In November 2026, expectations for global AI investment and valuations in capital markets have undergone a marked shift. In an updated model released on November 13, semiconductor research firm SemiAnalysis lowered its forecast for global AI accelerator capital expenditure in 2027 from USD 770 billion to USD 627 billion, a reduction of approximately 19%. Its forecast for 2028 was cut from USD 1.054 trillion to USD 758 billion, a decline of 28%. Meanwhile, the latest research by JPMorgan found that NVIDIA GPUs accounted for only approximately 28% of the announced computing capacity agreements signed by leading AI laboratories for deployments planned in 2026 and beyond. The remaining share was distributed among AMD, Google, Amazon and custom chip suppliers. Although this measurement differs from NVIDIA's previously reported 80% to 90% share of AI chip revenue, the trend towards diversification in computing capacity supply is clear. After several years of sustained investment enthusiasm, the AI sector is beginning to face fresh assessments of demand and a changing competitive landscape. However, these developments do not in themselves prove that the AI industry is in decline. The key questions the market now needs to answer are what level of AI investment is economically justified and how much adjustment is required in the valuations of AI stocks, which have risen rapidly in recent years.
However, the global capital markets have already reacted quite sharply to this, well before anyone could reach a reasonable conclusion.
Public filings show that Duquesne Capital, the family office controlled by Stanley Druckenmiller, nicknamed “Stan the Man”, completely liquidated its positions in Advanced Micro Devices, SanDisk and Google during the third quarter. Its reduction in Taiwan Semiconductor Manufacturing Company (TSMC) was particularly substantial, with its holdings falling from 589,700 shares to just 32,700, a reduction of 94%. In addition, George Soros's family office, along with investment firms belonging to other billionaires around the world, reduced their holdings in NVIDIA, locking in gains before Jensen Huang's dramatic departure from the highly regarded chip design company. According to data compiled by Bloomberg from quarterly regulatory filings, fund managers overseeing Soros's investments reduced their NVIDIA holdings by nearly one-fifth in the quarter ended September 30. This is the largest reduction in the company's disclosed NVIDIA holdings this year, following an earlier increase in its position. Michael Platt's BlueCrest Capital Management and the investment company of Abu Dhabi's ruler, Sheikh Mohamed bin Zayed Al Nahyan, also reduced their NVIDIA holdings during the same period.
According to research released by Citadel Securities in early November, the forward price-to-earnings (P/E) ratio of the S&P 500 semiconductor sector fell to 13.8 times, down 45% from its January 2026 peak. Another set of figures from JPMorgan showed that the P/E ratio of the US semiconductor index, based on earnings over the following 24 months, had fallen from a June peak of 22.5 times to 11.4 times, almost halving. Share prices have also retreated significantly. On November 10, Wells Fargo reported that semiconductor stocks had fallen approximately 38% from their June highs. Reuters' figures on November 17 showed that the Philadelphia Semiconductor Index had declined approximately 34% from its June peak. Although the figures differ across indices and reporting dates, a correction of 30% to 38%, together with valuation multiples in some segments approaching half their previous levels, is sufficient to demonstrate a major shift in market expectations.
Does this adjustment represent a rational correction of overinvestment, or does it already incorporate a significant degree of panic selling?
Making this assessment requires first distinguishing between the long-term economic value of AI technology and the current capital returns of AI companies. As it stands, AI is not an ordinary consumer electronics product, but a general-purpose technology capable of transforming how production and knowledge work are carried out, possessing strong and distinct infrastructure attributes. Its applications can continue expanding into fields such as industrial manufacturing, scientific research, finance, healthcare, and public services. Even if some computing power investments fail to generate expected returns in the future, that does not mean AI as a technological infrastructure lacks development prospects. The long-term value of a technological revolution and the scale of investment at any given stage do not inherently have to remain aligned at all times.
The history of the steam engine offers a useful reference point for understanding this distinction. Research by economic historian Nicholas Crafts found that the greatest contribution of steam power to British economic growth emerged gradually, approximately a century after Watt's invention. This suggests that a groundbreaking technology often requires a considerable period of time to move from invention and adoption to a comprehensive transformation of the production system. However, as a technology becomes widespread, manufacturers that initially enjoyed high profits because of the scarcity of equipment may not be able to sustain their early valuation premiums over the long term. As the technology matures, equipment supply increases and market competition intensifies, the productivity gains enjoyed by society can continue to grow, while technology suppliers' excess profits may gradually decline.
Using a 100-year as observation period, if the valuation premium established during the early stages of a general-purpose technology is assumed to eventually contract by a factor of 30 to 50, the implied average annual decline in valuation multiples would be approximately 3.3% to 3.8% on a compounded basis. This assumption does not represent a historical statistical measure of the actual decline in steam-engine-related stocks. Rather, it serves as a long-term analytical scenario illustrating the relationship between technological diffusion and valuation convergence. Modern technological development and capital markets' responses to future earnings prospects are clearly faster than during the Industrial Revolution. Even if the diffusion and evolution of AI technology are assumed to occur at ten times the speed of previous technological revolutions, a century-long change in valuations cannot be mechanically translated into short-term stock price movements. However, provided that corporate earnings expectations have not deteriorated significantly, a 3% to 5% decline can be used as an initial sensitivity range for normal valuation adjustments, while 10% can serve as a moderate stress scenario incorporating some increase in the risk premium. These assumptions provide a framework for assessing whether market reactions have deviated significantly from underlying fundamentals. These ranges should not be interpreted as objective pricing benchmarks derived from the historical experience of steam engines. Instead, they can be seen as analytical assumptions that require continuous reassessment and adjustment in light of actual earnings data.
This touches an often-overlooked issue in the valuation of AI-related investments. While technological progress typically drives down the cost per unit of a product or service, such a decline does not necessarily diminish the total economic value of an entire industry. For instance, as steam power became more affordable, industrial firms were able to integrate mechanical power into a broader range of production processes, thereby fueling more extensive industrialization. The same logic applies to AI. As algorithmic efficiency improves, competition among chipmakers intensifies, and the costs of model inference fall, the barriers to adopting AI will likely continue to lower, which might create even greater demand for its application. While this could be a positive development for the economy as a whole, it forces equipment suppliers who have historically relied on the scarcity of computing power to secure high profit margins to re-evaluate their future profit potential.
Such are the competitive changes faced by NVIDIA. Leading AI laboratories are beginning to increase their purchases of AMD chips, Google's TPUs, Amazon's Trainium chips and other custom-designed processors. This does not mean that NVIDIA will immediately lose its technological advantage. Rather, it indicates that customers are no longer willing to depend on a single supplier over the long term. In the past, computing equipment could command substantial premiums because of supply constraints and strong hardware and software ecosystem barriers. Suppliers will have to contend with a wider range of alternatives. The resulting price declines may reduce revenue growth for chip suppliers, while simultaneously improving the cost structures of cloud service providers, model developers and businesses using AI applications. Profits in the AI industry will not necessarily disappear altogether. Instead, they may be redistributed among different segments as the technology becomes more widespread and competition intensifies.
The appropriate scale of investment must also be reconsidered. Over the past few years, AI investment has been driven primarily by rapidly growing demand for computing capacity and competition to secure technological leadership. Companies have tended to build data centers in advance, purchase chips and secure electricity resources ahead of actual demand. As supply increases and commercial revenues gradually materialize, investment decisions must take greater account of equipment utilization rates, depreciation cycles, financing costs and the amount of revenue that actual applications can generate. If some data center projects fail to produce sufficient cash flow, those investments may lack economic justification even if overall AI demand continues to grow. Reducing such investments with relatively low marginal returns can ease financial pressure and improve the average efficiency with which the remaining capital is deployed. Therefore, slower investment growth does not necessarily indicate a deterioration in industry performance. In some circumstances, it may actually help improve returns on capital.
However, AI investment has not yet undergone an across-the-board contraction.
Although SemiAnalysis has lowered its forecasts, its revised model still projects that global AI accelerator expenditure will increase from approximately USD 392 billion in 2026 to USD 627 billion in 2027, before reaching USD 923 billion in 2030. A November report by Pacific Investment Management Company (PIMCO) showed that the market's combined forecast for 2027 capital expenditure by the five major cloud service providers, namely Amazon, Google, Microsoft, Meta and Oracle, remained close to USD 1.17 trillion, rising to approximately USD 1.30 trillion in 2028. Different institutions use different statistical scopes and methodologies, hence their figures cannot be directly compared. Nevertheless, they collectively indicate that the market is reassessing the pace and composition of future investment growth.
With this in mind, the current scale of the stock market correction deserves closer examination. Citadel Securities has observed that, since the end of June, forward earnings expectations for the S&P 500 semiconductor sector have risen by approximately 30%, while the forward price-to-earnings (P/E) ratio has fallen by about 35% over the same period. This combination of improving earnings expectations alongside a significant contraction in valuation multiples suggests that the market is not merely adjusting its near-term performance forecasts, but is also re-evaluating the sustainability of future growth and the associated risks. While this shift is reasonable, investors who interpret a slowdown in the growth of AI-related capital expenditures as an imminent collapse in industry-wide earnings may be underestimating the long-term demand driven by the diffusion of technology.
The distinction between a valuation adjustment and a panic-driven pullback can be illustrated through a simplified pricing scenario. Suppose a certain category of AI stocks reaches a high of 200 points, if the companies' future earnings expectations remain largely stable, a standard valuation adjustment would typically be in the range of 3% to 5%. Even accounting for a decline in risk appetite with a 10% drop as a "stress scenario", the price would still stand at 180 points. However, if the market price has already fallen to 150 points, which is a 25% decline, this clearly exceeds the scope of such a scenario. Should the price subsequently recover to 180 points, it would represent a 20% rebound from the 150-point level. This 20% figure serves as an estimate of potential valuation recovery based on specific assumptions, rather than a definitive prediction of future stock prices. Such an analysis is only meaningful provided that there are no significant downward revisions to corporate earnings and that the underlying assumptions regarding reasonable valuation remain valid.
In reality, the most important task is to verify whether these assumptions hold. The future market share, product prices and profit margins of AI chip suppliers may decline as customers develop their own chips and competition across the industry intensifies. Although cloud service providers may benefit from lower computing costs, they may also have to bear substantial equipment depreciation and long-term financing costs. There are also considerable differences in whether model companies can generate sufficient revenue growth to cover their training, inference and research and development expenditure. Even though the semiconductor sector as a whole has already experienced a correction of more than 30%, this does not mean that all companies are clearly undervalued. Capital markets may simultaneously contain companies whose valuations have been adjusted excessively and others whose earnings risks have not yet been fully reflected in their share prices.
It should also be noted that the concentration of capital in AI infrastructure is not at odds with the widespread adoption of AI applications. The production of steam engines required specialized firms, technical talent, and capital investment, yet steam power ultimately served a vast array of industries. Similarly, while AI necessitates the centralized development of high-performance computing infrastructure, it can enable a broader range of enterprises and individuals to access such technological capabilities by lowering inference costs and providing standardized services. If the "democratization of AI" is interpreted as requiring all market participants to possess foundational models and computing infrastructure of equal scale, it clearly runs counter to the economic logic of capital-intensive industries. However, if it refers to a continuous lowering of the barriers to using such technology, it could well serve as a crucial catalyst for expanding market demand for AI. One must not conflate the way capital is organized with the scope of technological adoption.
The global AI industry has moved beyond simply pursuing computing capacity expansion and is beginning to reassess investment efficiency and commercial returns. Investment in computing capacity may continue to increase, but the pace of growth, supplier structure and distribution of profits will all change. For capital markets, AI's long-term technological prospects remain important, but they are insufficient on their own to justify any particular valuation level. The sharp contraction in semiconductor valuation multiples in recent months already reflects a substantial repricing of risk. Some high-quality companies may have been oversold. Whether the market can achieve a stable recovery in valuations will still depend on the relationship between actual demand, earnings and capital expenditure.
Taken together, it is reasonable to conclude that AI stocks have recently entered a phase of panic-driven downward adjustment. Although overall AI stock valuations have room for correction from previously elevated levels, the current sell-off displays clear characteristics of panic and has gone far beyond what would be required for a normal adjustment. Given the current decline of 35%, or potentially even greater losses in the future, AI valuations have room for a recovery of at least approximately 20%.
Final analysis conclusion:
A moderate adjustment in AI-related investment is logical. Reducing inefficient spending and improving capital returns could even help this technological revolution transition into a more sustainable phase of commercialization. What must be avoided is the misinterpretation of a slowdown in capital expenditure growth as the evaporation of AI’s long-term value, which could trigger panic selling. Such a reaction runs entirely counter to the historical inevitability of technological revolutions; the broader development of social progress does not support such alarmist conclusions. Consequently, a degree of pullback in AI-related stocks is to be expected.
