The AI Bubble Will Deflate, Not Pop

I lived through the dot-com boom and the bust. I was at Microsoft during both - riding the wave up and watching $5 trillion evaporate on the way down. This AI moment feels familiar, but the ending will be different.

I understand why teams adopt this approach—it solves real problems.

From its peak in March 2000 to its trough in October 2002, the NASDAQ lost 78% of its value, wiping out more than $5 trillion in market value. Pets.com, Webvan, and hundreds of other companies disappeared. Colleagues with stock options that looked life-changing suddenly held worthless paper. "Internet" became a dirty word for investors. The conventional wisdom was that the whole thing had been a mirage.

It wasn't. The infrastructure got built. The survivors became giants. But it took a decade for the hype to match reality again. And for those of us who were there, it was a formative lesson in how markets disconnect from reality - and eventually reconnect.

I think AI is in a similar spot - but with one crucial difference. The bubble will deflate, not pop. Here's why.

The Dot-Com Pattern

Here's what I watched happen in 2000 - and I was in the middle of it, running a software company in Redmond:

Overinvestment in infrastructure. Companies laid fiber optic cable everywhere. The "dark fiber" statistic became famous: by some estimates, 85-95% of installed fiber was unused after the bust. Billions of dollars sitting in the ground, unlit.

Business models that didn't work. Pets.com sold dog food online with negative margins and spent millions on Super Bowl ads. The unit economics were impossible. Growth for growth's sake. I watched clients at Core Logic make the same mistake - chasing valuation instead of revenue, assuming the next round would always come.

Valuations detached from reality. Companies with no revenue had billion-dollar market caps. "Eyeballs" and "mindshare" replaced profit as metrics. Nobody asked "but how does this make money?"

The correction was brutal. Not a correction - a collapse. Companies that were worth billions became worthless. Engineers who had been wooed with stock options found themselves unemployed and holding paper.

The AI Similarities

Sound familiar? Look at AI in 2024-2026:

Massive infrastructure investment. NVIDIA's market cap exploded. Hyperscalers are spending tens of billions on AI infrastructure. Everyone is building data centers full of GPUs.

Questionable business models. Many AI startups are wrappers around foundation models with no defensible moat. They're spending more on compute than they're making in revenue. The unit economics are underwater. Meanwhile, most AI pilots never make it to production.

Valuations detached from reality. Companies with thin technology layers are valued at billions. "AI" in the name adds multiples to valuations. Nobody asks "but how is this different from calling the API directly?" A data-driven comparison between the AI and dot-com bubbles shows that today's AI valuations, while high, are backed by stronger fundamentals than the dot-com era.

The hype is unsustainable. Every company claims to be an "AI company." Every product adds "AI features." The term has become meaningless through overuse.

Why Deflation, Not Pop

Here's where the analogy breaks down - in AI's favor.

The dot-com crash happened because the underlying technology wasn't ready. Broadband wasn't ubiquitous. Mobile didn't exist. The infrastructure was ahead of the use cases.

AI is different. The technology works. GPT-4 is genuinely useful. Image generation creates real value. Code assistants improve developer productivity. These aren't vaporware demos - they're production systems used by millions.

The bubble isn't in whether AI works. It's in how much value gets captured by whom, and how quickly.

That means the correction will be a repricing, not a collapse. The companies building real value will survive and grow. The companies riding hype will fail. But "AI" won't become a dirty word like "internet" did in 2001.

Who Survives

Based on the dot-com pattern - and what I learned running Core Logic Software through that crash - here's who makes it through:

Infrastructure providers. Amazon survived the crash and built AWS. NVIDIA, the hyperscalers, and infrastructure companies will survive this. Picks and shovels always win in a gold rush.

Companies with real moats. OpenAI has billions in training compute that competitors can't match. Anthropic has safety research depth. Foundation model companies with genuine differentiation will endure.

Companies solving real problems. If your AI product saves customers money, makes them money, or does something impossible without AI - you're probably fine. The value is real.

Companies with sustainable unit economics. If you can serve customers profitably at current scale, you can weather the downturn. If you're burning cash hoping to find a model later, you won't make it.

Who Doesn't

Wrapper companies. If your entire product is a nice UI on top of GPT-4, you have no moat. OpenAI can add your feature tomorrow. A teenager with an API key can clone you this weekend.

Companies dependent on cheap capital. If you need continuous fundraising to survive, the music stops when investor sentiment shifts. You need enough runway to reach profitability without another round.

Companies without differentiation. "We do X, but with AI" isn't a company - it's a feature. If your AI advantage can be replicated in a few API calls, it's not an advantage.

Companies optimizing for hype. If your strategy is press releases, conference talks, and partnership announcements rather than revenue and retention - you're building a story, not a business.

The Timeline

My guess - and it's only a guess - is that we're 12-24 months from the correction. As BlackRock's analysis notes, today's tech giants are self-financing their AI investments through retained earnings rather than debt, making them more resilient than dot-com era companies. Here's what I expect:

2025-2026: More AI startups fail quietly. "AI" fatigue sets in among buyers who got burned by overpromised products. Investor appetite decreases.

2026-2027: The repricing happens. Valuations come down 50-70% for most AI companies. Several high-profile failures make headlines. "AI winter" articles appear.

2027-2030: The survivors consolidate. Like Amazon emerging from the dot-com crash, the companies with real technology and sustainable businesses grow into the space.

2030+: AI becomes infrastructure, like cloud computing did. Nobody gets excited about "AI" anymore because it's everywhere. The hype is over, but the value is real.

How to Position

If you're building an AI company:

Get to profitability. Or at least get unit economics positive. You need to survive without external capital for 2-3 years.

Build defensible technology. Fine-tuned models on proprietary data. Specialized capabilities that can't be replicated with a prompt. Something that takes time and money to recreate.

Solve real problems. Not "AI for X" but "we save customers Y dollars" or "we enable Z that was impossible before." Concrete, measurable value.

Don't depend on the hype. If your sales strategy is "AI is hot right now," you'll have nothing when it's not. Build for the world where AI is table stakes.

If you're buying AI products:

Demand proof. Not demos, not pilots - production results. Does this actually work at scale? Does it actually deliver ROI? In my experience advising startups through Barbarians, the companies that ask these hard questions early are the ones that survive market corrections.

Consider vendor risk. Will this company exist in two years? Do they have sustainable economics? What happens to your integration if they fail?

Build capabilities internally. The companies that win long-term will have AI competency in-house. Vendors come and go; capabilities endure.

Deflation Resilience Audit

Score your AI company (or vendor) against the dot-com survival criteria:

"The bubble isn't in whether AI works. It's in how much value gets captured by whom, and how quickly."