Is an AI Bubble Creating a 2000 Repeat?
:format(webp)?_w=3840)
Investment Note #24 - 7th February 2025
Is an AI Bubble Creating a 2000 Repeat?
.
On the 29th of January, Key Capital held its Annual Market Outlook event, at which we discussed the possibility that the AI capital expenditure boom and the performance of US equity markets were creating similarities with the late 1990s/2000 period. This Investment Note will delve into this in more detail and add some additional AI observations related to DeepSeek.
At the bottom of this Note are links to edited videos of the two themes and the entire presentation.
Equity Valuations
US equity markets are expensive, and we are close to the peak reached in 2000. The CAPE ratio is a valuation measure that uses rolling 10-year real earnings to smooth out cyclicality over a business cycle. The current CAPE is 38 versus a peak of 44 in 2000, which compares to an average of 27 since the Financial Crisis (source: Schillerdata.com).
The 10 largest US stocks had a simple average Price/Earnings (P/E) ratio of 70 in Q1 2000. Currently (as of 24 January 2025), they have a simple average P/E of 58.
However, it may be reasonable to argue that the current valuation of the US stock market is heavily influenced by the tech sector and this sector has created a structural shift in US earnings relative to the long-term trend.
Historically, a rising US equity market dragged the rest of the World along, but currently, the US market is trading at a significant premium to the rest of the World, with other G7 equity markets currently averagely valued.
Market Concentration
The S&P 500 is significantly more concentrated than it was in 2000. The top 5 accounted for c. 18% in 2000 compared to 28% today, which exceeds the Nifty Fifty peak in the late 1960s/early 1970s when IBM, AT&T, Exxon, Eastman Kodak and GM accounted for c. 24% of the index. Even broadening it out and looking at the top 10% largest US stocks, the 2025 market is the most concentrated it's ever been, edging past the early 1930s.
Earnings
Strong earnings growth has propelled the US equity market to outperform other equity markets, but strong earnings growth has been uneven across companies and sectors.
Of the top 10 US stocks in Q1 2000, four make lower nominal earnings today than they did in 2000. These four companies are Intel, Citigroup, GE, and IBM. This highlights the inherent uncertainty over future earnings for individual stocks, even for index heavyweights. In comparison, Microsoft’s nominal earnings have increased 9.8 times since Q1 2000.
For the Mag-7, analysts forecast a range of 9% to 22% per annum earnings growth for the next 5 years. Long-term S&P earnings growth is c.6% p.a., but the Mag-7 has comfortably outperformed this over the last decade. The forecast analyst range for Apple, Meta, Alphabet, and Microsoft is 9 - 15%, and the range for Tesla, Amazon, and Nvidia is 19 - 22%.
Mag-7 More Like Countries
To a certain extent the Mag 7 are more like countries than companies with respect to size and profitability. Mag-7 last 12-month earnings were US$460bn versus the combined earnings of all public companies in France, Germany and the UK of US$436bn (country aggregates are a sum of LTM net profit of common and preference stocks domiciled in a particular country, excluding stocks with a market cap below $200m).
DeepSeek - Elevated Capital Expenditure Creates Risks
The market’s “shoot first, ask questions later” reaction to DeepSeek, highlighted the risk of the capital expenditure binge not delivering the expected earnings. There has been significant coverage of the cost of DeepSeek’s model but there has been some misinformation surrounding this point.
Commentary has focused on the $6m training cost for DeepSeek's model, but the company has acknowledged that this cost is just for the final pre-training run rather than the full cost to develop the model. SemiAnalysis, a semiconductor research firm, estimates that DeepSeek's server capital expenditure is almost $1.3bn, so this is not some scrappy start-up operating on a shoestring budget.
Nvidia’s Moat
Nvidia’s competitive moat is primarily a function of its CUDA software, its in-house programming language, which is the default for AI models and only works on NVIDIA chips. It has a significant lead in its ability to combine multiple chips in one large virtual Graphics Processing Unit (GPU) with industry-leading interconnect technology. However, several companies are attempting to design code that works on a variety of GPUs.
DeepSeek had to get creative with optimising the chips they could get, and the fact that they could achieve similar results to leading models is worrying for Nvidia.
Positive For Other Mag-7?
In the long run, the commodification of AI models combined with a lower cost to run queries on AI models is probably a positive for other Mag-7 companies.
Amazon, Apple, Microsoft, and Meta being able to provide their customers with AI at a much lower cost means spending less on GPUs and data centres. History has shown that mass adoption of new technologies is typically a positive for market leaders.
For Google, the picture is more complicated. Who knows how cheaper AI impacts their search business? Lower cost is still a positive for them, but given how dominant they already are here, any change is probably a net negative.
Summary
US fundamentals are strong, and the excesses of the 1990s are not present, but the market is discounting a very positive future. Right now, the US growth premium helps justify valuations. DeepSeek provided a mini-shock, and the reaction so far has been painful. If growth were to slow or another DeepSeek-type shock upends the current narrative around AI, that valuation premium may become a headwind.
The most relevant current comparison to the 1990s is that investing in new technologies can deliver eye-watering returns in the short term, but there is a lot of risk in crowning winners too early.
[Source (unless otherwise stated) is Deutsche Bank]
Annual Market Outlook Event video links:
AI Technology Stack. A discussion of the AI Technology Stack - 6 minutes.
US Equity Markets. A discussion of US equity markets - 8 minutes.
Full Outlook Presentation. The full presentation - 43 minutes.