Seventy Candles
Seventy years after a summer workshop at Dartmouth College gave the field its name, artificial intelligence sits at the centre of an investment and valuation boom its founders could not have priced. In AI at 70: 14 lessons from a lifetime of boom and bust, Deutsche Bank Research Institute thematic strategist Adrian Cox mines those seven decades for what he calls a treasure trove of clues. John McCarthy, who coined the term, proposed that ten researchers working together for two months could make significant headway. What followed instead was a lifetime of summers and winters, stunning breakthroughs and false dawns. Cox's framing, a play on the 2017 transformer paper that launched generative AI, is that "context is all you need."
Progress Does Not Move in Straight Lines
Cox opens with a visual joke that carries analytical weight. Training compute for notable AI systems plotted on a log scale climbs steadily since 1947. Plotted linearly, seven decades of effort look like a flat line until roughly 2020, then a vertical wall. Exponential growth is almost invariably underestimated, and the second lesson explains why the acceleration outruns intuition. Conventional computing capability has doubled every 18 to 24 months. Training compute grew 1.4 times per year through 2009 and roughly four times per year since 2010, because systems, memory and algorithms improved at once. To be clear, that is not Moore's law. It is Moore's law compounded by three other curves.
Today's Champion, Tomorrow's Footnote
The AI family tree is a graveyard as much as a lineage. Symbolic AI and rules-based logic dominated, then faded. Expert systems suffered a commercial collapse in 1987 and were mostly superseded. Recurrent networks were displaced by transformers. Cox's point is that large language models may in turn give way to something like world models, and the market is not positioned for that.
The browser wars make the same case in market-share terms. Internet Explorer overtook Netscape and was itself overtaken by Chrome. Today ChatGPT still leads generative AI platforms at roughly 5.6 billion monthly visits, but it peaked near 6.2 billion in late 2025 and has plateaued, while Google Gemini has climbed toward 3.0 billion. Early leads do not guarantee long-term winners.
The Pipeline, the Plant, and the Bottleneck
DeepSeek was not an overnight success. China's gross domestic R&D spend overtook the United States in 2024, AI patents granted to Chinese entities approached 100,000, and notable Chinese models kept rising even as the US count fell from its 2023 peak. Pipelines take years and they are visible in advance.
Cheaper compute has not reduced spending. GPU computation costs have fallen more than 99% since 2006, yet data centre electricity consumption is projected to roughly double between 2024 and 2030. Jevons Paradox is alive.
Follow the money and it lands on hardware first. Since the ChatGPT launch on November 30, 2022, median total returns in the Russell 1000 have been led by tech hardware, the four hyperscalers, construction and engineering, communications equipment and semiconductors. Software itself has badly lagged the index. Data centre investment as a share of GDP has gone near vertical. The bottleneck this time sits with producers, not consumers: hourly H100 rental prices fell to roughly $1.90 in early 2025 and have since climbed back above $2.70 as scarcity reasserted itself. US semiconductor imports have surged past $170bn while exports flatlined, which is cheapness and concentration risk in the same chart.
The Payoff Problem
General purpose technologies generate costs before results. Fewer than half of US workers report using AI at work, at a mean of 6% of their time, saving 2% of hours. Consumers adopted generative AI faster than any prior technology, yet monetisable enterprise use lags because reorganising a business takes years. For this readership the relevant number is that finance and insurance runs ahead of the national average, with roughly 37% of firms currently using AI and 44% expecting to within six months.
What the Market Is Already Paying For
The Shiller CAPE sits near 42, just below the dot-com peak, and Cox notes that the biggest valuation booms have mostly been tech booms: electrification, radio and autos, the tronics era, the internet. OpenAI and Anthropic are exceeding the revenue growth records set by Google, Meta, Nvidia and Alibaba at the same stage. This time genuinely is different. The uncomfortable corollary is that it will need to be even more different, because investor projections imply revenue trajectories several times steeper than any predecessor achieved.
Against that, the most sobering chart is the calmest. S&P 500 earnings per share have compounded at roughly 6.5% through wars and recessions since 1935, and US GDP per capita has tracked a 1.9% trend since 1870. An AI-boosted decade at 2.1% is barely visible on the page. Cox also plots the tails, from a benign singularity to extinction, and lets them sit there without adjudication. It is simply too early to tell which line the data will follow.
Five Takeaways for Advisors and Investors
- Own the constraint, not the headline. Returns have accrued to power, chips, construction and equipment. The bottleneck moves, and portfolios should follow it rather than the narrative.
- Efficiency gains are demand signals. Falling cost per token has raised aggregate spending, not lowered it. Treat cost collapses in AI as bullish for volume, not bearish for capex.
- Underwrite the fade risk in leaders. Dominant architectures and dominant platforms have both been displaced repeatedly. Concentration in today's winners is a bet on continuity that history does not support.
- Separate adoption from monetisation. Consumer uptake is record-fast, enterprise productivity is not. The J-curve implies costs land in earnings before benefits do.
- Anchor to trend. Earnings and GDP trends have absorbed every prior revolution. Positioning built on permanent deviation from trend carries the burden of proof.
The discipline here is not prediction. It is context. To borrow from the UK chart-topper of 1956 that Cox invokes, Que Sera, Sera, with the caveat that portfolios still have to be positioned in the meantime.
Footnote:
Cox, Adrian. AI at 70: 14 Lessons from a Lifetime of Boom and Bust. Deutsche Bank Research Institute, Aug. 2026, theideafarm.com/wp-content/uploads/2026/08/PROD0000000000637666.pdf.
