The capabilities of AI systems have continued to improve quickly in the last 18 months (Artificial Analysis (2025)).
AI has the potential to have a transformational impact on many sectors of the economy as capabilities improve further (Crafts (2021)). The potential impact of AI can already be observed. For example, the AlphaFold models developed by Google DeepMind have provided a breakthrough in predicting the 3D structure of proteins from their amino-acid sequences, solving a root-node problem that is now unlocking advances across diagnosis, protein engineering and drug discovery (Jumper et al (2021)).
The current level of AI model capabilities, and the rate of progress, has resulted in some AI industry leaders anticipating powerful models (AGI) this decade, which they expect to proliferate across society, transforming it in the process (Altman (2025); and Amodei (2024)). Additionally, several countries have made AI progress and adoption a core part of their national economic and security strategies, which is contributing to the accelerated pace of AI development (EU AI Continent Action Plan (2025); UK AI Opportunities Action Plan (2025); and US AI Action Plan (2025)).
AI stocksfootnote [1] now account for roughly 44% of the S&P 500 market capitalisation, with AI-impacted companies extending beyond the technology sector.
The AI asset price story is not just about the big AI model providers – understanding the full impact of AI-driven events on asset prices and financial stability requires an understanding of the full AI stack, including key dependencies. Hyperscalers and cloud service providers, AI model developers, specialist AI chip manufacturers, AI application developers, companies that specialise in networking, storage, and cooling systems for AI data centres, and data centre operators all derive their expected future earnings from the future trajectory of AI progress and adoption. For example, JP Morgan have developed an ‘AI stock’ index which contains 30 heavily AI-impacted S&P 500 constituents across multiple sectors including technology, real estate, utilities and consumer discretionary. Those stocks comprised roughly 26% of the S&P 500 in late 2022, compared to 44% in October 2025.
AI system training and inferencefootnote [2] also requires power, while the development of new AI data centres and energy infrastructure can also drive demand for a range of commodities – such as copper and uranium. Subsequently, the asset prices of utility companies, commodity producers and commodities can also be impacted by AI-driven events.
AI stocks have pushed some US stock valuation metrics to their highest level since the dot com bubble 25 years ago, though these metrics do not fully account for the high projected earnings growth of many AI-impacted companies.
The cyclically adjusted price to earnings (CAPE) ratio of the S&P 500 is close to its dot com peak – Chart 1. AI stocks are the key driver of this, and as of early October have a median forward 12-month price-to-earnings ratio of 31x, compared to 19x for the broader S&P 500 index. However, these valuation metrics do not capture the high future earnings growth projections that many AI-impacted companies have, with staff at the Bank of Italy showing that the valuation of these AI companies are justifiable if that earnings growth materialises (Albori et al (2025)). Whether these earnings projections will be realised – or even prove underestimates – is uncertain (see below).
Chart 1: Some US stock valuation measures are at the highest level since the dot com bubble
Footnotes
- Sources: Professor Robert J Shiller, Yale University and Bank calculations.
The build-out of infrastructure to improve AI capabilities further and meet growing projected adoption is forecast to require trillions of dollars of capital investment this decade, a significant part of which is expected to be financed by debt.
McKinsey estimate that data centres equipped to handle AI processing loads will require $5.2 trillion in capital expenditures by 2030 to keep up with AI ‘compute’ demand for model training and inference (McKinsey (2025)). While AI infrastructure capital expenditure (CAPEX) requirements have to-date largely been met by hyperscalers leveraging their strong internal cash flows, the speed and scale of the projected AI infrastructure buildout is expected to substantially increase the role of external financing, including debt. Morgan Stanley Research estimate that AI Infrastructure CAPEX between 2025 and 2028 will be $2.9 trillion, with $1.5 trillion expected to be met by external capital, including $800 billion from private credit – Chart 2 (Morgan Stanley (2025)).
Chart 2: There is a large projected role for external financing in AI infrastructure spending
Footnotes
- Source: Morgan Stanley Research.
This CAPEX on AI infrastructure is underpinned by the expectation of: (a) substantial adoption of AI (and therefore demand for inference ‘compute’); and (b) the necessity of using huge amounts of computational power to develop very powerful AI models. The latter expectation is underpinned by the observation that increasing the computational power of AI models with neural network architectures (eg ChatGPT 5) increases model performance – often referred to as the ‘Bitter Lesson of AI’ after a blog by Rich Sutton (Kaplan et al (2020); and Sutton (2019)). Recent advances in reasoning models and agentic AI have demonstrated that these ‘AI scaling laws’ apply during model inference as well as training, with models that ‘think for longer’ exhibiting higher levels of performance (Briski (2025)). These scaling laws continue to underpin the investment decisions of major AI companies: the training computational capacity of frontier AI models has grown at 5x a year since 2020, and frontier models requiring between 4 and 16 gigawatts of power to train are expected by 2030 – enough to power millions of homes (Epoch AI (2025a); and You and Owen (2025)).
The infrastructure spending figures outlined above also do not account for CAPEX on the development of energy infrastructure necessary to fuel these power-hungry AI data centres – which itself could be considerable. Goldman Sachs estimate that 60% of data centre power demand growth through 2030 will need to be met with new capacity (Goldman Sachs Research (2025)). The International Energy Agency estimates that electricity demand from data centres worldwide will more than double by 2030 in their central case, to a consumption level higher than Japan (IEA (2025)).
There are a range of developments that could trigger a re-evaluation of future earnings/project revenues and a subsequent fall (or rise) in AI-impacted asset prices.
These could include (but are not limited to) underwhelming speed of AI capability progress or user adoption of AI, or below-expectation ability of AI companies to monetise the users of their AI applications. The speed of AI progress and economic impact is highly uncertain, as seen in the wide range of estimates of the future impact of AI on productivity by research economists and timelines to very powerful AI models by AI experts (AAAI (2025); Epoch AI (2025b); and OECD (2024)). Several factors could also prove to be bottlenecks to AI progress, most likely power, but also including training data and AI chip production (Sevilla et al (2024)). Conversely, the realisation of AI with the transformational capabilities forecast by Demis Hassabis (‘10 times bigger than the Industrial Revolution, and maybe 10 times faster’), Leopold Aschenbrenner and some other AI experts could lead to these valuations proving underestimates (Aschenbrenner (2024); and Guardian (2025)).
For companies who depend on the continued demand for massive computational capacity to train and run inference on AI models, an algorithmic breakthrough or other event which challenges that paradigm could cause a significant re-evaluation of asset prices. A natural event study of this was seen in January 2025, where the introduction of the DeepSeek reasoning model triggered a fall in the stock price of many AI-impacted companies lower down the ‘AI stack’ – Chart 3.footnote [3]
Chart 3: The DeepSeek reasoning model announcement provided a natural event study
Footnotes
- Sources: Refinitiv Workspace from LSEG and Bank calculations.
Financial stability consequences of an AI-related asset price fall could arise through multiple channels. If forecasted debt-financed AI infrastructure growth materialises, the potential financial stability consequences of such an event are likely to grow.
The impact of asset price bubbles on systemic risk depends crucially on which actors are exposed and are greater when vulnerabilities such as leverage and liquidity mismatch exist which can amplify shocks and impose externalities on the rest of the financial system (Adrian et al (2014); Aoki and Nikolov (2012); and Bank of England (2023)). In that context, the nature of the AI ‘boom’ as primarily an equity story up until this year has meant that a fall in AI-related asset prices would not necessarily lead to severe financial stability consequences.
However, past episodes have demonstrated that hidden leverage can exist within the financial system. For example, the collapse of Archegos Capital Management in March 2021 after failing to meet margin calls on their equity total return swap positions of several technology companies demonstrated how leveraged equity positions can result in risks to systemic financial institutions through prime brokerage exposures when prices fall (ESMA (2022)).
AI could also impact financial stability through commodity markets. On top of the previously mentioned power impact, every megawatt of AI data centre power capacity is estimated to require 20–40 tonnes of copper (JP Morgan Research). Shocks to commodity prices can have spillover consequences for systemic institutions, as was seen in 2022 where large margin calls in LME Nickel futures markets and the subsequent threat of large defaults forced a suspension of trading by the CCP LME Clear (Heilbron (2024)). The potential for AI to drive similar dynamics was raised this year in a blog by a staff member at the Banque de France (Brousse (2025)). If power acts as a bottleneck to the operation of AI data centre projects, it can also weigh on their credit risk (Tsui et al (2025)).
A fall in AI-related asset prices could also adversely impact US economic growth, for example, through a fall in business investment and a consumption response through wealth effects. AI investment has been an outsized driver of US GDP growth in the first half of 2025. The dot com bubble did contribute to a mild US recession, driven by falling business investment (Lansing (2003)). A fall in AI-related asset prices would happen in a significantly different macroeconomic context to the early 2000s.
If the projected scale of debt-financed AI and associated energy infrastructure investment materialises over this decade, financial stability risks are likely to grow. Banks would be exposed to this directly through their credit exposures to AI companies, as well as indirectly through their provision of loans and credit facilities to private credit funds and other financial institutions which are exposed to AI-impacted asset prices. The degree of risk, like all credit, would depend on its size and quality.
This is a fast-evolving topic, and the future is highly uncertain. Bank of England staff will continue to monitor the financial stability risks from AI as they evolve.
This post was prepared with the help of Owen Lock and Andrew Walters.
Share your thoughts with us at BankOverground@bankofengland.co.uk
Defined as those stocks which appear in JP Morgan’s JPAMAIDE Equity Basket, which is comprised of 30 S&P 500 stocks which are particularly impacted by AI.
Inference refers to the use of the AI systems.
Hyperscalers include Microsoft, Alphabet, Meta and Oracle; AI Chipmakers include Nvidia, AMD, Broadcom and Micron; AI Application Providers include Tesla, Uber and Salesforce; Utilities include Constellation Energy and Vistra; Uranium Producers include Cameco, Uranium Energy and Denison; the Dow Jones Industrial Average (DJIA) is selected as a counterfactual due to its comparatively limited AI exposure.
