A video comparing today’s AI boom to the dot com crash climbed to 167 points on Hacker News, and the comment thread underneath it turned into the argument the whole industry keeps circling back to. Is this 1997 or 1999? According to Hacker News, the discussion split along familiar lines, and that split tells you more than the video does.
What stands out here is that the bubble question has moved from contrarian blog posts into the mainstream of engineering discourse. When the people building the systems start comparing notes with Pets.com, something has shifted in the mood.
The case that history is repeating
The bear argument rests on pattern matching, and the patterns are real:
- Capex is running ahead of revenue. Hyperscalers are spending tens of billions on data centers against AI revenue that’s growing fast but from a small base.
- Circular deals are everywhere. Chip makers investing in model labs that then buy chips. Cloud providers taking equity in startups that spend it on cloud credits. That flywheel looked great in 1999 too, when telecom vendors financed their own customers.
- Valuations price in perfection. A lot of current multiples only work if adoption curves stay vertical for years.
- The infrastructure gets built anyway. Fiber got laid in 1999 and sat dark for a decade. GPUs depreciate faster than fiber did, which makes the overbuild math worse, not better.
The case that it’s different
The other side of the Hacker News thread makes a point worth taking seriously: the dot com era burned money on companies with no revenue. Today’s AI leaders sit inside businesses that print cash.
Microsoft, Google, Amazon and Meta are funding this buildout from operating profit, not from junk debt and IPO proceeds. That’s a structurally different risk profile. If AI revenue disappoints, they cut capex and the stock drops. They don’t go bankrupt.
The usage is also real. Developers use these tools daily. Support teams route tickets through them. That wasn’t true of most 1999 web companies, which had traffic but no behavior change.
What both sides actually agree on
Here’s the interesting part. Almost nobody in the thread argues the technology is fake. The disagreement is about timing and distribution of returns, not about whether AI works.
That’s the same shape as the dot com aftermath. The internet won. The companies that bet on it in 1999 mostly lost. Amazon dropped roughly 90 percent and then became Amazon. Both things happened.
So the useful question isn’t “bubble or not.” It’s “which layer of the stack captures the value.” Right now capital is concentrated in compute. History suggests value migrates upward over time, toward applications and distribution, once the infrastructure gets commoditized.
What this means for you
If you’re building or buying AI right now, a few practical moves:
- Don’t architect around one vendor’s pricing. Model costs have fallen hard and will keep falling. Anything you build assuming today’s prices is either a bargain later or a lock-in trap.
- Measure the actual workflow change. If your AI feature doesn’t move a metric someone in finance cares about, it’s a demo, not a product.
- Watch hyperscaler capex guidance, not model releases. Capex commentary on earnings calls is the leading indicator. A single quarter of “we’re moderating spend” language moves the whole sector.
- Assume compute gets cheaper, not scarcer. Building a moat on GPU access is building on sand.
- Keep hiring flexible. If a correction comes, it hits AI-adjacent headcount first, the way it hit web agencies in 2001.
Why now
This debate is heating up because the first real payback questions are landing. Boards funded exploratory AI budgets in 2024 and 2025. Now they want returns, and the companies that can’t show them will quietly cut.
That’s not a crash. It’s the sorting phase. The dot com era had one of those too, about eighteen months before the actual top.
The full discussion and the arguments on both sides are worth reading at the original source.