AI power supply market grows ninefold in ten years: data center power architecture enters a restructuring phase

An important shift is taking place in the global AI infrastructure competition: from "scrambling for computing power" to "competing for electricity." The power supply units supporting GPU clusters are evolving from easily overlooked components inside servers into a key factor that determines whether AI projects can be launched on schedule. According to the latest "AI Power Supply Units (AI PSU) Market" report released by Fact.MR, the global AI power supply unit market will grow from $698 million in 2026 to $6.298 billion in 2036, at a compound annual growth rate of 24.6%. In other words, over the next decade this market will expand about ninefold, adding approximately $5.6 billion of absolute opportunity to the industry chain.

Power suppliers covered in the report include Delta Electronics, Lite-On Technology, Flex, Murata Manufacturing, and Advanced Energy Industries. Their products are shifting from standardized server power supplies to a new generation of power solutions adapted to high-density GPU clusters and hyperscale data centers.

Background: AI Data Center Power Supply Pressure Comes to the Fore

The chain reaction from surging AI computing demand is spreading from GPU chips to the power supply and distribution systems of entire data centers. Traditional cloud data centers are mostly CPU-oriented, with per-rack power typically ranging from single-digit kilowatts to over ten kilowatts, while high-density GPU racks in AI training clusters can require several times the power of traditional racks. This change means that existing power supply, distribution, and cooling systems must be upgraded as a whole.

Fact.MR's data reflects this upgrade from multiple dimensions. By output power, PSUs in the 1000W to 3000W range account for 46% of the market share, making them the mainstream configuration; high-power PSUs above 3000W occupy another 15%, mainly serving hyperscale data centers. Meanwhile, entry-level products from 500W to 1000W still account for 34%, driving parallel growth across all tiers of products.

In terms of form factor, hot-swappable PSUs dominate with a 43% share, modular PSUs account for 25%, and OCP open computing standard power supplies account for 15%. Hot-swappable and modular designs help operations teams replace or expand power supplies without interrupting tasks, which is crucial for AI workloads that often require continuous training for weeks or even months. The growth of OCP power supplies reflects that more and more hyperscale data centers want to escape vendor lock-in and adopt open-standard power systems.

On the cooling front, air-cooled PSUs still hold a 58% share, but liquid-cooled PSUs have rapidly risen to a 42% share. As the power consumption of individual AI servers continues to climb, traditional air cooling is increasingly unable to meet heat dissipation requirements in high-power-density environments. Incorporating power modules into liquid cooling loops has become a prominent technology direction for next-generation data centers.## Technical Analysis: Four Key Differences Between AI PSUs and Traditional Server Power Supplies

Unlike traditional power supplies that previously powered telecom or general-purpose servers, AI power supply units face loads with higher power, stronger fluctuations, and stricter availability requirements.

First is the power rating. Standard servers typically use one or two CPUs, so the power supply burden is limited; AI servers, however, often carry multiple GPUs, causing peak current and sustained power consumption to increase exponentially. Therefore, the output of a single module has risen from hundreds of watts in the past to thousands of watts today—the report's finding that 1000W–3000W units have become the market mainstream is proof of this.

Second is load management. Power supplies with multiple output rails can split electrical energy across different lines, avoiding single-point overloads and reducing the impact on other hardware when one line fails. Multi-rail PSUs' 62% market share in 2026 illustrates, from one angle, that AI servers demand far higher system stability than ordinary servers.

Third is availability design. The popularity of hot-swappable and modular form factors shows that data center managers no longer care only about whether a power supply "works," but also about "whether it can be replaced quickly after it fails." The cost of interrupting AI training tasks is extremely high, and power supply replacement time directly affects the effective utilization of GPU clusters.

Fourth is the cooling method. High-power power supplies still generate heat during energy conversion. When the power consumption of a single unit reaches several kilowatts, air-cooling fans not only consume more electricity, but also affect the temperature distribution in the machine room. Liquid-cooled power supplies use coolant to directly remove heat from components, forming a complete thermal management system together with GPU and CPU liquid cooling schemes. This is also one of the reasons why liquid-cooled PSUs hold a 42% share in the report.

In addition, high conversion efficiency has become a hard metric. Higher power conversion efficiency means less electricity is wasted as heat, and it also reduces the burden on UPS and cooling systems. The adoption of titanium-grade and more efficient power supplies will directly affect the PUE and overall electricity costs of AI data centers.

Enterprise Impact Analysis: Power Infrastructure Should Be Included in AI Financial Models

The AI PSU market is not particularly large relative to the entire IT infrastructure, but its growth curve—from approximately $560 million in 2025 to approximately $6.3 billion in 2036—is enough to prompt enterprises to rethink their AI deployment strategies.

For decision-makers building self-managed or customized data centers, the impact is first reflected in capital expenditures. If they plan to deploy AI compute capacity, they must reserve budget for high-power racks, liquid cooling piping, backup power, and intelligent power distribution. The power supply unit itself is not the largest cost item, but the associated power supply and cooling retrofits will significantly increase a project's initial investment.

Second, operating expenditures also need to be recalculated. High-power AI clusters will drive up energy consumption, and differences in power supply efficiency will amplify over time. Taking a campus-level data center as an example, even a single percentage point improvement in power efficiency could correspond to annual electricity cost differences of millions of yuan across tens of thousands of servers. Therefore, selecting high-efficiency power modules and rack-level power supply solutions should be incorporated into TCO calculations.Third, this imposes requirements on availability and operations/maintenance organization. Hot-swap, redundant N+1 or 2N configurations are no longer optional designs in AI business scenarios. If a power failure interrupts training tasks, it can waste hundreds of thousands of dollars in computing costs. Enterprise IT needs to redesign monitoring, alerting, and maintenance processes so that power systems have predictive maintenance capabilities.

Enterprises that heavily purchase public cloud AI services should not ignore this trend either. To control costs and meet sustainability goals, cloud service providers will gradually reflect the benefits of power efficiency improvements, or the capital pressure of related retrofits, in GPU instance pricing and regional availability. When selecting cloud regions, enterprises may need to assess local power supply and data center cooling capacity.

Market Competition: Who Benefits, Who Faces Pressure?

From a market landscape perspective, the growth of AI PSUs will first benefit leading suppliers with high-power power supply design and liquid cooling integration capabilities. Delta Electronics, Lite-On, Flex, Murata, and Advanced Energy, among others, have deep expertise in power electronics and manufacturing processes, giving them a better chance of securing long-term orders from hyperscale customers. New entrants, by contrast, must simultaneously master high-power-density design, thermal management, and intelligent manufacturing capabilities—the technical barrier has risen significantly.

By region, North America and Asia-Pacific remain the main markets. India and China, with expected growth rates of 28.9% and 27.6%, respectively, are among the fastest-growing markets globally. This indicates that the focus of global AI data center construction is gradually shifting from traditional mature markets to emerging markets with rapid population and digital economy growth. China’s continued investment in AI computing infrastructure, along with the expansion of multinational enterprises and foreign-invested data centers in India, will generate dense demand for power equipment over the next decade.

The pressure within the industry falls mainly on traditional small and medium-sized data center operators. These operators often lack sufficient capital and technical capability to retrofit high-density power supply and liquid cooling systems. If they cannot meet new AI hosting demand, they may lose bargaining power in the computing power competition. At the same time, the rise of OCP-standard power supplies with lower vendor lock-in will also impact the traditional closed server power supply supply chain.

Industry Trend Watch: AI Infrastructure Competition Is Moving from Chips to Energy

The rapid growth of the AI power supply market sends a clear signal to the market: the focus of competition in AI infrastructure is shifting from “who has more powerful GPUs” to “who can power and cool GPUs at lower cost.”

Over the past few years, enterprises have often focused on AI chips, interconnect bandwidth, and training frameworks when making technology choices. As the power density of GPU clusters continues to rise, rack-level power supply architecture, cooling methods, and energy procurement strategies are becoming new variables that determine the economics of computing power. It is foreseeable that AI data center planning will move closer to the logic of “energy infrastructure projects”: assessing local grid capacity, renewable energy supply, cooling resources, and physical space.This shift also means that the concept of the "AI-native cloud" will be extended. Future cloud infrastructure will not only be software-defined, but also energy-defined. Whoever can reduce energy consumption per unit of computing power through rack-level power optimization, liquid cooling, and open standards will hold greater pricing power in future AI cloud service competition.

For enterprise architects, now is a critical window to reserve higher power density, liquid cooling piping, and power redundancy capacity in data center design. Even if current business has not yet adopted GPUs at scale, the next wave of AI applications could change the compute workload structure in the short term. Power and cooling systems should not become the ceiling for future AI business growth.

CloudTechDaily Insight

The nearly ninefold growth of the AI PSU market over the past decade should not be simply understood as an undersupply of a server component. What it truly points to is that the bottleneck of the AI industry is shifting from chip manufacturing to power infrastructure. As GPU compute supply continues to expand, the efficiency and availability of power systems determine the real cost of computing power and the speed at which it can be brought online. If cloud service providers and large enterprises want to build long-term advantages in AI competition, they must treat power supply and cooling as first-class citizens of IT infrastructure, rather than retrofitting them after systems are deployed. The segment-level changes in the power supply market—high power, liquid cooling, modularization, and OCP standards—also provide a clear technology roadmap for the industry. The future AI data center will be more like a highly integrated power-and-computing system than a simple collection of servers. For enterprise technology decision-makers, the question that now needs to be answered is not "whether to use GPUs," but "what scale of AI can my power architecture support." Incorporating power budgets into strategic planning as early as possible will determine whether a company becomes a frontrunner in this round of the AI race, or a latecomer that passively pays higher costs.

---

Source: Fact.MR - AI Power Supply Units (AI PSU) Market