{"course":{"slug":"ai-actually-explained","title":"AI, Actually Explained","url":"https://addaly.com/learn/ai-actually-explained","level":"beginner","track":"start"},"lesson":{"slug":"the-stack-and-its-chokepoints","title":"Four layers and their chokepoints","url":"https://addaly.com/learn/ai-actually-explained/the-stack-and-its-chokepoints","position":83,"of":92,"minutes":9,"answer":"Every frontier model depends on one Dutch machine-maker, one Taiwanese fabricator, one American chip designer and a handful of clouds.","keyIdea":"Every frontier model depends on one Dutch machine-maker, one Taiwanese fabricator, one American chip designer and a handful of clouds, so control of AI is control of a few physical chokepoints that governments have already begun to use."},"body":{"format":"markdown","text":"## Follow the chip down\n\nThe history module told you a frontier model is something perhaps a dozen organisations can build. This lesson is about why the number is that small, and it comes from following one chip from the model back down to the sand.\n\n## Layer one: the machines that make the chips\n\nThe transistors on a modern AI chip are a few nanometres across, and drawing them requires a lithography machine that uses extreme ultraviolet light. One company in the Netherlands makes those machines. Nobody else does; nobody else has come close after two decades of trying. Each costs on the order of two hundred million euros, the newest generation nearer four hundred million, and the company ships a few dozen a year. That single firm is the narrowest point in the entire industry, and since 2019 its government has, under pressure from the United States, forbidden it to sell the machines to China.\n\n## Layer two: the factories\n\nOwning the machine is not enough; running a fabrication plant at the leading edge is a skill held by very few. One company in Taiwan makes roughly ninety per cent of the world's most advanced chips, including every AI accelerator that matters. A South Korean firm is a distant second. The United States, having lost this capability over thirty years, is spending tens of billions to rebuild a fraction of it, and India, which has none, is building its first plant at a technology level two decades behind the frontier. That most of the world's AI compute passes through one island is a fact that shapes foreign policy.\n\n## Layer three: the designs\n\nThe chips themselves are designed by a handful of firms and, for AI, overwhelmingly by one. The company from the graphics-chip lesson holds, by most estimates, four fifths or more of the market for AI accelerators, and its programming platform from 2007 is the reason: fifteen years of tools, libraries and trained engineers are built on it, and switching is expensive. Google designs its own chips and uses them internally; Amazon and Microsoft have followed; a US rival designs competitive hardware with a smaller software base. The high-bandwidth memory stacked beside every one of these chips comes from three firms, two of them Korean.\n\n## Layer four: the buildings and the labs\n\nThe chips are bought in tens of thousands and installed in data centres owned mostly by three American cloud companies and a few specialists, drawing power from grids the chip-makers did not plan for. On top of the buildings sit the labs: perhaps a dozen organisations, in the United States and China with one or two in Europe and the Gulf, that can pay for a frontier training run. Above those, the data (the web, the licensing deals, the human labour from the previous module) and above that, everything you use.\n\n> Figure: Four layers, and how few firms sit on each. Layers, top to bottom: Lithography machines (One firm in the Netherlands. Around 200 to 400 million euros each, a few dozen shipped a year. Barred from selling to China since 2019); Leading-edge fabrication (One firm in Taiwan makes roughly 90% of the most advanced chips. One South Korean firm is a distant second); Chip design (One company holds four fifths or more of the AI accelerator market, on a software platform it opened in 2007. Stacked memory comes from three firms); Data centres and labs (Three American clouds hold most of the buildings; perhaps a dozen organisations worldwide can pay for a frontier training run). Follow a frontier model down to the sand and the number of organisations shrinks at every step. A chokepoint that exists will be used: export controls on the top two layers began in October 2022 and have changed several times a year since. At every layer above the sand, India is a buyer.\n\n## How the chokepoints get used\n\nA chokepoint that exists will be used, and the first use came in October 2022. The United States banned the export of its most advanced AI chips to China, then tightened the rules a year later, then in early 2025 proposed a system of tiers for every country on earth (with India in a middle tier subject to caps) and then withdrew it four months later. In 2025 the rules for the chips sold to China changed several times in a year. A Chinese lab's efficient model of early 2025 was, in part, a response to being restricted to a deliberately weakened chip.\n\nThe point is not the particular rule. It is that the physical structure of the industry lets two or three governments decide who may compute, and they have started deciding.\n\n## Where India stands\n\nAt every layer above the sand, India is a buyer. It has no leading-edge fabrication and no chip design at the frontier. It has a national programme, funded in 2024 at around ten thousand crore rupees, to buy tens of thousands of accelerators and make them available to researchers and start-ups as shared compute, and it has a large share of the world's data-labelling and software labour. Its own language models are built on open weights from American and Chinese labs, adapted with Indian text. That is a reasonable position for a country to take and it is also a dependent one, and the open-weights lesson in the history module is the reason it is possible at all.\n\n## What this means for the rest of the module\n\nRegulation, copyright, energy and safety are all shaped by the fact that the makers are few. A law can reach a dozen labs more easily than a million users. A copyright settlement is negotiated with a handful of firms. An electricity grid is strained by the same handful of buildings. And a safety commitment is only as broad as the labs that sign it. Keep the four layers in mind; every lesson that follows is about who sits on one of them."},"check":{"question":"Why can two or three governments decide which countries may build frontier models?","options":[{"text":"Because they hold the patents on the transformer design.","correct":false,"tempting":"Locates control in an idea; the design is published and freely used, and control is in the physical supply chain."},{"text":"Because the machines, factories and chip designs each come from one or a few firms that answer to those governments.","correct":true},{"text":"Because they control the internet and can block access to training data.","correct":false,"tempting":"Overstates control over data; the web crawl is public and the constraint is on hardware."},{"text":"Because their labs employ nearly all the researchers who know how to train a frontier model, and the skill has not spread.","correct":false,"tempting":"Puts the chokepoint in people; the skills have spread widely, and the Chinese labs' results show it."}]},"prev":{"slug":"what-to-do-about-it","title":"What an early-career person can actually do","url":"https://addaly.com/learn/ai-actually-explained/what-to-do-about-it"},"next":{"slug":"the-laws-that-have-arrived","title":"The laws that have arrived","url":"https://addaly.com/learn/ai-actually-explained/the-laws-that-have-arrived"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'Four layers and their chokepoints', in AI, Actually Explained. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-actually-explained/the-stack-and-its-chokepoints (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"Four layers and their chokepoints","courseTitle":"AI, Actually Explained","module":{"index":9,"slug":"who-decides","title":"Who controls it, and what is unsettled"},"answer":null,"keyIdea":"Every frontier model depends on one Dutch machine-maker, one Taiwanese fabricator, one American chip designer and a handful of clouds, so control of AI is control of a few physical chokepoints that governments have already begun to use.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure 9.1","figure":{"type":"stack","title":"Four layers, and how few firms sit on each","layers":[{"label":"Lithography machines","detail":"One firm in the Netherlands. Around 200 to 400 million euros each, a few dozen shipped a year. Barred from selling to China since 2019"},{"label":"Leading-edge fabrication","detail":"One firm in Taiwan makes roughly 90% of the most advanced chips. One South Korean firm is a distant second"},{"label":"Chip design","detail":"One company holds four fifths or more of the AI accelerator market, on a software platform it opened in 2007. Stacked memory comes from three firms"},{"label":"Data centres and labs","detail":"Three American clouds hold most of the buildings; perhaps a dozen organisations worldwide can pay for a frontier training run"}],"caption":"Follow a frontier model down to the sand and the number of organisations shrinks at every step. A chokepoint that exists will be used: export controls on the top two layers began in October 2022 and have changed several times a year since. At every layer above the sand, India is a buyer."},"alt":"Figure: Four layers, and how few firms sit on each. Layers, top to bottom: Lithography machines (One firm in the Netherlands. Around 200 to 400 million euros each, a few dozen shipped a year. Barred from selling to China since 2019); Leading-edge fabrication (One firm in Taiwan makes roughly 90% of the most advanced chips. One South Korean firm is a distant second); Chip design (One company holds four fifths or more of the AI accelerator market, on a software platform it opened in 2007. Stacked memory comes from three firms); Data centres and labs (Three American clouds hold most of the buildings; perhaps a dozen organisations worldwide can pay for a frontier training run). Follow a frontier model down to the sand and the number of organisations shrinks at every step. A chokepoint that exists will be used: export controls on the top two layers began in October 2022 and have changed several times a year since. At every layer above the sand, India is a buyer."}},"blocks":[{"t":"h2","id":"follow-the-chip-down","runs":[{"text":"Follow the chip down"}]},{"t":"p","runs":[{"text":"The history module told you a frontier model is something perhaps a dozen organisations can build. This lesson is about why the number is that small, and it comes from following one chip from the model back down to the sand."}]},{"t":"h2","id":"layer-one-the-machines-that-make-the-chips","runs":[{"text":"Layer one: the machines that make the chips"}]},{"t":"p","runs":[{"text":"The transistors on a modern AI chip are a few nanometres across, and drawing them requires a lithography machine that uses extreme ultraviolet light. One company in the Netherlands makes those machines. Nobody else does; nobody else has come close after two decades of trying. Each costs on the order of two hundred million euros, the newest generation nearer four hundred million, and the company ships a few dozen a year. That single firm is the narrowest point in the entire industry, and since 2019 its government has, under pressure from the United States, forbidden it to sell the machines to China."}]},{"t":"h2","id":"layer-two-the-factories","runs":[{"text":"Layer two: the factories"}]},{"t":"p","runs":[{"text":"Owning the machine is not enough; running a fabrication plant at the leading edge is a skill held by very few. One company in Taiwan makes roughly ninety per cent of the world’s most advanced chips, including every AI accelerator that matters. A South Korean firm is a distant second. The United States, having lost this capability over thirty years, is spending tens of billions to rebuild a fraction of it, and India, which has none, is building its first plant at a technology level two decades behind the frontier. That most of the world’s AI compute passes through one island is a fact that shapes foreign policy."}]},{"t":"h2","id":"layer-three-the-designs","runs":[{"text":"Layer three: the designs"}]},{"t":"p","runs":[{"text":"The chips themselves are designed by a handful of firms and, for AI, overwhelmingly by one. The company from the graphics-chip lesson holds, by most estimates, four fifths or more of the market for AI accelerators, and its programming platform from 2007 is the reason: fifteen years of tools, libraries and trained engineers are built on it, and switching is expensive. Google designs its own chips and uses them internally; Amazon and Microsoft have followed; a US rival designs competitive hardware with a smaller software base. The high-bandwidth memory stacked beside every one of these chips comes from three firms, two of them Korean."}]},{"t":"h2","id":"layer-four-the-buildings-and-the-labs","runs":[{"text":"Layer four: the buildings and the labs"}]},{"t":"p","runs":[{"text":"The chips are bought in tens of thousands and installed in data centres owned mostly by three American cloud companies and a few specialists, drawing power from grids the chip-makers did not plan for. On top of the buildings sit the labs: perhaps a dozen organisations, in the United States and China with one or two in Europe and the Gulf, that can pay for a frontier training run. Above those, the data (the web, the licensing deals, the human labour from the previous module) and above that, everything you use."}]},{"t":"h2","id":"how-the-chokepoints-get-used","runs":[{"text":"How the chokepoints get used"}]},{"t":"p","runs":[{"text":"A chokepoint that exists will be used, and the first use came in October 2022. The United States banned the export of its most advanced AI chips to China, then tightened the rules a year later, then in early 2025 proposed a system of tiers for every country on earth (with India in a middle tier subject to caps) and then withdrew it four months later. In 2025 the rules for the chips sold to China changed several times in a year. A Chinese lab’s efficient model of early 2025 was, in part, a response to being restricted to a deliberately weakened chip."}]},{"t":"p","runs":[{"text":"The point is not the particular rule. It is that the physical structure of the industry lets two or three governments decide who may compute, and they have started deciding."}]},{"t":"h2","id":"where-india-stands","runs":[{"text":"Where India stands"}]},{"t":"p","runs":[{"text":"At every layer above the sand, India is a buyer. It has no leading-edge fabrication and no chip design at the frontier. It has a national programme, funded in 2024 at around ten thousand crore rupees, to buy tens of thousands of accelerators and make them available to researchers and start-ups as shared compute, and it has a large share of the world’s data-labelling and software labour. Its own language models are built on open weights from American and Chinese labs, adapted with Indian text. That is a reasonable position for a country to take and it is also a dependent one, and the open-weights lesson in the history module is the reason it is possible at all."}]},{"t":"h2","id":"what-this-means-for-the-rest-of-the-module","runs":[{"text":"What this means for the rest of the module"}]},{"t":"p","runs":[{"text":"Regulation, copyright, energy and safety are all shaped by the fact that the makers are few. A law can reach a dozen labs more easily than a million users. A copyright settlement is negotiated with a handful of firms. An electricity grid is strained by the same handful of buildings. And a safety commitment is only as broad as the labs that sign it. Keep the four layers in mind; every lesson that follows is about who sits on one of them."}]}],"check":{"question":[{"text":"Why can two or three governments decide which countries may build frontier models?"}],"options":[{"text":[{"text":"Because they hold the patents on the transformer design."}],"correct":false,"tempting":[{"text":"Locates control in an idea; the design is published and freely used, and control is in the physical supply chain."}]},{"text":[{"text":"Because the machines, factories and chip designs each come from one or a few firms that answer to those governments."}],"correct":true,"tempting":null},{"text":[{"text":"Because they control the internet and can block access to training data."}],"correct":false,"tempting":[{"text":"Overstates control over data; the web crawl is public and the constraint is on hardware."}]},{"text":[{"text":"Because their labs employ nearly all the researchers who know how to train a frontier model, and the skill has not spread."}],"correct":false,"tempting":[{"text":"Puts the chokepoint in people; the skills have spread widely, and the Chinese labs’ results show it."}]}],"order":[0,2,1,3],"why":null},"prev":{"slug":"what-to-do-about-it","title":"What an early-career person can actually do","minutes":9},"next":{"slug":"the-laws-that-have-arrived","title":"The laws that have arrived","minutes":10}}}