Artificial intelligence is often discussed through the companies that design chips or develop software. Those businesses matter, but neither layer can create economic value without the infrastructure that connects them.

This is Post 5 in our eight-part AI investment series, following the value chain from AI Semiconductor Supply Chain → AI Compute & Networking → AI Intelligence. The series anchor, “AI Is Bigger Than AI Stocks: The Eight-Layer Value Chain Every Investor Should Understand,” can be found in the Articles and News archive.

The AI Compute & Networking layer is the engine room beneath every model. It includes servers, storage systems, networking equipment, optical components, cloud infrastructure, and the data centers that house them. Semiconductors provide the processing capability. Compute infrastructure assembles that capability into usable systems. Software and intelligence layers then run on top of it.

That structure has an important investment implication: AI is a connected economic value chain. Every layer depends on the layers beneath it, and a portfolio can unintentionally count the same AI exposure multiple times.

Why Compute and Networking Are Necessary

A high-performance AI chip is not a complete product for a model developer or enterprise customer. It must be integrated into a server, connected to other servers, supplied with data, cooled, powered, and managed through specialized software.

Training and operating large models require thousands: or, in the largest deployments, many thousands: of processors working together. The system must move data rapidly between:

• Individual processors inside a server
• Multiple servers within a rack
• Server racks across a data-center cluster
• Separate data centers and regions
• Compute systems and high-capacity storage

This creates several infrastructure requirements.

Servers package processors, memory, power delivery, cooling, and management systems into deployable units. AI-optimized servers are materially different from conventional enterprise servers because they must support high-power accelerators, specialized interconnects, and dense thermal-management systems.

Networking equipment allows processors to communicate with one another. In AI workloads, network performance can directly influence how efficiently expensive compute capacity is used. A bottleneck in the network can leave processors waiting rather than calculating.

Storage holds the enormous datasets used to train and fine-tune models, along with model weights, checkpoints, logs, and customer data. Training systems need both high capacity and high throughput. Storage is not simply an archival function; it is part of the performance profile of the AI cluster.

Cloud infrastructure converts these physical assets into a service. Customers can access computing capacity through infrastructure-as-a-service, rather than purchasing and operating an entire data center themselves.

The result is a layered system in which chips, servers, networks, storage, power, cooling, and software must operate as a coordinated whole. The direct benefit is higher utilization, better reliability, and more predictable access to compute.

Demand Drivers: From Hyperscaler Capital Expenditure to Network Speed

The immediate demand cycle is being driven by substantial capital expenditures from hyperscalers and other large technology companies. Amazon, Microsoft, Alphabet, Meta Platforms, and Oracle are among the major providers investing in data centers, servers, networking, power systems, and cloud capacity.

The relevant question for investors is not simply whether capital expenditure is increasing. It is how that spending is allocated and whether it produces durable revenue, acceptable returns on invested capital, and recurring customer demand.

Several infrastructure trends are especially important.

AI-Optimized Server Demand

AI servers typically contain high-value accelerators, large memory configurations, advanced cooling systems, and high-speed network interfaces. Their dollar content can be substantially higher than that of general-purpose enterprise servers.

However, higher revenue does not automatically mean higher profitability. Server manufacturers face component costs, customer pricing pressure, supply-chain constraints, and rapid product refresh cycles. Investors should distinguish between reported shipment growth and sustainable margins.

Higher-Speed Networking

As clusters grow, the network must move more data with lower latency. The industry is transitioning through 400G and 800G Ethernet and optical connections, while 1.6-terabit technologies are beginning to appear in advanced road maps and deployments.

The exact mix depends on cluster size, architecture, distance, power constraints, and whether the customer uses Ethernet, InfiniBand, or a hybrid approach. In general:

• 400G remains important for many server-facing and existing deployments.
• 800G is increasingly central to new, large-scale AI fabrics.
• 1.6T represents the next step in bandwidth density for the highest-performance systems.

Optical transceivers, fiber, direct-attach copper cables, active cables, and related components become increasingly important as electrical signals face distance, power, and signal-integrity limits. NVIDIA’s LinkX interconnect overview provides a useful technical illustration of how cables, transceivers, InfiniBand, and Ethernet fit together within AI infrastructure.

Scale-Up and Scale-Out Architectures

AI infrastructure uses two complementary approaches.

Scale-up connects processors within a tightly integrated system so that they can share data rapidly. This approach emphasizes memory bandwidth, low latency, and specialized interconnects.

Scale-out connects many systems across a broader cluster. It emphasizes switching capacity, network topology, workload scheduling, reliability, and the ability to expand without rebuilding the entire environment.

The distinction matters because different suppliers benefit from different infrastructure decisions. A company focused on servers may have limited exposure to the optical or switching components required for scale-out. Conversely, a networking supplier may benefit from cluster expansion even when individual server configurations change.

Storage for Training Data

Large language models and other AI systems require extensive datasets, model checkpoints, and continuous data movement. High-performance solid-state storage, storage networking, data-management software, and capacity-oriented systems all participate in this demand.

Storage demand can be uneven. Customers may initially build performance storage for training and later add less expensive capacity for inference, archives, and backup. This creates opportunities across different storage categories but also introduces pricing pressure and inventory risk.

Potential Investment Opportunities Across the Layer

The Compute & Networking category includes several distinct business models. Representative public companies and exchange-traded funds can help investors identify areas for further research, but they are illustrative examples: not recommendations.

Server OEMs and ODMs

Server original equipment manufacturers and original design manufacturers assemble systems for enterprise, government, cloud, and specialized customers. Examples include Dell Technologies, Hewlett Packard Enterprise, and Super Micro Computer among branded or system-focused vendors. Quanta Computer and Wiwynn are examples of large-scale manufacturing and ODM exposure.

The investment case depends on more than server revenue. Investors should examine:

• Backlog quality and cancellation terms
• Customer concentration
• Gross-margin stability
• Working-capital requirements
• Dependence on a small number of accelerator suppliers
• The durability of AI-related demand after initial deployments

Networking Equipment and Optical Components

Networking exposure can include switch manufacturers, network silicon providers, optical-transceiver companies, fiber suppliers, and component makers.

Examples for research include Arista Networks, Cisco Systems, Broadcom, Coherent, and Lumentum. These companies do not represent identical exposures. Some sell equipment, some provide semiconductors or optical components, and some serve broader communications markets in addition to AI data centers.

Investors should separate genuine order growth and backlog visibility from broad AI-related narrative. A networking company may benefit from AI cluster expansion without being a pure-play AI business.

Cloud Providers and Infrastructure-as-a-Service

Cloud providers allow customers to rent compute, storage, networking, and related services. Examples include Amazon Web Services, Microsoft Azure, Google Cloud, and Oracle Cloud Infrastructure.

Cloud economics require careful analysis. Revenue growth can be strong while free cash flow is constrained by data-center construction, power commitments, equipment purchases, and depreciation. Investors should assess:

• Capital expenditure relative to cloud revenue
• Capacity utilization
• Contracted customer commitments
• Pricing and competitive intensity
• Depreciation schedules and asset lives
• The return expected from new data-center capacity

Cloud exposure can be strategically valuable, but it is also capital intensive.

Data-Center Owners and Infrastructure Providers

Data-center owners, colocation providers, and infrastructure landlords participate in the physical expansion of AI capacity. Equinix and Digital Realty are well-known examples of publicly traded data-center REITs. Other infrastructure businesses provide power, cooling, construction, equipment leasing, or specialized financing.

These companies may benefit from long-term demand for data-center capacity, but their risk profiles differ from those of chip or software companies. Leverage, interest rates, power availability, tenant concentration, lease terms, and construction costs are central considerations.

Principal Risks Investors Should Monitor

The infrastructure opportunity is substantial, but the risk profile is not passive or uniform.

Hyperscaler Customer Concentration

A small group of large technology companies accounts for a significant share of industry spending. Suppliers may report strong growth while remaining dependent on a few customers with considerable negotiating power.

A change in one customer’s capital-expenditure plans can affect orders, inventory, and revenue across multiple vendors.

Rapid Technology Refresh Cycles

AI infrastructure evolves quickly. A transition from one generation of processors, servers, network switches, or optical modules to another can create demand: but it can also make existing inventory less valuable.

Companies must manage a difficult balance between stocking enough components to meet demand and avoiding excess inventory when designs change.

Competition and Margin Pressure

Attractive markets invite competition. Large system vendors, specialized manufacturers, cloud providers, and internal hyperscaler design teams may all compete for portions of the same economic value.

Higher unit volumes do not guarantee higher earnings. Pricing pressure, customer customization, warranty costs, component shortages, and manufacturing complexity can reduce operating leverage.

Cloud Capital Intensity

Cloud providers must invest before all capacity is fully utilized. Data centers require land, power, cooling, networking, security, and long-lived equipment. If demand growth slows or customers shift workloads, the provider may face underutilized assets and lower returns.

Interest Rates and Financing Conditions

Data-center construction and equipment deployment require substantial financing. Higher interest rates can raise the cost of capital, reduce the value of long-duration infrastructure assets, and pressure highly leveraged businesses.

These risks make balance-sheet analysis and cash-flow review essential. The direct benefit of disciplined risk assessment is a clearer distinction between durable infrastructure demand and a temporary capital-spending surge.

Near-Term Buildout Versus Long-Term Secular Demand

The current AI infrastructure cycle has a near-term component and a long-term component.

The near-term opportunity is the physical buildout: new data centers, accelerator servers, high-speed switches, optical links, storage systems, and cloud capacity. This phase can produce rapid revenue growth for suppliers with strong backlogs and favorable customer relationships.

The long-term case is broader. AI workloads may become embedded in enterprise software, cybersecurity, customer service, scientific research, industrial processes, and public-sector operations. If usage expands, the need for compute, storage, and connectivity should continue beyond the initial construction cycle.

Those outcomes are not guaranteed. The durability of demand will depend on model efficiency, inference volume, customer willingness to pay, energy availability, and the ability of businesses to generate measurable productivity or revenue from AI applications.

The Compute & Networking layer also connects the semiconductor and software layers. Improvements in processor performance can increase demand for servers and networks, while better software efficiency can reduce the amount of hardware required for a given workload. A portfolio that owns companies in both areas may have legitimate diversification: or it may be counting the same spending cycle twice.

Portfolio Relevance: Avoiding Double-Counting AI Exposure

AI exposure should be evaluated across the entire portfolio, not security by security.

An investor may own a semiconductor company, a server manufacturer, a networking company, a cloud provider, and a data-center REIT. These businesses have different financial models, but they may still depend on the same hyperscaler capital-expenditure cycle. If that cycle slows, several holdings could decline together.

A disciplined review should consider:

• The percentage of revenue each company derives from AI-related demand
• Customer and supplier concentration
• Whether exposure is tied to hardware sales, cloud usage, or recurring software revenue
• Capital intensity and balance-sheet leverage
• Valuation relative to normalized cash flow
• Correlation among holdings during a technology-spending slowdown
• The role of each position within the household’s broader investment plan

The objective is not to avoid AI exposure. It is to understand the exposure being purchased and prevent overlapping holdings from creating an unintended concentration.

The Compute & Networking layer is indispensable because it turns semiconductor capability into usable infrastructure and provides the foundation on which AI intelligence and applications operate. For investors, the central question is whether a company is supplying durable infrastructure with credible economics: or simply participating in a crowded narrative.

Continue following the Articles and News archive for the next installment, which examines the AI Intelligence layer: the models, platforms, and software systems that use this infrastructure.

Disclosure

This article is provided for general informational and educational purposes and does not constitute investment, tax, or legal advice. References to companies, securities, technologies, or exchange-traded funds are for illustrative purposes only and are not recommendations to buy or sell any security. InSight Financial Planners is a Registered Investment Advisor. Investment involves risk, including possible loss of principal. Past performance does not guarantee future results. Investors should consider their objectives, risk tolerance, time horizon, tax circumstances, and overall financial plan before making investment decisions. Linked third-party materials are provided for reference, and InSight Financial Planners does not endorse or independently guarantee their accuracy.