Maybe you woke up to a 3x bigger AI bill on Monday.
Or
Maybe your AI hallucinated something really important.
Whatever straw broke the camel’s back, you’re on the other side now.
This is part of my coverage of AI marketing.
I’m an AI realist. But what do you do if you’re coming down the other side of the slope?
Maybe you feel left out or discombobulated because your boss is demanding crazy things from you.
First, I’m sending you a hug. I’m feeling it.
I’ve seen this movie
The advantage of being old is you’ve lived through bubbles before. I had a front seat to the last bubble working at a search engine. My coworkers were day trading. Crazed valuations. People begging you to hire them so they could get in on this thing called the internet.
Then the bubble popped, and I had a colleague who had to tell his wife that since he was trading on margin, they would need to sell their home.
He wasn’t alone. NYSE margin debt hit a record $278.5 billion in March 2000, the same month the Nasdaq peaked. Day-trading firms drove 10 to 15% of Nasdaq volume in 1999.
The Nasdaq closed at 5,048 on March 10, 2000. By October 2002 it was at 1,114, down 78%. It didn’t set a new record until April 2015. Fifteen years.
Search engines weren’t spared. Terra bought Lycos for $12.5 billion in stock in May 2000. Four years later Daum bought it for $95 million. Less than 1% of the price.
We are in a bubble and it will pop. I want you to be prepared for what’s on the other side.
The bubble, by the numbers
The spending:
- The five hyperscalers are on track to spend about $828 billion on capex in 2026. Amazon says 2027 will be “meaningfully more.”
- Bain says AI needs $2 trillion in annual revenue by 2030 to pay for the build, and the world is still $800 billion short.
- OpenAI’s run rate is near $50 billion. The Information reports it expects to burn $25 billion in 2026, $57 billion in 2027 and $85 billion in 2028, with no positive cash year until 2030.
The circle:
- Nvidia signed a letter of intent to invest up to $100 billion in OpenAI. OpenAI buys Nvidia GPUs. AMD gave OpenAI warrants for up to 10% of AMD, tied to OpenAI buying AMD chips. OpenAI signed a $300 billion compute deal with Oracle. The Register laid out the loop.
- OpenAI is valued at $852 billion. Anthropic at $965 billion.
The people saying it out loud:
- Sam Altman, asked if investors are overexcited about AI: “My opinion is yes.”
- Jeff Bezos called it “a kind of industrial bubble.”
- MIT’s NANDA study found 95% of enterprises got zero return on $35 to 40 billion of generative AI spend.
The realist’s counterpoint
This isn’t 1999.
Goldman Sachs argues the big players generate huge free cash flow and pay for most of this from it, which very few firms could do in 1999. Nvidia trades near 23.5x forward earnings. Cisco traded above 100x at its 2000 peak.
And the use I see is real. Ethan Mollick put it well: even if there’s a financial bubble, there’s no “use bubble.” Billions of people use AI every week.
The dot-com crash proves both can be true. Amazon’s stock fell about 94% from its peak. The company survived and the internet kept growing. The money was the bubble. The technology wasn’t.
The bubble-proof checklist
1. Do you have two AI providers in case one goes out of business?
Well-funded AI companies can disappear. Builder.ai was Microsoft-backed, raised over $450 million and was valued above $1 billion. In May 2025 it ran out of money.
Big companies already hedge. In a16z’s survey of Global 2000 CIOs, 81% use three or more model families, up from 68% a year earlier.
Do this: Pick a backup provider. Run one real workflow through it this week so you know it works. (I wrote about this in April: Your AI Has a Single Point of Failure.)
2. Is your data yours and not locked in Claude or ChatGPT?
Models get pulled. OpenAI retired GPT-4o on February 13, after pulling it once in August 2025 and bringing it back after a backlash. About 800,000 people were still using it.
The vendors know switching is a fight. In March, Anthropic added a memory import so you can bring what ChatGPT or Gemini knows about you into Claude.
Do this: Keep your prompts, brand docs, skills and client notes in files you own (obsidian, a Drive, a Git repo). The chat app should read them, not hold them.
3. If you had to pay the actual cost of your currently subsidized tokens, could you?
A $200 Claude plan can deliver about $11,700 a month in API-equivalent usage, per SemiAnalysis’s October test of every major AI subscription. That’s 58 times what you pay. When Anthropic added weekly limits in 2025, it said one user “consumed tens of thousands in model usage on a $200 plan.”
That gap is the subsidy. And the rules change fast:
- Cursor switched to usage-based pricing in June 2025, users hit surprise charges, and the CEO apologized and offered refunds.
- One Replit user went from $200 a month to $1,000 in a week after a pricing change.
- On September 29, OpenAI said it would cut its $200 Pro plan’s usage in half. A month later, after backlash, it reversed course.
Do this: Price your top three AI workflows at API rates. Then multiply by three. If that number breaks your budget, find out now.
4. If AI system X stopped working, would your business stop?
It doesn’t take a bankruptcy. On November 18, 2025, a bad config file at Cloudflare knocked ChatGPT, Claude and X offline for hours. A month earlier, AWS us-east-1 went down for about 15 hours.
Do this: Write a one-page manual fallback for your most important AI workflow. Who does it, and how, if the tool is gone for a day?
Watch this
- How Much Would an AI Crash Destroy? Patrick Boyle (Aug 22, 2026). The long, dry, smart version.
- Ken Griffin on whether AI is in a bubble Bloomberg Television (Oct 21, 2026). Two and a half minutes.
- AI Bust Fallout Would Be “More Significant” Than Dot-Com, Says George Noble Bloomberg Podcasts (Jul 10, 2026).
Worth reading on X
- Sam Altman (Oct 29, 2025): OpenAI has committed to about 30 gigawatts of compute, with a total cost of about $1.4 trillion.
- Ethan Mollick (Jan 19, 2026): “there is no ‘use bubble’.”
- Ed Zitron (Sep 29, 2026): “When will the AI bubble burst? When the money runs out.”
Fact vs. fiction
The hallucination bill is already coming due:
- A tribunal held Air Canada liable for a refund policy its chatbot made up.
- Deloitte refunded part of a A$439,000 government contract after its report cited sources that don’t exist.
- A public database now tracks 2,262 court cases involving AI hallucinations. In March, a federal appeals court issued sanctions around $116,000.
If your name is on it, a human checks it.
The bubble popping will only accelerate the change we are experiencing, because it’ll separate fact from fiction.
I want you to be on the fact side.