What Is the AI Supply Chain, From Chips to Power?

By Adam Patti, Chief Executive Officer, VistaShares | November 2026

The AI supply chain is the sequence of industries required to build and run artificial intelligence systems at scale. It runs from chip design and semiconductor manufacturing, through memory, networking and servers, into data centers, and finally to the electricity generation and grid equipment that power them. Each layer depends on the others, and a shortage in any one of them can limit how quickly AI capacity is built.

Why think about AI as a supply chain?

Public attention tends to focus on AI models and the applications built on them. Those models, however, run on physical infrastructure that has to be designed, manufactured, installed, cooled and powered. Each of those steps involves different companies, different technologies and different constraints.

Viewing AI as a supply chain helps explain why spending on AI reaches far beyond software. When a cloud provider adds AI capacity, it buys accelerators, memory, networking equipment and servers, builds or leases data center space, and secures power. Four major U.S. cloud providers have guided to roughly $730 billion of combined capital spending in 2026, according to company guidance compiled in our November 2026 report, “The Deployment Gap.”¹ That spending flows through every layer described below.

Demand for AI computing is also tied to usage. Alphabet has reported that the number of tokens, the units of text processed by its AI models, rose from about 9.7 trillion per month in May 2024 to more than 3.2 quadrillion per month in May 2026.² Token volume measures activity rather than revenue, but it illustrates why infrastructure requirements have grown.

What are the layers of the AI supply chain?

Chip design. AI workloads run mainly on accelerators, such as graphics processing units and custom chips designed for specific workloads. Chip designers create the architectures and the software tools developers use to program them. Many large cloud providers now design their own custom chips alongside those they buy from merchant suppliers.

Semiconductor manufacturing. Most designers do not manufacture their own chips. They rely on foundries that produce chips at advanced process nodes. Those foundries in turn depend on semiconductor equipment makers for lithography, deposition, etching and inspection tools, and on suppliers of specialty materials and chemicals.

Advanced packaging. Modern AI accelerators combine multiple pieces of silicon, including the processor and stacks of memory, in a single package. Advanced packaging techniques that connect these components at high density have become a distinct capacity requirement within manufacturing.

Memory. AI models require fast access to large amounts of data. High-bandwidth memory, which stacks memory chips vertically and places them next to the processor, is a specialized product with a small number of producers.

Networking. Training and running large models requires thousands of accelerators to work together. That depends on high-speed networking within servers, between servers in a rack, and across the data center, using switches, network interface cards, optical transceivers and cabling.

Servers and systems. Server manufacturers assemble accelerators, processors, memory, storage and networking into complete systems and racks, often designed specifically for AI workloads.

Data centers. The buildings that house these systems provide space, power distribution, cooling and connectivity. Our article on how data centers work describes their components in detail.

Power and grid equipment. Data centers require reliable electricity. That involves power generation, transmission and distribution lines, substations, transformers and switchgear, along with on-site backup systems.

Where are the constraints?

Different layers have been the limiting factor at different times. In the early phase of the AI buildout, accelerator supply was the main constraint. Over time, attention has shifted toward advanced packaging, high-bandwidth memory, and the time required to build data centers and connect them to the grid.

Electricity has become a central constraint. Grid connections can take years to approve and build, and equipment such as large power transformers has long lead times. These timelines are often longer than the time needed to construct the data center building itself. As a result, access to power increasingly shapes where and how quickly AI capacity can be added.

Because the constraints move, investors and analysts often focus less on any single company than on which layers are supply-limited and which have spare capacity at a given time.

Is the AI buildout only a U.S. story?

No. While U.S. cloud providers account for a large share of spending, governments elsewhere increasingly treat computing capacity as national infrastructure. Saudi Arabia’s HUMAIN has announced plans for data centers totaling about 2.2 gigawatts of capacity. The United Arab Emirates is developing a large AI campus in partnership with U.S. companies. The European Union’s InvestAI initiative targets €200 billion of investment, including funding for AI gigafactories, and Japan has pledged more than ¥10 trillion of public support for AI and semiconductors over seven years.³

Within the United States, policy is also evolving. In November 2026, the Senate voted 73 to 25 to advance the Bipartisan American Affordability and Jobs Act, legislation intended to speed approvals for power plants and transmission lines while requiring large data centers to pay more of the grid upgrades they require.⁴

The result is that data center capacity tends to go where power, land, permitting and policy support are available. When one location becomes more difficult, projects can move to another, within the United States or abroad. Some sovereign projects remain in early stages, and announced plans may change.

How do the layers differ for investors?

Each layer has its own economics. Chip designers and equipment makers are typically asset-light or research-intensive businesses. Foundries and memory producers are capital-intensive and cyclical. Data center developers and utilities often operate with long-term contracts and regulated returns, along with substantial borrowing needs. Power equipment manufacturers depend on order backlogs that can extend years into the future.

These differences mean that companies across the supply chain respond differently to changes in demand, interest rates and financing conditions. A slowdown in spending can affect suppliers with short-term orders sooner than those with long-term contracts. A tightening in credit markets can affect capital-intensive builders more than asset-light designers.

What are the risks in the AI supply chain?

Spending cycles. Infrastructure spending can rise and fall faster than end demand, leaving suppliers exposed if customers slow their purchases.

Concentration. Several layers rely on a small number of suppliers or locations, which can create bottlenecks and geopolitical exposure.

Financing. Building data centers and power infrastructure requires large amounts of capital. Changes in credit conditions can affect the pace of construction.

Technology change. Improvements in chip and model efficiency can change how much hardware is needed for a given workload.

Policy and trade. Export controls, tariffs, permitting rules and local opposition can all affect where and how infrastructure is built.

Valuation. Market expectations can run ahead of or behind actual results. Investments involve risk, including possible loss of principal.

How does the AI supply chain connect to other themes?

The AI buildout overlaps with several other areas. It draws on the same power and grid equipment needed for broader electrification. It relies on semiconductor and materials supply chains that also serve the defense industry, discussed in our article on the defense supply chain. And AI models are an essential component of advanced robotics, covered in our article on humanoid robots.

Frequently asked questions

What is the difference between AI chip designers and foundries?

Chip designers create the architecture of a processor and the software used to program it. Foundries manufacture the physical chips using advanced fabrication processes and equipment. Many designers outsource production to foundries, which allows each to specialize. A small number of foundries can produce chips at the most advanced process nodes.

Why is high-bandwidth memory important for AI?

AI models move very large amounts of data between memory and processors. High-bandwidth memory stacks memory chips and places them close to the processor, which allows data to move faster than with conventional memory. It is produced by a small number of manufacturers and requires specialized packaging.

Why has electricity become a constraint for AI?

Large data centers require substantial, reliable power. Connecting them to the grid can involve new generation, transmission lines, substations and transformers, many of which take years to permit and build. These timelines are often longer than construction of the data center itself, so power availability frequently determines project timing.

What are sovereign AI programs?

Sovereign AI programs are government-led or government-supported initiatives to build domestic computing capacity, often with national champions or international partners. Examples include projects in Saudi Arabia, the United Arab Emirates, the European Union and Japan. They add buyers of infrastructure whose decisions are driven partly by national policy.

Do tokens measure AI revenue?

No. Tokens measure the volume of text or data processed by AI models. Token counts can rise because more people use AI or because models reason at greater length. They indicate how much computing is being used, which drives infrastructure demand, but they are not a direct measure of revenue or profit.

This material is provided for educational purposes only. It is not investment, legal or tax advice, and it is not a recommendation or an offer to buy or sell any security. Information is drawn from sources believed to be reliable as of the date of publication, is not guaranteed as to accuracy or completeness, and may change without notice. Figures attributed to third parties are their estimates. Projections and forecasts are not guarantees of future results. Investors should consult their own financial, legal and tax professionals before making investment decisions. Investments involve risk, including possible loss of principal. Companies in a single industry or theme may be affected by common economic, regulatory or technological developments and may be more volatile than the broad market. References to companies, governments or programs are for illustration only.