The AI Bubble Skeptic Case: Sourced Bear Arguments You Should Not Ignore
Whether or not you think AI is a bubble, the bear case is worth reading because it names the specific places where a downturn would hit builders. Here are its strongest arguments, what the evidence says in November 2026, and what the other side points out.
TL;DR
- The serious bear case is about financing and concentration, not about whether the technology works.
- Strongest points: revenue is measured inconsistently, compute commitments are huge and mostly noncancelable, and growth leans on a few labs.
- Counter-evidence is real too: Anthropic's S-1 shows fast-growing revenue and a profitable quarter.
- Most pilots still fail to deliver measurable returns, and generated code carries security debt.
- Plan for a downturn without betting on one: stress-test margins, vendors and runway.
Update, November 2026: The sceptic case has been tested since this was written. Anthropic’s S-1 shows real revenue and a profitable operating quarter, but also $498 billion of noncancelable compute commitments. We weigh the best-known sceptic’s hits and misses in Ed Zitron and the AI bubble case.
What the bear case actually claims
The serious version does not say AI is useless. It says the money committed to AI infrastructure assumes revenue that does not yet exist, and that the financing structure makes a slowdown painful for many parties at once. That is a claim about capital, and it can be checked against filings and reporting.
Argument 1: revenue is measured loosely
AI companies talk in annualised run-rates, which multiply a recent period by twelve. In October, OpenAI told investors about $50 billion of annualised revenue, below the $68 billion figure widely reported earlier, because the higher number included partners’ gross revenue. Investing.com’s summary of the Financial Times report notes that Anthropic includes sales through AWS and Google Cloud in its run-rate and OpenAI does not. The same word covered two different rulers, and chip stocks fell between 3% and 8% in a day on the correction.
Argument 2: commitments dwarf current revenue
Anthropic’s public S-1, as summarised by Fortune, shows eight compute contracts worth $657 billion, of which $498 billion is noncancelable, roughly $70 billion a year. Its second-quarter revenue of $11.6 billion annualises to about $46 billion, which is our arithmetic. Revenue is growing very fast, but the bill is fixed.
Argument 3: the deals are circular
Chip makers invest in labs, labs buy chips and cloud capacity, and cloud providers invest in labs. In September 2025, Nvidia announced an intention to invest up to $100 billion in OpenAI, a proposed staged investment rather than a completed transfer. Circular flows are not fraud, but they make demand look larger than end-customer demand. See AI circular financing.
Argument 4: most enterprise pilots stall
MIT NANDA’s 2025 report found that roughly 95% of organisations saw no measurable return from generative AI initiatives. The methodology has been criticised, but the direction matches what many teams report: demos are easy and workflow integration is hard. Our summary is in the GenAI divide.
Argument 5: quality debt is piling up
Veracode’s 2025 study found AI-generated code introduced security flaws in 45% of test cases. Faster generation without review stores up cost that arrives later, as covered in AI cleanup after vibe coding.
What the other side says
| Bear point | Counter-evidence |
|---|---|
| Revenue is inflated | Anthropic’s S-1 reports audited figures: $16.3B in H1 2026 and a $974M operating profit in Q2 |
| Nobody pays | Enterprise customers paying over $100,000 a year rose to 11,900 from 950 a year earlier, per Fortune |
| Progress has stalled | Costs have fallen sharply; commentators such as Kelsey Piper argue that GPT-4-level capability now costs a fraction of its original price |
| It is all hype | Real usage: coding tools, support and document work show measurable time savings in many teams |
A fair reading is that the technology and the demand are real, while the financing is aggressive and concentrated. One well-known sceptic’s record shows the risk in calling timing: see our review of Ed Zitron’s hits and misses.
What a correction would look like
Not all scenarios are catastrophic. A mild one means slower fundraising, tougher terms and price increases from subsidised vendors. A sharp one means a major lab missing a commitment, with knock-on effects for chip makers, data-centre lenders and the startups built on cheap tokens. You cannot time it, but you can decide how exposed you are.
What to do about it
- Stress-test unit economics at token prices 3 times higher.
- Diversify models, so one vendor’s change does not stop you.
- Keep 18 months of runway where you can, since capital may tighten.
- Measure retention, not demos or run-rate.
- Track the financing news from filings, not social media.
The wider framework is in durable AI value versus speculation and the survival playbook.
Signals to track
Watch audited revenue growth against compute commitments, cancellation terms in new contracts, spreads on data-centre debt, startup gross margins and the pace of new enterprise customers. Note changes quarterly and compare them with your own plan.
Your next step
Pick one downside scenario, such as token prices tripling or your main model vendor changing terms, and write down what you would change in the first two weeks. Having that plan on paper turns an abstract debate into a decision you can make quickly.
Bottom line
The bear case is strongest where it is about money: loose revenue definitions, huge fixed commitments and circular deals. It is weakest where it claims the technology does not work. Take the questions seriously, check numbers against filings, and build so that you survive both a boom and a bust.
Frequently Asked Questions
Is AI a bubble?
Nobody knows, and confident answers on either side have a poor record. The evidence shows real adoption and revenue alongside very large financing commitments, so the sensible position is to plan for both outcomes.
What is the strongest bear argument?
That the build-out is financed by long-term commitments, partly with debt and circular deals, that only pay off if revenue keeps growing quickly. Anthropic's S-1 shows $498 billion of noncancelable compute commitments against about $46 billion of annualised quarterly revenue.
What is the strongest bull argument?
That adoption and revenue are real and growing fast. Anthropic reported $16.3 billion of first-half 2026 revenue and a profitable second quarter, and OpenAI told investors about $50 billion of annualised revenue.
How should a startup prepare?
Model margins at higher token prices, avoid single-vendor dependence, keep runway for a slower fundraising market, and measure value with retention rather than demos.
