Beneath
the gold rush
the gold rush
The uneven economy of AI, from labour
and global power to energy and resources
and global power to energy and resources
Renée Friedman, PhD
Global Head of Research
EXANTE
Global Head of Research
EXANTE
Beneath
the gold rush
the gold rush
The uneven economy of AI, from labour
and global power to energy and resources
and global power to energy and resources
Renée Friedman, PhD
Global Head of Research
EXANTE
Global Head of Research
EXANTE
AI is the catalyst, not the conquering hero.
It does not remove demographic, geopolitical and environmental constraints. It relocates and intensifies them: productivity depends on skills, sovereignty depends on interdependence and decarbonisation depends on physical infrastructure.
“In this report, we examine how AI is reshaping each theme, how they interact and what their convergence may mean for the global economy and investors.”
It does not remove demographic, geopolitical and environmental constraints. It relocates and intensifies them: productivity depends on skills, sovereignty depends on interdependence and decarbonisation depends on physical infrastructure.
“In this report, we examine how AI is reshaping each theme, how they interact and what their convergence may mean for the global economy and investors.”
Renée Friedman, PhD
Global Head of Research, EXANTE
AI is the catalyst, not the conquering hero.
It does not remove demographic, geopolitical and environmental constraints.
It relocates and intensifies them: productivity depends on skills, sovereignty depends on interdependence and decarbonisation depends on physical infrastructure.
“In this report, we examine how AI is reshaping each theme, how they interact and what their convergence may mean for the global economy and investors.”
It does not remove demographic, geopolitical and environmental constraints.
It relocates and intensifies them: productivity depends on skills, sovereignty depends on interdependence and decarbonisation depends on physical infrastructure.
“In this report, we examine how AI is reshaping each theme, how they interact and what their convergence may mean for the global economy and investors.”
Renée Friedman, PhD
Global Head of Research, EXANTE
Beyond the investment story
When most people think about AI, they think first about its technology: software, semiconductor chips, robotics and automation. Yet, as we noted in our Alpha Vibes report, The Architecture of Disruption, it is also a capital-intensive story, reshaping bond markets, power grids and equity leadership across the cycle.
Hyperscalers such as Meta, Alphabet, Microsoft, Amazon, Alibaba and Oracle are investing heavily in specialised chips, data centres and power infrastructure to capture what they see as a once-in-a-generation shift in the global economy. Google, Meta, Microsoft and Amazon alone are on track to invest more than $725bn in AI infrastructure in 2026. Goldman Sachs forecasts $1tn in AI-related investment globally in 2026, including $581bn in the US.
Hyperscalers such as Meta, Alphabet, Microsoft, Amazon, Alibaba and Oracle are investing heavily in specialised chips, data centres and power infrastructure to capture what they see as a once-in-a-generation shift in the global economy. Google, Meta, Microsoft and Amazon alone are on track to invest more than $725bn in AI infrastructure in 2026. Goldman Sachs forecasts $1tn in AI-related investment globally in 2026, including $581bn in the US.
$1tn
forecast global AI investment in 2026
Source: Goldman Sachs
$581bn
forecast US AI investment in 2026
[empty]
[empty]
[empty]
[empty]
The significance of this spending extends well beyond the technology sector. As S&P notes, it is becoming one of the defining forces shaping
the global economic outlook, influencing manufacturing activity, trade flows, productivity expectations and labour-market planning.
AI also brings risks beyond the investment case. OpenAI’s Sam Altman
and SpaceX’s Elon Musk have supported a call by Anthropic’s Dario Amodei for a slowdown in the pace of AI development. He cites the loss of control over AI systems, their misuse in cyberattacks or bioterrorism and serious economic disruption.
The surge in AI-related corporate bond issuance also poses risks to bond and credit markets. Higher borrowing costs can filter through to individuals via inflation, tougher mortgage markets, higher credit-card rates and potentially weaker pension portfolios.
The central question for investors, therefore, is not simply whether AI spending will generate returns above the cost of capital. It is whether AI can benefit all levels of society and improve human existence without causing irreversible environmental damage, weakening competition
and privacy or creating greater inequality.
the global economic outlook, influencing manufacturing activity, trade flows, productivity expectations and labour-market planning.
AI also brings risks beyond the investment case. OpenAI’s Sam Altman
and SpaceX’s Elon Musk have supported a call by Anthropic’s Dario Amodei for a slowdown in the pace of AI development. He cites the loss of control over AI systems, their misuse in cyberattacks or bioterrorism and serious economic disruption.
The surge in AI-related corporate bond issuance also poses risks to bond and credit markets. Higher borrowing costs can filter through to individuals via inflation, tougher mortgage markets, higher credit-card rates and potentially weaker pension portfolios.
The central question for investors, therefore, is not simply whether AI spending will generate returns above the cost of capital. It is whether AI can benefit all levels of society and improve human existence without causing irreversible environmental damage, weakening competition
and privacy or creating greater inequality.
The main question is whether AI can benefit all levels of society and improve human existence without causing irreversible environmental damage, weakening competition and privacy or creating greater inequality
[empty]
Three frontiers, one investment story
The labour
frontier
The land
grab
Prospecting
for gold
The labour frontier
Work, ageing and healthcare
In brief
- AI could offset the economic effects of ageing populations by improving productivity, expanding labour capacity and supporting overstretched healthcare systems.
- The evidence currently points more towards jobs being transformed than eliminated, although younger workers and routine entry-level roles face greater disruption.
- The gains will be uneven. Countries with strong digital infrastructure, skilled workforces and adaptable institutions are best placed to benefit.
Introduction
The evidence presented points in two directions at once, and the gap between them is where the investment questions sit. Investors worry about the impact AI will have on workers, productivity and company costs, as well as how deeply it will become integrated in day-to-day life. Will it really make the world more efficient? What about all the people it may replace? Could it destroy the economy?
History suggests that major technological advances have generally produced positive economic and social effects. One mechanism is the Jevons Paradox: when technology makes a resource cheaper or easier to use, demand for it tends to increase. Torsten Slok argues that a similar Jevons Employment Effect could shape AI’s impact on work. If AI makes workers more efficient and reduces the cost of completing tasks, companies may respond by expanding output. The resulting growth could increase, rather than reduce, overall headcount.
That outcome is not guaranteed. Regulation or other bottlenecks may prevent businesses from expanding enough to create new jobs. Where AI can replace human labour or intellectual input almost entirely, demand for workers may not recover at all. Excessive automation could therefore widen inequality and depress wages, particularly for people at the lower end of the educational and skills spectrum.
What is observable now is a softening entry-level jobs market and a wage premium for AI skills. What remains speculative is whether AI will deliver the productivity gains needed to justify the capital being spent. OECD modelling puts the potential gain at 0.4 to 1.3 percentage points a year, but that range assumes reskilling, organisational change and institutions that adapt. None of those is a given, and ageing societies tend to be weaker on all three.
The binding constraint, then, may not be the technology itself, but the human and institutional capacity to absorb it. Regulation may be the single biggest factor accelerating or constraining its spread. Meanwhile, the costs of the transition are already falling on identifiable groups: graduates entering the workforce, people in routine white-collar roles and low-income economies where diffusion has barely begun. The benefits are more dispersed and remain further out. That leaves two questions running through the paper: what would need to appear in the data before the productivity case is proven, and what would count as evidence that it has failed?
History suggests that major technological advances have generally produced positive economic and social effects. One mechanism is the Jevons Paradox: when technology makes a resource cheaper or easier to use, demand for it tends to increase. Torsten Slok argues that a similar Jevons Employment Effect could shape AI’s impact on work. If AI makes workers more efficient and reduces the cost of completing tasks, companies may respond by expanding output. The resulting growth could increase, rather than reduce, overall headcount.
That outcome is not guaranteed. Regulation or other bottlenecks may prevent businesses from expanding enough to create new jobs. Where AI can replace human labour or intellectual input almost entirely, demand for workers may not recover at all. Excessive automation could therefore widen inequality and depress wages, particularly for people at the lower end of the educational and skills spectrum.
What is observable now is a softening entry-level jobs market and a wage premium for AI skills. What remains speculative is whether AI will deliver the productivity gains needed to justify the capital being spent. OECD modelling puts the potential gain at 0.4 to 1.3 percentage points a year, but that range assumes reskilling, organisational change and institutions that adapt. None of those is a given, and ageing societies tend to be weaker on all three.
The binding constraint, then, may not be the technology itself, but the human and institutional capacity to absorb it. Regulation may be the single biggest factor accelerating or constraining its spread. Meanwhile, the costs of the transition are already falling on identifiable groups: graduates entering the workforce, people in routine white-collar roles and low-income economies where diffusion has barely begun. The benefits are more dispersed and remain further out. That leaves two questions running through the paper: what would need to appear in the data before the productivity case is proven, and what would count as evidence that it has failed?
AI’s productivity case depends on reskilling and changes to how work is organised. The proof will be gains in output and wages beyond the technology sector
[empty]
Demographic pressure and productivity
Across OECD economies, population ageing is tightening labour supply and increasing the need for productivity growth. The OECD reports
that the old-age dependency ratio rose from 19% in 1980 to 31%
in 2023 and is projected to reach 52% by 2060. Without compensating productivity gains, fiscal pressure will rise as labour markets tighten.
The demographic challenge looks very different in developing economies. The World Bank estimates that about 1.2 billion young people will reach working age between 2025 and 2035. Whether this becomes a demographic dividend or persistent underemployment will depend on job creation and access to global supply chains.
Slower population growth does not necessarily mean lower output.
An NBER working paper found no negative effect on aggregate GDP
or earnings over the past seven decades. Its authors argue that technology can compensate for a shortage of younger workers by increasing productivity gains.
Professor Andrew Scott claims that economies have repeatedly adapted through technology, institutional change and investment in human capital However, he also warns that rapid demographic change may not be completely offset by technological adaptation.
OECD modelling suggests AI could add 0.4–1.3 percentage points to annual labour productivity growth over the next decade in high-exposure economies. The gains are expected to be strongest in economies specialising in AI-exposed, knowledge-intensive sectors such as finance and ICT, and where adoption is more widespread.
Even so, the gains will not be evenly distributed. Governments will need
to capture the benefits of AI while protecting employment and maintaining social stability.
The result is a sharply uneven starting point: advanced economies are ageing, emerging markets have younger workforces and patterns of urbanisation are changing. This forms the backdrop against which AI will redefine labour markets, social patterns, and even healthcare delivery.
that the old-age dependency ratio rose from 19% in 1980 to 31%
in 2023 and is projected to reach 52% by 2060. Without compensating productivity gains, fiscal pressure will rise as labour markets tighten.
The demographic challenge looks very different in developing economies. The World Bank estimates that about 1.2 billion young people will reach working age between 2025 and 2035. Whether this becomes a demographic dividend or persistent underemployment will depend on job creation and access to global supply chains.
Slower population growth does not necessarily mean lower output.
An NBER working paper found no negative effect on aggregate GDP
or earnings over the past seven decades. Its authors argue that technology can compensate for a shortage of younger workers by increasing productivity gains.
Professor Andrew Scott claims that economies have repeatedly adapted through technology, institutional change and investment in human capital However, he also warns that rapid demographic change may not be completely offset by technological adaptation.
OECD modelling suggests AI could add 0.4–1.3 percentage points to annual labour productivity growth over the next decade in high-exposure economies. The gains are expected to be strongest in economies specialising in AI-exposed, knowledge-intensive sectors such as finance and ICT, and where adoption is more widespread.
Even so, the gains will not be evenly distributed. Governments will need
to capture the benefits of AI while protecting employment and maintaining social stability.
The result is a sharply uneven starting point: advanced economies are ageing, emerging markets have younger workforces and patterns of urbanisation are changing. This forms the backdrop against which AI will redefine labour markets, social patterns, and even healthcare delivery.
AI’s impact on labour markets
AI could ease labour shortages and support growth, but outcomes will vary by country. Reaping the benefits will require reskilling, upskilling and labour-market reallocation. That transition may prove harder in economies where complex regulation constrains business dynamism, R&D investment is low or corporate culture is risk-averse. Ageing societies may face additional barriers because business dynamism and innovation can already be weaker.
Debate over whether and how AI will replace workers has intensified. Varsha Bansal notes “the tension between the stark prediction that artificial intelligence advancements would wipe out jobs en masse.” Anthropic CEO Dario Amodei warned in May 2025 that half of all entry-level white collar jobs would vanish, while OpenAI’s CEO, Sam Altman said that AI will end certain job categories. Companies are increasingly citing AI in layoffs.
Those warnings have not yet been matched by aggregate evidence. Widespread AI-driven job losses remain difficult to identify, even as business adoption rose from 33% in 2023 to nearly 80%, according to McKinsey. Public use is also high, with Sensor Tower estimating that ChatGPT has reached 1bn monthly active users.
It is not yet clear that productivity growth and job growth will go hand-in-hand. KPMG notes that copilots, automation, generative AI and increasingly autonomous agents are being deployed at pace. Goldman Sachs has found slower growth in job openings since the second half of 2022 in industries with greater exposure to automation, particularly in Germany, Australia and the US. Employment growth in information and communication services has also slowed across most major developed economies.
The question remains how generative AI driven automation will affect both labour supply and demand. The growing use of AI could lead to skill displacement, but, as suggested by the McKinsey Global Institute, there may be greater resiliency. Although their research estimates generative AI could automate up to 30% of hours currently worked in the US by 2030. Even so, highly exposed occupational categories, including STEM roles, could continue to add jobs over the same period.
The disruption will not be evenly distributed. People in the two lowest wage quintiles are up to 10 and 14 times more likely, respectively, than the highest earners to need to change occupations by the end of the decade.
The aggregate data is less clear-cut. Analysis by researchers at the Institute for Economic Policy Research (SIEPR) at Stanford University found that unemployment among the most AI-exposed fifth of workers rose by 0.77 percentage points since 2022, compared with 0.85 points among the least exposed. That points to a broadly softening labour market rather than widespread job losses caused by AI.
Younger workers appear to face the greatest pressure. PwC found that, across 2.4 million US entry-level jobs, roles with the highest AI exposure were seven times more likely to require traditionally senior, human-intensive skills such as leadership, creativity or face-to-face interaction. Openings for these ‘seniorised’ entry-level roles have grown by 35% since 2019, while other entry-level roles have shrunk by 10%.
Debate over whether and how AI will replace workers has intensified. Varsha Bansal notes “the tension between the stark prediction that artificial intelligence advancements would wipe out jobs en masse.” Anthropic CEO Dario Amodei warned in May 2025 that half of all entry-level white collar jobs would vanish, while OpenAI’s CEO, Sam Altman said that AI will end certain job categories. Companies are increasingly citing AI in layoffs.
Those warnings have not yet been matched by aggregate evidence. Widespread AI-driven job losses remain difficult to identify, even as business adoption rose from 33% in 2023 to nearly 80%, according to McKinsey. Public use is also high, with Sensor Tower estimating that ChatGPT has reached 1bn monthly active users.
It is not yet clear that productivity growth and job growth will go hand-in-hand. KPMG notes that copilots, automation, generative AI and increasingly autonomous agents are being deployed at pace. Goldman Sachs has found slower growth in job openings since the second half of 2022 in industries with greater exposure to automation, particularly in Germany, Australia and the US. Employment growth in information and communication services has also slowed across most major developed economies.
The question remains how generative AI driven automation will affect both labour supply and demand. The growing use of AI could lead to skill displacement, but, as suggested by the McKinsey Global Institute, there may be greater resiliency. Although their research estimates generative AI could automate up to 30% of hours currently worked in the US by 2030. Even so, highly exposed occupational categories, including STEM roles, could continue to add jobs over the same period.
The disruption will not be evenly distributed. People in the two lowest wage quintiles are up to 10 and 14 times more likely, respectively, than the highest earners to need to change occupations by the end of the decade.
The aggregate data is less clear-cut. Analysis by researchers at the Institute for Economic Policy Research (SIEPR) at Stanford University found that unemployment among the most AI-exposed fifth of workers rose by 0.77 percentage points since 2022, compared with 0.85 points among the least exposed. That points to a broadly softening labour market rather than widespread job losses caused by AI.
Younger workers appear to face the greatest pressure. PwC found that, across 2.4 million US entry-level jobs, roles with the highest AI exposure were seven times more likely to require traditionally senior, human-intensive skills such as leadership, creativity or face-to-face interaction. Openings for these ‘seniorised’ entry-level roles have grown by 35% since 2019, while other entry-level roles have shrunk by 10%.
Are workers right to be worried?
PwC and the World Economic Forum estimate that 37% of young workers globally are in occupations with medium to high exposure to AI-driven task change. The proportion reaches 75% in Eastern Asia, 69% in Northern America and 63% in Europe.
37%
of young workers globally are in occupations with medium to high
exposure to AI-driven task change
[empty]
[empty]
[empty]
[empty]
[empty]
[empty]
[empty]
Exposure is highest in financial services, information and communication, professional services, science and education. Agriculture, construction and food services are less affected. PwC also reports that jobs requiring AI skills are growing by 69%, compared with 9% for the wider jobs market, while the average wage premium for AI skills has reached 62%.
Given such statistics, of course workers, particularly young graduates, are worried about their ability to enter their chosen professions. Whether valid
or not, this worry has an impact on the economy. If people are nervous about jobs then you see the impact on consumer spending in the short term. However, over the medium term, when the data is clearer, it may show that job losses are not as severe or as widespread as may have been assumed.
Given such statistics, of course workers, particularly young graduates, are worried about their ability to enter their chosen professions. Whether valid
or not, this worry has an impact on the economy. If people are nervous about jobs then you see the impact on consumer spending in the short term. However, over the medium term, when the data is clearer, it may show that job losses are not as severe or as widespread as may have been assumed.
Gains from AI could be twice
as large in advanced economies
[empty]
Uneven gains across economies
World Bank research, using data from the Anthropic AI Usage Index, shows that generative AI tools are being adopted first by technologically-savvy ICT professionals, then gradually spreading to other occupations. While the diffusion process is well underway in high-income countries and beginning in middle-income countries, it has barely started in low-income contexts.
An IMF working paper projects that AI could increase global GDP by up to 4% over the next decade, but it won’t be evenly distributed. Gains could be twice as large in advanced economies as in developing ones. Countries with strong digital infrastructure, capable institutions and skilled workforces are likely to benefit most, while those with limited access or readiness could fall further behind.
AI will also reshape international specialisation. Automation may help older societies offset labour shortages. In younger markets, it could narrow traditional manufacturing-led development pathways before sufficient jobs have been created. The result may be a geopolitical realignment of labour, with countries’ areas of international specialisation changing as digital services and remote knowledge work are redistributed globally.
As AI embeds into robotics and the physical economy, governments – and potentially the largest hyperscalers – may be forced to develop policy that supports skill retraining for affected roles. This in turn will require a rethink in education and educational policies.
An IMF working paper projects that AI could increase global GDP by up to 4% over the next decade, but it won’t be evenly distributed. Gains could be twice as large in advanced economies as in developing ones. Countries with strong digital infrastructure, capable institutions and skilled workforces are likely to benefit most, while those with limited access or readiness could fall further behind.
AI will also reshape international specialisation. Automation may help older societies offset labour shortages. In younger markets, it could narrow traditional manufacturing-led development pathways before sufficient jobs have been created. The result may be a geopolitical realignment of labour, with countries’ areas of international specialisation changing as digital services and remote knowledge work are redistributed globally.
As AI embeds into robotics and the physical economy, governments – and potentially the largest hyperscalers – may be forced to develop policy that supports skill retraining for affected roles. This in turn will require a rethink in education and educational policies.
How workers perceive AI’s impact
A King’s College London study found that 69% of people are worried about the economic impact of job losses caused by AI and concern is consistent across all groups, with employers (63%) almost as worried as workers (69%). Fifty-seven per cent of the public expect widespread unemployment, while 17% think AI will create at least as many jobs as it removes; 48% of employers take the latter view.
Nevertheless, 55% of workers remain relatively unconcerned about being replaced by AI, although 19% of graduate workers are very worried. This may be because history shows that it takes time for new business models to be created for companies to finally reorganise around the new technology.
In the US, Goldman Sachs estimates that AI reduced monthly payroll growth by about 16,000 jobs over the past year and raised the unemployment rate by 0.1 percentage point.
An IFO survey found that German companies using AI expect the greatest pay pressure among less experienced workers. About half anticipate lower salaries for employees with under five years’ experience, compared with around 40% for more experienced staff. Depending on qualifications, one-third to one-half expect salaries to remain stable.
PwC describes a two-track jobs market. Jobs ‘professionalised’ by AI are growing twice as fast as those ‘democratised’ by it and have recorded 42% faster wage growth since 2021. PwC also finds that companies using AI to amplify human performance can unlock growth and enter new markets. Wages are rising faster at the most AI-exposed companies, suggesting that workers share some of the gains.
World Bank and IMF research points to a risk of technological exclusion as AI use remains concentrated in high-income countries. Without public and private investment in digital infrastructure and skills, low- and middle-income countries could become more marginalised. Improvements in readiness alone are unlikely to close the gap.
JPMorganChase researchers warn that widespread commercial deployment could displace jobs, fuel populism and strengthen calls for protectionism. Governments will have to balance adoption with employment and social stability.
However, regulators may have limited ability or even political willingness to assess adequately the benefits and risks of AI tools in the workplace. Fragmented compliance landscapes, regulatory disconnects and delays could limit the effectiveness of regulation. So could the higher costs of complex legal discovery involving audit logs and vendors’ trade-secret information.
These constraints could weaken attempts to mitigate bias and discrimination in hiring, protect privacy and ensure oversight of algorithmic dismissals or promotions. At the current rate of AI implementation, it is still unclear whether AI will cause mass job losses, a new jobs boom or both simultaneously.
Nevertheless, 55% of workers remain relatively unconcerned about being replaced by AI, although 19% of graduate workers are very worried. This may be because history shows that it takes time for new business models to be created for companies to finally reorganise around the new technology.
In the US, Goldman Sachs estimates that AI reduced monthly payroll growth by about 16,000 jobs over the past year and raised the unemployment rate by 0.1 percentage point.
An IFO survey found that German companies using AI expect the greatest pay pressure among less experienced workers. About half anticipate lower salaries for employees with under five years’ experience, compared with around 40% for more experienced staff. Depending on qualifications, one-third to one-half expect salaries to remain stable.
PwC describes a two-track jobs market. Jobs ‘professionalised’ by AI are growing twice as fast as those ‘democratised’ by it and have recorded 42% faster wage growth since 2021. PwC also finds that companies using AI to amplify human performance can unlock growth and enter new markets. Wages are rising faster at the most AI-exposed companies, suggesting that workers share some of the gains.
World Bank and IMF research points to a risk of technological exclusion as AI use remains concentrated in high-income countries. Without public and private investment in digital infrastructure and skills, low- and middle-income countries could become more marginalised. Improvements in readiness alone are unlikely to close the gap.
JPMorganChase researchers warn that widespread commercial deployment could displace jobs, fuel populism and strengthen calls for protectionism. Governments will have to balance adoption with employment and social stability.
However, regulators may have limited ability or even political willingness to assess adequately the benefits and risks of AI tools in the workplace. Fragmented compliance landscapes, regulatory disconnects and delays could limit the effectiveness of regulation. So could the higher costs of complex legal discovery involving audit logs and vendors’ trade-secret information.
These constraints could weaken attempts to mitigate bias and discrimination in hiring, protect privacy and ensure oversight of algorithmic dismissals or promotions. At the current rate of AI implementation, it is still unclear whether AI will cause mass job losses, a new jobs boom or both simultaneously.
It is still unclear whether AI will cause mass job losses, a new jobs boom,
or both simultaneously
[empty]
For businesses, governments and educators, the immediate task is to ensure that AI improves productivity, wages and hiring while preserving the value of empathy, judgement, creativity and leadership.
This question is ever more critical as we consider that in many large economies, the demographic challenges posed by aging populations are only expected to worsen. AI usage and adaptation may not only become critical for improving labour productivity, economic growth and ensuring fiscal sustainability, but it will also be used far more in keeping us alive.
This question is ever more critical as we consider that in many large economies, the demographic challenges posed by aging populations are only expected to worsen. AI usage and adaptation may not only become critical for improving labour productivity, economic growth and ensuring fiscal sustainability, but it will also be used far more in keeping us alive.
AI, demography and the transformation of healthcare
By 2050, nearly one in three EU citizens will be aged 65 or older, according to the European Commission’s third report on demographic transformation. The resulting labour shortages will add to pressure on healthcare systems. The Commission sees AI automating administrative work, supporting clinical decisions and expanding access to online consultations.
Economist Dambisa Moyo argues that AI could improve preventive diagnosis, speed up drug discovery and support more personalised care. It may also reduce the human and economic costs of cancer and neurodegenerative diseases, help predict the spread and mutation of communicable diseases and lower costs across healthcare. However, it also carries the risk that algorithmic biases could perpetuate health disparities and increase rates of misdiagnosis.
Economist Dambisa Moyo argues that AI could improve preventive diagnosis, speed up drug discovery and support more personalised care. It may also reduce the human and economic costs of cancer and neurodegenerative diseases, help predict the spread and mutation of communicable diseases and lower costs across healthcare. However, it also carries the risk that algorithmic biases could perpetuate health disparities and increase rates of misdiagnosis.
From consumer tools to regulation
Generative AI is reshaping consumer health care experiences. According to a 15-nation survey by BCG, 57% of consumers are already using AI for personal health, with younger patients, those in higher-income brackets and consumers in emerging markets leading the way. However, significant concerns remain. Privacy and reliability were cited by 62% and 59% of respondents respectively as their greatest concerns about using AI for health. These concerns show where regulators may need to step in.
57%
use AI for personal health
Source: BCG
62%
cite privacy concerns
[empty]
59%
cite reliability concerns
[empty]
The US Food and Drug Administration (FDA) is now exploring how generative AI-enabled healthcare including the use of AI-enabled medical devices should be regulated. The agency issued a discussion and feedback paper in August 2026 highlighting issues such as risk assessment, premarket evaluation, postmarket monitoring and the unique challenges created by generative AI systems that may produce variable outputs.
In the UK, the government has set up a National Commission into the Regulation of AI in Healthcare. It has tasked that commission with setting out a regulatory framework to address both the opportunities and the risks of AI in healthcare. That report has found that regulation must support safe, effective and trusted use throughout the lifecycle of a device, from development and deployment through to monitoring, updating and learning from real-world use. It suggested more continuous monitoring of these products, once deployed, to provide ongoing assurance across the lifecycle. In Europe, the European Commission is focusing on the reform of the EU pharmaceutical legislation and aligning this with the European Artificial Intelligence Act.
These are major signals. Generative AI is not just another layer of traditional medical software. A tool that drafts reports, summarises clinical information or generates recommendations can behave differently from a fixed algorithm. It may therefore require different evaluation methods, stronger monitoring after deployment and clearer rules governing when human review is required. For AI-assisted healthcare, the FDA discussion marks a shift from asking whether generative AI can be useful to asking what evidence and oversight should be required before it enters clinical workflows.
In the UK, the government has set up a National Commission into the Regulation of AI in Healthcare. It has tasked that commission with setting out a regulatory framework to address both the opportunities and the risks of AI in healthcare. That report has found that regulation must support safe, effective and trusted use throughout the lifecycle of a device, from development and deployment through to monitoring, updating and learning from real-world use. It suggested more continuous monitoring of these products, once deployed, to provide ongoing assurance across the lifecycle. In Europe, the European Commission is focusing on the reform of the EU pharmaceutical legislation and aligning this with the European Artificial Intelligence Act.
These are major signals. Generative AI is not just another layer of traditional medical software. A tool that drafts reports, summarises clinical information or generates recommendations can behave differently from a fixed algorithm. It may therefore require different evaluation methods, stronger monitoring after deployment and clearer rules governing when human review is required. For AI-assisted healthcare, the FDA discussion marks a shift from asking whether generative AI can be useful to asking what evidence and oversight should be required before it enters clinical workflows.
AI’s impact on the pharmaceutical sector
Ageing populations will increase demand across healthcare and pharmaceuticals. Drugmakers and regulators are likely to use AI in research and workflows, although accountable human judgement will remain essential.
AI is also changing how medical data is collected, interpreted and used. Generative and predictive models combined with biotechnology could reduce research and development costs and shorten clinical trials. The US and China are competing in this area as each links AI innovation to bio-industrial policy.
Larger health-data ecosystems could connect genomic, behavioural and environmental information, supporting population health management and longevity-focused industries.
AI is also changing how medical data is collected, interpreted and used. Generative and predictive models combined with biotechnology could reduce research and development costs and shorten clinical trials. The US and China are competing in this area as each links AI innovation to bio-industrial policy.
Larger health-data ecosystems could connect genomic, behavioural and environmental information, supporting population health management and longevity-focused industries.
It is still unclear whether AI will cause mass job losses, a new jobs boom,
or both simultaneously
[empty]
AI is likely to be used throughout drug discovery, candidate design and clinical development, as well as in documentation and regulatory review.
It can also optimise production cycles and assess supplier risk.
Research described in Nature offers one example of the potential. A machine-learning algorithm developed by researchers at Harvard Medical School, Dana-Farber Cancer Institute and Massachusetts General Hospital can predict the likelihood of 348 diseases using routinely collected electronic health records and information about genetic risk.
It can also optimise production cycles and assess supplier risk.
Research described in Nature offers one example of the potential. A machine-learning algorithm developed by researchers at Harvard Medical School, Dana-Farber Cancer Institute and Massachusetts General Hospital can predict the likelihood of 348 diseases using routinely collected electronic health records and information about genetic risk.
Pharma + AI: early adoption, uneven returns
Pharmaceutical companies are already applying AI across research, regulation and production. Amgen uses a generative biology platform, AstraZeneca applies AI to pathology analysis and regulatory submissions, Sanofi uses simulations and Pfizer uses AI in production. Merck applies it across drug discovery, clinical trials, internal workflows and manufacturing.
Share-price performance does not yet show a clear and consistent AI premium. Year to date, Merck, Amgen and AbbVie have outperformed both the S&P 500 and the S&P 500 Health Care index. Pfizer and Sanofi have beaten the healthcare index but trailed the broader market, while Roche, Eli Lilly, Novartis, GSK and AstraZeneca have lagged both.
AstraZeneca’s particularly weak performance has coincided with late-stage trial failures, legal investigations, patent concerns and merger uncertainty. Across the group, drug pipelines, patent cycles and demand for products such as GLP-1 drugs remain important performance drivers. AI may improve research productivity and manufacturing efficiency over time, but the market data does not yet prove the commercial case.
Share-price performance does not yet show a clear and consistent AI premium. Year to date, Merck, Amgen and AbbVie have outperformed both the S&P 500 and the S&P 500 Health Care index. Pfizer and Sanofi have beaten the healthcare index but trailed the broader market, while Roche, Eli Lilly, Novartis, GSK and AstraZeneca have lagged both.
AstraZeneca’s particularly weak performance has coincided with late-stage trial failures, legal investigations, patent concerns and merger uncertainty. Across the group, drug pipelines, patent cycles and demand for products such as GLP-1 drugs remain important performance drivers. AI may improve research productivity and manufacturing efficiency over time, but the market data does not yet prove the commercial case.
Performance of Pharma and the S&P 500 YTD
YTD to 11 September 2026
Source: FactSet and EXANTE research
Past performance is not a reliable indicator of future results
Past performance is not a reliable indicator of future results
Care systems and workforce capacity
Ageing societies such as Japan and South Korea are already using AI-enabled robots and therapeutic companions to lift patients, help prevent bedsores, change incontinence briefs and monitor sleep. Europe may also turn to AI-assisted diagnosis and robotics as its healthcare workforce shrinks.
The European Commission, in its study for the deployment of AI in healthcare, noted that twenty EU countries reported a shortage of doctors in 2022 and 2023, while fifteen countries reported a shortage of nurses. Based on minimum staffing thresholds for universal health coverage (UHC), EU countries had an estimated shortage of approximately 1.2 million doctors, nurses and midwives in 2022 alone.
Mamta Murthi and Sarvesh Suri at the World Bank argue that AI-powered clinical decision tools could deliver faster, more accurate diagnoses and reduce medical errors. Their use, however, requires specialist knowledge that may not yet be widely available. Health workers at every level will need training. Regulators and administrators must also manage the safety risks associated with using AI in clinical settings and ensure practitioners understand the potential harm to patients from AI hallucinations: confidently stated but factually incorrect outputs that could influence clinical decisions.
As these World Bank specialists have stressed, AI-enabled health systems depend on wider ecosystems, including laboratories, data scientists, medical device manufacturers, logistics providers, telehealth platforms and digital infrastructure specialists. As AI tools advance and adoption grows, workforce development must keep pace – creating new jobs in diagnostics, digital health, and manufacturing supply chains.
AI may help ageing societies offset labour shortages and ease pressure on healthcare, but the gains will depend on skills, infrastructure and access. Those dependencies make AI a geopolitical issue.
The European Commission, in its study for the deployment of AI in healthcare, noted that twenty EU countries reported a shortage of doctors in 2022 and 2023, while fifteen countries reported a shortage of nurses. Based on minimum staffing thresholds for universal health coverage (UHC), EU countries had an estimated shortage of approximately 1.2 million doctors, nurses and midwives in 2022 alone.
Mamta Murthi and Sarvesh Suri at the World Bank argue that AI-powered clinical decision tools could deliver faster, more accurate diagnoses and reduce medical errors. Their use, however, requires specialist knowledge that may not yet be widely available. Health workers at every level will need training. Regulators and administrators must also manage the safety risks associated with using AI in clinical settings and ensure practitioners understand the potential harm to patients from AI hallucinations: confidently stated but factually incorrect outputs that could influence clinical decisions.
As these World Bank specialists have stressed, AI-enabled health systems depend on wider ecosystems, including laboratories, data scientists, medical device manufacturers, logistics providers, telehealth platforms and digital infrastructure specialists. As AI tools advance and adoption grows, workforce development must keep pace – creating new jobs in diagnostics, digital health, and manufacturing supply chains.
AI may help ageing societies offset labour shortages and ease pressure on healthcare, but the gains will depend on skills, infrastructure and access. Those dependencies make AI a geopolitical issue.
The productivity dividend is not a given
The labour story is being told on two timescales, and investors are pricing one while living in the other. Sentiment surveys show workers expecting job losses and unrest. The hard data shows a jobs market cooling for reasons that AI does not yet explain. Consumer confidence responds to the first. Corporate earnings, over time, respond to the second. Anyone assessing consumer-facing sectors should treat the gap between perceived and measured disruption as a variable in its own right, not noise.
The productivity dividend is a claim to be tested, not a base case to be assumed. The test is simple: does the gain show up in national accounts and in wage growth outside the technology sector, or only in vendor case studies? Until it does, the strongest evidence of durable value sits with organisations that pair AI spend with visible changes to how they hire, train and structure work. Spend on its own is a cost line.
Healthcare presents a different test: the constraint has shifted from capability to evidence. The FDA discussion paper of August 2026 signals that generative tools face a longer and more demanding route into clinical use than earlier medical software. That lengthens timelines for revenue, not for research. A useful distinction is between applications that change discovery and manufacturing, where adoption is already measurable, and those that touch patient decisions, where the regulatory path is still being written.
The productivity dividend is a claim to be tested, not a base case to be assumed. The test is simple: does the gain show up in national accounts and in wage growth outside the technology sector, or only in vendor case studies? Until it does, the strongest evidence of durable value sits with organisations that pair AI spend with visible changes to how they hire, train and structure work. Spend on its own is a cost line.
Healthcare presents a different test: the constraint has shifted from capability to evidence. The FDA discussion paper of August 2026 signals that generative tools face a longer and more demanding route into clinical use than earlier medical software. That lengthens timelines for revenue, not for research. A useful distinction is between applications that change discovery and manufacturing, where adoption is already measurable, and those that touch patient decisions, where the regulatory path is still being written.
The land grab
Sovereignty, security and the AI stack
In brief
- The global AI economy is fragmenting into competing ecosystems, led by different approaches in the US, China, Europe and the Middle East.
- Complete technological self-sufficiency is unlikely. Countries will instead seek control over critical capabilities while forming selective international partnerships.
Introduction
Sovereignty is the word every government now uses. Here, we treat it as a claim to be tested rather than a policy to be admired.
Two things are already measurable. Capital is flowing into computing capacity, or compute, at rates with no recent precedent: Korinek and McKelvey put US AI compute spending growth above 140% a year in both 2024 and 2025. And the rules governing that capital are hardening, from US export controls to the EU definition of technological sovereignty. What remains unproven is whether control of one layer of the stack produces a durable national advantage, or simply raises the cost of every layer above it.
The binding constraints are physical. Chips, energy and about 40 critical minerals sit beneath every model, and their supply chains do not respect borders. That leaves middle-sized economies with a choice they did not ask for: pay to build capacity they may never fully use, or accept dependence on a supplier whose terms can change.
The cost of fragmentation lands first on companies operating across blocs, which face duplicated infrastructure and divergent standards. The benefit accrues to whoever becomes indispensable at a single chokepoint.
The most useful way to read this report is as a map of chokepoints rather than a list of national ambitions.
Two things are already measurable. Capital is flowing into computing capacity, or compute, at rates with no recent precedent: Korinek and McKelvey put US AI compute spending growth above 140% a year in both 2024 and 2025. And the rules governing that capital are hardening, from US export controls to the EU definition of technological sovereignty. What remains unproven is whether control of one layer of the stack produces a durable national advantage, or simply raises the cost of every layer above it.
The binding constraints are physical. Chips, energy and about 40 critical minerals sit beneath every model, and their supply chains do not respect borders. That leaves middle-sized economies with a choice they did not ask for: pay to build capacity they may never fully use, or accept dependence on a supplier whose terms can change.
The cost of fragmentation lands first on companies operating across blocs, which face duplicated infrastructure and divergent standards. The benefit accrues to whoever becomes indispensable at a single chokepoint.
The most useful way to read this report is as a map of chokepoints rather than a list of national ambitions.
The question is, which of these positions is currently being priced, and which is not?
[empty]
AI sovereignty and strategic power
AI is now a central axis of geopolitics and economic statecraft. The global system is fragmenting into AI blocs, driven by competing standards, divergent governance models and strategic resources. Governments are using trade barriers, export controls and restrictions on technology transfers to protect domestic capabilities and limit rivals. Advantage increasingly rests on access to compute, advanced semiconductors, cloud infrastructure, proprietary data and cutting-edge models. AI governance is increasingly shaped by national power and technological dominance, not simply commercial competition or collaborative science.
Sovereign AI is a new strategic priority
[empty]
AI hard power still rests on physical, infrastructure-heavy assets.
The technology may render some traditional boundaries less relevant,
but it turns energy grids and chip supply chains into economic and military battlegrounds. Because sovereignty is now tied to industrial policy, countries must identify where they are strong, where they are exposed
and which dependencies they can tolerate.
McKinsey defines “sovereign AI” as a nation’s or organisation’s ability to develop and control its own AI capabilities to ensure strategic independence and alignment with domestic values and laws.
McKinsey identifies four components:
The technology may render some traditional boundaries less relevant,
but it turns energy grids and chip supply chains into economic and military battlegrounds. Because sovereignty is now tied to industrial policy, countries must identify where they are strong, where they are exposed
and which dependencies they can tolerate.
McKinsey defines “sovereign AI” as a nation’s or organisation’s ability to develop and control its own AI capabilities to ensure strategic independence and alignment with domestic values and laws.
McKinsey identifies four components:
-
Territorialwhere data and compute physically reside - Operationalwho manages and secures data and compute
- Technologicalwho owns the underlying stack and intellectual property
- Legalwhich jurisdiction governs access and compliance
The current US administration’s support for AI data centres, chip development and compute capacity, alongside limits on access to US technology, reflects this push for sovereign AI. Legislation now before Congress would align semiconductor export controls with allies through the STRIDE Act, while the MATCH Act targets Chinese chipmakers. This is not solely a Republican agenda. There is bipartisan support for wider restrictions on AI technologies, including semiconductor equipment, and for extending the extraterritorial reach of US controls through allied governments.
>200%
annual growth in raw compute capacity
Source: PIIE
>2,000%
annual growth in quality-adjusted AI output
[empty]
[empty]
[empty]
[empty]
According to PIIE estimates, raw AI compute capacity in the US grew by more than 200% and quality-adjusted AI output by more than 2,000% in both 2024 and 2025. These growth rates reflect three compounding forces: expanding data-centre capacity, continued improvements in chip efficiency and rapid algorithmic progress.
The US has been accused of practising ‘weaponised interdependence’, a term coined by Henry Farrell from Johns Hopkins University and Abraham Newman from Georgetown University. It describes the use of control over key global infrastructure to punish countries seen as resisting American interests.
Cristina Vanberghen at Yerevan State University notes that US restrictions on frontier AI models are presented as a security measure, based on concerns that advanced systems could help actors conduct or intensify cyberattacks. But nationality-based restrictions may create strategic costs that outweigh their security benefits. The political and diplomatic costs of such restrictions may actually undermine historical security partnerships such as NATO and ASEAN, creating an even more fragmented global power structure.
Europe has developed its own concept of “technological sovereignty”, which the European Commission’s digital strategy defined as the capacity to develop, control and regulate critical digital and technological infrastructure within European jurisdiction. Whether the security case is proportionate remains contested.
Compute spending, meanwhile, has grown rapidly worldwide. The push for AI sovereignty is encouraging countries and companies to build their own capabilities, reinforcing demand for open-weight models and the infrastructure needed to train, host and deploy AI at scale.
China offers the clearest example of the drive to self-reliance. It has publicly pursued global AI leadership for about a decade while promoting a multipolar world order intended to counter US influence. JPMorgan Chase research notes that China has publicly proclaimed its ambition to lead the world in AI for about a decade. President Xi Jinping has called on China to “leverage national capabilities” and seize the initiative to “win superiority in the field of AI”. The government’s strategy is framed as a “new whole-of-nation” approach to building self-reliance.
The US has been accused of practising ‘weaponised interdependence’, a term coined by Henry Farrell from Johns Hopkins University and Abraham Newman from Georgetown University. It describes the use of control over key global infrastructure to punish countries seen as resisting American interests.
Cristina Vanberghen at Yerevan State University notes that US restrictions on frontier AI models are presented as a security measure, based on concerns that advanced systems could help actors conduct or intensify cyberattacks. But nationality-based restrictions may create strategic costs that outweigh their security benefits. The political and diplomatic costs of such restrictions may actually undermine historical security partnerships such as NATO and ASEAN, creating an even more fragmented global power structure.
Europe has developed its own concept of “technological sovereignty”, which the European Commission’s digital strategy defined as the capacity to develop, control and regulate critical digital and technological infrastructure within European jurisdiction. Whether the security case is proportionate remains contested.
Compute spending, meanwhile, has grown rapidly worldwide. The push for AI sovereignty is encouraging countries and companies to build their own capabilities, reinforcing demand for open-weight models and the infrastructure needed to train, host and deploy AI at scale.
China offers the clearest example of the drive to self-reliance. It has publicly pursued global AI leadership for about a decade while promoting a multipolar world order intended to counter US influence. JPMorgan Chase research notes that China has publicly proclaimed its ambition to lead the world in AI for about a decade. President Xi Jinping has called on China to “leverage national capabilities” and seize the initiative to “win superiority in the field of AI”. The government’s strategy is framed as a “new whole-of-nation” approach to building self-reliance.
How quickly countries and companies can build such capacity will shape future economic strength and geopolitical influence. But some countries may pursue access rather than self-reliance. Isabella Wilkinson, a senior research fellow at Chatham House in the UK, suggests they could offer public data, energy resources or preferential treatment to US or Chinese AI companies in return for cutting-edge capabilities or support in building domestic infrastructure.
For some countries, alignment with a single superpower or alliance is unacceptable because it threatens sovereignty; for others, it may bring security and domestic benefits.
For some countries, alignment with a single superpower or alliance is unacceptable because it threatens sovereignty; for others, it may bring security and domestic benefits.
From fragmented globalisation
to “tech-sovereign multipolarity”
to “tech-sovereign multipolarity”
According to the JPMorgan Chase Center for Geopolitics, the US and China dominate global AI, but along divergent institutional models: the US emphasises private-sector innovation and defence integration, whereas China advances state-led self-reliance and diffusion of low-cost AI systems abroad. Europe seeks tech sovereignty, balancing privacy, regulation and competitiveness, while the Middle East uses AI investment as a diversification strategy tied to sovereign wealth and energy surpluses.
It is highly unlikely that one state can fully control every layer of the AI stack. Yet a state may be able to exert outsized influence, regardless of its physical size, by becoming indispensable in specific parts of the AI ecosystem.
States are not the only actors with influence. As JPMorgan Chase notes, the private sector has a decisive role in scaling innovation, securing critical supply chains and ensuring that AI advances support the principles and prosperity of a rules-based market economy. Companies that understand this can protect their own businesses while strengthening the foundations of long-term national power.
Yet disagreement over regulatory standards is already eroding interoperability. The result is a fragmented global AI ecosystem in which cross-border data flows, compute sovereignty and chip supply chains define new power alignments.
As a result, control over semiconductor production, critical minerals and energy grids has become paramount. These assets are replacing traditional oil routes as chokepoints of strategic influence, echoing the energy geopolitics of the 20th century.
It is highly unlikely that one state can fully control every layer of the AI stack. Yet a state may be able to exert outsized influence, regardless of its physical size, by becoming indispensable in specific parts of the AI ecosystem.
States are not the only actors with influence. As JPMorgan Chase notes, the private sector has a decisive role in scaling innovation, securing critical supply chains and ensuring that AI advances support the principles and prosperity of a rules-based market economy. Companies that understand this can protect their own businesses while strengthening the foundations of long-term national power.
Yet disagreement over regulatory standards is already eroding interoperability. The result is a fragmented global AI ecosystem in which cross-border data flows, compute sovereignty and chip supply chains define new power alignments.
As a result, control over semiconductor production, critical minerals and energy grids has become paramount. These assets are replacing traditional oil routes as chokepoints of strategic influence, echoing the energy geopolitics of the 20th century.
The rise of “net states” and early adopters
Countries that master the AI stack, from data control and compute capacity to regulatory frameworks, will emerge as “net states” wielding influence disproportionate to their size. Smaller but digitally advanced nations such as South Korea, Japan, Singapore and the UAE exemplify this: each invests strategically across education, data sovereignty and public-private R&D ecosystems to maintain AI competitiveness.
Competition for influence may also explain China’s efforts to promote open-source AI platforms and data-sharing initiatives. JPMorgan Chase’s analysts argues that China aims to accelerate industry progress and circumvent potential export controls on proprietary technology.
Competition for influence may also explain China’s efforts to promote open-source AI platforms and data-sharing initiatives. JPMorgan Chase’s analysts argues that China aims to accelerate industry progress and circumvent potential export controls on proprietary technology.
Sovereignty should not mean isolation, but strategic agency through interdependence
[empty]
They also suggest that China’s strategy combines more than $100 billion in public funding for AI and semiconductors with regional incubation zones like Hangzhou’s “Dream Town”, designed to attract young start-ups. Researchers at the Tony Blair Institute for Global Change also point to the Digital Silk Road, which offers affordable, turnkey AI infrastructure that lowers adoption costs while deeply embedding Chinese technology and standards within national digital architectures.
The US, meanwhile, retains its lead in frontier innovation, driven by private capital, defence demand and an open research ecosystem. India and the ASEAN bloc are focused on how quickly they can adopt and deploy AI at scale.
AI sovereignty, however, need not mean isolation. Some researchers define it instead as strategic agency through interdependence. Countries must choose for themselves where to build strength, whether in talent, data or energy, and where to partner globally. Complete self-sufficiency is not realistic. Selective sovereignty offers a more practical goal: securing access to frontier models while building domestic capacity.
The US, meanwhile, retains its lead in frontier innovation, driven by private capital, defence demand and an open research ecosystem. India and the ASEAN bloc are focused on how quickly they can adopt and deploy AI at scale.
AI sovereignty, however, need not mean isolation. Some researchers define it instead as strategic agency through interdependence. Countries must choose for themselves where to build strength, whether in talent, data or energy, and where to partner globally. Complete self-sufficiency is not realistic. Selective sovereignty offers a more practical goal: securing access to frontier models while building domestic capacity.
Drones account for 70–80% of battlefield casualties in Ukraine
Source: The New York Times
AI, national defence and minerals
Control of the AI stack is only one dimension of national power. Countries must also confront its integration into modern warfare. AI defence technologies fall into two broad categories: kinetic instruments, such as missiles, drones and automated guns, and non-kinetic instruments, including electronic warfare, cyber operations and cognitive data fusion. Legacy defence contractors and software-first contractors are aggressively deploying AI across both domains.
The use of AI in drones and autonomous systems has reshaped how wars are being fought. In Ukraine, drones account for 70–80% of battlefield casualties. AI-powered systems are also playing a growing role in the Iran war, from accelerating mine clearing in the Strait of Hormuz to tracking submarines and gathering intelligence. These technologies are redefining the capabilities required for effective deterrence and defence.
Defence companies such as Teledyne, RTX, BAE Systems, L3Harris Technologies, Elbit Systems and Parsons Corporation already incorporate AI into a range of kinetic systems. Applications include precision targeting and missile guidance, swarming aerial munitions that coordinate in flight and computer-vision systems that detect incoming tank rounds or anti-tank missiles and launch explosive countermeasures.
Those capabilities also create significant risks. Anthropic PBC has reported that its AI model Claude has been misused in attempts to develop a wide range of military applications, including kamikaze drone swarms and missile navigation systems. It also said that there were five cases that involved users in nations that it prohibits from accessing its models, including Russia, China and Iran.
The use of AI in drones and autonomous systems has reshaped how wars are being fought. In Ukraine, drones account for 70–80% of battlefield casualties. AI-powered systems are also playing a growing role in the Iran war, from accelerating mine clearing in the Strait of Hormuz to tracking submarines and gathering intelligence. These technologies are redefining the capabilities required for effective deterrence and defence.
Defence companies such as Teledyne, RTX, BAE Systems, L3Harris Technologies, Elbit Systems and Parsons Corporation already incorporate AI into a range of kinetic systems. Applications include precision targeting and missile guidance, swarming aerial munitions that coordinate in flight and computer-vision systems that detect incoming tank rounds or anti-tank missiles and launch explosive countermeasures.
Those capabilities also create significant risks. Anthropic PBC has reported that its AI model Claude has been misused in attempts to develop a wide range of military applications, including kamikaze drone swarms and missile navigation systems. It also said that there were five cases that involved users in nations that it prohibits from accessing its models, including Russia, China and Iran.
Defence +AI: A changing paradigm of 21st century warfare
Defence expenditure is expected to grow as NATO countries in particular aim to meet the agreed spending targets.
Europe offers stronger budgetary growth, while the US provides greater technological depth. Both converge around sensors, communications, software, autonomy and precision effects.
Europe offers stronger budgetary growth, while the US provides greater technological depth. Both converge around sensors, communications, software, autonomy and precision effects.
Military expenditure by country as percentage of government spending
YTD to 11 September 2026
Source: FactSet and EXANTE research
Past performance is not a reliable indicator of future results
Past performance is not a reliable indicator of future results
Non-kinetic applications
Companies
ISR and Earth observation
Electronic warfare
Cybersecurity
Data fusion & decision support
C5ISR and battlefield networks
Navigation & mission systems
Space-based intelligence
Electronic warfare
Cybersecurity
Data fusion & decision support
C5ISR and battlefield networks
Navigation & mission systems
Space-based intelligence
Defence companies and other manufacturers are also developing non-kinetic AI weapon systems. These include AI-driven electronic jammers, electro-optical systems and battlefield management tools. By fusing signals and applying cyber techniques, such systems can blind enemy radar and communications, force down enemy drones or neutralise swarms without physical impact.
These technologies depend on critical rare-earth minerals and metals.
The materials are indispensable across key economic sectors, yet their supply chains are vulnerable to disruption. Shifting production networks and addressing vulnerabilities across the AI stack are therefore becoming more important to both economic and military security.
As Josefina Lehnen and Lorenzo Cotula at the IIED note, much of the debate over copper, graphite, lithium, rare earths and nickel has focussed on the energy transition. Defence and AI are now adding to aggregate demand, creating an emerging energy-defence-AI convergence. Modern weapons systems depend on about 40 critical minerals, making access to these materials increasingly a national security issue. Investors therefore need to consider the pace at which energy, defence and AI are converging when assessing the medium- to long-term implications for their portfolios.
These technologies depend on critical rare-earth minerals and metals.
The materials are indispensable across key economic sectors, yet their supply chains are vulnerable to disruption. Shifting production networks and addressing vulnerabilities across the AI stack are therefore becoming more important to both economic and military security.
As Josefina Lehnen and Lorenzo Cotula at the IIED note, much of the debate over copper, graphite, lithium, rare earths and nickel has focussed on the energy transition. Defence and AI are now adding to aggregate demand, creating an emerging energy-defence-AI convergence. Modern weapons systems depend on about 40 critical minerals, making access to these materials increasingly a national security issue. Investors therefore need to consider the pace at which energy, defence and AI are converging when assessing the medium- to long-term implications for their portfolios.
Beyond kinetic warfare
The potential for terrorist groups or individuals to use AI-generated sequences to create dangerous pathogens casts a shadow. Researchers at RAND note that these actors operate across multiple points in the AI-biology ecosystem, probing different models, querying several synthesis providers and sourcing materials from separate vendors. Individual actions may appear ambiguous in isolation but can form identifiable patterns when analysed collectively.
Research from the OECD shows that synthetic biology is increasingly integrating with AI tools. Their dual-use nature means that governments need to actively engage mitigation measures to prevent the introduction of AI enabled biological weapons. One response, argues the Bipartisan Commission on Biodefense, is to invest in AI-enabled disease surveillance and diagnostics, medical countermeasures, microbial forensics and attribution, model evaluation and safeguards, and adaptive nucleic-acid synthesis screening.
Research from the OECD shows that synthetic biology is increasingly integrating with AI tools. Their dual-use nature means that governments need to actively engage mitigation measures to prevent the introduction of AI enabled biological weapons. One response, argues the Bipartisan Commission on Biodefense, is to invest in AI-enabled disease surveillance and diagnostics, medical countermeasures, microbial forensics and attribution, model evaluation and safeguards, and adaptive nucleic-acid synthesis screening.
Future models trained on the genetic
code of dangerous pathogens could
be used to design more harmful viruses
[empty]
The risk is no longer wholly theoretical. Bacteriophages, viruses that infect bacteria and are used to treat persistent infections, have reportedly been created by Dr Brian Hie, a chemical engineer at Stanford University, using AI. Future models trained on the genetic code of dangerous pathogens could be used to design more harmful viruses.
Professor Tom Inglesby and Dr Mori Hanke at the Center for Health Security at Johns Hopkins University make the governance gap clear: “Although this is promising for life sciences applications, it also raises urgent biosafety and biosecurity questions. The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not.”
Professor Tom Inglesby and Dr Mori Hanke at the Center for Health Security at Johns Hopkins University make the governance gap clear: “Although this is promising for life sciences applications, it also raises urgent biosafety and biosecurity questions. The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not.”
The cyber threat
The most obvious defence threat from AI is from cybersecurity risk. Models with advanced cyber capabilities are sparking global cybersecurity concerns. In June 2026, US president Donald Trump signed an executive order on AI and cyber defence asking AI developers to voluntarily allow US government agencies to test their models for 30 days before release.
In May, the US and China said they would reopen intergovernmental dialogue on AI safety, paused since 2024.
Yet the same models could also be used to strengthen cyber defence.
The networks and systems on which digital societies depend contain persistent vulnerabilities. Models such as Mythos may be able to repair them autonomously and faster than human teams.
But they are not without risk. In the wrong hands and with sufficient resources, powerful AI could supercharge cyberattacks. One obvious target is the supply chain. Networks of manufacturers, logistics providers, software platforms and data services now depend on shared systems, APIs and cloud infrastructure. AI-powered decision engines have accelerated this integration by automating planning, procurement, forecasting and execution, increasing the potential attack surface.
Energy companies that power this digital infrastructure are also exposed to AI-enabled cyberattacks. The digitalisation and interconnection of networks for power generation, transmission and distribution have created a broad attack surface for hackers.
In May, the US and China said they would reopen intergovernmental dialogue on AI safety, paused since 2024.
Yet the same models could also be used to strengthen cyber defence.
The networks and systems on which digital societies depend contain persistent vulnerabilities. Models such as Mythos may be able to repair them autonomously and faster than human teams.
But they are not without risk. In the wrong hands and with sufficient resources, powerful AI could supercharge cyberattacks. One obvious target is the supply chain. Networks of manufacturers, logistics providers, software platforms and data services now depend on shared systems, APIs and cloud infrastructure. AI-powered decision engines have accelerated this integration by automating planning, procurement, forecasting and execution, increasing the potential attack surface.
Energy companies that power this digital infrastructure are also exposed to AI-enabled cyberattacks. The digitalisation and interconnection of networks for power generation, transmission and distribution have created a broad attack surface for hackers.
Sovereignty depends on interdependence
Full self-sufficiency is not available to any state, so the relevant question for a portfolio is not which country wins, but which inputs are scarce, who controls them and how quickly a substitute could appear. Export controls, minerals and grid capacity are policy variables as much as market ones. That makes the regulatory calendars of the US, the EU and China part of the risk assessment for any AI-exposed holding, not a footnote to it.
Fragmentation is a cost before it is an opportunity. Companies that operate across blocs face duplicated data centres, incompatible standards and screening rules that can change with an election. Investors assessing multinational technology and industrial holdings should ask where the second-largest market sits and what it would cost to serve it under a stricter regime. The answer is rarely disclosed and rarely priced.
Defence, biosecurity and cyber are where the thesis is least mature and the exposure most asymmetric. The demand signal is real. The governance is not yet written. Regulatory and reputational risks can reprice positions in these areas faster than changes in their fundamentals, which argues for treating them as scenario exposures rather than growth stories.
Fragmentation is a cost before it is an opportunity. Companies that operate across blocs face duplicated data centres, incompatible standards and screening rules that can change with an election. Investors assessing multinational technology and industrial holdings should ask where the second-largest market sits and what it would cost to serve it under a stricter regime. The answer is rarely disclosed and rarely priced.
Defence, biosecurity and cyber are where the thesis is least mature and the exposure most asymmetric. The demand signal is real. The governance is not yet written. Regulatory and reputational risks can reprice positions in these areas faster than changes in their fundamentals, which argues for treating them as scenario exposures rather than growth stories.
Prospecting for gold
Energy, minerals and the climate cost
In brief
- AI presents a sustainability paradox: it can improve efficiency and reduce emissions, but its infrastructure consumes substantial energy, water and raw materials.
- Its overall climate impact will depend on whether grid decarbonisation and economy-wide efficiency gains can outpace growing demand from data centres.
- AI is also improving climate modelling, energy management, resource discovery, precision agriculture and environmental monitoring.
Introduction
This report starts from a number rather than a promise. Allianz estimates that, without grid decarbonisation, data-centre emissions could more than double to 643 MtCO₂ by 2030. Against that sits the IEA estimate that AI applied across the economy could cut 1.4 Gt a year by 2035. The first figure projects infrastructure already under construction. The second projects applications most of which do not yet exist at scale. Those two numbers are not equally certain and should not be weighted equally.
The constraint that decides the outcome is the pace of power-sector decarbonisation, and it is not in the hands of the companies buying the compute. Hyperscaler nuclear agreements address the optics of the problem faster than they address the grid. Even on an ambitious pathway, the footprint moves rather than disappears. By 2030 roughly half of it would sit in the embodied emissions of servers and semiconductors, the part least visible to current ESG reporting.
The costs are local and immediate: water demand, grid connection capacity and the planning decisions of the communities hosting data centres. The benefits are global and deferred. That asymmetry is the political risk in the sustainability story, and this paper examines where it is most likely to surface first.
The constraint that decides the outcome is the pace of power-sector decarbonisation, and it is not in the hands of the companies buying the compute. Hyperscaler nuclear agreements address the optics of the problem faster than they address the grid. Even on an ambitious pathway, the footprint moves rather than disappears. By 2030 roughly half of it would sit in the embodied emissions of servers and semiconductors, the part least visible to current ESG reporting.
The costs are local and immediate: water demand, grid connection capacity and the planning decisions of the communities hosting data centres. The benefits are global and deferred. That asymmetry is the political risk in the sustainability story, and this paper examines where it is most likely to surface first.
AI’s sustainability paradox
AI’s relationship with sustainability is paradoxical. It promises radical efficiency but requires vast amounts of energy and compute capacity. The rapid expansion of data centres is increasing electricity demand worldwide, with data centre electricity consumption growing by around 12% per year since 2017. AI can improve energy efficiency, help integrate renewable power and optimise electricity networks. Yet according to one study, it is also increasing net annual carbon emissions.
Warnings about AI’s impact on grids and carbon emissions are not new. What has changed is the question investors are now asking: do the climate targets set by companies using AI still hold? Analysts at Morningstar Sustainalytics argue that the issue is especially acute for technology companies as data-centre electricity demand rises.
In the past, these companies could reasonably commit to net-zero pathways based on assumptions of steady business growth and continuing efficiency gains. The scale and pace of AI-driven expansion have smashed that assumption. Electricity demand from data centres risks outstripping improvements in efficiency and renewable-energy procurement, opening a gap between long-term net-zero commitments and short-term operational realities.
Investors may see this as a retreat from earlier commitments. Partly to reassure them and avoid a backlash over missed ESG targets, hyperscalers including Meta, Alphabet and Amazon have committed to cleaner energy sources, including nuclear power, to help run their data centres. Yet this does not remove the CO₂ emissions produced when increased electricity demand is met by existing grid energy sources.
Allianz models two sharply different outcomes. Without sufficient grid decarbonisation, it estimates that the annual climate damage associated with data centres could rise from $68bn today to $154bn by 2030.
AI workloads alone already account for around $13bn and could exceed $50bn by the end of the decade. Under a more ambitious decarbonisation pathway, emissions could be held to around 329 MtCO₂ and the associated damage limited to $79bn, despite continued growth in computing demand. The contrast suggests that the carbon intensity of the power supply, rather than improvements in computing efficiency alone, will determine AI’s near-term climate impact.
Yet, as Allianz puts it, “the footprint would not disappear; it would move up the supply chain”. Scope 2 emissions could fall from more than 70% of the total footprint today to around half by 2030, leaving emissions embedded in servers, semiconductors and other infrastructure to account for almost 50%. The focus would therefore move upstream, making hardware production and digital supply chains increasingly important to any credible assessment of AI’s environmental cost.
Warnings about AI’s impact on grids and carbon emissions are not new. What has changed is the question investors are now asking: do the climate targets set by companies using AI still hold? Analysts at Morningstar Sustainalytics argue that the issue is especially acute for technology companies as data-centre electricity demand rises.
In the past, these companies could reasonably commit to net-zero pathways based on assumptions of steady business growth and continuing efficiency gains. The scale and pace of AI-driven expansion have smashed that assumption. Electricity demand from data centres risks outstripping improvements in efficiency and renewable-energy procurement, opening a gap between long-term net-zero commitments and short-term operational realities.
Investors may see this as a retreat from earlier commitments. Partly to reassure them and avoid a backlash over missed ESG targets, hyperscalers including Meta, Alphabet and Amazon have committed to cleaner energy sources, including nuclear power, to help run their data centres. Yet this does not remove the CO₂ emissions produced when increased electricity demand is met by existing grid energy sources.
Allianz models two sharply different outcomes. Without sufficient grid decarbonisation, it estimates that the annual climate damage associated with data centres could rise from $68bn today to $154bn by 2030.
AI workloads alone already account for around $13bn and could exceed $50bn by the end of the decade. Under a more ambitious decarbonisation pathway, emissions could be held to around 329 MtCO₂ and the associated damage limited to $79bn, despite continued growth in computing demand. The contrast suggests that the carbon intensity of the power supply, rather than improvements in computing efficiency alone, will determine AI’s near-term climate impact.
Yet, as Allianz puts it, “the footprint would not disappear; it would move up the supply chain”. Scope 2 emissions could fall from more than 70% of the total footprint today to around half by 2030, leaving emissions embedded in servers, semiconductors and other infrastructure to account for almost 50%. The focus would therefore move upstream, making hardware production and digital supply chains increasingly important to any credible assessment of AI’s environmental cost.
Without grid decarbonisation, global data-centre emissions could more than double to 643 MtCO₂ by 2030
Source: Allianz
The water needed to cool server farms is already straining local ecosystems in Northern Virginia, Ireland, Singapore and parts of Europe, according to the Washington Post. Small wonder that the expansion of AI is increasingly bound up with the renewable-energy transition, critical-mineral supply and carbon management.
Applied across the economy, AI could cut global CO₂ emissions by around 1.4 Gt a year by 2035, more than offsetting the emissions generated by its own infrastructure and creating net savings of roughly 750 MtCO₂. The IEA attributes these reductions to efficiency gains, optimisation and better resource management across energy, industry, buildings and transport. They would be equivalent to around 2.6% of current global emissions.
But this outcome is not guaranteed. Most AI applications remain at an early stage, so the ultimate climate impact will depend on whether these economy-wide benefits can scale faster than the infrastructure needed to support them.
Applied across the economy, AI could cut global CO₂ emissions by around 1.4 Gt a year by 2035, more than offsetting the emissions generated by its own infrastructure and creating net savings of roughly 750 MtCO₂. The IEA attributes these reductions to efficiency gains, optimisation and better resource management across energy, industry, buildings and transport. They would be equivalent to around 2.6% of current global emissions.
But this outcome is not guaranteed. Most AI applications remain at an early stage, so the ultimate climate impact will depend on whether these economy-wide benefits can scale faster than the infrastructure needed to support them.
AI’s projected climate savings depend on applications that have yet to reach scale. Its infrastructure costs are arriving now
AI and mining
Mining is where every part of the AI supply chain becomes physical. The world cannot build solar panels, wind turbines, electric vehicles or data-centre hardware without large quantities of lithium, copper, nickel and cobalt. The energy transition was already raising demand for all four. AI adds a second source of demand through the servers, chips and grid infrastructure examined earlier in this paper.
The sector carries costs that rarely appear in a technology company's disclosures. Traditional extraction relies on fossil fuels and heavy machinery. Open-pit mining causes deforestation and biodiversity loss, and the process is water-intensive in regions where supply is often already scarce.
Child labour and hazardous working conditions persist across supply chains, operations displace Indigenous communities and the work carries high rates of workplace fatalities and chronic illness. Governance adds a further layer: who holds the rights to resources, how anti-corruption law is enforced and who is accountable for monitoring safety and emissions.
The sector carries costs that rarely appear in a technology company's disclosures. Traditional extraction relies on fossil fuels and heavy machinery. Open-pit mining causes deforestation and biodiversity loss, and the process is water-intensive in regions where supply is often already scarce.
Child labour and hazardous working conditions persist across supply chains, operations displace Indigenous communities and the work carries high rates of workplace fatalities and chronic illness. Governance adds a further layer: who holds the rights to resources, how anti-corruption law is enforced and who is accountable for monitoring safety and emissions.
Efficiency gains at the mine face do not reduce the volume the AI build-out requires. They make it cheaper to extract more
[empty]
AI is presented as part of the answer. AI-powered mapping and exploration tools could speed up discovery, reduce drilling time and modernise operations, lowering the footprint per tonne extracted. Some of these tools are already in use. But efficiency gains at the mine face do not reduce the volume the AI build-out requires. They make it cheaper to extract more.
The distinction matters for anyone assessing ESG exposure. A hyperscaler's reported footprint stops at its purchased electricity and its suppliers' declared emissions. The mining costs listed above sit further down the chain, in jurisdictions with weaker reporting and in commodities whose prices do not yet reflect them. The scope 3 emissions, i.e., all the indirect greenhouse gas emissions that occur in a company's wider value chain, including both upstream suppliers and downstream, are not being properly assessed. The greater environmental and social costs are paid by communities and workers far from the data centre.
The distinction matters for anyone assessing ESG exposure. A hyperscaler's reported footprint stops at its purchased electricity and its suppliers' declared emissions. The mining costs listed above sit further down the chain, in jurisdictions with weaker reporting and in commodities whose prices do not yet reflect them. The scope 3 emissions, i.e., all the indirect greenhouse gas emissions that occur in a company's wider value chain, including both upstream suppliers and downstream, are not being properly assessed. The greater environmental and social costs are paid by communities and workers far from the data centre.
AI, food security and sustainability
AI adoption in agriculture is accelerating. The "AI-agriculture nexus" links directly to climate and ESG objectives, turning sustainability into a growth engine. As noted by research from the Center for International and Strategic Studies, AI-assisted market forecasting tools and agricultural advisory tools can help farmers manage risks throughout a growing season, anticipate crop prices, and plan what crops to grow and when to sell them.
Global Market Insights projects that the AI agriculture market will grow from more than $4.7bn to $46.6bn by 2034, and the demand behind that figure is not speculative. The average US farmer was 58.1 years old in the 2022 Census of Agriculture. In the EU the average is 57 and only 12% of farmers are under 40, according to the European Commission. Smart tractors and sprayers that run 24 hours a day are an answer to a workforce that is ageing out.
Global Market Insights projects that the AI agriculture market will grow from more than $4.7bn to $46.6bn by 2034, and the demand behind that figure is not speculative. The average US farmer was 58.1 years old in the 2022 Census of Agriculture. In the EU the average is 57 and only 12% of farmers are under 40, according to the European Commission. Smart tractors and sprayers that run 24 hours a day are an answer to a workforce that is ageing out.
One compromised model or corrupted dataset has the reach of a weather event
[empty]
Each layer of automation removes labour and adds dependence. Syngenta chief executive Jeff Rowe describes platforms that combine GPS, machine-learning models and satellite imagery to guide applications of fertiliser or pest control to the nearest metre. The efficiency gains are real. So is the consequence: a farm run this way cannot easily revert to manual operation if the platform, the connection or the data behind it fails. Where a handful of platforms serves growers across a region, one compromised model or corrupted dataset has the reach of a weather event.
That risk is unevenly distributed. Large operators can negotiate terms and audit their suppliers. Smaller farms and the regions that depend on them must accept the terms they are given. Researchers at the University of Cambridge argue that concern about centralisation should be less about who owns the technology than about who is exposed when it fails.
The same tools can also strengthen food-system resilience. Researchers at the Center for Strategic & International Studies note that AI can improve early-warning systems by processing more data from more sources, more frequently. Yield prediction, for example, is a key component of sustainable farming. Precision agriculture reduces fertiliser use, optimises irrigation and improves resilience to climate shocks. Whether the resilience gained outweighs the fragility added depends on decisions about redundancy and supplier concentration that few food systems have yet made.
That risk is unevenly distributed. Large operators can negotiate terms and audit their suppliers. Smaller farms and the regions that depend on them must accept the terms they are given. Researchers at the University of Cambridge argue that concern about centralisation should be less about who owns the technology than about who is exposed when it fails.
The same tools can also strengthen food-system resilience. Researchers at the Center for Strategic & International Studies note that AI can improve early-warning systems by processing more data from more sources, more frequently. Yield prediction, for example, is a key component of sustainable farming. Precision agriculture reduces fertiliser use, optimises irrigation and improves resilience to climate shocks. Whether the resilience gained outweighs the fragility added depends on decisions about redundancy and supplier concentration that few food systems have yet made.
AI in the energy sector
As Arno van den Haak at the Atlantic Council notes, AI is already operating across the energy sector. At Saudi Arabia’s Khurais oil field, it processes data from more than 40,000 sensors in real time and adjusts production without a human in the loop. The same architecture is used on Shell’s deepwater platforms, in SLB’s autonomous drilling operations off Brazil and across ADNOC’s refineries in Abu Dhabi.
40,000+ sensors
feed AI systems that adjust production in real time at Saudi Arabia’s Khurais oil field,
without a human in the loop
[empty]
[empty]
[empty]
[empty]
[empty]
[empty]
[empty]
These systems combine three layers of AI:
- Generative AI interprets incoming data;
- Agentic AI builds a situational picture and decides how to respond;
- Edge AI executes that response at the wellhead, substation or compressor station.
The ability to integrate all three layers will help determine which companies lead in AI-enabled energy operations.
AI’s mixed impact for ESG investors
AI is also redefining corporate governance and regulatory expectations. ESG scoring, traditionally backward-looking, is becoming more forward-looking as providers apply AI analytics. EU regulators, aware of the biases AI can reproduce or amplify, are now requiring ESG rating providers to disclose their use of AI-based tools, supporting transparency, good governance and clear audit trails.
Dependence on AI in ESG reporting also creates risks. The law firm Norton Rose Fulbright notes that, while AI can accelerate reporting speed, it exposes compliance teams to governance risks involving algorithmic bias, data accuracy and auditability. Rules also lack uniformity because financial regulators are at different stages of developing their ESG frameworks.
Investors are now considering how AI changes the ESG case for major technology companies. LSEG has reported that their clients are asking providers of sustainable indices and ESG scores how AI’s impact on energy consumption and water usage is changing the profile of companies long viewed as ESG stalwarts. AI is therefore becoming both a tool for measuring ESG performance and a source of ESG risk.
Dependence on AI in ESG reporting also creates risks. The law firm Norton Rose Fulbright notes that, while AI can accelerate reporting speed, it exposes compliance teams to governance risks involving algorithmic bias, data accuracy and auditability. Rules also lack uniformity because financial regulators are at different stages of developing their ESG frameworks.
Investors are now considering how AI changes the ESG case for major technology companies. LSEG has reported that their clients are asking providers of sustainable indices and ESG scores how AI’s impact on energy consumption and water usage is changing the profile of companies long viewed as ESG stalwarts. AI is therefore becoming both a tool for measuring ESG performance and a source of ESG risk.
AI is also an infrastructure position
Energy access has become a competitive variable for anyone building or using AI at scale, and it is not evenly distributed. The pace of grid decarbonisation, not the ambition of corporate net-zero targets, determines whether the emissions gap closes. Investors reviewing those targets should ask what share of a company's compute growth is covered by new clean power capacity rather than existing grid supply, and over what period. Where the answer is unclear, the target is an aspiration, not a plan.
The footprint is moving up the supply chain, and reporting has not followed it. If embodied emissions from semiconductors and servers approach half the total by 2030, the material ESG exposure extends to hardware manufacturers and their suppliers, rather than sitting only with the operators of data centres. Current disclosure regimes were not built for that, so the risk is present in portfolios before it is visible in the data.
The same physical inputs connect the three papers in this series: minerals for defence and for AI hardware, power for compute and for electrification, water for cooling and for agriculture. Because these inputs serve competing uses, disruption or surging demand in one area can constrain capacity elsewhere. AI exposure therefore cannot be assessed as a technology allocation alone. It is also an energy, minerals and infrastructure position, and it should be examined as one.
The footprint is moving up the supply chain, and reporting has not followed it. If embodied emissions from semiconductors and servers approach half the total by 2030, the material ESG exposure extends to hardware manufacturers and their suppliers, rather than sitting only with the operators of data centres. Current disclosure regimes were not built for that, so the risk is present in portfolios before it is visible in the data.
The same physical inputs connect the three papers in this series: minerals for defence and for AI hardware, power for compute and for electrification, water for cooling and for agriculture. Because these inputs serve competing uses, disruption or surging demand in one area can constrain capacity elsewhere. AI exposure therefore cannot be assessed as a technology allocation alone. It is also an energy, minerals and infrastructure position, and it should be examined as one.
This article is provided to you for informational purposes only and should not be regarded as an offer or solicitation of an offer to buy or sell any investments or related services that may be referenced here. Trading financial instruments involves significant risk of loss and may not be suitable for all investors. Past performance is not a reliable indicator of future performance.
While every effort has been made to verify the accuracy of this information, EXT Ltd. (hereafter known as “EXANTE”) cannot accept any responsibility or liability for reliance by any person on this publication or any of the information, opinions, or conclusions contained in this publication. The findings and views expressed in this publication do not necessarily reflect the views of EXANTE. Any action taken upon the information contained in this publication is strictly at your own risk. EXANTE will not be liable for any loss or damage in connection with this publication.
Marketing content. Not investment research.
While every effort has been made to verify the accuracy of this information, EXT Ltd. (hereafter known as “EXANTE”) cannot accept any responsibility or liability for reliance by any person on this publication or any of the information, opinions, or conclusions contained in this publication. The findings and views expressed in this publication do not necessarily reflect the views of EXANTE. Any action taken upon the information contained in this publication is strictly at your own risk. EXANTE will not be liable for any loss or damage in connection with this publication.
Marketing content. Not investment research.