Every Bubble Believes It's Different

The $539 billion bet that forgot about impermanence

The AI boom made the world’s 500 wealthiest people $2.2 trillion richer in 2025. Bill Gurley, one of Silicon Valley’s most respected investors, looked at those gains and told CNBC in March: “One day we’re going to have an AI reset, because waves create bubbles, because interlopers come in.” Ray Dalio warned in January that AI is “in the early stages of a bubble” and to “watch out for 2026.” Nobel laureate Joseph Stiglitz says the bubble will hurt the macroeconomy and workers will bear the cost.

These aren’t doomsayers on Twitter. These are people who manage the money.

Here’s the number that should keep you up at night: Goldman Sachs projects $539 billion in AI capital expenditure for 2026. American consumers spend $12 billion a year on AI services. That’s a 45-to-1 ratio between what’s being built and what’s being bought. Ninety-five percent of organizations investing in generative AI are reporting zero return. Elizabeth Yin of Hustle Fund estimates most AI startups will be bankrupt within 18-24 months.

The Convergence

The dot-com boom - crypto winter or the railway mania of the 1840s. Every bubble follows the same arc: real technology creates real value, capital floods in, hype overshoots reality, collective delusion peaks, then gravity wins.

The technology underneath is always real. The internet was real in 2001. Blockchain is real now. AI is genuinely transformative. That’s not the question. The question is whether you’re building on the technology or building on the hype. One has a foundation. The other is standing on air.

What makes this cycle feel different is the scale of conviction. Microsoft down 26% from its peak. Oracle down 28%. IBM down 20%. AI stocks leading the S&P 500’s 7% pullback. And yet the spending accelerates. Companies that burned through $200 million can’t pivot because the narrative is too heavy to put down. “We’re so close” is the mantra. The sunk cost doesn’t feel like a fallacy when the board deck still says “generational opportunity.”

There’s an older word for this pattern. Not from finance. From psychology that’s been pressure-tested for millennia. The word is attachment. Not wanting something, but refusing to let go of it even when the ground shifts beneath you. The Upadana Sutta uses fire as its central metaphor: clinging is fuel. The thing that keeps the fire burning past the point where it should have gone out.

Watch how it plays out. Clinging to pleasure: the dopamine hit of a funding round, the high of a viral demo. Clinging to views: the conviction that transformers will scale infinitely, that AGI is five years away. Clinging to rituals: adding “AI” to your product name, hiring more PhDs than you need, building features nobody asked for because the roadmap says so.

Then there’s the deepest trap: “this time it’s different.” Every bubble believes this about itself. Railway investors believed steam would eliminate distance. Dot-com founders believed the internet would eliminate scarcity. AI founders believe scaling will eliminate the need for understanding. It’s not lack of information. It’s active misperception. Seeing permanence where there’s only change.

Here’s what makes impermanence useful rather than depressing: it’s the most accurate forecasting model available. Not because it predicts doom, but because it predicts change. The companies that survive corrections aren’t the ones that built for permanence. They’re the ones that built for adaptation. Amazon survived 2001 because Bezos built infrastructure, not hype. The AI equivalent would be companies building genuine capability, not impressive demos that collapse under real-world conditions.

---

The Sunk Cost Trap

Venture capital has a structural problem that no one in the industry likes to name: fund lifecycles demand returns on a timeline that doesn’t respect technological reality.

A VC fund typically has a 10-year horizon. Partners raise on thesis, deploy in years 1-3, and need markups by years 4-6 to raise the next fund. When 60% of all US venture capital flows into a single sector, the incentive isn’t to find the best companies. It’s to not miss the wave. So investors double down on narratives rather than fundamentals. They’re not stupid. They’re structurally incentivized to keep the fire burning.

The sunk cost fallacy is just attachment wearing a suit. You’ve spent $50 million on an approach. Walking away feels like death. So you spend another $50 million, not because the evidence supports it, but because accepting the loss is emotionally unbearable. This is how 95% of genAI investments produce zero return and the checks keep coming.

The companies that thrive after every correction are the ones that treated the boom as a temporary resource advantage, not a permanent state. They built infrastructure while everyone else built hype. They hired for capability while everyone else hired for credibility.

The prescription is counterintuitive: assume the correction now. Don’t wait for it to teach you. The AI companies worth watching aren’t the ones raising the biggest rounds. They’re the ones building modular architectures that assume today’s approach will be obsolete. They’re treating current models as stepping stones, not destinations.

Thought Exercise: What Would You Build If It Had to Die?

Pick the project you’re most invested in right now. The one you’d defend in any meeting. The one that feels permanent.

Now give it a three-year death sentence. Not a slow decline. A hard stop. In 36 months, it’s gone. Whatever you built, whatever you learned, whatever community formed around it — dissolved.

Sit with that for a moment. Notice what resists.

Now ask yourself three questions:

1. What would you stop building? Which features exist because “we might need them someday”? Which optimizations serve the roadmap more than the user? Cut those. They’re attachment disguised as planning.

2. What would you start building? If the product dies but the knowledge transfers, what would you want to have learned? Build that. Skills, relationships, and understanding survive corrections. Code doesn’t.

3. What would you give away? If it’s all temporary anyway, what are you hoarding that could help someone else right now? Open-source it. Publish the findings. Write the post-mortem before there’s a mortem.

   
   

I believe the best technology isn't permanent. It isn't disposable either. It's useful, honest, and designed to make the next thing better. Your project will end. The question is whether it ends having contributed something, or having just consumed resources defending its own existence.

---

Glossary

Attachment — Skt: upadana / Pali: upadana. Clinging or fuel. The mental grasping that sustains suffering by refusing to accept change.

Impermanence — Skt: anitya / Pali: anicca. The universal characteristic that all conditioned phenomena are transient.

---

Signal & Noise

Michael Pollan Wants You to Rethink Consciousness — Pollan challenges materialist assumptions about consciousness just as AI companies assume it will emerge from scaling. Both are hitting the limits of “just add more.”

The Attention Economy and the Right to Attention — The Journal of Buddhist Ethics asks who owns your attention in an economy designed to capture it. When $539 billion chases your eyeballs, this stops being philosophy.

Inside Our Approach to the Model Spec — OpenAI publishes its alignment framework. The interesting question isn’t what’s in it. It’s whether ethical frameworks built during a bubble survive the correction.

Stroke Survivors’ Brains Rejuvenate to Compensate for Injury — Damaged brains reverse their aging to rewire around injury. The mind’s capacity for renewal after collapse is deeper than we imagined.