AI Infrastructure Beyond GPUs: Why Power, Cooling, and Networking May Create the Next Opportunities

Generative AI may look like a software revolution, but its expansion increasingly depends on physical infrastructure. Training a powerful model is only the first step. Running that model at commercial scale requires GPUs, high-bandwidth memory, advanced packaging, high-speed networks, data centers, cooling equipment, and a reliable supply of electricity.

NVIDIA’s fiscal 2026 results show how quickly this infrastructure cycle has expanded. The company’s full-year data center revenue reached $193.7 billion, up 68% year over year. Demand is no longer limited to individual processors; customers are purchasing integrated computing systems built around chips, networking hardware, memory, and software. As MEXC’s analysis of AI as a capital expenditure cycle explains, the next stage of the AI market may be determined not only by who develops the best model, but also by who can secure enough compute, electricity, network capacity, and deployment space.

The GPU Is Only One Link in the Compute Chain

Owning thousands of GPUs does not automatically create usable AI capacity. Large clusters require accelerators to exchange data at extremely high speeds. Information must move efficiently between processors, high-bandwidth memory, storage systems, and networking equipment. The chips must also be packaged into systems that can operate continuously without overheating or losing performance.

Advanced packaging has consequently become an important part of the AI supply chain. In its 2025 annual report, Taiwan Semiconductor Manufacturing Company said that demand for its CoWoS advanced packaging services had grown rapidly since 2023, driven by AI applications. CoWoS enables processors and high-bandwidth memory to be integrated more closely, improving the bandwidth and energy efficiency required by large AI systems.

A shortage of advanced packaging capacity can therefore delay a finished GPU from becoming part of an operational cluster. The same applies to high-bandwidth memory and networking equipment. A company may announce a large chip order, but the purchased hardware cannot generate revenue until the full system has been assembled, installed, connected, cooled, and made available to paying customers.

Investors assessing AI infrastructure demand should look beyond GPU shipments. Relevant indicators include advanced packaging capacity, memory pricing, networking revenue, system delivery times, and the amount of computing capacity that customers have actually activated. MEXC’s guide to NVIDIA’s data center business and key operating metrics also shows why NVIDIA’s position extends beyond chips. CUDA software, networking products, and complete systems create switching costs that cannot be measured through GPU sales alone.

Power Could Become a Harder Constraint Than Chips

Electricity may be one of the most difficult AI inputs to expand quickly. The International Energy Agency estimated in its 2025 Energy and AI report that global data center electricity consumption could rise from approximately 415 terawatt-hours in 2024 to around 945 terawatt-hours by 2030.

The problem is not simply whether the world can produce enough electricity in total. Data centers need large amounts of continuous power in specific locations. A suitable site also requires transmission lines, substations, transformers, grid connections, backup capacity, and regulatory approval. These physical systems can take much longer to build than a new server hall.

The IEA estimated that around 20% of planned data center projects could face delays if grid bottlenecks are not addressed. In advanced economies, a major transmission line can require four to eight years to complete, while demand for transformers, cables, and generation equipment has increased. A developer may have financing, land, and chips but still be unable to begin operations because the project has not secured a timely grid connection.

This constraint expands the AI opportunity beyond semiconductor companies. Potential beneficiaries may include producers of transformers and electrical equipment, grid construction companies, energy storage providers, natural gas generators, nuclear power developers, and data center operators that already control land with available power.

However, a proposed data center or electricity agreement should not be treated as confirmed revenue. Investors need to distinguish between an early development plan and a project with financing, regulatory approval, a signed interconnection agreement, a construction schedule, and a creditworthy customer. The most valuable asset may not be the land itself, but the ability to deliver power to that land within a commercially useful timeframe.

Cooling and Networking Determine Whether Compute Can Be Monetized

AI servers consume more power and produce more heat than conventional enterprise servers. As rack density increases, some facilities must replace or supplement traditional air cooling with liquid-cooling systems. That transition may require changes to racks, pipes, pumps, power distribution, and the overall design of the data center.

Cooling is not simply an operating expense. It directly affects how much computing equipment can be deployed in a building and how efficiently that equipment can run. If a cooling system cannot handle the required power density, the operator may need to reduce processor performance or leave part of the facility’s theoretical capacity unused. In that situation, expensive chips are installed but cannot generate their expected level of revenue.

Networking creates a similar problem. A large AI cluster is valuable only when its processors can communicate without spending excessive time waiting for data. Slow connections reduce the utilization of costly GPUs, particularly during model training, where thousands of accelerators may need to coordinate the same workload. This is why high-speed switches, optical components, interconnects, and data-processing units have become increasingly important parts of AI infrastructure spending.

Cooling and networking suppliers may benefit from the buildout, but the strength of their exposure varies. An investor should examine how much revenue has already been recognized, whether orders are cancellable, how concentrated the customer base is, and whether higher sales are producing stronger margins and free cash flow. MEXC’s analysis of which AI infrastructure companies are converting capital expenditure into revenue emphasizes measurable indicators such as recognized revenue, backlog, long-term commitments, and actual capacity expansion.

How to Tell Whether the “Picks and Shovels” Trade Still Works

MEXC’s view is that rising AI infrastructure spending does not make every upstream company an automatic beneficiary. The strongest opportunities are more likely to appear where demand is measurable, supply takes time to expand, and the supplier can maintain pricing power as revenue grows.

Six indicators can help test that thesis: capital expenditure by large cloud companies, revenue from paid AI services, data center occupancy, GPU utilization, grid-connection timelines, and supplier order margins. These measures connect spending with actual use. A rising capital expenditure budget is more convincing when customers are also consuming more AI services and operators are reporting high utilization.

The opposite combination should prompt greater caution. If capital expenditure continues rising while AI revenue, equipment utilization, operating margins, and free cash flow weaken, infrastructure spending may be creating excess capacity instead of durable returns. Large technology companies can afford to invest ahead of demand for a period, but suppliers that expand production on the assumption of uninterrupted growth may be more exposed if orders slow.

The thesis also has clear invalidation conditions. It would weaken if more efficient models sharply reduced the computing power required for common tasks, if enterprises cut their AI budgets, if cloud providers postponed data center construction, or if supply-chain expansion began to outpace actual usage. Improved efficiency does not necessarily reduce total electricity or compute demand because lower costs can encourage more applications, but investors should not assume that this rebound effect will always be large enough to absorb new capacity.

Risk Warning

AI infrastructure projects typically require large upfront investments, long construction periods, and dependence on a small number of major customers. Grid delays, higher financing costs, rapid equipment obsolescence, contract cancellations, or spending reductions by hyperscale cloud providers can directly affect project returns.

Growth in data center electricity consumption also does not prove that any particular semiconductor, utility, cooling, networking, or data center company will become more profitable. Company-level analysis should still consider contract quality, capacity utilization, customer concentration, debt, margins, and free cash flow. AI infrastructure may remain a powerful investment theme, but the economic value will not necessarily be distributed evenly across every company associated with it.