AI Power Industry Transformation: From AI "Supporting Component" to Future "Computing Throat"

```

According to LP Information data, in the AC-DC power supply segment for AI servers alone, the global market size is expected to grow from $2.089 billion in 2025 to $9.521 billion in 2032, with a CAGR as high as 21.1%.

The AI power industry is standing at a dual inflection point of "technological revolution" and "market expansion." On the technology side, as single-cabinet power moves toward the megawatt level, traditional AC UPS power architectures are approaching their physical limits. New-generation power supply technologies, represented by 800V high-voltage DC (HVDC), solid-state transformers (SST), and vertical power delivery (VPD), are moving from concept to large-scale application.

1. What’s happening? AI Power Revolution Imminent

Recently, feedback from both mainland manufacturing chains and core Taiwanese power companies shows that HVDC is progressing faster. It is expected that ±400V HVDC will begin mass production in Q3, and see large-scale adoption in 2027. Technical pathways are evolving from ±400V to 800V and then to ±800V. HVDC will rapidly inflate the value created for power supply enterprises, from the previous 2 RMB/W for UPS, to 5-6 RMB/W for HVDC. Currently, only 4-5 companies have actual HVDC prototypes, so industry leaders can fully enjoy the cyclical dividends.

We believe the AI power industry is at a dual inflection point of "technological revolution" and "market expansion." On the technology side, as single-cabinet power moves toward the megawatt level, traditional AC UPS supply architectures are reaching physical limits. New technologies like 800V HVDC, SST, and VPD are moving from concept to mass deployment. On the market side, according to LP Information, the global AI server AC-DC power supply market will expand from $2.089 billion in 2025 to $9.521 billion in 2032, CAGR 21.1%. If HVDC rack power, BBU, DC-DC converters and broader systems are included, the market potential is even larger.

The explosive growth of AI computing power is pushing the power density of data centers to unprecedented levels. At the chip level, Nvidia GPUs’ TDP has surged from 250W in the Maxwell era to over 1000W for the B200 chip, and will keep rising. At the rack level, designs like NVL72 (Nvidia), Superpod (Google), and others mean single-rack power is moving swiftly from tens of kilowatts toward 100kW or even the megawatt level. Gartner predicts that, by 2027, 40% of existing AI data centers will have operations restricted due to insufficient power supply.

Traditional data centers widely use AC UPS power systems, which operate on a typical "AC-DC-AC" double-conversion. Utility power is rectified to DC (to charge batteries), then inverted back to AC to supply the load. This framework was effective when power density demands were low, but it has shown three core deficiencies in the AI era:

First, efficiency ceiling: Double conversion produces heat losses, and more importantly, to ensure reliability, UPS systems usually adopt N+1 redundancy, so individual load rates are often limited to 35-53%. Real conversion efficiency is often below 80%.

Second, unsustainable space and cost: For example, for equipment with NVIDIA GB200 NVL72, if 54V DC distribution is used, copper busbars alone would require up to 200 kg. For a 1 GW data center, rack busbars would require up to 500,000 tons of copper.

Third, 48V busbars can no longer support megawatt-level racks. At MW power levels, using 48V DC busbars would mean currents of tens of thousands of amperes, not only creating shocking line losses, but power architectures would occupy most cabinet space, leaving nowhere for compute equipment.

2. Why Is It Important? Technology Keeps Evolving

To address the above bottlenecks, NVIDIA and OCP (Open Compute Project) have proposed a clear iteration roadmap:

Phase 1 (current mainstream): 415/480VAC AC bus + 50V DC output. Suitable for single-rack power <72kW, and still the common configuration for most data centers.

Phase 2 (medium-term transition, 2026-2027): Introduction of HVDC sidecar architecture. Power supply system is separated from the IT rack and placed in side power racks, interconnected with IT cabinets via an 800V high-voltage DC bus. This architecture has already been adopted in NVIDIA’s Kyber rack system.

Phase 3 (long-term goal, 2027-2028): Facility-level HVDC + Solid-State Transformer (SST). All AC-DC conversions are completed centrally at the facility level, converting 10kV medium-voltage AC directly into 800V DC, sent via high-voltage DC bus to each rack, with DC-DC conversion at the end directly powering chips. This plan greatly reduces conversion stages and improves efficiency.

The AI power market consists of several sub-tracks, including AC-DC server power supplies, HVDC rack power, DC-DC converter modules, BBU/CBU, etc. According to LP Information, in the AC-DC server power segment alone, global market size is expected to grow from $2.09 billion in 2025 to $9.52 billion in 2032, CAGR 21.1%.

QYResearch data confirm this trend, expecting the global AI PSU market to grow from $2.185 billion in 2025 to $5.532 billion in 2032, CAGR 14.4%. Over the next seven years, the AI power market will maintain a compound growth of over 15~20%.

Traditional server power supplies are highly standardized products with fierce price competition. But in the AI server field, the value of power supplies is being redefined.

First, single module power improvement drives up single unit price. Oulu Tong's high-power server power supplies (2000W-4000W) sell for 1,000~4,000 RMB each, while standard server supplies (800W-2000W) are only 400~1,000+ RMB [citation: Shenwan Hongyuan]. As module power increases to 5.5kW, 8kW, 12kW, prices may rise further.

Second, 80Plus Titanium certification brings a premium. AI servers have extremely high energy efficiency demands, generally using Titanium-level (96%+ efficiency) power supplies, whose design and component costs are much higher than regular power supplies.

Third, HVDC and SST schemes’ system-level value. HVDC sidecar racks, SST, and so on are not just single power modules, but system-level products integrating distribution, conversion, backup and monitoring, with values reaching hundreds of thousands of RMB or more.

TrendForce analysis points out that currently, IT rack power systems account for about 3-5% of data center capital expenditures. As high-voltage DC is introduced, this proportion is expected to rise.

According to Precedence Research’s statistics and forecasts, new global data center IT-side installed capacity will surge from 10.5GW in 2024 to 40.3GW in 2030. By 2028, North America's new AI data center capacity may reach 71GW.

3. What’s Next? Who Can Reshape the Power Landscape

To quantify the investment logic, we simulated the relationship between AIDC power demand and power module value growth using Python.

①HVDC

High-voltage DC (HVDC) is not a new concept—China started exploring 240V HVDC for telecom as early as 2007. But in the AI era, HVDC’s role is transforming—from an "energy-saving alternative" to "the only feasible solution supporting megawatt-level cabinets."

NVIDIA 800V ecosystem: NVIDIA has clearly stated that new-generation AI factories will gradually transition from AC distribution to 800V DC distribution [citation: Tianfeng Securities]. Its Kyber rack architecture uses 800V DC distributed directly to each node; single-stage conversion brings higher space utilization and electrical efficiency, with rack volume reduced by 26% and energy consumption down about 8% in tests.

In October 2025, NVIDIA officially announced its full list of 800V DC power architecture partners, covering the full industry chain from chips and power components to system integration.

②SST: "Chip-ization" of Power Systems

Solid-state transformer (SST) is seen as the ultimate HVDC solution. Using power electronics, it transfers and converts energy, converting 10kV medium-voltage AC directly into 800V DC, providing electrical isolation, voltage transformation, and reactive compensation.

SST’s core advantages are:

1) Extreme integration: Compressing multiple stages of traditional power links into a single device, saving much space

2) High efficiency: Full-chain efficiency improved by over 3%

3) Material savings: A DC800V solution saves 1.4 tons of transformer copper and two-thirds of cable copper versus traditional AC systems.

SST represents the main direction for data center DC power systems. Its key R&D difficulties are the power conversion module (PSU) and high-frequency isolation transformer.

③VPD: Solving the Last 100 Meters of Power at the Chip Level

At the chip level, traditional lateral power delivery (LPD) is hitting physical limits. As processor working currents are expected to rise to 10,000A in the next decade, resistance and inductance effects within power delivery networks (PDN) lead to sharply rising power losses.

Vertical Power Delivery (VPD) places the voltage regulator module (VRM) directly on the back of the PCB directly beneath the processor, vertically delivering power through the PCB, effectively shortening power delivery distance.

At this year’s CES, NVIDIA confirmed the Rubin architecture will adopt VPD. Intel and Google have also begun trying VPD.

In summary, we propose three core logics for investing in the AI power industry:

First, technological revolution drives value re-rating. From UPS to HVDC to SST, the technical threshold, product forms, and value of power systems are all rising. This will reshape competition, and companies with core technology reserves may gain premium valuations.

Second, high certainty of market space. AI computing investment remains on an upward trajectory. As the "must-have" infrastructure, power demand is rigidly certain. Even if AI chip iteration slows, existing data centers still have energy efficiency upgrade needs.

Third, broad space for domestic substitution. In AC-DC power supply, mainland vendors still have low market share but have begun entering NVIDIA’s supply chain; in HVDC, domestic companies like Zhongheng Electric have accumulated deep experience and should be able to expand abroad.

The AI power industry is experiencing a revaluation from "supporting component" to "strategic choke point." The underlying driver is the exponential growth in computing density; the technical direction is systemic restructuring from AC to DC; and the market space is a blue ocean growing by hundreds of billions. In this process, companies with core technology reserves, deep binding with leading customers, and precise product positioning are likely to enjoy dual dividends from both rising volumes/prices and re-rating of valuations.

Risk Warning and DisclaimerThe market has risks, investment should be cautious. This article does not constitute personal investment advice and does not take into account specific users’ individual investment objectives, financial situation, or needs. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Investments are made at your own risk based on this information. ```