Cambridge Finds a "Blind Spot": 33% of the AI Supply Chain Is Not Mentioned in the Policies of 43 Countries
Governments write AI strategies about chips, data and models, but they forget what a chip cannot be made without. Cambridge broke AI production down into 14 layers and 137 elements and checked them against 117 documents from 43 countries. More than 33% of the elements, including chip design software and die packaging, are not mentioned anywhere. Companies with expensive AI infrastructure should assess risks more broadly than export controls on GPUs.
AI-processed from TechXplore AI; edited by Hamidun News
When governments talk about AI, they usually name three things: compute, data and models. Researchers at Cambridge decided to test how complete that picture is. They broke AI production into 14 layers (from raw materials to a trained model) and identified 137 specific elements within it, such as gallium or hyperscale cloud. They then checked this list against 117 official AI policy documents from 43 countries and territories, the EU and the OECD. It turned out that just over 33% of the elements, meaning a third of the chain, are not mentioned in any of the documents. This is the "zone of invisibility": links without which no data center works, but about which officials do not write.
How the Map Is Built
The AI and Geopolitics project (AIxGEO) at the Bennett School of Public Policy described the chain as a whole. It includes minerals and ultra-pure gases for semiconductors, silicon wafers, chips, servers and trained models. Standards for governance and sea routes run across all the layers.
Within the 137 elements, the researchers identified another 351 sub-elements: technical specifications that determine who can supply a given component at all. Nearly 800 companies from 56 countries and territories hold positions in the chain. The key bottlenecks are often not in the US or China.
For comparison, according to the document analysis, policymakers focus on models, data centers, funding and regulation. Factories, chemistry, chip design software and packaging technologies are practically absent from them.
Where Exactly the Blind Spot Lies
Most of the invisible elements sit in the physical middle of the chain: chip design, packaging and testing, along with some of the hardware and software around computing, including the assembly of AI servers.
The first example in the report is EDA, the electronic design automation software used to design microchips. Chips from Nvidia, AMD and Google rely on it, and so do all the models trained on those chips. Three American companies share the market: Synopsys, Cadence and Siemens EDA. Yet policymakers barely discuss this layer, even though they regulate the chips and LLMs themselves.
The second example is packaging: combining several dies into a single powerful AI processor. It is concentrated in Asia, primarily in Taiwan, Vietnam and the Philippines. According to lead author Xia Liu, policymakers look at GPUs and overlook the operations that turn dies into finished AI hardware.
Even the well-known bottlenecks depend on other countries. The Dutch company ASML, the leader in EUV lithography, relies on critical components from German and American firms.
Unexpectedly Strong Countries
The least obvious conclusion concerns Japan. By the number of positions in the chain, it ranks second after the US and ahead of China: 101 Japanese companies hold positions in 69 elements, from silicon wafers to semiconductor testing. Yet Japan's national documents mention only three of these elements. For example, Ajinomoto controls more than 95% of the world's supply of build-up film, the insulating film used to manufacture modern AI chips.
The UK participates in 41 of the 137 elements, but only three are named in its policy. The country produces almost no hardware, yet it influences the chain through Arm (chip design) and Linde (clean gases for semiconductor plants). Belgium is strong in chemical processing for chips and hosts the research center imec. Ireland produces nothing, yet it manages AI capacity through energy and planning: Google and AMD's European data center bases are there. Qatar supplies a significant share of the world's helium.
What Follows
The authors consider full "AI sovereignty", where a country controls the entire chain itself, impossible. But understanding one's own position gives weight in negotiations. For governments, the team has built an audit toolkit that helps decide where to diversify suppliers, and where to exit dependence on an external player and build domestic capacity.
For practitioners, the conclusion is straightforward. If your product runs on expensive GPUs, the risks are not only in export restrictions on chips. A failure in one upstream link, whether gas, film or packaging, can cascade into a shortage of compute and hit markets and investment. For companies with long-term contracts for AI infrastructure, it is worth asking suppliers where these components come from.
There are limitations. This is a map of dependencies and a comparison with document texts, not a forecast of specific disruptions. The absence of a mention in a strategy does not prove that the state is doing nothing: it may be working with industry outside public documents. But as a starting point for assessing supply risks, the report is useful.
- Nov 21, 2026Amazon Offers AWS Quick AI Agent with Claude for Free, Plus Tier at $25 a Month
- Nov 21, 2026GigaChat: 7th in the World on a Blind Russian-Language Ranking, but 78 Elo Points Behind the Leader
- Nov 21, 2026Anthropic ends first week on the stock market at a $1.6 trillion valuation, but falls 8.5% in a single day
- Nov 21, 2026Gemini 4 Hydrogen on xhigh Rises from 3.99 to 23.61 Cents per Request
Key Questions Answered
How the Map Is Built?
The AI and Geopolitics project (AIxGEO) at the Bennett School of Public Policy described the chain as a whole. It includes minerals and ultra-pure gases for semiconductors, silicon wafers, chips, servers and trained models. Standards for governance and sea routes run across all the layers. Within the 137 elements, the researchers identified another 351 sub-elements: technical specifications that determine who can supply a given component at all.
Where Exactly the Blind Spot Lies?
Most of the invisible elements sit in the physical middle of the chain: chip design, packaging and testing, along with some of the hardware and software around computing, including the assembly of AI servers. The first example in the report is EDA, the electronic design automation software used to design microchips. Chips from Nvidia, AMD and Google rely on it, and so do all the models trained on those chips. Three American companies share the market: Synopsys, Cadence and Siemens EDA.
What are the key details on Unexpectedly Strong Countries?
The least obvious conclusion concerns Japan. By the number of positions in the chain, it ranks second after the US and ahead of China: 101 Japanese companies hold positions in 69 elements, from silicon wafers to semiconductor testing. Yet Japan's national documents mention only three of these elements. For example, Ajinomoto controls more than 95% of the world's supply of build-up film, the insulating film used to manufacture modern AI chips.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.