An Investor’s Guide to the AI Infrastructure Stack

Inside the AI ecosystem

Few innovations have evolved as quickly as AI, and the language surrounding its ecosystem has expanded just as fast. To help demystify the AI value chain, Allspring Global Investments has built a useful framework for identifying where investment dollars are being spent and where attractive growth opportunities are emerging.

Most people interact with AI through end-user applications like ChatGPT, Claude, Gemini, or Copilot. These apps are developed by AI model providers, sometimes referred to as “AI labs,” that, despite their novelty, generate revenue much like a cellular provider. Some users pay a monthly subscription while others are billed based on users’ token consumption, which represents the compute power required to perform a task. Yet, every AI interaction, no matter how complex or simple, depends on a scaffolding of infrastructure, with different participants capturing economic value at each step.

Who supplies the technology that enables the AI models? The key layers of the value chain, or infrastructure “stack,” are below:

Hyperscalers: Supplying cloud capacity

Cloud-computing services are the first ingredient required for training and deploying AI models. As AI adoption accelerates, demand for cloud-computing increases on a massive scale.

• Hyperscalers are large tech companies supplying the compute and physical infrastructure required to deliver AI. These companies are fighting to win the largest share of cloud-computing business to become the “toll roads” for AI and they are investing heavily to do so. The largest hyperscalers are Microsoft, Amazon, Google, Meta, and Oracle.

• “Circular” relationships: Many model providers maintain strategic relationships with these hyperscalers. For example, OpenAI has ties to Microsoft Azure and Oracle, while Anthropic uses both Amazon Web Services and Google Cloud. These relationships are viewed as circular because the hyperscalers are some of the largest investors and customers of the model providers themselves. These alliances are critical to the ultimate deployment of AI applications to consumers and enterprises alike.

• Neoclouds are similar to hyperscalers in that they supply cloud-computing services but are smaller and more specialized. Neoclouds specialize in building pathways for efficient AI traffic so that models can be trained and run at scale. Leading neoclouds are currently CoreWeave, Nebius, Hut 8, and DigitalOcean, though each occupies a slightly different niche (be it model training, physical power, or deployment for enterprises) in the industry.

Semiconductors: Processing AI workloads

Semiconductors play a critical role in the AI ecosystem, and technical experts often discuss this layer in terms of “racks.” Racks are large systems packed with dozens of processing and memory chips (GPUs) and networking equipment that work together as a single processing unit to develop large-scale AI models.

This layer captures a disproportionate share of hyperscalers’ spend because it is arguably the most important in the stack. To illustrate its magnitude, a single rack costs several million dollars and data centers often house tens of thousands of racks. The rest of the AI infrastructure stack exists primarily to support and scale “compute” or processing capacity. Key components include:

• Graphics processing units (GPUs): Often called “accelerators,” GPUs are high-performance chips that power AI by accelerating massive mathematical calculations, integrating multiple processing or “logic” chips into a single unit through what’s called “advanced packaging.”

• NVIDIA captures a significant portion of the economics of this layer because its GPUs are the default for training and deploying AI models. Advanced Micro Devices also designs processing chips and is a primary competitor to NVIDIA.

• ASICs and foundries: Application-specific integrated circuits (ASICs) are specialized processing chips designed to perform specific AI tasks with greater speed and efficiency and lower power consumption than general-purpose GPUs. Similar to GPUs, most ASIC companies operate on a fabless model, meaning they design chips but rely on third-party foundries to manufacture them. Broadcom, Marvell, and Cerebras design ASICs and rely on foundries including TSMC, Samsung, and Intel for manufacturing.

• Memory and storage: Dynamic random-access memory (DRAM) and ultra-fast high-bandwidth memory (HBM) chips integrated with GPUs supply data for active processing, while NAND flash storage sits outside the GPU and holds data until needed. Leading memory providers include Micron, SK hynix, and Samsung, while leading storage providers include Sandisk, Western Digital, and Seagate.

Data centers: The physical infrastructure

Data centers house all the physical equipment required to host AI workloads. Companies supplying technologies for this foundational infrastructure layer benefit from the hyperscalers’ massive investment in increased AI capacity. Examples of such technologies include:

• Power: AI requires massive amounts of power, and by 2030, it is estimated that data centers will consume around 1,000 terawatt-hours of electricity annually—roughly the same amount as Japan’s total annual electricity consumption today¹. To meet this demand, a massive investment to modernize and expand U.S. electrical infrastructure will be needed, with power companies like Quanta Services benefiting from custom services and labor. Data centers are also increasingly adopting “behind the meter” solutions or specialized generators from providers like Bloom Energy or GE Vernova to power buildings without needing to rely on the long and cumbersome process of connecting to the existing electrical grid.

• Cooling and thermal management: Data centers generate substantial amounts of heat that must be continuously managed to maintain reliable operations. Industrial-scale HVAC systems (from companies like Comfort Systems and EMCOR) help maintain optimal temperatures, while key mechanical components such as the pumps, valves, and fluid-management devices (from firms like Johnson Controls and Modine Manufacturing) are crucial for regulating temperatures inside the buildings. The actual racks themselves also have their own liquid cooling solutions, and Vertiv Holdings is one of the key suppliers in this area.

• Interconnect: This refers to the cables that transmit the power and signals that make AI possible. Cabinets and containment systems provide organized structure for efficient power delivery and airflow within dense compartments of equipment. In terms of connection materials, copper is widely used and highly effective but heat intensive. Fiber-optic connectors offer greater efficiency but require more precise manufacturing, installation, and testing. Credo, Astera Labs, and Amphenol have historically offered copper interconnect but are increasingly offering optical as well. Viavi Solutions provides the testing and measurement needed to validate high-speed optical networks at scale.
• Colocation: Real estate providers such as Equinix, Digital Realty, QTS, and CyrusOne develop and operate data center facilities. They often lease space and provide services to hyperscalers and other customers rather than owning AI equipment themselves. Within the data centers, automation technologies include the sensors and control systems that optimize power and facility operations with Schneider Electric and Siemens supplying these technologies.

Allspring Growth team perspective: Go deeper into the stack

The scale of the AI infrastructure buildout is unlike anything this country has seen in modern times, only potentially rivaled in U.S. history by the railroad boom in the 1800s. In the first stage of the AI investment cycle, hyperscalers and mega-cap tech giants such as Amazon, Microsoft, Alphabet, Micron, and NVIDIA dominated headlines and captured the lion’s share of investment returns.

But an important shift is now underway, with investors increasingly rewarding the recipients of AI capital spending rather than the spenders themselves. The focus has broadened from the companies funding the buildout to the companies enabling it, including infrastructure providers that are helping scale capacity and address key constraints in areas such as power and storage.

Next-generation data centers are measured in terms of gigawatts, or the amount of compute capacity generated by 1 billion watts of power. A single gigawatt is similar to the amount of power consumed by a metropolitan area like Indianapolis, Indiana with around 800,000 households². For context, the amount of capital required to stand up a 1-gigawatt data center is estimated to be around $35 billion and some believe nearly 100 gigawatts in data centers will be added by the end of the decade³.

Linked to this power buildout, businesses in the industrials sector are increasingly in focus. These companies often have order backlogs extending several years into the future, helping create durable and visible earnings growth, and easing investors’ fears around speculative excess.

Furthermore, the market is broadening and beginning to reward small and medium-sized companies with innovative solutions to modern challenges. One area our research is focused on is the power and cooling of massive data centers. Temperature control can take up as much as 40% of data center energy usage overall, so there is an immense financial incentive to come up with cutting-edge solutions to become more efficient⁴.

Ultimately, investing in AI today involves searching for opportunity beyond the headline names. That’s why we believe investors who look deeper into the infrastructure stack may be positioned to benefit during the next stage of AI growth. Active managers can uncover compelling opportunities among the recipients of AI-related capital spending through fundamental research across industrial, power, technology, and infrastructure companies. As the AI buildout evolves, we are increasingly focused on areas where demand is outpacing available capacity. These bottlenecks often create attractive opportunities for the businesses supplying the infrastructure, equipment, and services needed to unlock additional scale.

¹Source: International Energy Agency (IEA), Energy and AI, April 2025

²Source: Empirical Research Partners, July 2026

³Source: Jefferies, September 2025

⁴https://www.energy.gov/articles/doe-announces-40-million-more-efficient-cooling-data-centers

⁵https://energy.mit.edu/current-initiatives/data-center-power-demand/

The securities listed in this article should not be considered a recommendation to purchase or sell any particular security

This material is provided for informational purposes only and is intended for retail public distribution in the United States. Use outside the United States is for professional/qualified investors only.

CFA® and Chartered Financial Analyst® are trademarks owned by CFA Institute.

ALL-09302026-guxnsrdl