Deutsche Bank's latest research report maps the evolution of artificial intelligence since its birth in 1956, extracting 14 critical insights from historical patterns to guide investors in assessing the trajectory of the current AI boom.
August this year marks exactly 70 years since the 1956 Dartmouth Summer Workshop, widely recognised as the birthplace of artificial intelligence. According to the report, AI's seven-decade journey has been marked by alternating periods of exuberance and contraction, and the current wave of investment and valuation enthusiasm is replaying technological revolution paradigms seen numerous times throughout history.
The report argues that understanding the context is crucial for discerning AI's future direction. From non-linear growth and infrastructure bottlenecks to the inflation and deflation of valuation bubbles, historical signals are clearly discernible. While some may claim "this time is different," the data from the past 70 years offers an alternative frame of reference—for investors betting on the AI track, these lessons directly impact asset allocation logic and risk assessment.
Growth Is Non-Linear and Often Severely Underestimated
The report opens by highlighting a core feature of AI progress: non-linearity. When training compute data is plotted on a logarithmic scale, it becomes evident that compute power used to train major AI systems has grown across dozens of orders of magnitude since 1956, while a linear chart visually obscures this trend almost entirely. Exponential growth is intuitively easy to underestimate—this is the first cognitive hurdle to understanding the AI wave.
Closely related is the fact that AI's pace of progress has surpassed Moore's Law. Traditional computing power doubles every 18 to 24 months, but since the deep learning era began, annual compute growth has reached approximately 4x, well above the pre-deep-learning rate of around 1.4x annually. This is driven by simultaneous improvements across system scale, memory enhancement, and algorithmic optimisation, creating compounding effects.
Technology Routes Continuously Evolve; Today's Leaders May Not Be Tomorrow's Winners
The report presents a 70-year technology lineage map: from symbolic logic and expert systems to statistical machine learning, deep learning, and the currently dominant large language models. Each generation of mainstream technology has experienced cycles of rise and displacement, with some routes (such as recurrent neural networks) already surpassed while others continue evolving in parallel. The report notes that large language models may eventually give way to new paradigms like "world models," and technology generational shifts do not bend to the will of current leaders.
Historical market share changes reinforce this point. Internet Explorer once overwhelmed Netscape, only to be displaced by Chrome itself. In the current competitive landscape of generative AI platforms, ChatGPT leads in monthly visits, but Google Gemini, DeepSeek, and Claude are all rapidly catching up. Early advantages do not equal durable moats.
R&D Accumulation Determines Competitive Dynamics; DeepSeek's Rise Is No Accident
The sudden emergence of Chinese AI models—represented by DeepSeek—appears superficially as an overnight success but is actually the result of years of accumulated R&D investment. Data shows that China surpassed the United States in total R&D spending in 2024, and its pace of catching up in the number of major AI models is equally significant. In AI patent grants, China's growth curve also leads other economies substantially. For investors, this means competitive shifts often build beneath the surface for years before becoming visible.
Cost Declines Don't Shrink Demand—They Expand It
The report cites Jevons Paradox to illustrate that a dramatic drop in AI usage costs does not reduce overall spending but instead stimulates a surge in demand. Since 2006, GPU compute costs have fallen by over 99%, yet according to International Energy Agency projections, global data centre electricity consumption is set to double from 2024 to 2030. Lower marginal costs translate into more application scenarios and higher total demand.
Software Revolutions Require Hardware First; Current Bottlenecks Are on the Supply Side
Every technological revolution has depended on large-scale hardware investment, and AI is no exception. Since ChatGPT launched in November 2022, data centre, hardware, and chip-related industries within the Russell 1000 index have delivered far higher total returns than the software sector. Unlike historical constraints where consumer device adoption was the limiting factor, the current bottleneck lies more on the supply side of AI chips: the hourly rental price of Nvidia's H100 GPU has fluctuated consistently over recent quarters, reflecting structural tension between supply and demand.
Globalisation Lowers Technology Costs but Also Seeds Supply Chain Risks
As the AI boom intensifies, US semiconductor imports have risen sharply. Meanwhile, global supply chain concentration has increased, with products becoming significantly more dependent on single sources. The report points out that while globalisation makes technology cheaper, it also exposes supply chains to higher geopolitical and concentration risks.
Technology Dividends Take Time to Materialise; Commercial Deployment Remains Early Stage
Technology revolutions typically follow a productivity "J-curve": costs appear before benefits. The report cites data showing that fewer than half of US employees currently use AI to complete work tasks, spending on average 6% of working hours on AI-related activities, saving only about 2% of work time.
Consumer adoption has already set historical records, with generative AI spreading faster than the internet and personal computers. However, directly monetisable enterprise applications are progressing noticeably slower—downloading an app takes minutes, while restructuring corporate operations around new technology takes years. Information, professional services, and financial insurance sectors lead in AI adoption, while manufacturing lags behind.
Valuations Are at Historic Highs; Revenue Expectations Need to Be "More Different" to Deliver
On the Shiller Cyclically Adjusted Price-to-Earnings ratio, current S&P 500 valuations are near the peak of the 2000 internet bubble, sitting in the highest range in 150 years. Historically, every valuation peak has corresponded to a transformative technology wave: electrification in 1899, radio and automobiles in 1929, the electronics wave in 1966, the internet in 2000—and AI in 2026.
On the revenue front, OpenAI and Anthropic's current growth rates have already surpassed the historical peak levels of comparable companies in the same period. However, the report also notes that to justify the forward revenue expectations embedded in current valuations, both companies' growth curves would need to be "more different" compared with the historical trajectories of tech giants like Google, Meta, and Nvidia.
Long-Term Trends Are Resilient, but Extreme Scenarios Have Entered the Discussion
The report concludes with the long-term trend of S&P 500 earnings per share, showing that this metric has grown at approximately 6.5% annually since 1935, maintaining stability through numerous wars and recessions. At the GDP level, the report outlines three AI scenarios: a baseline trend, an AI-driven moderate acceleration (around 2.1% average annual growth over the next decade), and two extremes under the "singularity" scenario—from technological utopia to human extinction. The report maintains a neutral stance but reminds investors: between historical trends and extreme scenarios, where current market pricing sits on this spectrum is a core question worth continuous monitoring.
Comments