A third of the global AI supply chain is ‘invisible’ to governments, study finds
Researchers identify 137 supply chain elements for global AI production. They compare these to government documents from dozens of countries and find 33% of the elements, mainly from the “physical middle” of AI production (such as wafers, packaging and networking), are largely absent from top-level AI policy worldwide.
“AI sovereignty” is an impossibility for any one nation, including the US and China, and countries such as Japan and the UK have far more influence over the supply of AI than they realise, say experts. A country-by-country report is available alongside the main report.
As much as a third of the AI supply chain receives no attention from national and international policymaking and regulation, according to a new Cambridge report that reveals a vast “zone of invisibility” in the middle of global AI production.
A University of Cambridge team found vulnerabilities across 14 supply-chain layers, with 137 elements produced or convened in over fifty countries. Many critical “chokepoints” sit outside the AI powerhouses of China and the US.
“The US and China are central to the AI story, but this research shows us the extent to which critical technologies and materials across the AI stack have dependencies that extend far beyond their borders,” said Verity Harding, Director of the AI and Geopolitics project. “It shows that no one country can control the AI supply chain alone.”
One Japanese firm has a 95% market share of a material needed for AI chips, the report shows. Even notorious chokepoints depend on other nations: the Dutch firm dominating EUV lithography (ASML) relies on essential parts controlled by German and US companies.
In fact, the study suggests that countries such as Japan, Germany, and some so-called middle powers – from Belgium to South Korea – have far more leverage over global AI production than even their own governments realise.
Researchers describe an “unexpected mismatch” between a nation’s top AI policy documents and its actual positions in the supply chain, calling it a “universal pattern”.
“Most policy discussions reduce artificial intelligence to just a few inputs, such as compute, data and models. This is not an accurate understanding of the complexity of the AI supply chain,” said lead author Dr Siyi Liu.
“Many governments are unaware of the strengths and vulnerabilities they hold at the heart of the global AI supply chain. Understanding this will be the basis of credible bargaining power in AI governance.” she added.
The UK is involved in 41 of the 137 supply-chain elements researchers identify, yet its AI policies name just three, ignoring UK specialisms in chip design and industrial gases. Japan is second only to the US in terms of supply-chain positions, yet its AI policy documents recognise only three.
Cambridge researchers analysed 117 AI-related policy documents from 43 countries and territories, the EU and OECD, and found a fixation on models and datacentres, as well as funding and governance, while the factories, chemicals, chip-design software and packaging technologies vital to the AI supply chain are completely overlooked.
Disruption to any one of these “upstream” components could trigger a domino effect that leads to AI supply shortages, potentially destabilising stock markets, investment and productivity around the world.
Experts from the AI & Geopolitics Project (AIxGEO) at Cambridge’s Bennett School of Public Policy say that total “AI sovereignty”, a nation autonomously operating its own AI supply chain – a goal many nations hint at pursuing – is “impossible”.
The research team created a toolkit for governments to audit their place in the chain. They say it will help policymakers understand where and when to diversify across AI element providers, who they depend on, or when to “exit” entirely from an outside supplier and build national capacity.
For the latest study, the AIxGEO team comprehensively mapped the global AI supply chain to reveal 14 distinct layers, from the raw minerals and the ultra-pure gases needed for semiconductors, to the wafers, chips and servers, and eventually the trained models. Cross-cutting layers include governance standards and shipping routes.
Breaking it down, researchers identified 137 specific “elements”, such as gallium or hyperscale cloud platforms, with a further 351 “sub-elements”: tech specs that determine who can supply it. Nearly 800 firms identified hold supply-chain positions across 56 countries and territories.
However, the team’s analysis of AI policy and regulation documents from dozens of countries as well as intergovernmental organisations shows that just over 33% of the identified supply-chain elements are simply not mentioned.
Most of this “invisibility” is in the “physical middle” of the chain. This includes chip design, packaging and the testing of chips, and some of the hardware and software behind “compute” such as making AI servers and “merchant accelerators” (platforms that allow businesses to adopt AI).
An example of a hidden chokepoint highlighted by the report is the EDA (Electronic Design Automation) software used to design AI chips. It is essential for companies like NVIDIA, AMD and Google.
This software is dominated by three firms – Synopsys, Cadence and Siemens EDA in the US – yet receives almost no policy attention, unlike the chips and LLMs that wouldn’t exist without it.
Another major “blind spot” is advanced packaging: connecting multiple semiconductor chips into high-performance AI processors essential for frontier models. This packaging is overwhelmingly concentrated in Asian countries and territories such as Taiwan, Vietnam and the Philippines.
“Policymakers focus on finished chips rather than the ecosystem behind them. They focus on GPUs but ignore the packaging operations needed to turn those chips into usable AI hardware,” said Liu.
Japan emerges as the world’s most underrated AI power, with the second largest number of supply chain positions, directly ahead of China. The report shows 101 Japanese firms hold positions across 69 elements, from silicon wafers to semiconductor testing. However, Japan’s own policy documents feature only three of these elements.
The researchers point out that a Japanese company, Ajinomoto, controls more than 95% of the global supply of build-up film, an insulating material needed for advanced AI chip manufacturing.
“Japan’s strength comes from equipment, and without it there's no way you can produce the hardware,” said Liu. “The supply chain is sequential, so the trickle down could affect computing capacities of the top frontier models.”
The UK has little AI-related manufacturing, but holds leverage in chip design IP through Cambridge company Arm (even with majority Japanese SoftBank ownership). It is a base for research and expertise, particularly in AI security, and home to valuable health and genomic datasets.
The report highlights absences from reviewed UK policy where it has footholds in the supply chain, such as chip design, and the pure gases needed for semiconductor plants via Linde, one of the world's largest industrial gas companies.
Some of the smaller nations with more clout than they might realise include Belgium, which is strong in recycled semiconductor chemicals and has the major nanoelectronics R&D centre Interuniversity Microelectronics Centre (imec).
Ireland has no production supply-chain position, yet it can affect AI capacity through energy and planning as it provides a European base for big US tech e.g. Google and AMD, making it a “host and infrastructure state”. While Qatar produces much of the world’s helium, a vital supply-chain raw material, and is also a huge potential AI investor.
Earlier this year, Nvidia’s former-CEO Jensen Huang described the AI supply chain as a “five-layer cake” (energy, chips, infrastructure, models, and applications).
Verity Harding added: “The AI supply chain is far more complex, and power over it far more diffused, than some simplified narratives might have us believe. Many governments hold more cards than they realise.”
Notes:
The 14 supply chain layers identified by the report: minerals that must be scooped and refined; ultra-pure gases and chemicals required by a fabrication plant or facility; wafers; machines that print the circuits; chips; memory and packaging around chips; servers; data centres; power and cooling that run data centres; cloud platforms; data; models.
About the AI & Geopolitics project (AIxGEO)
AIxGEO, led by Verity Harding, exists to interrogate the framing of AI development as a zero-sum arms race between two superpowers, and to shift the policy debate towards a more cooperative future so the technology can deliver the broad benefits it promises.
About the Bennett School of Public Policy
The Bennett School is a leading school of public policy in the UK and wider world, for public policy thinking, research, teaching and engagement with practitioners from policy, business and civic organisations. Drawing on Cambridge’s broad talent base, it supports effective decision-making through interdisciplinary research, training, and collaboration.
Article Title
The invisible AI supply chain
Article Publication Date
18-Nov-2026
The internal metaphysics of an LLM
Pulling on an AI model reveals how it represents physical relationships, and how we can train better AI models that understand physics from first principles
image:
A spring illustrates the mechanics-inspired idea: a system’s behavior is revealed by observing how it responds to a controlled change. In the language model, paired scientific questions produce changes in internal representations, depicted as interconnected networks. Targeted interventions then test how selected representations influence the model’s answer preferences. The figure is a conceptual illustration, not a visualization of measured data. Image credit: Markus J. Buehler/MIT.
view moreCredit: Markus J. Buehler/MIT
The challenge
Large language models are increasingly used to explore scientific knowledge, analyze materials and generate ideas. Yet a correct answer alone does not reveal how a model arrived at it. The answer might reflect a physical relationship, a familiar phrase or a statistical shortcut.
Physical laws describe relationships between quantities. An increase in one quantity can increase a material's response under one law and decrease it under another. Testing whether an AI model represents that distinction requires more than examining its response to a single question.
The Solution
Markus J. Buehler, the Jerry McAfee Professor of Engineering at MIT, investigated this problem using three open-weight models from Google's Gemma family. Inspired by mechanics, he compared matched questions in which a physical condition changed while the surrounding context was controlled. He then measured how the model's hidden states, the internal numerical representations it creates while processing a question, moved in response.
Across 60 materials-science laws, the resulting changes tracked the ordering of direct, neutral and inverse relationships with a rank correlation of 0.910. They correctly oriented 39 of 40 directional laws. When equations and prescribed reasoning instructions were removed, the models still oriented 38 of 40 laws correctly at the smallest tested scale and all 40 at the two larger scales.
The study also tested whether scientific concepts could be read from intermediate hidden states when their names were omitted from the prompt. Some concepts remained recoverable, while others did not. This established a distinction between information that can be read from a model and information that is used in a decision.
To test causal use, the author modified selected internal representations without retraining the model or adding verbal hints. In one controlled test, a fixed direction associated with grain size shifted answer preferences appropriately across 12 grain-refinement and grain-coarsening cases. Further interventions across 24 relationships showed selective causal contributions across all three model sizes.
The effects were context-dependent: the same grain-size intervention did not transfer successfully to every new wording or physical regime tested. The experiments therefore distinguish three levels of evidence: readability, representation and causal use.
The Future
The next step is representation-aware scientific AI: training and evaluating models not only on their answers, but also on whether their internal representations respond consistently to changes in physical conditions. This opens a path to testing models against scientific relationships and identifying where they rely on shortcuts.
The present work establishes the measurements and interventions for selected, known relationships in controlled tasks. Extending them to new scientific problems and model architectures is the next challenge, including applications in developing enhanced training rewards for AI models that foster broader and deeper generalizations of physics principles. This can lead to new first-principles based AI models that think and reason from foundational concepts and that may be far more efficient and powerful than current models.
The Impact
By treating an AI model's internal representations as objects that can be measured and experimentally altered, this research brings an approach inspired by the physical sciences to the study of AI itself. It gives researchers a more precise way to ask what a model represents, what it uses and where its scientific reasoning breaks down.
Publication: M.J. Buehler, "Reading and Steering Representations of Materials-Science Mechanisms in an Open-Weight Language Model," AI for Science 2, 045001 (2026). https://doi.org/10.1088/3050-287X/aeb27b
Video: https://youtu.be/XtSQ7IXDQl8
The internal metaphysics of an LLM [VIDEO]
Journal
AI for Science
Method of Research
Computational simulation/modeling
Article Title
Reading and steering representations of materials-science mechanisms in an open-weight language mode
Article Publication Date
15-Nov-2026
Artists have an intuitive sense for what is memorable
Why do some artworks stick in our minds, while others fade away soon after we view them? A new study published in Psychological Science suggests that it’s partly because the artists themselves can imbue memorability into their works.
Researchers at the University of Chicago tested this ability by conducting a nationwide art contest, asking artists across the country to create works that were intended to be either memorable or forgettable. They found that not only were artists successful in completing this task, but that a deep learning model was most adept at predicting memorability. These findings provide insights into how artists may intuitively understand memory and how technology can predict it.
“Our study suggests that artists can offer valuable insights into memory,” said Yifei Chen, a coauthor of the new study. “They seem to have an intuitive sense of what makes artworks actually stick in people’s minds.”
Chen and her colleagues collected art contest submissions over the course of three months for the study. They received a total of 86 submissions that adhered to the contest’s rules and first tested the memorability of the artworks with a group of online participants. Overall, the researchers found that the works submitted to the “memorable” category were more likely to be remembered, and that “forgettable” works tended to slip the mind. Analyses also showed that simpler artworks, as well as ones with fewer repetitive elements and less symmetry, had more memorability.
The team then chose 20 of both the most memorable and most forgettable artworks from these subsets. They asked the artists to ship their art to Chicago for a month-long exhibit in an in-person gallery. When the exhibition opened, visitors could freely move through the gallery, after which their memory of the artworks was tested. “The same pattern held in the gallery: Artworks that were more memorable online were also better remembered in person,” Chen said.
“It was a very risky experiment for our lab,” noted Wilma Bainbridge, senior author of the paper and Chen’s advisor, as the gallery only had 40 pieces and had an overarching theme of memory. The process of the experiment was also captured in a short documentary.
“But even still, they forgot the forgettable pieces and remembered the memorable ones, which shows how strong this effect really is,” she said. “It works on the order of just a few minutes, even in highly motivated people in this real-world setting.”
Finally, the researchers tested if memory of artworks could be predicted by artists, online laypeople without any art expertise, or a neural network model called ResMem that was developed by researchers to predict image memorability. In the new study, ResMem performed best at distinguishing the most and least memorable artworks, followed by artists’ submission categories, then ratings by laypeople.
“ResMem had only been trained on photographs,” Chen explained, making these new, original works a test of its ability. And yet, it did “much better than any human can do,” she said, highlighting that “what people think is memorable is different from what actually is.”
Although artists weren’t as proficient as AI in predicting memorability, their success in designing for memory suggests they could be valuable scientific collaborators. Studying their creative choices could help researchers understand which visual features support memory and how those features might be deliberately shaped. “This gives us so much potential for tapping into what drives memorability,” Chen said.
Refining models like ResMem could also help artists find tangible ways to make their works stand out or aid exhibit designers with how to arrange galleries. Bainbridge adds that ResMem could also be applied in a broader educational context, such as finding out what kinds of visuals may help children, patients, or other groups learn information that is difficult to remember. “Our preliminary findings are that if you have a memorable picture with all this information, they learn that additional information better, too,” Bainbridge said.
These findings contribute to a growing body of work examining how the properties of images and the real-world contexts they are encountered in can shape what we remember, Chen said. Because memory research has often focused on processes within the viewer, studying the relationships of visuals and memory in everyday settings is, as Chen says, “a direction that the whole field could explore further in the future.”
Reference
Davis, T. M., Chen, Y. E., & Bainbridge, W. A. (2026). Artists can design memory, and AI can predict it. Psychological Science, 37(11), 774–790.
Journal
Psychological Science
Article Title
Artists Can Design Memory, and AI Can Predict It

No comments:
Post a Comment