Historical Parallels
What the Dot-Com Bust Actually Teaches About AI
Everyone says AI is the next dot-com bubble. They're asking the wrong question. The dot-com era didn't punish the internet. It punished companies without revenue. That distinction matters now.
The Wrong Lesson
Most People Misread the Dot-Com Crash
The popular version goes like this: tech stocks got crazy, the bubble popped, everyone lost money. Be careful with AI stocks because the same thing will happen again.
That version is wrong in a useful way. The NASDAQ fell 78% from its March 2000 peak. But it didn't fall evenly. Companies with real revenue and defensible business models dropped and recovered. Companies burning cash on Super Bowl ads and "eyeball" metrics went to zero and stayed there.
The crash didn't prove that the internet was overhyped. The internet went on to become the most important economic force of the next two decades. The crash proved that investors were bad at distinguishing real companies from stories.
78%
NASDAQ peak-to-trough decline, March 2000 to October 2002
The index took 15 years to regain its 2000 peak. But Amazon, which fell 94%, eventually returned 200x from its low. The index number hides the divergence between survivors and casualties.
Survivors vs. Casualties
What Separated Amazon from Pets.com
Look at the companies that survived the crash versus those that didn't. The pattern is clear. Survivors had revenue. Casualties had stories.
Dot-Com Survivors vs. Casualties
Revenue at the time of the crash (annualized, in millions) and outcome
| Company |
2000 Revenue |
Peak Decline |
Outcome |
| Amazon |
$2.8B |
-94% |
Survived. Now $600B+ revenue. |
| eBay |
$431M |
-80% |
Survived. Profitable by 2001. |
| Priceline |
$1.2B |
-99% |
Survived. Now Booking Holdings. |
| Pets.com |
$110M |
-100% |
Bankrupt. Spent $12 on marketing per $1 of revenue. |
| Webvan |
$178M |
-100% |
Bankrupt. Built $1B warehouse network before proving demand. |
| Kozmo.com |
~$30M |
-100% |
Bankrupt. Free delivery on every order. Negative unit economics. |
Amazon had $2.8 billion in annual revenue in 2000. It was losing money, but customers were buying things. The revenue was real. When the crash came, Amazon cut costs, reached profitability by Q4 2001, and kept growing. The stock dropped 94% and then returned 200x.
Pets.com had revenue too, but it spent $12 acquiring each dollar of revenue. The unit economics never worked. Revenue alone isn't the filter. Revenue that can become profitable is.
The question was never "is the internet real?" The question was "does this specific company generate revenue that could become profitable?" The same question applies to AI companies today.
The Pattern
Three Traits That Predicted Survival
Across dozens of dot-com companies, survivors shared three traits. Two out of three wasn't enough.
01
Real Revenue from Real Customers
Paying customers, not registered users. Amazon had millions of people actually buying things. eBay had transaction volume generating commission fees. Pets.com had revenue, but each sale lost money after shipping. The test is revenue that comes from a transaction a customer chose to repeat, not a one-time curiosity purchase subsidized by free shipping.
02
A Path to Positive Unit EconomicsWhether each transaction can eventually make money. Losing money per sale means scaling just loses more.
Could the company make money on each transaction if it scaled? Amazon could. Its negative margins came from infrastructure investment, not from the core business model being broken. Webvan could not. Grocery delivery at the cost structure Webvan built required order density that didn't exist. The infrastructure expense was structural, not temporary.
03
A Defensible Position Against Incumbents
The internet gave new companies distribution, but incumbents could build websites too. Amazon succeeded partly because traditional retailers were slow to go online. eBay had a network effect that couldn't be replicated. Kozmo.com had no defense. Any pizza chain could deliver faster and cheaper. A defensible position means something about your business is hard to copy even after the incumbents wake up.
Revenue, unit economics, defensibility. All three, or the crash eventually finds you. Two out of three means you survive the bust but get acquired at a discount.
INVESTING IN THE AGE OF AI
Data-driven AI investment analysis, delivered weekly. No hype, no tips—just the numbers behind the thesis.
You're in. The data starts next week.
The AI Scorecard
Applying the Framework to 2026
Use the three survival traits to score today's major AI companies. This is not a buy or sell recommendation. It's a framework for asking better questions.
AI Company Survival Scorecard
Scored on the three traits that predicted dot-com survival (Strong / Moderate / Weak)
| Company |
Revenue |
Unit Economics |
Defensibility |
| NVIDIA |
Strong. $130B+ annual revenue. Data center growing 90%+ YoY. |
Strong. 75% gross margins. |
Strong. CUDA ecosystem lock-in. |
| Microsoft |
Strong. $245B+ revenue. Azure AI growing rapidly. |
Strong. Cloud margins expanding. |
Strong. Enterprise distribution. |
| OpenAI |
Strong. $5B+ annual revenue run rate. |
Weak. Heavy compute costs. Net losses in billions. |
Moderate. Brand lead but commoditizing fast. |
| Anthropic |
Moderate. ~$1B+ run rate and growing quickly. |
Weak. Same compute cost problem as OpenAI. |
Moderate. Safety positioning. Enterprise traction. |
| AI Startups (avg) |
Weak. Most under $100M revenue. |
Weak. Burning cash on model training. |
Weak. API wrappers with no moat. |
NVIDIA and Microsoft look like Amazon circa 2000. Real revenue, improving economics, strong competitive positions. They'll get hit in any correction, but the business survives.
OpenAI looks like a more interesting case. Revenue is growing fast, but the unit economics are brutal. Compute costs eat most of the revenue. If they can bring inference costs down and hold market share, the economics fix themselves. If a competitor matches their capability at lower cost, the current revenue doesn't protect them.
The average AI startup wrapper looks like Kozmo.com. A thin layer on top of someone else's infrastructure, with no structural advantage and no path to margins.
The Timeline Problem
Right Thesis, Wrong Timing
The most painful lesson from the dot-com era: being right about the technology didn't save you from the crash. Amazon was the future of commerce. It still fell 94%.
94%
Amazon's peak decline, 1999 to 2001
If you bought Amazon at $107 in December 1999, you watched it fall to $6. You were right about the company. You were wrong about the timing. It took until 2009 to break even.
10 yrs
Time for Amazon to regain its 1999 peak price
A decade of being right and still underwater. Most investors sold during that decade. The ones who held made generational wealth. Conviction and cash reserves determined the outcome.
This is the risk with AI stocks today. NVIDIA might be Amazon. The thesis might be right. But if the market corrects 40-50%, can you hold through it? The dot-com lesson says the technology wins but the timing kills portfolios.
If You're Early
Position sizing matters more than stock selection. If you put 5% of your portfolio in NVIDIA and it drops 50%, you lose 2.5% of your total wealth. Painful but survivable. If you put 40% in and it drops 50%, you might sell at the bottom out of fear. The dot-com survivors who benefited were the ones who could afford to hold.
If You're Late
Buying NVIDIA after a 900% run is different from buying Amazon in 1997. The early money already made the easy gains. What remains is the question of whether growth continues faster than the market expects. Check the forward P/EPrice-to-earnings ratio based on next year's estimated profits. A high number means the market expects big growth ahead.. If the market already prices in 35% growth, you need to believe in more than 35% to outperform.
Your Checklist
Before You Buy Any AI Stock
Run every AI investment through the dot-com filter. Four questions, five minutes. Use the AI Exposure Calculator to assess your current portfolio exposure.
Q1
Does it have real revenue?
Not "projected revenue" or "total addressable market." Actual revenue from paying customers, growing quarter over quarter. If the revenue is under $500M annually, it's still in the story phase.
Q2
Can the unit economics work?
Does each transaction have a path to profit at scale? For AI companies, this means: can inference costs decline fast enough that the margin per query turns positive? If the model is "grow now, figure out margins later," that's a Webvan playbook.
Q3
What's the moat?
CUDA ecosystem lock-in. Enterprise distribution. Proprietary data. Network effects. If the answer is "better AI model," that's not a moat. Models improve on 6-month cycles. What survives the next model release?
Q4
Can you hold through a 50% drawdown?
This isn't about conviction. It's about position sizing. If a 50% drop in this stock would cause you to sell everything at the bottom, you own too much. The dot-com survivors rewarded patience, but patience requires being able to eat while you wait.
4 Questions
Revenue, unit economics, defensibility, and position sizing. The dot-com bust didn't punish the technology. It punished the businesses that couldn't answer these questions. The same filter works today.
How I Built This
Historical data from SEC filings, company annual reports, and NASDAQ historical price data. Here are the key assumptions and limitations.
Dot-Com Revenue Figures
Annual revenue at time of crash (fiscal year 2000)
Revenue figures are rounded from public filings and contemporary reporting. Pets.com's $110M includes both product sales and advertising revenue. Webvan's figure reflects its peak revenue year before shutdown. Kozmo.com's revenue is estimated from press reports as the company was private when it closed.
Survival Framework
Three-trait model derived from post-hoc analysis
This framework is a simplification. Survival also depended on access to capital during the downturn, quality of management decisions, and luck. Amazon nearly ran out of cash in 2001 and would have died without a timely convertible bond offering. The framework identifies patterns, not laws.
AI Company Assessments
Qualitative, based on public financials and reporting
NVIDIA and Microsoft revenues from latest earnings. OpenAI and Anthropic revenue estimates from press reporting, not audited financials. The "average AI startup" row is a generalization that does not apply to every startup. Some AI startups have strong moats and unit economics. The row reflects the median, not the best case.
Historical Analogy Limitations
The AI market in 2026 differs from the internet in 2000
The dot-com era featured many unprofitable companies with no revenue path. Today's AI leaders already generate massive profits. The comparison is about framework, not prediction. The next AI correction may be shallower, deeper, or different in character. History rhymes but doesn't repeat.
Jesse Walker has been an individual investor for 30 years. Before that, he was a poker professional, which is where he learned that the best decision and the best outcome aren't always the same thing. He writes about financially navigating the uncertainties of AI.