What rhymes with 1999, and what doesn’t.
The 1999 comparison: what rhymes, what doesn’t, and which tactics not to repeat.
From the summary: “None is new to the AI era, and each has a pre-AI twin.” Back to the summary
The comparison to 1999 comes up in every conversation about this data, usually as a way of dismissing it. It is worth doing properly, because the parts that rhyme and the parts that don’t point in opposite directions.
What rhymes
- Capital behavior
- Money is moving faster than evidence, valuations are being set on multiples of a revenue figure whose basis varies between companies, and the infrastructure layer is absorbing enormous capital expenditure funded partly by debt. The picks-and-shovels vendor is the market’s favorite stock. All of that is familiar.
- Circular revenue
- Vendors investing in customers who then buy their product was a documented feature of the telecom build-out, and versions of it exist now between model providers, cloud vendors, and application companies.
- Narrative substituting for measurement
- A company that reports an impressive annualized figure without a basis, a margin, or a retention number is asking to be valued on a story. That happened at scale in 1999.
- Land-grab pricing
- Below-cost pricing to win share, on the theory that economics follow scale. Sometimes they do. Often the plan gets throttled instead, and the reviews follow.
What doesn’t
- Customers are paying
- The fast companies in this dataset have paying customers, in volume, often from month one. Lovable, Cursor, ElevenLabs, Gamma, and Replit all built revenue before they built a sales organization. Many prominent dot-com failures had traffic without a durable revenue model or unit economics.
- Unit costs can improve faster
- Webvan and Kozmo had to build physical fulfillment networks before they had enough demand density to support them. AI applications have real variable inference costs, but the cost per unit of computation has fallen rapidly. That creates a plausible margin path; it does not guarantee one, because usage and demand for higher-end models can rise just as quickly. Photoroom reached $20M ARR on about $2M invested, and Midjourney took no outside money.
- Distribution already exists
- In 1999 a company had to build the audience’s access to the internet along with the product. Now the buyer has a browser, a card, and a colleague who already uses the thing.
- This is not a 1999-style IPO wave
- There have been a handful of AI IPOs against hundreds in 1999. Exposure is concentrated in private markets and in the balance sheets of large public companies, which distributes the risk differently and makes the retail comparison weaker.
My take: The application-layer companies studied here differ from the stereotypical dot-com failure in one important respect: they have paying customers and a working revenue model. That does not settle the AI-bubble question. Application companies still carry retention, margin, platform-dependency, and capital-cycle risk.
For a marketing leader, the useful test is local: does this company have improving unit economics, customers who stay or expand, and a revenue number the leadership team can defend? Those questions are answerable with data the leadership team can request now.
Tactics not to replicate
Five patterns that worked in 1999 until they didn’t, each with its current form.
| Pattern | Then | Now | The tell |
|---|---|---|---|
| Buying growth ahead of retention | Acquire users at any cost, monetize later | Paid acquisition scaled before cohort retention is known | The company cannot tell you what a cohort looks like at 90 days |
| Making the number the marketing | User counts and traffic as proof of a business | Annualizing a strong month and publishing the result | Jasper, Bolt.new, 11x, Cluely. In two of those the number itself was the story |
| Hiring the organization you aren’t yet | A national sales force before a repeatable sale | A full demand-generation and field org on a Supernova plan before the loop exists | Cursor hired its first salesperson in late 2024. Bolt ran under 10 in go-to-market to $40M |
| Renting distribution | Portal deals and paid placement | Paying platforms and creators for reach without cohort data underneath | It works while the payments continue. Gamma’s creator spend was on performance terms, which is the counter-example |
| Pricing to win share, fixing it later | Below-cost delivery | “Unlimited” plans on subsidized inference | Replit improved reported gross margin from -14% to +23% in a year. Higgsfield has never disclosed one |
What actually carries over
The durable lesson from the dot-com era is not that speed is dangerous. eBay, Amazon, and Google all grew extremely fast through the same window and are still here. The companies that endured combined speed with improving economics, customers who returned, and numbers that held up. Many of the failures did not.
Those three are measurable early by the leadership team, and Marketing can insist that they be visible beside the growth story.
On growth frameworks: the SaaS-era shorthand was T2D3, triple then double, articulated by Neeraj Agrawal at Battery Ventures. Bessemer’s Shooting Star path runs faster than that on similar margins, and its Supernova path is not a multiple sequence at all.