For those of us who experienced the “AI Winter” of the late 1980s, or the collapse of the dot com bubble in the early 2000s, current developments in the AI market are concerning. The possibility of another impending AI Winter has arisen as investment levels soar, frontier model vendors engage in an escalating technological arms race, and many organisations struggle to translate AI enthusiasm into measurable business returns.

Several major technology companies are committing significant investment in AI infrastructure, businesses are rushing to deploy AI in an environment lacking substantial standardisation and legislative protection, while venture capital continues to flow into startups seeking to capitalise on AI's promise. Yet beneath the optimism lies a familiar pattern that should concern businesses, investors, executives, and policymakers alike.

In the past, transformative and disruptive technological developments have often demonstrated cycles of exuberance, disappointment, correction, and eventual maturation. Artificial intelligence has already undergone multiple such cycles in the past.

The so called “Second AI Winter” of the late 1980s and early 1990s followed the collapse of the expert systems boom, while the dot-com crash of the early 2000s demonstrated how more general technological revolutions can result in speculative bubbles and subsequent collapses.

There is no doubt that AI is a revolutionary and transformative technology, which can deliver significant benefit to an economy. However, the question is whether current expectations, valuations, and investment levels reflect AI’s true potential.

The Expert Systems Boom

Following the first AI Winter of the 1970s, enthusiasm returned as researchers shifted away from ambitions of creating general intelligence and instead focused on specialised systems designed to mimic the decision-making processes of human experts. Systems such as MYCIN for medical diagnosis and XCON for configuring computer hardware demonstrated that AI could deliver useful commercial outcomes within narrowly defined domains. Governments and corporations responded with substantial investment, particularly after Japan's Fifth Generation Computer Systems project signalled national commitment to AI.

For a brief period, expert systems appeared poised to revolutionise business. Companies invested heavily in expert systems, knowledge engineering teams, and large-scale deployments. AI became a corporate priority for many organisations.

However, limitations soon became apparent. Expert systems with thousands, or even tens of thousands of rules, were expensive to build, maintain and validate. Capturing and encoding expert knowledge proved far more difficult than anticipated, and even the application of machine learning to the knowledge engineering task did not effectively address this issue. Expert systems performed poorly outside narrow domains, often providing sub-standard results when presented with input that was not anticipated during programming (a phenomenon called “brittleness”) and it was difficult to adapt them to changing environments. Confidence in the technology fell and investment dried up. The expert systems market sector collapsed, and many expert systems projects were abandoned. This contributed directly to the Second AI Winter from the late 1980s to the early 1990s

Now compare this to today's AI landscape.

Modern foundation models are undeniably more capable than expert systems. Nevertheless, many organisations are discovering that deploying AI effectively remains difficult, expensive, and highly dependent on quality data, change management, data governance and organisational redesign. Just as with expert systems, there was a promise of intelligent automation but the reality was often significant costly and complexity.
Equally concerning is that many AI initiatives today remain trapped in pilot stages without producing commensurate economic returns.

The AI Arms Race

Today's frontier model vendors appear locked in increasingly expensive competition. OpenAI, Microsoft, Google, Anthropic, Meta, Amazon and others are investing enormous sums in training larger models, aggressively acquiring novel training data, building data centres, and acquiring scarce GPU resources. In the meantime, expected returns on investment remain highly speculative.

Compare this to the expert systems era, the telecommunications infrastructure boom and investment overheating that preceded the dot-com crash. In each case, firms feared being left behind and in turn invested aggressively en-masse, placing disproportionate emphasis on speculation ahead of sustainable profitability.

Several warning signs are emerging:

•    Capital expenditure continues to rise faster than proven revenue streams,

•    The rate of frontier model improvement has fallen in recent times (less revolution and more evolution),

•    Performance of and differentiation between leading models appears to be narrowing, and

•    Many enterprise customers appear uncertain about long-term return on investment.

At the same time, adoption is progressing unevenly. While some sectors exhibit high uptake of the technology, regulated and safety-critical sectors such as healthcare, government, insurance, and banking continue to face governance, compliance, cybersecurity and reliability challenges that slow deployment. Studies of enterprise AI adoption frequently show significant experimentation and pilot studies but significantly less success in deployments and meaningful business outcomes.

We have seen such expectation/realisation mismatches before. Most of the disruptive technology bubbles of the past have undergone a period when investment growth substantially outpaced economic adoption and demonstrable returns.

The Dot-Com Bubble

During the late 1990s, investors correctly recognised that the internet would fundamentally transform commerce, communications, and society. What they misjudged was the timeline. Many companies attracted enormous valuations despite lacking sustainable business models, profits, or competitive advantages. The NASDAQ rose dramatically before losing a large proportion of its value between 2000 and 2002.

Thousands of startups disappeared. However, there were winners as well. Amazon survived. Google emerged. E-commerce established and eventually, cloud computing transformed enterprise IT.

The internet did indeed reshape the global economy, just not at the pace investors expected.

Today's AI boom shares several characteristics with the dot-com period:

1.    Speculative investment driven by fear of missing out,

2.    Valuations based on speculative future potential rather than demonstrable and sustainable shorter-term earnings,

3.    Infrastructure buildout occurring pre-emptively, and ahead of demonstrated demand, and

4.    A degree of disregard for traditional investment diligence.

However, there is an important distinction. During the dot-com bubble, many companies lacked revenue entirely. Today's AI leaders already possess substantial businesses, with diverse recurrent income and established customer bases. Microsoft, Google, Amazon, and Meta are funding AI expansion from positions of financial strength, and in turn can weather and survive a significant loss with their respective AI ventures should this eventuate.

While this reduces systemic risk to the AI sector overall, other smaller organisations involved in the AI boom, which do not have the benefit of significant and diverse income streams, are more exposed to a future collapse or correction.

AI Winter or AI Autumn?

The most likely outcome is not a repeat of the AI Winter of the late 1980s, nor a collapse on the scale of the dot-com crash.

Instead, we may be approaching what could be described as anAI Autumn.

In contrast to a dramatic collapse, an AI Autumn could involve:

•    Reduced investor enthusiasm and concomitant increase in diligence,

•    Consolidation among AI startups, or acquisition of such startups by the bigger players,

•    A slowdown in growth for AI spending, and

•    Greater focus on profitability, rationalisation of infrastructure investment and practical, safe application.

Historically, technological revolutions eventually transition from disruption to maturity. Railways, air travel, electricity, automobiles, the internet, and mobile computing all experienced periods of excessive optimism followed by correction. The correction did not destroy the technology. Instead, the application of the technology evolved into a sustainable, practical and profitable model.

Artificial intelligence seems to be following this pattern.

The expert systems of the 1980s exhibited inflexibility and lack of compute horsepower. Both of these issues have been addressed for modern AI. And unlike dot-com startups, (also for most of the current startups) the leading AI organisations possess enormous cash reserves and reliable income streams from other activities. 

In a sense, modern AI is “too big to fail”. But not immune to a correction.

In Conclusion

The current AI boom exhibits many of the warning signs associated with previous periods of technological exuberance: massive capital inflows, escalating competition, inflated expectations, and uncertainty about long-term returns.

The expert systems boom and subsequent collapse demonstrates that technical success alone does not guarantee economic sustainability. The dot-com bubble demonstrates that disruptive technologies can still experience significant market corrections.

Nevertheless, it appears that we are likely heading toward an AI Autumn than another AI Winter. A correction is probable, some vendors will fail, investments will be written down and expectations will moderate. But the technology will remain.

A key factor is that, unlike previous AI cycles, the underlying technology is already producing real value across a wide range of industries and there is a relatively higher level of adoption as a result. The likely outcome is not the freezing of innovation, but rather a transition from excitement to pragmatism. 

In that sense, the near future may resemble the post-dot-com recovery: less exuberant, more disciplined, and ultimately more productive.