Britain built more railway than it needed in the 1840s. Investors were right that the technology would transform the country, yet many lost money because prices had already discounted decades of growth. Today’s artificial-intelligence build-out presents the same challenge: separating confidence in the technology from confidence in the price.
Morgan Stanley estimates that global AI-infrastructure spending could exceed $1.3 trillion by 2027. Oracle has reported an AI cloud backlog of $664 billion, with customer prepayments funding roughly 40% of its latest quarter’s capital expenditure. These are substantial commitments, but contracted demand, cash received and revenue earned are not the same thing.
Anthropic provides a live test. It is reportedly considering a Nasdaq listing at a valuation of up to $2.3 trillion, while Nvidia may invest up to $10 billion as an anchor investor. Anthropic’s annualised revenue run rate reportedly rose from about $9 billion at the end of 2025 to more than $65 billion by July, with $190–200 billion projected by 2028. A valuation close to $2 trillion would require investors to pay today for much of that future growth arriving on schedule.
On 12 September, Anthropic chief executive Dario Amodei called for the industry deliberately to slow improvements in model capabilities, citing systems that help build the next generation of AI and an incident involving autonomous attacks on Hugging Face systems. Open AI’s Sam Altman and Tesla’s Elon Musk publicly agreed with the broad concern.
Semiconductor shares fell over the next two trading sessions, with the sector benchmark down close to 6% as investors reconsidered how much near-term chip demand assumes uninterrupted capability growth. Fitch Ratings separately published a stress scenario in which an AI-led correction could reduce US equities by about a third and push the economy into recession. These are illustrations, not forecasts; our task is to understand how much of a portfolio would be affected if the optimistic scenario faltered.
We cannot predict when a setback will arrive, but we can control exposure. Over the summer we sold Scottish Mortgage Investment Trust, whose substantial positions included Nvidia, Amazon, ASML, Meta and Broadcom. We also replaced Jupiter Asian Income, whose largest holdings included Samsung, SK Hynix, MediaTek and Taiwan Semiconductor, with a broader emerging-market dividend ETF.
The holdings we reduced had performed strongly over the previous two years, so our strategies participated in the advance. We did not step away from equities, technology or Asia; we reduced concentration in the area most sensitive to valuation.
The comparison with a global tracker is instructive. Nvidia represented 5.56% of the MSCI World Index at the analysis date. Its nine largest technology-related holdings accounted for 26.6%, while technology represented 29.8% of the index. A tracker investor cannot reduce this concentration when valuations become uncomfortable; the index grows with its largest constituents.
Our portfolios currently hold roughly a third of the MSCI World’s exposure to this theme. That is a deliberate decision, not an accidental outcome. We expect AI to be one of the next decade’s most consequential economic developments, just as railways were in the 1840s and the internet in the 1990s. Yet being early and being right about the technology did not protect investors from paying too much in either boom. Believing in AI and accepting any price for exposure are different decisions.
