The dot-com lessons AI leaders are ignoring
A quarter century after the internet bust wiped out trillions in market value, the AI boom is repeating old mistakes at the edges.
In markets, history never quite repeats itself, but it has a habit of leaving its notes on the table.
Around the turn of the millennium, the dot-com bubble pushed the Nasdaq Composite up roughly 600 percent from 1995 to its peak in March 2000, before it collapsed about 75–78 percent by late 2002. Trillions in market value evaporated, thousands of companies disappeared, and it took the index 15 years to reclaim its high.
Today’s AI boom is nowhere near that kind of collapse. Yet the questions being asked in boardrooms hark back to the time the bubble burst: are valuations out of line with reality, are we overbuilding infrastructure, and how much of our strategy now assumes that this particular wave of technology will pay for everything?
The good news for executives is that the AI cycle is not a simple replay of 1999. The bad news is that the places where it differs are not always the places decision-makers think.
What really broke in 2000 —
The dot-com bubble is often remembered as a morality tale about hype. The reality at the time was far less clean-cut, and can arguably be better seen as a chaotic cascade of events: capital got too cheap, discipline eroded, and governance failed to keep up.
From 1995 to early 2000, the Nasdaq soared on the back of anything with “.com” in its name. Venture capital flooded into startups with little revenue and no clear path to profitability, while traditional valuation metrics were played down or dismissed entirely.
When interest rates rose, growth slowed, and a few high-profile accounting scandals surfaced, the air went out of the trade. By October 2002, the Nasdaq had fallen to about 1,140 — roughly three-quarters below its peak — and total losses from the bust were estimated at around $5 trillion in market cap.
Underneath the wreckage, however, something more subtle happened. The internet did transform business models and productivity, but on a slower and messier timeline than markets had priced. The platforms with genuine economics — Amazon, eBay, Google — rebuilt from much lower valuations, while the long tail of “eyeballs over earnings” plays simply disappeared.
That pattern is worth keeping in mind when looking at AI.
AI-related stocks have led global equity indices higher, and a small cluster of mega-cap names dominates index returns. The “Magnificent Seven” — Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia, and Tesla — returned more than 75 percent in 2023, while the broader S&P 500 gained about 24 percent. Again in 2024, they delivered an average total return of around 60 percent, while the S&P 500 returned roughly 25 percent including dividends. Over the two years from the start of 2023 to the end of 2024, the group rose more than 150 percent as a whole, compared with under 60 percent for the index and about 25 percent for the other 493 constituents.
More recently, the AI trade has broadened a little to include companies such as Broadcom, Oracle, and Palantir, but concentration remains extreme. AI leaders and their closest peers still account for over a third of S&P 500 market capitalisation, and a large share of its gains in 2025.
While this is happening, the funding cycle around AI looks increasingly speculative at the margins:
In 2024, AI companies attracted almost a third of global venture capital investment, and AI funding grew more than 80 percent year on year even as overall VC barely rose.
In the first half of 2025 alone, AI startups raised roughly $116 billion, already surpassing the full-year 2024 total, with another $45 billion added in the third quarter.
The largest rounds are concentrated in foundation model and infrastructure players — Anthropic, xAI, Mistral AI — with valuations running into tens of billions of dollars for companies that are still in the early stages of monetisation.
However, we must note that there are important differences from 1999.
Most leading AI stocks sit inside large, profitable incumbents. An iShares analysis published in November 2025 argues that while tech valuations have risen sharply, they remain well below dot-com extremes, and today’s AI spending is largely funded by profits and strong balance sheets rather than by speculative IPOs.
A separate review from Janus Henderson puts numbers on that contrast: in 2000, technology traded at more than double the valuation of the broader equity market; today, the tech sector’s multiple is closer to 1.3 times the rest of the index.
So the AI boom, at its core, is anchored in highly profitable platforms with real demand. It’s only around the edges that it exhibits many of the behaviours that made dot-com so fragile.
Lesson one: cheap capital still hides bad ideas —
In the dot-com era, easy capital made it too simple to fund weak business models. Venture money poured into loss-making companies whose plans rested on capturing users first and worrying about profit later. When capital conditions tightened, many of those stories simply could not survive.
AI is not immune to the same dynamic.
On one side, a series of official reports — notably the IMF’s April and October 2025 Global Financial Stability Reports and the October 2025 World Economic Outlook — now explicitly flag that risk asset prices are “well above fundamentals”, that valuations and concentration in the Magnificent Seven and AI-related stocks are at historical highs, and that a disappointing AI productivity payoff could trigger an abrupt repricing of tech shares and an end to the AI investment boom.
On the other, the physical infrastructure behind AI is absorbing extraordinary amounts of capital. The Bank of England has noted that global AI infrastructure — data centres, specialised chips, and power — may require “trillions of dollars” of investment over the next five years, with a significant share financed by debt rather than by retained earnings.
Corporate and investor spending on AI reached an estimated $252 billion in 2024, up 44.5 percent from the previous year, with a substantial portion of that funded through borrowing. Analysts point to Oracle as a case study, with some estimates suggesting it is taking on roughly $25 billion in new debt annually to fund AI contracts, contributing to a total debt load in the tens of billions.
The pattern should feel familiar. Capital is flowing freely into a hot narrative, and more risk is migrating onto balance sheets on the assumption that future economics will justify it.
Boardroom question —
If your organisation applied pre-AI hurdle rates and a normalised cost of capital, which AI projects and partnerships would still make the cut?
Lesson two: concentration risk is not diversification —
Before the dot-com bust, the biggest names in technology came to dominate major indices. When they fell, they dragged millions of supposedly diversified portfolios down with them. Some estimates put the losses from the broader 2000–2002 downturn at over $5 trillion in equity market value.
The AI era has its own version of that story in that, as outlined, the “Magnificent Seven” and their AI-adjacent peers account for an unusually high share of S&P 500 capitalisation and returns. Recent analysis from Apollo shows that the top ten companies in the index — dominated by AI and cloud leaders — now account for well over half of total gains in market cap since 2021, with average trailing price/earnings ratios around 50.
A Capital Group review of market breadth highlights how rare it has become for other S&P 500 stocks to outperform this cohort; in several recent six-month windows, fewer than one in twenty non-“Magnificent Seven” names beat the group’s weighted return.
Goldman Sachs, in a report titled “AI: In a Bubble?,” acknowledges the growth potential but warns that the “circularity” of AI spending — AI leaders selling to each other and to companies whose own valuations are propped up by AI optimism — leaves the market more vulnerable to an unwind if expectations slip.
The core names may be better businesses than their dot-com predecessors, but the portfolio math has not changed. For many companies and pension schemes, “owning the market” has become “owning a leveraged view on a handful of AI leaders”.
Boardroom question —
Do your treasury, pension, and employee savings plans have an explicit view on AI concentration risk, or is it being taken by default via market-cap-weighted benchmarks?
Lesson three: hype still compresses due diligence —
In the late 1990s, the rush to participate in the internet boom shortened investment processes and hollowed out the checks that usually slow capital down. Investment banks hustled half-ready companies onto the market, while boards signed off “dot-com” projects largely because their closest rivals had just done the same. The post-crash inquiries all tell a similar story: under the pressure to look modern, governance gave way, and the habit of asking hard questions was quietly suspended.
The AI cycle has its own version of due-diligence compression.
On the private side, venture and growth funds continue to write very large cheques into foundation models and AI infrastructure, sometimes at valuations that assume a winner-takes-most outcome. CB Insights’ data shows AI funding on track to double its 2024 record of about $108 billion, despite a broader environment of more cautious capital.
On the corporate side, AI has become a default line item in strategic plans. In many sectors, internal capital committees are seeing a wave of proposals that lead with “AI” rather than with a problem statement, and that rely on relatively optimistic assumptions about adoption, pricing power, or productivity.
The risk here is not that every AI project is flawed. It is that the label itself becomes a shortcut that allows otherwise standard governance processes to be relaxed.
Boardroom question —
Would this project still look compelling if you removed the word “AI” from the deck and evaluated it strictly on the numbers?
Lesson four: platforms endure, passengers don’t —
The dot-com crash did not mark the end of the internet, obviously. It did, however, mark the end of a long tail of weak, undifferentiated businesses that had been funded on the assumption that “being online” was, in itself, a moat.
Platform-type companies with network effects and genuine business models survived and ultimately dominated — Amazon, eBay, Google among them. Many narrower plays disappeared or were absorbed at fractions of their peak valuations.
AI’s structure looks similar. At the top sit a small number of hyperscale clouds and model providers, with access to capital, proprietary data, and distribution. Around them, a crowded ecosystem of startups and corporate ventures builds point solutions, thin application layers, and “AI-enabled” features.
Recent comparative work on the AI boom and the dot-com era suggests this is where the resemblance is strongest. There is a rapid proliferation of copycat offerings, fierce competition for a limited number of defensible niches, and significant capital flowing into companies that are, effectively, features on someone else’s platform. In such an environment, platform economics and bargaining power matter more than the novelty of any individual AI capability.
Boardroom question —
In your AI roadmap, how much of your value creation depends on owning the platform, and how much depends on being a passenger on someone else’s?
Lesson five: productivity arrives on its own schedule
One of the more uncomfortable lessons from the dot-com boom is that investors were right about the internet’s long-term impact and still, collectively, lost a great deal of money.
Academic and policy work on the period shows that the biggest productivity gains from digitalisation showed up years after the bubble burst, once firms had redesigned processes, integrated systems, and re-trained staff. Markets, however, had priced in that future far too quickly. AI carries the same risk of temporal mismatch.
Central banks and international institutions have broadly acknowledged AI’s potential to raise productivity, but they underline the uncertainty around timing and distribution. The IMF’s financial stability work stresses that AI’s impact on earnings and valuation could be highly uneven, with scenarios where adoption is slower or less profitable than markets currently assume.
The Bank of England’s recent focus papers echo the point that AI could boost efficiency, but it also introduces new operational and market risks, and its economic benefits may take longer to materialise than the current wave of capex and hiring implies. In other words, AI may well transform company economics. It just may not do so on the timescale implied by today’s share prices or internal investment cases.
Boardroom question —
How exposed is your strategy to an outcome where AI’s benefits arrive, but two to three years later than your plan assumes?
From history lesson to governance checklist —
The comparison between dot-com and AI works best when it is precise:
Valuations: Today’s AI leaders are expensive, but they are not trading at the 60-plus forward earnings multiples that characterised many 1999 darlings, and their profits are real.
Funding: AI is attracting an outsized share of global VC and corporate capex, similar to dot-com, but from a much stronger base of incumbent profitability.
Risk: The long tail of speculative AI and the debt-financed infrastructure build-out look more fragile than the cash-generating core.
For leaders, the practical challenge is to enjoy the upside of this cycle without sleepwalking into the same traps that defined the last one.
The internet era did not punish the companies that experimented. It punished the ones that mistook a powerful technology for a guarantee of economic timing. The AI cycle will almost certainly reward boldness again, but the rewards are unlikely to go to those who ignore the last time markets thought they had seen the future.