Who Controls the AI Supply Chain

When you ask a chatbot a question, you interact with a software layer that sits atop one of the most concentrated, geographically fragile, and politically contested supply chains ever constructed. Seven layers of infrastructure connect minerals in the ground to the model on your screen. At each layer, a small number of actors control what gets built, where, and for whom.

Understanding this chain is not an exercise in industrial trivia. It is a map of who holds power over AI — and who doesn't.

The Seven Layers

The AI supply chain runs from raw materials to end-user applications. Each layer depends on everything below it. Each layer concentrates control in fewer hands than you might expect.

Layer 1: Raw materials. Silicon, cobalt, lithium, rare earth elements. The physical inputs that every chip, battery, and data center requires.

Layer 2: Chip design. The architectures that define what hardware can do. Nvidia, AMD, Google, Apple, and Qualcomm design the processors that power AI.

Layer 3: Chip fabrication. Turning designs into physical chips. A handful of foundries — overwhelmingly one — manufacture the world's advanced semiconductors.

Layer 4: Equipment. The machines that make the machines. One company in the Netherlands monopolises the most critical step in chip manufacturing.

Layer 5: Cloud and compute. Data centers where chips are assembled into clusters and rented to the world. Three companies dominate.

Layer 6: Models. The AI systems trained on that compute — large language models, image generators, reasoning engines.

Layer 7: Applications. Products built on top of models, serving end users. The layer most people see; the layer with the least structural power.

Each layer extracts value. But as you move down the stack — toward the physical, the scarce, the hard to replicate — profit margins concentrate and barriers to entry grow. The communities that mine cobalt in the Democratic Republic of Congo see almost none of the value that flows from their labour through the chain to a Silicon Valley API endpoint.

As Kate Crawford documents in Atlas of AI, this material infrastructure is not incidental to AI. It is constitutive. The "cloud" is not ethereal. It is mines, refineries, factories, and power plants, connected by supply chains that span continents and reproduce patterns of extraction that predate digital technology by centuries (Crawford, 2021).

Layer 1: The Mineral Foundation

Advanced semiconductors require dozens of elements in precise combinations. Silicon forms the substrate. Rare earth elements — neodymium, dysprosium, gallium — go into magnets, substrates, and components. Cobalt and lithium power the batteries in data centers and the devices that access AI services.

China dominates this layer. It controls roughly 60% of rare earth mining and around 90% of rare earth processing (IEA, 2024). This is not primarily geology — the minerals exist elsewhere. It is the result of decades of strategic investment in refining capacity while other countries chose not to compete.

The Democratic Republic of Congo supplies over 70% of the world's cobalt. Much of it is mined under conditions that bear no resemblance to the sleek marketing of the companies that ultimately depend on it. The gap between the extraction site and the application layer is not just geographic. It is economic, political, and moral.

Building alternative mineral supply chains is possible but slow. Australia, Canada, and several African nations have resources. Processing capacity takes years and billions to build. In the interim, the mineral layer remains a chokepoint — controlled by a small number of countries and subject to export restrictions that can reshape entire industries overnight.

Layer 2: Who Designs the Chips

Chip design is where physics meets architecture, and where competitive moats are deepest.

Nvidia dominates AI hardware. Their GPUs — the H100, B100, and subsequent generations — are purpose-built for the matrix multiplications that neural networks require. But Nvidia's real advantage is not silicon. It is CUDA, their proprietary software framework. As Chris Miller documents in Chip War, CUDA has become the default programming environment for AI development (Miller, 2022). Millions of lines of code, thousands of research papers, and entire educational curricula assume CUDA. Switching costs are enormous. This is classic platform lock-in, operating at the hardware level.

AMD offers competitive alternatives. Their MI300 series matches Nvidia on paper specifications. But specifications are not the bottleneck — ecosystem is. AMD's ROCm software stack is less mature, less documented, and less integrated with the tools researchers already use.

Google designs Tensor Processing Units (TPUs), custom chips optimised for AI workloads. TPUs are powerful and cost-effective. They are also available only on Google Cloud. You cannot buy a TPU. Google uses them as a competitive advantage for their cloud platform, not as a product in their own right.

Apple builds its own silicon (the M-series) that makes on-device AI inference viable on consumer hardware. Qualcomm is pushing AI capabilities into mobile devices through its Snapdragon platform. Both represent a different model: AI at the edge rather than in the data centre.

The design layer is concentrated but not monopolised. The deeper problem is that every one of these designs must pass through the same manufacturing bottleneck.

Layer 3: The Fabrication Bottleneck

Every advanced AI chip — regardless of who designed it — is manufactured by Taiwan Semiconductor Manufacturing Company. TSMC fabricates over 90% of the world's most advanced semiconductors, those built on process nodes below 7 nanometres (CSET, 2021).

This is extraordinary concentration. The AI industry's multi-trillion-dollar ecosystem depends on facilities located in a single geography, operated by a single company, on an island 130 kilometres from a mainland that claims sovereignty over it.

Samsung and Intel are the only plausible alternatives for advanced fabrication. Samsung trails TSMC on yield rates and reliability. Intel, once the world leader, fell behind on process technology and is spending tens of billions attempting to catch up through its IDM 2.0 strategy and the CHIPS Act subsidies.

The CHIPS and Science Act (2022) allocated $52 billion for US semiconductor manufacturing. TSMC is building fabs in Arizona. Samsung is building in Texas. But replicating TSMC's manufacturing capability is not simply a matter of capital. It requires thousands of specialised engineers, years of process refinement, and supply chains for hundreds of chemicals and materials that are themselves concentrated. These fabs will take until the late 2020s to reach full production — and even then, they may not match TSMC's Taiwanese operations.

The fabrication layer is where the AI supply chain is most fragile. A disruption in Taiwan — whether from natural disaster, geopolitical conflict, or industrial accident — would halt the production of advanced AI chips for years.

Layer 4: ASML and the Lithography Monopoly

Beneath TSMC sits an even more extreme concentration. ASML, a Dutch company headquartered in Veldhoven, is the sole manufacturer of extreme ultraviolet (EUV) lithography machines. These are the instruments that print circuitry at the nanometre scale required for cutting-edge chips.

An EUV machine costs approximately $200 million. It contains over 100,000 parts sourced from hundreds of suppliers. It weighs 150 tonnes and requires multiple cargo planes to ship. It uses a tin-plasma laser to generate light at a wavelength of 13.5 nanometres — a feat of engineering that took decades and billions in R&D to achieve.

ASML has no competitors. Canon and Nikon produce older-generation lithography equipment, but neither has come close to EUV. This makes ASML arguably the most critical single point of failure in the global technology supply chain. No EUV machines, no advanced chips. No advanced chips, no frontier AI.

When the US imposed export controls on advanced semiconductor technology to China in 2022, blocking ASML's EUV shipments was the most consequential action. Without these machines, China cannot manufacture chips at the scales required for frontier AI training — regardless of how much money it invests.

Langdon Winner's question — do artifacts have politics? — finds its most literal answer here (Winner, 1980). A single machine, built by a single company, determines which nations can participate in cutting-edge semiconductor manufacturing. The political choices embedded in export control policy for EUV equipment shape the global distribution of AI capability more directly than any algorithm.

Layer 5: Cloud and Compute

The chips, once manufactured, are assembled into clusters in data centres operated by a small number of cloud providers.

Amazon Web Services, Microsoft Azure, and Google Cloud Platform collectively control the majority of AI training and inference compute available to the market. Their dominance is self-reinforcing: they have the capital to buy GPUs in bulk, the engineering talent to operate them at scale, and the customer relationships that make switching costly.

Specialised providers have emerged. CoreWeave, Lambda Labs, and Together AI offer GPU clouds tailored for AI workloads, often at lower prices than the hyperscalers. But they depend on the same chip supply chain and compete for the same limited allocation of Nvidia's latest hardware.

The vertical integration at this layer is striking. Microsoft invested $13 billion in OpenAI and provides its compute through Azure. Amazon invested $4 billion in Anthropic and hosts it on AWS. Google both operates GCP and runs DeepMind internally. The companies that sell compute are also the companies that invest in and sometimes control the labs that consume the most of it. As Meredith Whittaker and the AI Now Institute have argued, this structural entanglement between cloud providers and AI labs raises questions that conventional competition policy is ill-equipped to address (Whittaker, 2023).

Layer 6: Models — and the Oligopoly That Builds Them

The model layer — OpenAI, Anthropic, Google DeepMind, Meta, Mistral, DeepSeek — is where most public attention concentrates. These organisations train the large language models and multimodal systems that power the current wave of AI applications.

But this layer is entirely dependent on everything below it. A frontier model requires rare minerals, chip designs, fabrication capacity, lithography equipment, and massive compute clusters. The organisations that train frontier models do not control most of these inputs. They are, in an important sense, tenants in an infrastructure built and owned by others.

This dependency shapes behaviour. Labs form exclusive partnerships with cloud providers partly because they must — training runs costing hundreds of millions of dollars require guaranteed compute access. The deals are as much about securing supply chain position as about investment capital.

The open-weight movement — Meta's Llama, Mistral's models, and DeepSeek's contributions from China — represents an attempt to distribute the outputs of this layer more broadly. But open weights do not open the supply chain. You can download a model freely. Running it still requires hardware that flows through every chokepoint described above.

Layer 7: Applications and the Value Capture Problem

The application layer is where most users encounter AI: chatbots, coding assistants, image generators, search engines, enterprise tools. It is also the layer with the least structural power.

Application developers build on top of model APIs. They do not control the models, the compute, the chips, or the minerals. When a model provider changes pricing, deprecates a version, or alters terms of service, application builders adapt or exit. The relationship is fundamentally asymmetric.

This mirrors patterns in earlier technology supply chains. App developers on Apple's App Store are subject to Apple's rules and revenue share. Websites built on Google's search traffic are subject to algorithm changes. The application layer captures value only to the extent that the layers below permit.

The redistribution implications are stark. The further up the chain — toward the user, toward the application — the less leverage any individual actor has. Value flows downward, toward scarcity. The companies that control chokepoints at layers 3 and 4 capture disproportionate margins regardless of what happens at the application layer.

The Redistribution Angle

Each layer of the AI supply chain extracts value. But the distribution of that value is sharply uneven.

ASML's gross margins exceed 50%. TSMC's hover around 55%. Nvidia's have reached 75% in recent quarters. These are the margins of monopoly or near-monopoly, earned by controlling chokepoints that no competitor can easily replicate.

Contrast this with the mineral layer: cobalt miners in the DRC earn subsistence wages. Rare earth processing workers in China face environmental health risks for modest pay. The communities that provide the physical foundation of AI capture almost none of its economic value.

This pattern — value extraction at the base, value accumulation at the chokepoints — is not unique to AI. It echoes the oil supply chain, where resource-rich nations often remained poor while refiners and distributors accumulated wealth. It echoes agricultural supply chains, where farmers capture pennies while processors and retailers capture dollars.

But the AI supply chain intensifies the pattern through extreme geographic concentration. Oil is distributed across many countries. Advanced chip fabrication is in one. This makes the AI supply chain simultaneously more fragile and more susceptible to political control than almost any comparable industrial system.

What Diversification Would Require

Breaking these concentrations would require action at multiple layers simultaneously.

Alternative chip architectures. RISC-V, an open-source instruction set architecture, offers a path away from proprietary chip designs. Companies like SiFive and Tenstorrent are developing RISC-V-based AI accelerators. If these mature, they could weaken the design-level lock-in that Nvidia's CUDA ecosystem creates. But maturity is years away.

Distributed manufacturing. The CHIPS Act and EU Chips Act represent initial efforts to geographically diversify fabrication. Success is uncertain and slow. Even optimistic timelines put meaningful diversification in the 2030s.

Open hardware initiatives. Projects like the Open Compute Project (founded by Meta for data centre design) and open-source chip design efforts could reduce barriers to entry. But manufacturing scale economies are punishing. Being open does not make you competitive if your competitor's fab cost $20 billion.

Public compute infrastructure. The EU's EuroHPC initiative, the UK's AI Research Resource, and similar programmes could provide compute access outside commercial cloud providers. These are early-stage efforts that face the same supply chain constraints as everyone else — they still need Nvidia's chips, fabricated by TSMC, using ASML's machines.

Mineral supply chain diversification. Building rare earth processing capacity outside China is a decade-long project. Australia, Canada, and Brazil have resources. Building the refining infrastructure requires sustained investment and political commitment that has historically been lacking.

None of these alternatives is sufficient alone. The supply chain is a system. Diversifying one layer while leaving others concentrated accomplishes less than it appears. But progress does not require solving every layer simultaneously. The CHIPS Act is a genuine start at the fabrication layer. Open-weight models like Llama and Mistral are making progress at the model layer. Edge AI and small language models are reducing dependence on cloud infrastructure. Each initiative is partial, but together they constitute a real movement toward diversification. The semiconductor industry itself was once concentrated in the US; the geographic spread that followed was the product of deliberate industrial policy by Japan, South Korea, and Taiwan — a reminder that supply chain diversification has actually succeeded before, given sustained political commitment.

Who Controls the Chain Controls AI

The AI supply chain is a study in how power concentrates at chokepoints. ASML controls lithography. TSMC controls fabrication. Nvidia controls the GPU ecosystem. Three cloud providers control most available compute. A handful of labs train frontier models.

Each chokepoint is a point of leverage — over who can build AI, where it can be deployed, and on what terms. Export controls on chips and lithography equipment have already demonstrated that governments understand this: controlling the supply chain is controlling AI.

For the redistribution question — when AI creates value, who gets it? — the supply chain provides an uncomfortably clear answer. Value flows to whoever controls the scarcest layer. Right now, that means a small number of companies in a small number of countries. The communities that provide raw materials, the workers who assemble hardware, and the users who generate the data that trains models all participate in the supply chain. They do not control it.

Changing this would require treating the AI supply chain not as a series of private markets but as public infrastructure — with the obligations, investment, and governance that implies. Whether any government has the will to do so remains an open question. The supply chain, as it exists, was not designed. It emerged from decades of market decisions, industrial policy choices, and geopolitical accidents. Redesigning it is a project that matches the scale of the technology it supports.

Related

• Hardware Bottleneck
• Compute Is the New Oil
• Frontier Lab Oligopoly
• AI Chip Export Controls
• Magnificent Seven
• Nvidia $3T Question
• Atlas of AI
• Environmental Cost of AI
• Data Center Energy
• Edge AI

Sources

• Crawford, Kate. Atlas of AI. Yale University Press, 2021. Publisher
• CSET (Center for Security and Emerging Technology). "The Semiconductor Supply Chain." Khan, Saif M., 2021. CSET
• IEA (International Energy Agency). "Critical Minerals Market Review." 2024. IEA
• Miller, Chris. Chip War: The Fight for the World's Most Critical Technology. Scribner, 2022. Publisher
• Whittaker, Meredith. "The Steep Cost of Capture." Interactions, 2023. ACM
• Winner, Langdon. "Do Artifacts Have Politics?" Daedalus, 1980. JSTOR