The AI Bubble Is About to Pop Like 2000

I've been building software for long enough to recognize the pattern we're in right now. The Super Bowl ads this weekend are going to be packed with AI companies spending millions per 30-second spot. Companies burning billions in losses will pay millions to tell you their product is revolutionary. This is not normal behavior. This is end-stage bubble behavior, and if you've seen it before, you know exactly what comes next.

Twenty-six years ago, almost to the day, I watched the same thing happen with dot-coms. January 2000 felt exactly like January 2026 feels right now. An internal essay called "bubble.com" had leaked through the startup community, laying out in detail why tech valuations were insane and a crash was inevitable. Things felt tenuous. People were worried. Then came the Super Bowl with so many dot-com commercials they got their own Wikipedia page. The most memorable was E-Trade's monkey dancing in a garage with the tagline, "Well, we just wasted $2 million. What are you doing with your money?"

Two months later, the NASDAQ peaked. Over the next eighteen months, it lost three-quarters of its value. It wouldn't recover those highs for fifteen years. The Super Bowl ads weren't the cause of the crash, but they were the warning sign. When companies with unsustainable business models spend absurd sums on advertising, they're not building for the future. They're desperately trying to stave off the inevitable.

The Uncomfortable Parallels

The current AI situation has the same discordant feel. We know LLMs have plateaued. Industry insiders are admitting it's time to return to research mode rather than pretending we're six months from AGI. The valuations don't make sense. The revenue forecasts are implausible. OpenAI is losing billions. Anthropic is losing billions. The entire foundation model ecosystem is subsidized by venture capital that expects returns that simple math says can't happen.

There's a perfect example of the desperation I'm talking about. Anthropic ran ads mocking OpenAI for - ironically - running ads and using paid marketing to prop up their user base. Anthropic positioned themselves as the principled alternative, the company that wouldn't resort to such desperate measures. Except that positioning itself is the desperate measure.

And yet, a social network exclusively for AI agents to communicate is going viral. Moltbook is fascinating as a technical experiment, but it's also a symptom of how disconnected the AI hype is from economic reality. We're building infrastructure for AI agents to socialize before we've figured out how to make the agents reliably useful, let alone profitable.

This weekend's Super Bowl will feature a surge of AI commercials. Companies hemorrhaging money will try to convince you they're the future. It's the same playbook as 2000, with a similar likely outcome. The crash may take months to unfold, but the direction is clear.

What Happened After The Dot-Com Crash

Here's what most people forget about the post-2000 period. The crash was terrifying while it was happening, but what came after was actually worse for developers in the short term. Not only did VC funding dry up and startups fold, but a massive wave of offshoring hit at the same time. Companies that survived the crash looked for ways to cut costs, and moving development overseas was the obvious choice.

For a couple of years, it felt like software development jobs in the US were just gone. Companies were closing positions because the money had dried up, and the positions that did exist were being moved to offshore vendors. It was a double hit that made the job market brutal. Many developers left the field entirely during this period.

But then something interesting happened. The offshore code started coming back. It turned out that, for various reasons—immature foreign firms, communication barriers, high turnover at overseas providers—the code quality wasn't up to par. US companies were getting deliverables that didn't meet requirements, had serious bugs, or couldn't be maintained. They needed developers to clean up the mess.

Some of us made substantial money on those cleanup projects. The offshore experiment had failed not because the developers overseas were incompetent, but because the model of shipping requirements across twelve time zones and expecting production-quality code back was fundamentally flawed. Complex software needs context, communication, and iteration. You can't offshore that effectively when the cost savings depend on minimal interaction.

AI Code Generation Is Just Offshoring 2.0

The parallels to AI code generation are striking. AI coding agents are another iteration of the same pattern—offloading development tasks to cheaper, less experienced resources with inconsistent quality. Claude Code, Cursor, and all the others are producing code faster and cheaper than human developers. But they're also producing bugs, security vulnerabilities, and architectural decisions that don't account for context humans would have caught.

I've been using Claude Code in production for months now. It's genuinely useful for certain tasks. Large-scale refactoring where consistency matters more than cleverness. Boilerplate generation. Migrating from one library to another. But every piece of AI-generated code requires human review, and a non-trivial percentage requires substantial fixes. The agent writes code that compiles and looks plausible but breaks subtle invariants or introduces security holes that won't surface until production.

We're going to generate a massive amount of substandard code over the next year. Companies are already pushing AI-generated features to production with minimal review because the productivity gains are too tempting. When the crash comes and budgets tighten, those same companies will need developers to audit and fix all that code. There will be work. It won't be the work we want, nobody dreams of debugging AI-generated spaghetti, but it will be necessary and lucrative.

The Tools That Actually Made Developers More Productive

Back in the late nineties, we didn't have Stack Overflow. We didn't have GitHub. We didn't have npm, pip, or any modern package manager. We barely had Google—it was still smaller than Yahoo. When you ran into a problem, you were digging through mailing list archives or hoping someone on IRC knew the answer. Building software took longer because you were constantly reinventing wheels.

There was real fear when these tools emerged that they'd make developers so productive that companies wouldn't need as many of us. If one developer with Stack Overflow and npm could do the work that previously required five developers without those tools, obviously, the demand for developers would crater.

The opposite happened. Those tools made it economically feasible to solve problems with software that previously weren't worth the development cost. The productivity gains didn't reduce the number of developers needed. They expanded the number of problems we could tackle, creating greater demand for developers. The field grew because the accessible problem space grew.

LLMs Will Probably Do The Same Thing

I'm building side projects with Claude Code that I never would have attempted on my own. Not because I couldn't write the code, but because the time investment wouldn't have been justified for a weekend experiment. When the activation energy for trying an idea drops low enough, you try more ideas. Some of them turn into real products. Some of them solve problems that were always there but never worth addressing.

This is the actual value proposition of AI coding tools, and it's not what the marketing departments are selling. They're selling "replace your developers" because that's what executives want to hear. The reality is more like "enable your developers to tackle problems that weren't previously economically feasible." That's a harder sell but a more sustainable business model.

The crash is going to clear out the companies selling fantasy and leave the ones building actual tools. When the funding dries up and the billion-dollar losses become unsustainable, we'll see which AI companies were solving real problems versus which ones were just burning money to acquire users who'll leave the moment the free tier disappears.

What Comes After

If your company's future depends on AI hype, the coming year will be tough. But skeptical developers should be fine.

There will be a correction. Valuations will come back to earth. Companies will fold. The ones that survive will be the ones with sustainable unit economics and actual product-market fit. The field will contract for a while, then stabilize, then grow again as we figure out what LLMs are actually good for.

The dot-com crash was brutal, but we came out the other side with sustainable internet companies instead of unsustainable dot-coms. The offshore bubble was painful, but we figured out which development models work and which don't. The AI bubble will pop, probably starting around the time those Super Bowl commercials air, and we'll go through a painful adjustment period. Then we'll build something real with the leftover pieces.

In the meantime, there's work to be done. Code to review. Bugs to find. Security vulnerabilities to patch. Architectural decisions to validate. The AI tools aren't going away, but the fantasy that they'll replace human judgment is. Enjoy the commercials this weekend. Some of them will be genuinely funny. And try not to panic when the crash comes. We've survived worse, and the opportunities on the other side are probably more interesting than what we're working on right now anyway.