Edition 12
An AI answer arrives on a screen with no visible machinery behind it. It is the final product of one of the most concentrated industrial supply chains ever built.
A question goes in, an answer comes back, and almost everything between those two moments is hidden. The experience feels closer to software than to industry. There is no factory floor, no shipment arriving at a warehouse, no fuel being loaded, no raw material waiting at a port.
Yet an AI answer begins in a physical world. It depends on semiconductor equipment precise enough to manufacture features measured in nanometres, on foundries capable of producing leading-edge chips at scale, on specialised memory placed close enough to processors to keep data moving, and on advanced packaging that can bring several pieces of silicon together. It depends on networking, storage, data centres, cooling systems, transformers, substations and electricity. Behind those industries stands another layer of suppliers providing optics, chemicals, gases, metals, substrates, cables, pumps and thousands of components most AI users will never hear about.
What appears to be intelligence on demand is the final product of an industrial supply chain. Parts of that chain are held by very few hands.
AI is moving from experiment to infrastructure. Businesses are placing customer service, software development, research, operations, security and internal knowledge systems on top of it. Governments are thinking about national AI capacity. Investors are financing data centres on the expectation of continuing demand. Model providers are planning around enormous quantities of future compute.
We are beginning to depend on intelligence before we have understood how it is supplied.
The Prompt
Few people outside the industry know what it takes to produce a single advanced AI accelerator. Long before it reaches a data centre, somebody has to design the chip. Manufacturing equipment has to exist that can turn the design into silicon. A foundry has to allocate capacity. The wafer needs materials of extraordinary purity. Layers have to be deposited, patterned, etched and inspected repeatedly. The finished dies need to be tested and packaged alongside memory capable of feeding them quickly enough. Each of those steps contains another supply chain.
Near the start of that chain sits ASML, the Dutch company formerly known as Advanced Semiconductor Materials Lithography. Extreme ultraviolet lithography, usually shortened to EUV, is used for the most intricate layers in leading-edge semiconductor manufacturing, and ASML states in its 2025 annual report that it remains the world’s only manufacturer of EUV systems. That is an unusual industrial position. ASML itself depends on thousands of suppliers. Some components in its systems are single-sourced or available from only a limited number of vendors, because the engineering involved is specialised enough that duplicating supply is difficult both technically and economically.
Taiwan Semiconductor Manufacturing Company, or TSMC, is the next constraint. Its 2025 annual report describes continuing expansion of leading-edge manufacturing and advanced packaging capacity in response to AI demand. Advanced process technologies accounted for 74 percent of its wafer revenue in 2025. Its chip-on-wafer-on-substrate packaging, known as CoWoS, has become an important part of high-performance AI systems, because processors and high-bandwidth memory need to operate together at enormous speed.
AI systems do not run on processors alone. Large models need huge quantities of data moved rapidly between compute and memory. That has made high-bandwidth memory, or HBM, strategically important, with SK hynix, Samsung and Micron all racing to increase capability and capacity.
Memory makers have limited fabrication capacity, and every wafer allocated to HBM is a wafer not making conventional memory. The same is true of foundry capacity and advanced packaging. As the industry has tilted toward AI, ordinary computing has started paying for it. Memory module prices have risen sharply, some more than doubling within a year, and buyers of laptops, servers, phones, cars and industrial equipment are competing for supply that was reallocated to somebody else’s data centre. Shortages have become real enough that manufacturers outside AI have delayed products and revised prices.
This rarely appears in AI cost discussions. The price of intelligence is not only what a company pays its model provider. Part of the cost is borne by every organisation buying computing hardware for something else entirely, and by consumers who will never knowingly use a frontier model.
Thousands of accelerators then need high-speed networking. Servers need storage. Racks need power. Equipment generates heat that must be removed. Data centres need connections to an electrical grid capable of supplying power continuously and handling extraordinary concentrations of demand. At that point the AI supply chain becomes an energy supply chain.
The International Energy Agency expects global electricity consumption by data centres to rise from roughly 485 terawatt-hours in 2025 to around 950 terawatt-hours in 2030, with AI-focused facilities growing considerably faster than data centres overall. This does not mean AI is about to consume the world’s electricity. Even at those levels, data centres would account for only a few percent of global demand. The difficulty is concentration. A large AI data centre does not draw its electricity evenly across the planet. It needs substantial power in one place, connected at one time, with enough transmission, transformers, generation, cooling and backup infrastructure to support it.
The software can scale quickly. The physical world keeps its own schedule.
The Mirage
The cloud taught a generation of businesses to stop thinking about hardware. Much of what people paid for was the freedom to ignore it. A company no longer had to order servers months in advance to launch an application. Compute became something requested from a console, storage became a service, capacity appeared when needed, and most of the machinery disappeared from the customer’s view. Computing had become a utility, drawn from a socket like power, water or gas, with the plant that produced it somebody else’s concern.
AI inherits that. An enterprise signs up for a model through an application programming interface and begins using intelligence almost immediately. Another provider is added. A third model is tested. The market appears broad and competitive. But diversification at the application layer does not mean diversification underneath it.
A previous edition of this newsletter looked at how the commercial layers of the AI industry have become entangled, with investors, suppliers and customers appearing repeatedly in the same transactions. The manufacturing layer has a different character. Commercial concentration is a matter of choice and contract. Industrial concentration is a matter of physics, capital and time. A financing arrangement can be restructured in a quarter. A fabrication plant cannot.
Switching a model provider is a procurement decision. Switching the plant that manufactures your accelerators is not a decision anybody at your company can make. Neither, in most cases, is it a decision your provider can make, because the alternative capacity may not exist yet at the required process node.
None of this points to a badly designed industry. The work at these layers is simply hard enough that few organisations manage it at all. A leading-edge fab costs enormous sums and takes years to build. Advanced lithography emerged from decades of research and an extraordinary engineering ecosystem. High-bandwidth memory requires deep manufacturing expertise, and advanced packaging has become difficult enough to be a competitive technology in its own right. Duplicating those capabilities merely to have a second supplier is not straightforward.
Abstraction hides the physical layer until something interrupts it. A company that can move between three model providers has genuine protection against an outage or a commercial dispute. It has almost none against a shortage of HBM, which reaches all three providers in the same quarter. Vendor risk is written down somewhere, reviewed at renewal, argued over by lawyers. Supply risk of this kind appears in no contract the organisation has ever signed.
This is the paradox of the intelligence supply chain. AI can look abundant at the interface while remaining scarce at critical points underneath it.
The Reality Check
Supply-chain risk is easy to turn into a geopolitical discussion, and there are reasons for that. Semiconductor manufacturing, advanced packaging, memory, equipment and critical materials are distributed unevenly around the world. Governments have noticed, and are spending heavily to expand domestic capacity. But geopolitics is only one source of interruption. Factories lose power. Specialised equipment fails. Suppliers have quality problems. Ports close. Earthquakes and floods interrupt production. Cyberattacks disrupt industrial systems. Energy projects slip. Export rules change. A critical supplier can simply find that demand has grown faster than its ability to manufacture.
ASML’s own risk disclosures note that many components and subassemblies come from single or limited suppliers, and that disruption can arise from causes ranging from energy shortages and infrastructure problems to cyberattacks and natural disasters.
Power is the constraint I would watch most closely. A data centre can move from planning to operation much faster than a new transmission line, substation or power plant can be built. The International Energy Agency has pointed to exactly this mismatch: data-centre development takes a few years, while the energy infrastructure supporting it works on longer planning and construction cycles.
For much of the software era, extra usage was an engineering and economic problem, because capacity could usually be bought somewhere. AI pushes harder against physical limits. More intelligence requires more accelerators, which require more memory, more networking, more power, and more cooling and electrical infrastructure behind that.
Efficiency is improving quickly. New chips perform more work per unit of energy. Models become more efficient. Specialised hardware reduces unnecessary compute. Smaller models handle tasks once reserved for much larger ones. But efficiency does not reduce total demand if usage expands faster still. The history of computing is full of examples where cheaper computation produced more computing rather than less, and AI may follow the same path.
The constraint keeps moving. At one point it is advanced chips, then memory, then packaging, then networking, then transformers, then electricity, then suitable land, then cooling. A bottleneck does not have to stop AI to matter. It only has to become the slowest part of the chain.
The Leadership Question
Most enterprises do not need to become semiconductor supply-chain experts. They do need to know where their dependencies actually end, which is a different exercise from listing which AI vendors the company uses.
Imagine a business has adopted four AI systems. One supports developers, one answers internal questions, one helps customer service, and one performs document analysis. The organisation has four vendor relationships and may believe risk is distributed. Now map the dependencies one level deeper. Which foundation models power those products? Where are those models hosted? Which cloud regions are involved? What happens when capacity becomes constrained? Are alternative models genuinely interchangeable, or would switching require new prompts, testing, evaluation and workflow changes?
Then go further down. Are those providers relying on similar accelerator infrastructure? Would a broader compute shortage affect all of them at once? Does the organisation have any workload that could run on smaller models, alternative hardware or local infrastructure if external capacity became constrained?
No company needs a second AI supply chain waiting in reserve. Resilience costs money, and most of that spending would be wasted. The useful thing to know is which dependency the business can live with and which one has quietly become critical.
Timing is the part most organisations underestimate. Businesses can change software quickly. Industrial supply chains cannot move at that speed. If an organisation discovers today that an AI provider no longer meets its requirements, it may be able to test an alternative next week. If the industry discovers it needs twice as much advanced packaging, HBM, transformer capacity or electrical generation, the response takes years. That difference applies when leaders hear predictions about AI capability rising continuously. Model capability may improve quickly. Availability, cost and scale depend on the rest of the chain keeping up.
Questions for the next review
- Which workloads genuinely require frontier capability, and which are being sent there out of habit?
- Which parts of our AI cost depend on scarce physical resources rather than on software margins that can keep falling?
- What happens to the business if demand is rationed, prices move or a particular region becomes constrained?
- If our main provider could supply only half our current volume next quarter, what stops working first?
- Which of our AI dependencies has never appeared in a business continuity plan?
Most organisations have not reached that last question. We already plan around telecom networks, cloud providers, payment systems, logistics partners and critical software. Intelligence is joining that list, and in most companies nobody has yet written it down.
The Closing Signal
Everything this edition has described stands underneath a single text box.
What the user sees, and what carries it
Procurement can negotiate layers 01 and 02. Nobody in the building can negotiate layers 04 to 08.
Compressing that into something anybody can use is one of the great achievements of the technology. The risk is that the simplicity of the interface becomes our mental model for the system underneath it.
AI does not float above the physical economy. It operates firmly inside it. An advanced model depends on machines only a handful of companies can build, chips only a limited number of facilities can manufacture at the required level, specialised memory and packaging, enormous data centres, electrical equipment with its own constrained supply chains, and grids that cannot expand simply because software demand rose last quarter.
The build-out may well succeed. The industrial response to it is already enormous. Semiconductor capacity is expanding. Packaging capacity is expanding. Memory makers are investing. Data centres are being constructed. Power companies are planning around new loads. Governments are working to strengthen domestic supply chains. The system is adapting. What matters is understanding what kind of system it is adapting into.
We speak about AI as a race in models, algorithms and intelligence. It is also a race in factories, memory, packaging, transformers, power plants, cooling systems and construction schedules. The next breakthrough may come from a research laboratory. Putting that breakthrough into the hands of hundreds of millions of people will depend on an entirely different set of organisations. That makes the supply chain part of the intelligence.
For businesses beginning to depend on AI, the question is no longer only what the model can do. It is this: how many layers stand between your organisation and the intelligence it depends on, and how many of them can you actually see?
The Closing Signal
Intelligence arrives as software. It is delivered by industry.