubble? — Five Reasons It Is and Five It’s Not

Overview

This episode of the AI Daily Brief (hosted by Nathaniel Whittemore, podcast and video series covering major AI news and analysis) tackles the recurring question of whether the current AI investment boom constitutes a financial bubble. The episode is prompted by a wave of renewed bubble discourse in early October 2025, triggered by The Information’s reporting on Oracle’s thin cloud profit margins and xAI’s $20 billion funding round. The host structures the analysis around five (ultimately six) arguments for and against the bubble thesis, then adds his own editorial perspective and a watchlist of indicators to monitor.

Source video: URL not provided in submission.

Prerequisites

Basic familiarity with financial concepts: gross margin, price-to-earnings (P/E) ratios, depreciation schedules, vendor financing, commercial debt markets, and credit default swaps
General awareness of the AI industry landscape: OpenAI, Anthropic, NVIDIA, Oracle, xAI, Microsoft Azure, Google, Meta, CoreWeave
Familiarity with historical tech market events: the dot-com bubble (1999–2001), the 2008 Global Financial Crisis (GFC), the DeepSeek moment (early 2025)
Basic knowledge of startup revenue metrics (ARR — Annual Recurring Revenue)
Understanding of what large language model inference is and why compute demand matters

Main Points

Background: The Bubble Debate Has Been Continuous Since Late 2022

ChatGPT launched November 2022; bubble calls followed within weeks (Josh Brown of Ritz-Holz Wealth Management, February 2023)
Goldman Sachs countered in September 2023, arguing stocks had room to run despite 60% gains
Bubble discourse peaked in August 2025 (GPT-5 disappointment, the widely-circulated but contested MIT “95% of AI pilots failing” study, Sam Altman’s public comments)
Deutsche Bank (September 30, 2025) declared the bubble narrative itself had burst, with web searches for “AI bubble” falling 85% from their August 21 peak
Within two weeks, searches tripled again, driven by Oracle’s margin data and xAI’s funding round

The Triggering Events (October 2025)

The Information’s Oracle exposé: Oracle’s AI cloud gross margin was ~14% in the three months ending August 2025 ($125M profit on ~$900M revenue) — below Walmart’s 25% and far below Azure’s 69%
Oracle stock fell 6.9% on the news before partially recovering
xAI’s $20 billion raise: A mixed debt/equity round to fund the Colossus II data center in Memphis; NVIDIA invested $2 billion on the equity side, which critics flagged as circular investment
OpenAI–AMD deal: Non-traditional structure in which OpenAI could own 10% of AMD if milestones are hit, also viewed as bubbly by some observers

Five (Six) Reasons AI Is a Bubble

1. Circular Investment

NVIDIA invests $2B in xAI → xAI buys NVIDIA chips → revenue appears on NVIDIA’s books three months later
OpenAI’s deals with AMD and Oracle follow similar patterns
Critics (e.g., Stanfield Capital) argue: “None of these circular AI deals would be happening if there were genuine cash demand for chips at list price”
The concern is that this artificially inflates reported demand across the ecosystem

2. Risk of Infrastructure Overbuilding

Trillions of dollars in AI data center commitments are planned over the next five years
Bain & Company forecast that $2 trillion in annual revenue would be needed to justify AI compute spending by 2030; they estimate an $800 billion shortfall
Ares Capital CEO Kip Devere warned that historically, large capital surges in a sector produce marginal overbuilds
Hyperscalers themselves, however, continue to insist their primary risk is underbuilding, not overbuilding

3. Echoes of the Dot-Com Bubble

Senior Wall Street figures had formative experiences in the dot-com era and see structural parallels: overbuilt infrastructure (90% of fiber optic cable lay dormant post-burst), circular revenues (dot-com ad spend largely flowed between dot-com companies)
OpenAI chairman Bret Taylor stated: “I think there’s a lot of parallels to the internet bubble… I think we’re also in a bubble and a lot of people will lose a lot of money”
The concern is not that AI is useless — it is that prices are detached from near-term fundamentals, as they were in 1999–2000

4. Stretched Valuations

The Shiller CAPE (Cyclically Adjusted Price-to-Earnings) ratio for the S&P 500 hit its highest level in September 2025 since the year 2000
AI investments have accounted for an estimated 40% of U.S. GDP growth and 80% of U.S. stock gains in 2025 (per Rockefeller International Chair Rishi Sharma writing in the Financial Times)
Sharma characterizes AI as a “magic fix” narrative that U.S. economic policy is now deeply dependent on, making a potential AI bust more systemically dangerous

5. Speaking the Bubble Into Existence (Narrative Risk)

Bubbles are partly narrative constructs and unwind when results fail to meet expectations
Sam Altman has made repeated public comments acknowledging boom-bust cycles, “dumb capital allocations,” and that AI stocks reacting to OpenAI mentions on stage is “weird”
These unguarded comments, compared to the tighter messaging of Jensen Huang or Larry Ellison, are seen as contributing to market nervousness

6. (Bonus) AI Has Become “Too Big to Fail” — and Too Big to Ignore

J.P. Morgan reported AI companies now represent 14% of the investment-grade debt market ($1.2 trillion in commercial bonds), overtaking U.S. banks
Rishi Sharma’s argument: the sheer concentration of AI exposure in U.S. economic growth makes the sector simultaneously “too big to fail” and wildly speculative

Five (Six) Reasons AI Bubble Talk Is Overblown

1. Revenues Are Real

NVIDIA’s earnings growth has tracked its stock price closely — unlike Cisco in 1998–2000, where valuation detached sharply from earnings
NVIDIA generated $46 billion in sales in a single quarter; Pets.com had $9 million in peak annual revenue at the height of dot-com
OpenAI has consistently beaten its revenue forecasts each year since ChatGPT launched; Anthropic is now reportedly growing even faster
Top 100 AI startups reach $1M ARR in 11.5 months vs. 15 months for non-AI SaaS (per Stripe data)
OpenAI and Anthropic’s ~$20B combined revenue comes from direct product purchases — not advertising or structured circular deals

2. Low Leverage

Google and Amazon are funding data center construction largely from their own balance sheets
Microsoft has net cash and trades at a 5 basis point premium to U.S. Treasuries — effectively more creditworthy than the U.S. government
The AI debt market is described by J.P. Morgan as “cash-rich, not highly levered” with bonds trading tight for sound reasons
Historical pattern: bull markets are killed by credit crunches, not by age or skepticism; no credit crunch is currently visible

3. Demand Is Real and Accelerating

OpenAI is serving approximately 3 quadrillion tokens per year (as of Dev Day 2025)
Google reported a 100x increase in monthly tokens served between May 2024 and July 2025, with a 100% increase in just the May–July 2025 window alone
A group of 30 power users across enterprises have individually consumed more than 1 trillion tokens
Token demand growth suggests the risk of underbuilding may be greater than overbuilding

4. Wall Street Can No Longer Afford to Be Bearish (“F You, I’m Buying”)

Repeated AI drawdowns (DeepSeek, Goldman’s “too much spend” note) were consistently bought by retail investors while institutional players missed out
AI now represents 60% of all VC dollars deployed — greater concentration than internet companies at peak dot-com (40% in 1999)
Wells Fargo chief equity strategist: “Outside of AI, I’m not really excited about anything”
Every major venture round for frontier AI companies remains upsized and oversubscribed; valuations show no sign of plateauing

5. GPU Depreciation Is Longer Than Bears Assume

A key bear argument is that GPUs depreciate far faster (2–3 years) than companies are booking (5–6 years)
The Information’s own Oracle data revealed that NVIDIA’s Ampere chips (released 2020) are still generating fat profit margins for Oracle — providing real-world evidence of a longer useful life
The host characterizes this as “probably the most bullish data point on actual versus theoretical useful life of NVIDIA silicon I’ve seen anywhere”
The difference between 2-year and 5-year GPU depreciation is worth hundreds of billions of dollars to the industry

6. (Bonus) The Watched Pot Doesn’t Boil

Bubbles historically do not burst when everyone is watching for them; bubble concern peaked in August 2025 and remains high — a historically atypical precondition for a burst
Bank of America chief strategist Michael Hartnett: “Every bubble in history has been popped by central bank tightening” — no tightening is imminent
Upcoming catalysts not yet priced in: OpenAI IPO, gigawatt-scale data centers, energy infrastructure, and embodied AI / industrial robotics (Figure’s O3 robot debuting imminently)
Jeff Bezos: industrial bubbles, unlike financial/banking bubbles, can be net positive for society even when they burst, because the underlying inventions persist

The Host’s Own Analysis

The “Big Short Generation”: Social media amplifies the historical cachet of being a contrarian bubble-caller; this distorts how frequently bubble warnings are issued
The Rearview Fallacy: Human cognition anchors future possibility to historical experience, making it genuinely difficult to accept that transformative technology could produce numbers as large as those now being reported
Systematic underestimation of revenue and growth: Most bubble analysis does not account for (a) cloud hyperscaler AI revenue, (b) “shadow AI revenue” (e.g., Meta’s improved ad performance from AI), or (c) the coming wave of enterprise AI spend — KPMG’s 2025 CEO Outlook found 83% of CEOs at companies with $500M+ revenue plan to spend 10–40% of their budget on AI in the next 12 months
Embodied AI is not yet priced in: Industrial robotics inference demand is not factored into any current supply/demand models
Circular deals are about “cheating time”: The motivation is not to artificially inflate demand but to accelerate the timeline to a future the companies are confident in — using future resources to act today

What to Watch: Leading Indicators of Trouble

Indicator	What to Look For
Enterprise AI adoption	Active withdrawal of spend; CEOs rewarded for dismissing AI
Funding market sentiment	OpenAI failing to raise; IPO stock dropping hard
Energy/infrastructure	AI blamed for major blackouts; data center growth outpacing energy supply
Revenue growth	Tapering of enterprise AI spend; slower-than-forecast adoption
Credit markets	Most important historically: Oracle credit default swap spreads widening; difficulty refinancing debt; a credit crunch spreading from one default

Morgan Stanley Wealth Management CIO Lisa Shalit: “Every morning, the opening screen on my Bloomberg is what’s going on with credit default swap spreads on Oracle debt… probably not [popping] in the next nine months, but possibly over the next 24.”

Key Concepts

Circular investment / vendor financing: A pattern in which Company A invests in Company B, which then uses those funds to purchase Company A’s products, artificially inflating reported revenues and demand across the ecosystem
Shiller CAPE ratio: The Cyclically Adjusted Price-to-Earnings ratio, a standard valuation metric that smooths earnings over 10 years to reduce cyclical distortion; a proxy for whether the overall stock market is overvalued
Gross margin: Revenue minus cost of goods sold, divided by revenue; used here to evaluate the profitability of renting GPU compute (Oracle’s AI cloud: ~14%)
GPU depreciation schedule: The accounting timeline over which data center GPUs are written down in value; industry standard is 5–6 years; bears argue real useful life is 2–3 years
Ampere (NVIDIA chip generation): NVIDIA GPU architecture released in 2020; the fact that Oracle is still generating strong margins from these chips is presented as evidence that GPU useful life is longer than bears assume
Token demand: The volume of AI inference requests processed; used as a proxy for real end-user consumption of AI services (OpenAI: ~3 quadrillion tokens/year as of late 2025)
Credit default swap (CDS) spread: The cost of insuring against a company’s debt default; widening spreads signal rising market concern about creditworthiness; flagged here as an early-warning indicator
ARR (Annual Recurring Revenue): A standard SaaS metric representing annualized subscription revenue; used to compare startup growth velocity across eras
Narrative fracture: The point at which real-world results diverge sufficiently from market expectations to undermine the story supporting asset valuations — the mechanism by which bubble psychology unwinds
Industrial bubble (vs. financial bubble): Jeff Bezos’s framing: an industrial bubble, where capital builds real infrastructure and technology that persists after a bust (railroads, fiber optic cable, potentially AI chips), is less damaging than a purely financial bubble (e.g., 2008 mortgage crisis)
The “Big Short Generation”: The host’s term for the social incentive structure — amplified by social media — that makes calling a bubble culturally prestigious, potentially inflating the frequency and confidence of bubble predictions
Rearview fallacy: The host’s term for the cognitive bias of constraining future possibility to historical precedent, which he argues causes systematic underestimation of genuinely transformative technological and economic change

Summary

The episode argues that while there are legitimate, non-trivial reasons to scrutinize the current AI investment environment — particularly circular investment structures, potential infrastructure overbuilding, stretched equity valuations, and historical parallels to dot-com — the weight of evidence does not yet support a bubble diagnosis. Revenues at the frontier companies are real, growing faster than any prior startup era, and backed by direct consumer and enterprise purchases rather than advertising or financial engineering. Leverage across the major players is low, demand for AI inference is accelerating rather than plateauing, and historical precedent suggests bubbles are killed by credit crises, not by skepticism — and no credit crunch is currently visible. The host identifies additional structural factors the bear case underestimates: “shadow” AI revenue embedded in hyperscaler and enterprise financials, the pending wave of CEO-directed AI budget increases, and the entirely unmodeled inference demand from embodied AI and industrial robotics. The most important indicators to monitor are not stock prices or media narratives, but Oracle’s credit default swap spreads, enterprise AI spending commitment rates, and any signs that the major hyperscalers are encountering difficulty servicing or refinancing debt.