Summary of Key Points
The training of large AI models has led to a surge in data center power consumption, making the traditional 54V low-voltage DC power supply system insufficient to meet the demands. The existing system faces issues such as limited space for power supplies, heavy copper materials, low conversion efficiency, and potential disruptions to grid stability. NVIDIA has introduced an 800V high-voltage DC (800VDC) solution that converts alternating current to 800V DC at the data center level, reducing conversion steps and offering benefits such as increased efficiency, reduced copper usage, and lower maintenance costs. This new industry chain involves three main categories of companies: power semiconductors (SiC/GaN), power modules, and data center infrastructure. It is expected to be widely adopted by 2027 as an industry-wide response to the growing demand for AI computing power.
Detailed Analysis
#### 1. Why Can't Traditional Power Supplies Keep Up with AI's High Energy Demands?
Data centers have traditionally used 54V low-voltage DC, with the power supply process being: "grid AC → multiple AC/DC conversions → 54V to servers." While this system is mature, it cannot accommodate the high energy demands of AI:
- Surging Power Consumption: NVIDIA's GPU power consumption has increased by 75% from the H100 to the Blackwell model, and the power density of a rack with 72 GPUs has tripled. In the future, Kyber racks will host 576 GPUs, consuming nearly 1 megawatt of power (equivalent to 1,000 household air conditioners running simultaneously).
- Limited Space: Traditional systems require 64U power modules for rack sizes of only 42U, leaving little space for other components.
- Heavy Copper Usage: Converting 1 megawatt of power at 54V requires 200 kilograms of copper; a gigawatt-level data center would need 200,000 kilograms (200 tons) of copper, which is impractical in engineering terms.
- Low Efficiency and Heat Generation: Multiple conversions result in energy loss, and the accumulated heat increases the risk of failures.
- Grid Instability: AI training processes cause power fluctuations, which can be problematic for grids handling large amounts of power.
The traditional system has reached its physical limitations, making the adoption of 800VDC inevitable.
#### 2. The Benefits of 800VDC
NVIDIA's 800VDC solution focuses on reducing intermediate conversion steps, leading to three significant advantages:
- 5% Efficiency Improvement: With fewer conversions, less energy is wasted. For example, a data center consuming 1 billion kWh per year could save 50 million kWh (equivalent to the annual power consumption of 50,000 households).
- 45% Reduction in Copper Usage: Higher voltage allows thinner wires to transmit the same amount of power, requiring only three wires (compared to four in traditional AC systems), significantly reducing copper costs and weight.
- 70% Reduction in Maintenance Costs: Fewer conversion points mean fewer maintenance issues and lower total cost of ownership (TCO).
The 800VDC solution also addresses grid stability by using supercapacitors to absorb short-term spikes and batteries on the DC bus to buffer longer-term fluctuations.
#### 3. New Industry Chain Distribution
The widespread adoption of 800VDC requires collaboration across the entire chain, which can be divided into three main segments:
- Power Semiconductors: High-voltage tolerance is required, so silicon-based chips are insufficient; carbonized silicon (SiC) and gallium nitride (GaN) are needed. Companies like STMicroelectronics and NVIDIA collaborate on power boards, while Texas Instruments provide complete solutions, and Infineon covers all material routes.
- Power Modules: These modules must safely deliver 800V DC to the racks and be designed for specific rack configurations. Delta is a key supplier, offering 55kW and 90kW modules; Wistron and Megmeet are also involved in this collaboration.
- Data Center Infrastructure: This segment ensures that 800VDC is successfully implemented from the grid entrance to the racks. Companies like Vertiv provide comprehensive solutions, Eaton drives industry standards, and giants like ABB and Siemens are also participating.
These companies will be direct beneficiaries of this "power revolution."
#### 4. Implications for the Industry
- Inevitable Trend: The International Energy Agency predicts that global data center power consumption will double by 2030 (reaching 945 terawatt-hours, equivalent to Japan's annual electricity usage). The efficiency improvements offered by 80VDC are essential.
- Industry-wide Adoption: This is not just NVIDIA's initiative but a collective response from the entire industry, with all segments from chips to infrastructure making adjustments.
- Tight Time Frame: 2026 is the year when products will be ready for market, and mass production will begin in 2027. Suppliers must act quickly to secure their place in this emerging market.
The energy costs associated with AI computing power are becoming increasingly significant, and 80VDC is the necessary solution to these challenges. It will also shape the new industry landscape.
In One Sentence
The high energy demands of AI have spurred a revolution in power supply technology, with NVIDIA leading the way. Various industry segments are competing to benefit from this change, and widespread adoption by 2027 is essential for data centers to meet future computing power requirements.
(End of translation)