# What Goes Into an AI Data Center? The Bill of Materials Explained Source: https://www.vistashares.com/ai-data-center-bill-of-materials/ Published: 2026-11-13 | Updated: 2026-11-13 By Adam Patti, Chief Executive Officer, VistaShares | November 2026 An AI data center's bill of materials is the full list of physical inputs needed to turn land and electricity into computing capacity: the building, the grid connection and electrical gear, cooling systems, accelerator chips, memory, networking equipment and storage. Each layer has its own suppliers, lead times and bottlenecks, which is why spending on AI infrastructure reaches far beyond chip designers. For investors, the bill of materials is a practical map of where AI capital spending flows. ## Why think about a data center as a bill of materials? A bill of materials, or BOM, is a manufacturing term for every component and subassembly that goes into a finished product. Applying it to an AI data center shows that the facility is less a single product than a stack of industries, each with different economics. The scale of spending makes the exercise relevant. Four large U.S. cloud and platform companies spent a combined $201.9 billion on property and equipment in the July to September 2026 quarter, roughly double the year-earlier figure, according to their quarterly filings.[1] Gartner forecasts worldwide AI spending of $2.67 trillion in 2026, with infrastructure accounting for most of the increase in its recent revisions.[2] Our article on [the AI supply chain from chips to power](https://www.vistashares.com/ai-supply-chain-chips-to-power/) traces how that money moves between industries. ## What are the main layers? **Land, shell and site work.** The building, foundations, roads, security systems and fiber connections. These are construction inputs, often sourced locally, and they set the physical footprint for everything else. **Power.** The grid interconnection, high-voltage transformers, switchgear, power distribution units, uninterruptible power supplies, batteries and backup generators. Increasingly, sites also add on-site generation. Power equipment has some of the longest lead times in the stack; Wood Mackenzie reported in 2025 that lead times for large power transformers had reached about 128 weeks.[3] **Cooling.** Chillers, cooling towers, heat exchangers, coolant distribution units, cold plates and piping. As rack power rises, cooling shifts from air toward liquid. Our article on [data center liquid cooling](https://www.vistashares.com/data-center-liquid-cooling/) covers this layer in detail. **Compute.** Accelerator chips such as graphics processing units (GPUs) and custom AI chips, plus central processors, circuit boards and the servers that hold them. In most AI builds, this is the costliest layer per megawatt. **Memory and storage.** High-bandwidth memory stacked next to each accelerator, conventional server memory and solid-state storage. Memory has become a constraint in its own right: suppliers have described 2026 high-bandwidth memory output as effectively sold out.[4] **Networking.** Switches, network interface cards, copper cables for short links and optical transceivers and fiber for longer ones. AI training clusters connect thousands of accelerators, so networking spending grows with cluster size. **Software and operations.** Management software, security, monitoring and the people who run the facility. These are ongoing costs rather than one-time purchases. ## How has AI changed the mix? Traditional enterprise data centers were built around racks drawing modest power. The Uptime Institute's 2026 survey put the most common rack density at about 11 kilowatts.[5] Current AI rack designs can draw well over 100 kilowatts, and Nvidia's reference architecture for one current system calls for up to 142 kilowatts per rack.[6] Higher density changes the bill of materials in three ways. First, it concentrates more spending on compute and memory per square foot. Second, it pushes cooling from fans and air handlers toward liquid systems. Third, it raises the electrical load each site must secure, making grid access and power equipment a gating factor for when a facility can open. ## Which layers tend to be bottlenecks? Bottlenecks shift over time, and the binding constraint in one year may ease in the next. In 2025 and 2026, commonly cited constraints have included: **Power availability.** Grid connection queues and the time needed to build transmission lines. **Electrical equipment.** Transformers, switchgear and gas turbines, where manufacturers report multiyear backlogs. **Advanced packaging and memory.** The specialized processes that join accelerators with high-bandwidth memory. **Optical components.** Lasers and transceivers for high-speed links. Our articles on [grid transformers](https://www.vistashares.com/grid-transformers-ai-bottleneck/) and [high-bandwidth memory](https://www.vistashares.com/high-bandwidth-memory-explained/) look at two of these constraints. ## What does the bill of materials mean for investors? A BOM view helps investors see that AI infrastructure exposure can come from many directions: chip designers and foundries, memory makers, networking and optics suppliers, electrical equipment manufacturers, cooling specialists, construction firms, utilities and owners of data center real estate. Each layer has a different relationship to AI spending. Some, such as accelerators, depend heavily on AI demand. Others, such as transformers, also serve utilities, factories and electric vehicles. The view also highlights risks. Spending is concentrated among a small number of large buyers, so a change in their plans can ripple through every layer. Component prices can rise quickly in shortages and fall quickly when supply catches up. Technology shifts, such as a move from pluggable optics to co-packaged optics, can change which suppliers capture value. Our guide on [how to evaluate a thematic ETF](https://www.vistashares.com/how-to-evaluate-a-thematic-etf/) discusses how fund holdings map to these layers. ## Frequently asked questions ### What is the most expensive part of an AI data center? In most AI builds, compute equipment, meaning accelerator chips, servers and their memory, is the costliest layer per megawatt. Power and cooling infrastructure and the building itself make up much of the rest. The exact split varies by design, location, chip generation and whether the owner builds the facility or leases space from a developer, so published estimates differ widely. ### Why does power equipment take so long to get? Large transformers, switchgear and turbines are custom engineered, use specialized materials such as electrical steel and copper, and are produced on a limited number of factory lines. Demand from data centers, utilities, renewable projects and electrification has risen faster than factory capacity. Wood Mackenzie reported transformer lead times of more than two years in 2025. ### Is a bill of materials the same as capital spending? Not exactly. A bill of materials lists the physical inputs to a facility. Capital spending, as reported by companies, also includes land purchases, construction labor, capitalized software and equipment for other purposes. Companies also lease data center capacity, which may appear as lease obligations rather than capital spending. The two measures overlap heavily but are not identical. ### Does every AI data center use the same components? No. Designs differ by purpose. Training clusters favor tightly linked accelerators, high-bandwidth networking and liquid cooling. Inference sites may use smaller clusters spread across more locations, sometimes with different chips. Climate, local power sources, water availability and the owner's preferences also shape the design, so component mixes vary from site to site. ### How do thematic ETFs use this framework? Some thematic ETFs organize their holdings around stages of the AI supply chain, such as semiconductors, networking, power and cooling. Others focus on a single layer. Reading a fund's methodology and holdings shows which layers it emphasizes. Investors can compare that mix with their own view of where spending and constraints are likely to sit, and with what they already own. ### Can the bill of materials change quickly? Yes. Each new chip generation can change power draw, cooling needs and networking requirements. A shift toward liquid cooling or optical links changes demand for specific components. Prices also move with supply and demand. These changes mean a company's share of AI spending can rise or fall as designs evolve, which is one reason thematic exposures carry concentration risk. ## NOTES - Microsoft, Alphabet, Amazon and Meta Platforms, Forms 10-Q for the quarter ended September 30, 2026, purchases of property and equipment; combined total as compiled by Walnut, updated November 8, 2026. - Gartner, worldwide AI spending forecast, as reported by RCP Magazine, September 21, 2026. - Wood Mackenzie, U.S. power transformer and distribution transformer supply analysis, second quarter 2025. - Samsung Electronics, first-quarter 2026 earnings call, and SK hynix, third-quarter 2025 earnings call, as compiled by IntuitionLabs, September 5, 2026. - Uptime Institute, 16th Annual Global Data Center Survey, 2026, as reported by Data Center Knowledge. - Nvidia, GB300 NVL72 AI factory reference architecture, as reported by Data Center Knowledge, 2026. ## IMPORTANT INFORMATION 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.