The year is 1997. Netscape went public two years ago at $28 and closed at $75 on day one. Amazon is a bookstore that hasn’t made money yet. Cisco is selling routers to companies building the internet’s nervous system. Investors are calling it a bubble. They are right that valuation has separated from reality. They are wrong about what is coming next. By 2000, the NASDAQ has risen 572% from its 1995 base. By 2002, it has crashed 78%. By 2015, it has exceeded its 2000 peak. The internet was not a bubble. It was the most important infrastructure build of the 20th century, wrapped inside a speculative cycle that punished those who bought the wrong names at the wrong time.

We are in an analogous moment. AI is real. Some AI companies are priced for perfection in a world that has never delivered perfection. Both of those statements are true simultaneously. The error most analysts make — and most HNW investors are currently making — is treating them as contradictory. They are not. They are a map.

This article does three things. First, it runs the dotcom comparison with actual data — not narrative. Second, it identifies where AI 2026 is structurally different and where it is structurally the same. Third, it gives you the framework for deciding where to position: which layer of the AI stack, at what valuation, with what confirmation signal, and what would change your thesis.

The most dangerous thing about historical comparisons is that people use them selectively. Bear analysts cite the NASDAQ crash. Bull analysts cite the internet’s ultimate triumph. Both are using the same history to reach opposite conclusions. The data deserves more precision than that.

The Netscape IPO on August 9, 1995 is the common marker for the beginning of the internet cycle. ChatGPT’s launch on November 30, 2022 is the equivalent marker for AI. From those starting points, the NASDAQ gained 115% in the 36 months following ChatGPT’s launch — nearly identical to the 110% the index gained in the same window after Netscape’s IPO. The pattern match is not metaphorical. It is numerical.

The comparison reveals something important: AI is tracking the internet pattern almost perfectly on market performance metrics. But the underlying business fundamentals are structurally different — and understanding exactly how different is what determines whether you are in the camp that gets burned or the camp that builds wealth through the cycle.

The number that bubble-callers point to is the capex-to-revenue gap. It deserves a serious look, not a dismissal.

The five largest hyperscalers — Amazon, Microsoft, Google, Meta, Oracle — will commit roughly $725 billion to AI infrastructure in 2026 alone. That is a 77% increase from 2025’s record of $410 billion. Goldman Sachs projects total cumulative AI capex over the coming years will exceed $1 trillion. The infrastructure being built is real. The demand is real. Hyperscalers are supply-constrained, not demand-constrained. But the revenue flowing back against that investment is, at this stage, a fraction of the spend.

The 95% enterprise ROI figure from MIT deserves specific attention. It is not evidence that AI doesn’t work. It is evidence that most enterprises haven’t yet figured out how to deploy it in ways that show up on a P&L. That is a deployment problem, not a technology problem — and it is exactly the problem that characterized enterprise internet adoption from 1996 to 2001. The companies that solved it first — Amazon, Google, Salesforce — became the defining businesses of the next two decades. The companies that didn’t either failed or became irrelevant.

The internet cycle’s most instructive lesson is not about valuations. It is about layers. The companies that built the physical infrastructure — fiber, routers, servers — mostly went bankrupt or never recovered. The companies that built the platforms on top of that infrastructure — Google, Amazon, eBay — became the most valuable businesses in the world. The companies that built applications on top of those platforms — Salesforce, Workday, LinkedIn — delivered the next wave. Each layer had a different risk profile, a different timeline, and a different valuation discipline required.

AI has the same three-layer structure. The error is treating them as one investment category.

The internet’s infrastructure layer — Cisco, Lucent, the fiber companies — experienced the crash hardest. Cisco fell 88% and never recovered to its 2000 peak. Not because its business failed. Because it was priced at 472× earnings in a world where the internet was real but the valuation had already borrowed against 20 years of future growth.

Nvidia at 56× trailing P/E with 50% net margins and supply-constrained demand is a different animal. It is not 1999 Cisco. But the lesson from Cisco is not irrelevant: being the essential infrastructure supplier does not protect you from valuation gravity. It only delays it. When the upgrade cycle moderates — when hyperscalers have enough GPUs, or when a competitive alternative scales — the multiple will compress. The business survives. The multiple does not.

The people who got wealthy through the internet cycle were not the ones who called the bubble correctly in 2000. They were the ones who identified the right companies, entered at disciplined valuations, and held through the noise. Jeff Bezos’s net worth was not built by avoiding tech in 2000. It was built by running a company that was solving a real problem with a scalable model, even when the stock fell 93% between 1999 and 2001.

The same framework applies to AI in 2026. The question is not bubble or no bubble. The question is: which companies are solving real problems, at what valuation, with what moat, and what would you need to see to know the thesis is broken?