AI 發展的電力能源新架構：解析 NVIDIA 800 V HVDC - 2025.05.25

AI Data Center 的能源議題

AI 的快速發展正在推動 Data Center 的電力需求升級，從千瓦級（kW）躍升至兆瓦級（MW）；傳統 54 V DC 電力架構已無法應對這種挑戰。NVIDIA 推出的 800 V HVDC 電力架構，為下一代 AI Data Center 提供了解決方案，預計將於 2027 年全面應用，重新定義 Data Center 電力基礎設施。 The rapid development of AI is driving a major upgrade in power demands for data centers, shifting from kilowatt (kW) levels to megawatt (MW) levels. The traditional 54V DC power architecture can no longer meet this challenge. NVIDIA's introduction of an 800V HVDC power architecture offers a solution for next-generation AI data centers, expected to be fully implemented by 2027—redefining the foundational infrastructure of data center power systems.

傳統 Data Center 電源架構面臨哪些問題?
Challenges Faced by Traditional Data Center Power Architectures

以目前的 AI Data Center 來說皆使用 54 V DC 電力架構，包括GB200 NVL72/GB300 NVL72，但隨著技術及算力升級，單一機櫃(1 Rack)功率將超過 200 KW，電力供應將會面對物理空間限制、銅使用量和 AC/DC 轉換效率不佳等問題。 Currently, most AI data centers rely on a 54V DC power architecture, including systems such as the GB200 NVL72 and GB300 NVL72. However, as AI technology and computing power rapidly advance, a single rack (1 Rack) may soon require over 200 kW of power. This surge in demand introduces significant issues related to physical space limitations, copper consumption, and inefficient AC/DC power conversion.

1. 物理空間限制 
   Physical Space Limitations
   
   在 GB200或GB300 系統中，普遍採用 NVL72/36 機型，其設計將 Power Shelf、Compute-node、NVLink 交換器與液冷單元（CDU）緊密整合於單一機櫃內。
   目前的機櫃配置有四分之三的空間已被上述單元占滿，幾乎未留下可用空間來增加額外電力元件或電力備援裝置（BBU），以目前的電力供應技術，顯然不適用於未來的超大型 AI Data Center。
   

   In systems such as the GB200 or GB300, commonly adopting NVL72/36 configurations, the design tightly integrates power shelves, compute nodes, NVLink switches, and liquid cooling units (CDUs) within a single rack.

   Currently, around three-quarters of the rack's internal space is already occupied by these critical modules. This leaves almost no room to add additional power components or backup power units (BBUs). Under the current 54V power delivery design, this becomes unsuitable for future ultra-large-scale AI data centers.

   
   
2. 銅使用量暴增
   Surge in Copper Consumption
   
   目前的54 V DC 電力架構設計來看銅使用量： 1MW Rack 需約 200 公斤的銅匯流排 1GW Data Center 則需高達 50 萬噸的銅使用量 (500,000,000公斤)
   

   Under the current 54V DC power architecture, the amount of copper required is staggering:
   1 MW rack requires approximately 200 kg of copper busbars. 
   1 GW data center could demand up to 500,000 metric tons of copper (500,000,000 kg).

   
   

3. 轉換效率不佳
   Inefficient Power Conversion
   
   目前的54 V DC 電力架構是先降壓在變壓至54V DC，每一次的降壓或變壓會造成效率降低且會增加故障點。
   
   The existing 54V DC system typically involves stepping down voltage followed by transformation to 54V DC. Each voltage conversion stage results in reduced energy efficiency and introduces additional points of failure. These inefficiencies accumulate, making the architecture unsustainable for the future scale of AI workloads.
   
   

Figure 1. Current data center power architecture

NVIDIA 800 V HVDC 解決方案
NVIDIA’s 800V HVDC Power Solution

800V HVDC 加大電壓等級、增加保護裝置、減少銅使用量為核心，打造全新供電系統。

NVIDIA’s 800V HVDC power architecture introduces a transformative approach to power delivery for AI data centers, focusing on increased voltage levels, enhanced protection, and reduced copper usage to build a new generation of efficient power systems.

1. 空間優化 Space Optimization
   消除機架內 AC/DC 轉換，釋放空間用於更多運算資源。
   By eliminating AC/DC conversions inside the rack, more internal space is freed up for compute resources, maximizing performance density.
   
   
2. 可擴展性 Scalability
   支援從 100 kW 到 1 MW 以上的機架，適應未來需求。
   The system supports rack power demands ranging from 100 kW to over 1 MW, making it future-ready for the growing needs of AI workloads.
   
   
3. 轉換效率 Efficiency
   800 V HVDC 端對端效率提升 5%，將有效改善能源使用效率。
   The 800V HVDC architecture improves end-to-end power conversion efficiency by 5%, significantly enhancing overall energy utilization.
   
   
4. 可靠性 Reliability 集中式電力轉換減少 PSU 故障，維護成本降低 70%。 Centralized power conversion architecture reduces PSU (Power Supply Unit) failures, cutting maintenance costs by 70%.
5. 銅使用量減少 Reduced Copper Usage 相同功耗下 800V HVDC 架構能降低電流，減少 45% 銅使用量，亦降低熱損耗。 With higher voltage, the same power can be delivered at a lower current. This results in a 45% reduction in copper usage, which also reduces thermal losses.

P(功率)=I(電流)×V(電壓)

※電壓越高，電流越小，功率損耗越小，使用的導體（銅或鋁）就越少，成本也就越低。

所以世界各國電力系統都盡量將電壓升到超高壓等級，減少線路損耗與導體使用量。 Increasing voltage lowers current for the same power, which means less conductor (copper/aluminum) is required and power losses are minimized.This is why national electrical grids worldwide adopt ultra-high-voltage transmission to reduce energy loss and lower infrastructure costs.

Figure 2. NVIDIA 800 V HVDC architecture minimizes energy conversions.

在基礎設施部署 800V HVDC系統前，將帶來安全性、標準制定與人員訓練等多方面的新挑戰。NVIDIA 及其合作夥伴正積極研究傳統變壓器架構與固態變壓器solid-state transformer （SST）兩種方案在資本支出（CapEx）、營運支出（OpEx）及安全性方面的影響，以促進這項轉型。

NVIDIA 預計於 2027 年推出Kyber Rack系統，屆時將同步上線800V HVDC 電力供電系統，確保能支援兆瓦級（MW） AI Data Center，以因應日益龐大且複雜的 AI 模型需求。

Deployment Challenges and Future Outlook

Before 800V HVDC systems can be widely adopted in infrastructure, new challenges must be addressed—including safety protocols, standardization, and personnel training.

NVIDIA and its partners are actively evaluating both traditional transformer architectures and solid-state transformers (SST) in terms of capital expenditures (CapEx), operational expenditures (OpEx), and safety implications, to enable a smooth transition.

NVIDIA plans to launch the Kyber Rack system in 2027, which will include native support for 800V HVDC power delivery. This upgrade will ensure support for megawatt-scale (MW) AI data centers, accommodating the exponentially increasing demands of next-generation AI models.

NVIDIA 800V HVDC 相關合作供應鏈

• Silicon providers: Infineon, MPS, Navitas, ROHM, STMicroelectronics, Texas Instruments
• Power system components: Delta(台達電 2308.TW), Flex Power, Lead Wealth, LiteOn(光寶科 2301.TW), Megmeet
• Data center power systems: Eaton, Schneider Electric, Vertiv

資料來源:

NVIDIA 800 V HVDC Architecture Will Power the Next Generation of AI Factories | NVIDIA Technical Blog