A comparison of the late-1990s dot-com boom and today’s AI surge, exploring how financial exuberance, over investment, and slow productivity diffusion distort genuine technological revolutions.
Introduction
The late-1990s dot-com bubble and the present-day surge in artificial intelligence (AI) investment are separated by a quarter-century and vastly different technologies. Yet the two share a familiar rhythm of excitement, excess, and overreach. Each began as a genuine revolution – transformative and world-changing – but both became distorted by financial expectations that ran ahead of real economic change.
Technological revolutions rarely unfold at the pace financial markets initially imagine. The internet reshaped commerce, but not in the five years between 1995 and 2000, when investors priced it as though entire industries would digitise overnight. The same illusion of immediacy now surrounds AI. Its potential is real, but so is the lag between technological promise and economic transformation.
The Dot-Com Era: A Chronology of Euphoria
1995 marked the turning point. Netscape’s initial public offering (IPO) closed 108 per cent higher on its first trading day, setting off a frenzy. Venture capital poured into start-ups promising to ‘reinvent’ everything from shopping to news. The NASDAQ index rose from 1,000 in 1995 to over 5,000 by March 2000.
By late 1998, the Federal Reserve’s rescue of Long-Term Capital Management injected further liquidity, fuelling speculation. Between 1999 and early 2000, more than 300 internet firms went public, many without revenue. The rhetoric was utopian: “the new economy” had rendered traditional valuation metrics obsolete.
The NASDAQ peaked on 10 March 2000 at 5048.62. Between 13-15 March, the market fell by 9 per cent in three days. It fell a further 9 per cent on 14 April (roughly 34 per cent below the March peak). By mid-2000, the cracks were unmistakable: Profit warnings multiplied, debt-financed telecom expansion stalled, and by 2001 the US entered recession. Corporate scandals at Enron and WorldCom eroded trust further. By October 2002, the NASDAQ had fallen 78 per cent from its peak. About \$5 trillion in market value had evaporated. It only regained its former peak in March-April 2015, around 15 years later.
Among the casualties, Pets.com – an online pet-supply and pet-food retailer – became the era's most emblematic failure. The company went public in February 2000, raising \$82.5 million despite a business model that lost money on most transactions – shipping costs for pet food and supplies exceeded already-thin margins. Undeterred, it spent \$1.2 million on a Super Bowl advertisement featuring a sock-puppet dog, hoping brand recognition would somehow solve the firm's underlying problems. By November 2000, less than nine months after its IPO, the company shut down. The sock puppet survived as the era's most recognisable icon: a triumph of marketing over mathematics.
The dot-com legacy, however, was not purely destructive. The infrastructure built during that bubble – data cables, software expertise, consumer habits – laid the foundations for the digital economy that emerged a decade later. Innovation endured; valuations did not.
The Rise of Artificial Intelligence: Promise and Overreach
In 2022, OpenAI’s release of ChatGPT ignited the fastest diffusion of new technology in memory. Within months, major firms declared themselves “AI-first.” Venture and corporate investment surged into model developers, chipmakers, and data-centre builders. By 2025, AI spending accounted for one of the largest new categories of global capital formation.
Financial markets price not the present, but the future. A company’s valuation reflects the profits investors expect it to earn, discounted back to today. The market’s enthusiasm is extraordinary. By mid-2025, Nvidia’s market capitalisation exceeded \$3.5 trillion. OpenAI’s valuation was reported near \$300 billion on annualised revenues of roughly \$5-6 billion. Anthropic’s valuation surpassed \$60 billion against perhaps \$1 billion in sales. These ratios rival or exceed those seen at the height of the dot-com boom.
| Company | Est. 2025 Revenue (USD) | Valuation (USD) | Approx. Price-to-Sales |
|---|---|---|---|
| OpenAI | \$5–6 billion | \$300 billion | 50–60× |
| Anthropic | \$1 billion | \$60 billion | 60× |
| xAI | negligible | \$24 billion | - |
| Nvidia | \$120 billion | \$3.5–4 trillion | 25–30× |
| Note: Approximate valuations and revenue, mid-2025. | |||
At face value, such valuations imply immense future profits – profits that, if realised, must stem from vast economy-wide productivity improvements and lower labour costs. But labour markets and organisational structures do not evolve at that pace. Firms take years to integrate new technologies, train staff, and redesign workflows. History shows that productivity gains diffuse gradually, not in a single investment cycle.
AI’s productivity performance has been patchy – although notable in routine text and data tasks, it is inconsistent in precision-dependent work. Recent events in Australia have underscored the risks of relying on generative AI for professional work. In October 2025, Deloitte refunded part of a \$440,000 government contract after an internal review found that sections of its report on a welfare-compliance system contained fabricated references and non-existent citations produced with an AI tool. The firm maintained that its conclusions remained sound but conceded that quality controls had failed. Two months earlier, in August 2025, a Victorian lawyer apologised to the Supreme Court after filing murder-case submissions that included invented case law generated by ChatGPT; another lawyer faced disciplinary action for similar conduct in a civil matter first uncovered in July 2024.
AI has proven valuable for summarisation, text generation, and data sorting, but remains unreliable in coding, reasoning, or factual accuracy. Accordingly, many AI tasks still require significant human oversight. The result is only partial efficiency gains, not the automation revolution implied by market valuations.
These multiples noted above suggest that investors are anticipating decades of high growth and monopoly-like margins – outcomes that may never materialise. They also imply that AI will drive a surge in productivity comparable to the industrial revolution itself. Yet global productivity data so far show little acceleration. OECD labour productivity growth averaged just 0.6 per cent in 2023 and 0.4 per cent in 2024 – well below the long-term average of 1.4 per cent from 2001-2019. US productivity of 1.6 per cent in 2023 and 2.3 per cent in 2024, though better, still lags its long-term average of 2.1 per cent.
Capital Deepening vs. True Productivity
Economists distinguish between capital deepening – investing in more or better equipment per worker – and total factor productivity (TFP), which measures how efficiently labour and capital together produce output. The first can raise costs without necessarily improving efficiency; the second reflects genuine progress.
Much of today’s AI boom falls into the first category. Firms are spending heavily on infrastructure – chips, servers, cloud platforms – hoping it will yield future efficiency. But if AI systems remain fallible or expensive to operate, the return on this capital will be modest. Better machines do not automatically mean better output.
In macroeconomic terms, AI investment has so far produced a surge in capital formation rather than a surge in measured productivity. This mirrors the late 1990s, when telecom and fibre-optic networks were built years before demand caught up. The capital was real; the profits were deferred.
Parallels and Differences Between Two Booms
- Valuations detached from fundamentals: Both eras priced technology as if transformation were immediate and universal.
- Infrastructure overbuild: Fibre cables then; data centres and GPUs now.
- Narrative excess: The “new economy” then, the “intelligence revolution” now.
- Concentration risk: A handful of firms – Cisco, Yahoo!, and Amazon in 2000; Nvidia, Microsoft, and OpenAI today – dominate market capitalisation.
- Speculative psychology: Retail and institutional investors alike chase momentum, assuming exponential growth is self-fulfilling.
Each similarity points to a mismatch between financial imagination and economic friction. The dot-com boom collapsed when earnings lagged promises; the same dynamic could yet test the AI narrative.
There are, however, important distinctions. The internet in 1999 was barely commercial; AI in 2025 already generates real revenue and demonstrable use-cases. Corporate governance is tighter, capital markets more diversified, and regulators better informed. Moreover, AI hardware and infrastructure have residual value – servers and chips can be repurposed, unlike many dot-com ventures whose assets vanished entirely.
Yet these differences only mitigate rather than eliminate risk. The dot-com collapse was not caused by fake technology but by unrealistic timing. The internet did transform the world – it simply took longer than many investors in 1995 or 1999 expected. AI may follow the same arc: a slow burn disguised as a bonfire.
AI’s advocates correctly note that it differs from dot-com ventures: it is a general-purpose technology affecting cognitive work itself, comparable to electricity or steam power. Yet this parallel undermines rather than supports current valuations. Electricity took four decades to become the dominant industrial power source after Edison’s first station in 1882, and productivity gains did not appear in economic data until the 1920s. General-purpose technologies follow extended diffusion curves precisely because they require wholesale reorganisation of work processes, capital equipment, and workforce skills. If AI is indeed comparable to electricity, today’s market is pricing a 40-year transformation as though it will arrive in five.
Market Structure and Model Convergence
A further risk lies in the industry’s competitive structure. Dozens of firms are now developing frontier AI models, yet these systems are rapidly converging toward similar architectures, training data, and performance benchmarks. In such conditions, only a handful of platforms are likely to achieve durable scale or pricing power. The rest will face escalating costs without commensurate revenue, as inference prices fall and customers gravitate to the largest and most reliable providers. Economic history offers a familiar pattern: in every general-purpose technology – from railways to semiconductors to search engines – the initial proliferation of competitors gives way to concentration. Most of the capital invested in marginal players is never recovered. Even as AI succeeds as a technology, much of the current investment will not survive its commercial sorting.
Economic and Behavioural Lessons
The recurring pattern is that technological revolutions are over-financed relative to their short-term returns. Markets pull future value into the present, mistaking early adoption for mass transformation. However, organisational inertia, regulatory adjustment, and skill retraining slow diffusion.
The gap between capital markets and the real economy creates volatility. When the expected profits fail to appear, valuations contract sharply. Yet the long-term trajectory of technology continues upward. The aftermath of the dot-com bust paved the way for today’s digital platforms; the aftermath of an AI correction could likewise produce a more sustainable ecosystem.
For policymakers, the lesson is that bubbles are not aberrations but side-effects of innovation. For investors, it is that valuation should not measure hope, but execution.
Conclusion
The dot-com bubble taught a lesson that every generation of investors must relearn: technological revolutions are real, but financial bubbles misprice them.
AI will transform work, communication, and knowledge creation. But it will do so gradually, unevenly, and with setbacks – not in the compressed timeframe that current valuations demand. When OpenAI must justify its price by generating \$15-20 billion in annual profit, or when Nvidia's growth rate inevitably slows, the market will rediscover an old truth: even revolutionary companies must eventually trade on earnings, not on narrative.
Current valuations imply that AI companies will achieve profit margins and market dominance exceeding those of any technology in history – and do so within 5-10 years. The historical record offers little support for such compressed timelines. The infrastructure is being built too fast, the productivity gains are materialising too slowly, and the gap between the two will not close in a single investment cycle.
The correction need not match the severity of 2000-2002. Better governance, real revenue, and tangible use-cases provide ballast that most dot-coms lacked. But some repricing is inevitable, and it will be substantial. A sustainable AI sector likely values OpenAI closer to \$50-80 billion than \$300 billion, and prices chipmakers on earnings multiples closer to 15× than 30×. These are still remarkable numbers – but they are not miracle numbers.
The intelligent position is not to bet against AI's long-term success – that success is nearly certain. It is to bet against the timeline. Markets are pricing AI as though the future has already arrived and profits are merely a formality. The future has not arrived. And when reality forces a reckoning between valuation and execution, many of today's prices will look as absurd as Pets.com did in 2001.
The revolution is coming – just not as fast as current market optimism demands, and not without casualties among those who priced perfection into the present.
No comments:
Post a Comment