Artificial intelligence is seventy years old, at least by the conventional reckoning that begins with the Dartmouth Summer Research Project of 1956. A Deutsche Bank Research Institute paper by Adrian Cox uses that long history to ask a useful question: what can the repeated booms, winters, breakthroughs and disappointments of AI tell us about the present surge?

The report offers fourteen lessons. Taken together, they suggest that the current AI boom is both genuinely unusual and recognisably part of a much older pattern. Capability can improve astonishingly quickly, but economic benefits arrive more slowly; software depends on enormous amounts of physical infrastructure; cheaper computing can increase total demand rather than reduce it; and neither today’s dominant technology nor today’s dominant companies are guaranteed to remain dominant.

Progress is not a straight line

One of the report’s strongest points is visual. Plot the computation used to train notable AI systems on an ordinary linear chart and almost the entire history looks flat until very recently. Plot the same data logarithmically and the underlying exponential trend becomes obvious.

Deutsche Bank cites Epoch AI data showing that training compute accelerated from roughly 1.4 times per year before the deep learning era to about four times per year after 2010. Epoch AI’s own published analysis puts the more recent trend for notable models at roughly 4 to 5 times per year. The important lesson is not that exponential growth will continue forever, but that our intuition is poor at understanding compounded technological change.

Today’s leading AI may not be tomorrow’s

The history of AI is a family tree rather than a straight road. Symbolic AI, expert systems, recurrent neural networks, reinforcement learning, deep learning, transformers, large language models, computer vision and robotics have all waxed and waned in importance. Technologies that once appeared dominant have sometimes become niche tools or been superseded altogether.

That matters when we talk about ChatGPT, Claude, Gemini and the current generation of large language models. Their importance today is beyond dispute, but seventy years of AI history offers little reason to assume that the architecture currently at the centre of attention will remain so indefinitely.

There is no cloud: there are buildings, chips and electricity

AI often feels like pure software to the person using it, but the report repeatedly returns to the physical infrastructure underneath it. Semiconductor manufacturing, data centres, networking equipment, cooling systems and electricity supply have become essential parts of the AI economy.

The International Energy Agency supports this point rather dramatically. Its 2025 Energy and AI report projects global data centre electricity use rising from around 460 TWh in 2024 to roughly 945 TWh by 2030. Accelerated servers, largely driven by AI, account for a substantial share of that growth. In other words, the apparently weightless world of AI has a very considerable appetite for concrete, copper, silicon and power.

Cheaper AI may mean more AI, not less spending

Deutsche Bank applies the Jevons paradox to artificial intelligence. When a resource becomes much cheaper to use, consumption can increase so much that total expenditure rises rather than falls. GPU computation costs have fallen spectacularly since the mid 2000s, yet demand for computing capacity and electricity continues to grow.

The same principle may apply at the individual level. If AI makes a task quicker and cheaper, we do not necessarily perform the same number of tasks for less effort. We may attempt things that previously were not worth doing at all. Cheaper intelligence can create demand for more intelligence.

Consumers can move quickly; businesses cannot

Perhaps the most immediately useful lesson for ordinary AI users is the difference between personal adoption and organisational adoption. A consumer can download an application and start experimenting within minutes. A company may have to change procurement, security, workflows, governance, training, data systems and management practices before it can obtain the same benefit at scale.

Deutsche Bank’s August 2026 data show generative AI being adopted by individuals considerably faster than earlier general purpose technologies such as computers and the internet. Yet business use remains uneven across sectors, with information and professional services well ahead of retail and manufacturing.

This creates an intriguing possibility for small businesses. A one person firm can often change its working method immediately. Large organisations have more resources, but they also have more organisation. In the early phase of a technological shift, having fewer layers of permission can be an advantage.

Productivity benefits may arrive late

The report invokes the productivity J curve: major technologies can generate investment costs, disruption and organisational change before measurable productivity benefits appear. Computers, the internet and smartphones all took time to reorganise working practices.

That warning is particularly important when judging AI by current national productivity statistics. A technology can be transformative for some individuals while remaining almost invisible in aggregate economic data. The interesting measure may not always be hours saved. It may be the number of useful things that become practical to attempt.

Fast growth does not guarantee permanent winners

Deutsche Bank compares today’s generative AI market with earlier technology battles. Netscape was overtaken by Internet Explorer; Internet Explorer was later overtaken by Chrome. Early leadership matters, but history is full of technology companies that appeared unassailable until the market moved underneath them.

The report makes the same point about valuations. Some of the largest stock market valuation peaks have coincided with transformative technologies, including electrification, radio, automobiles, electronics and the internet. A technology can be genuinely revolutionary while its financial valuation still becomes excessive. The two ideas are perfectly capable of being true at the same time.

The larger lesson

The most useful message from AI at 70 is therefore neither optimism nor pessimism. It is historical perspective. AI has repeatedly progressed faster than expected, stalled where enthusiasts expected immediate success, changed technological direction and required far more infrastructure than the word “software” might suggest.

The present boom may turn out to be larger than anything that came before it. Yet the history surveyed by Deutsche Bank advises some modesty about forecasts. Capability may continue to accelerate, adoption may continue to spread and AI may become deeply embedded in ordinary work. Even so, the eventual winners, technologies, costs and economic effects are far from settled.

For those of us actually using these tools, perhaps the most interesting question is smaller and more practical: what happens when the cost of turning an idea into a first attempt becomes extremely low? That may be where AI’s most immediate effect is being felt long before economists can see it clearly in the productivity statistics.