HomeArticle

The Prehistory of AI Power Supply: How Lighting, Automotive and Industrial Electronics Integrate into Computing Power Infrastructure

哈佛商业评论2026-11-21 09:05
"Hybrid evolution" is the underlying backdrop of technological development.

During the 2008 Beijing Olympic Games, the National Aquatics Center, known as the "Water Cube", turned into a huge glowing blue body against the night sky. The sci-tech Olympic landscape was supported by light-emitting diode (LED) lighting, which was still in the stage of rapid industrialization at that time. For audiences, it was a striking urban landscape; for the industry, it was an application verification answer submitted by China's semiconductor lighting industry: the gallium nitride (GaN) materials, device fabrication, packaging and heat dissipation, reliability verification and control systems for LEDs all transformed from laboratory indicators to manufacturing technologies at this moment.

No one at that time would have linked this beam of blue light to today's artificial intelligence (AI) data centers, nor was there a pre-designed technical route between the two. It was only later that the gallium nitride material and device technologies continuously accumulated by the lighting industry, the silicon carbide (SiC) high-voltage power electronics technologies driven by new energy vehicles, and the silicon (Si)-based power technologies iterated by the industrial and server sectors over a long period of time, happened to converge again in the power supply systems of AI computing infrastructure. The AI power supply that looks very new today is actually the result of the mixed evolution of capabilities built by earlier industrial development.

Why have these old capabilities suddenly become important again today? The most direct reason lies in the power supply demand for AI computing power. As the power consumption of AI chips and the power density of cabinets rise simultaneously, the power consumption of a single cabinet is increasing from tens of kilowatts to hundreds of kilowatts, and the next stage will even reach the megawatt level. At this power level, power supply efficiency is no longer just an issue of electricity bills: how many GPUs can be deployed in the same computer room space, and how much of the input power can finally be effectively delivered to the chips, will all be restricted by the power supply architecture.

One of the key technologies supporting this system is power semiconductors and their power conversion systems. In the process of electricity being directly delivered from the grid to GPUs in the data center, it needs to go through rectification, backup power supply, computer room power distribution, server power supply and multi-level DC conversion, before finally being converted into low-voltage and high-current power required by the chips. In the past when the power of servers was not high, even after multiple conversions in this link, the power loss and heat generation were acceptable; when the cabinet power approaches the megawatt level, the problems of current, copper consumption, loss and heat dissipation in the original architecture will be amplified at the same time.

To deliver power to a 1 megawatt (MW) cabinet, if 48 volt (V) DC transmission is still used, the theoretical current is about 20833 amperes (A); if the voltage is increased to 800 V, the current is about 1250 amperes. When the power is the same, the higher the voltage, the smaller the current, and the resistance loss on the wire is related to the square of the current. Texas Instruments (TI), a US semiconductor enterprise, estimates that if a 48V architecture is used to distribute power to a 1MW cabinet, the copper consumption will be close to 200 kg just to keep the loss within a reasonable range. This number is only an engineering problem for a single cabinet, but when applied to a large-scale AI park, it will turn into problems of cables, space, heat dissipation and huge capital expenditure.

This is also the reason why the power supply voltage of data centers keeps increasing. Google publicly released a ±400V DC solution that supports 1MW cabinets in 2025; NVIDIA is promoting the 800V high-voltage DC architecture, and further promoted related industrial coordination with Google, Microsoft and the Open Compute Project (OCP) in 2026. The core idea is not complicated: use a relatively high voltage as much as possible to complete high-power transmission on the cabinet side, and then drop to low voltage at the position close to the chip, so as to reduce copper loss and space burden caused by large current.

After the voltage rises, the importance of power semiconductors also increases accordingly. Gallium nitride is suitable for high-frequency, high-efficiency and high-power density conversion links; silicon carbide has more advantages in voltage resistance and high power; silicon has low cost and mature industry, and still undertakes a large number of functions of control, protection, low-voltage conversion and power management. Therefore, AI power supply will not be dominated by a single material, but will re-divide work according to different voltages, power densities and switching frequencies. To understand this division of labor, we need to trace back the path of how these three types of devices formed their respective industrial capabilities.

The gallium nitride industry left behind by a beam of blue light

The industrial history of gallium nitride is first and foremost a history of the development of semiconductor lighting. In the 1960s, red and green LEDs had already appeared, but high-efficiency blue light was absent for a long time. Without blue light, it was difficult to form a complete set of three primary colors of red, green and blue, nor could high-efficiency white light be obtained by exciting phosphors with blue light. This problem has plagued the industry for decades, and the core obstacle lies in the growth of high-quality gallium nitride materials and p-type doping (which refers to making the material conduct positive charges by incorporating specific impurities).

Three scientists, Isamu Akasaki, Hiroshi Amano and Shuji Nakamura, successively solved the key problems among them. In 1993, Nichia Corporation launched high-brightness blue LEDs, and in 2014, the three scientists won the Nobel Prize in Physics for this achievement. The combination of blue LEDs and phosphors made high-efficiency, long-life white LEDs a reality, opening up a new lighting industry.

It takes a long process of engineering to go from scientific discovery to large-scale industry. The Japan Science and Technology Agency (JST) supported Toyoda Gosei and Akasaki's team to promote the commercialization of blue LEDs through the entrusted development mechanism as early as 1987 to 1990, with an investment of about 550 million yen. By 2013, the related technologies had generated a total of about 5.6 billion yen in licensing revenue returned to JST. Public funds took early technical risks, universities provided original technologies, and enterprises completed manufacturing and market verification. Japan took the first step to move blue LEDs from the laboratory to the industry.

China then pushed this technology to a larger industrial scale. In 2003, the Institute of Semiconductors of the Chinese Academy of Sciences reported suggestions on developing nitride LEDs and semiconductor lighting to the Ministry of Science and Technology; the Ministry of Science and Technology launched the "National Semiconductor Lighting Project" in June of that year; afterwards, the "11th Five-Year Plan" National High Technology Research and Development Program (863 Program) continued to invest. The 2008 Beijing Olympic Games allowed LEDs to be intensively displayed in large-scale projects such as the "Water Cube"; in 2009, the "Ten Cities, Ten Thousand Lamps" initiative further turned urban lighting into a large-scale application site, which was later expanded to 37 cities; financial subsidies for semiconductor lighting products were implemented in 2012. By 2015, the scale of China's semiconductor lighting industry had grown from about 90 billion yuan in 2003 to 4.245 trillion yuan.

Hundreds of millions of LEDs need to be produced stably, which requires continuous solving of a series of problems such as epitaxial growth, crystal defect control, yield improvement, packaging and heat dissipation design, and reliability. The expansion of production capacity has created huge demand for Metal-Organic Chemical Vapor Deposition (MOCVD) equipment, substrate materials, testing technologies and engineering talents. The lighting market provided a long manufacturing learning cycle for gallium nitride, allowing technologies that originally existed in a few laboratories to gradually become engineering capabilities that the entire industrial chain can repeatedly manufacture and continuously reduce costs.

Of course, gallium nitride LEDs and gallium nitride power devices are not the same type of device, and the production lines cannot be simply interchanged. However, the understanding of gallium nitride materials, epitaxy and defect control capabilities, equipment ecology, engineering talents, and enterprises' experience in organizing the manufacturing of compound semiconductors can be migrated and evolved. After that, gallium nitride entered radio frequency power devices and communication base stations. The fifth-generation mobile communication (5G) expanded high-power radio frequency applications, and fast charging for consumer electronics allowed gallium nitride power devices to undergo larger-scale market verification in high-efficiency, miniaturized power supplies.

Today, this industrial chain continues to extend to AI power supply. Texas Instruments launched high-density power devices integrating 650V gallium nitride power transistors in 2025; Infineon released a high-voltage intermediate bus conversion reference design based on 650V gallium nitride in 2026, targeting ±400V and 800V DC AI server architectures. Chinese enterprises have also entered the forefront of global technology competition this time. Innoscience cooperates with NVIDIA to promote 800V DC AI data center power supply, and has demonstrated gallium nitride solutions covering 15V to 1200V.

If we look further upstream of the industrial chain from the devices, China's gallium nitride industry today has a complete engineering foundation. Epitaxial manufacturers such as Jingcan Semiconductor have long focused on gallium nitride materials, enabling China to have the capability to produce 12-inch (300mm) gallium nitride wafers as early as 2021. Further downstream, Shanghai New Micro Semiconductor relies on the state-owned assets and the achievement transformation system of scientific research institutes to develop professional compound semiconductor manufacturing, and is building a gallium nitride power foundry platform covering low, medium and high voltages. In addition, China has also deployed and built a public pilot R&D platform. Shenzhen Pinghu Laboratory is the Shenzhen comprehensive platform of the National Third-Generation Semiconductor Technology Innovation Center, and it is also an industry-oriented open power semiconductor R&D and pilot platform, which has laid out 8-inch silicon carbide and gallium nitride chip R&D and shared capabilities from materials, manufacturing to packaging and testing. China is no longer just a recipient of the application market and manufacturing capabilities, and key links have entered the global first echelon.

For China, the history of developing semiconductor lighting is particularly worthy of attention. The United States, Japan and Europe have their own scientific research, national defense, communication and industrial accumulation. However, the large-scale gallium nitride industry established by China around semiconductor lighting in the past 20 years has provided ready-made talents, equipment, supply chains and engineering experience for later entry into the radio frequency and power electronics fields, thus lowering the threshold for China to enter this round of gallium nitride industry competition.

What new energy vehicles leave behind

Is not just automobiles

Silicon carbide power devices had entered the industrial market before the popularization of new energy vehicles. Around 2000, silicon carbide Schottky diodes began to achieve commercial application. Around 2011, Cree (now Wolfspeed) further launched commercial SiC MOSFET (Metal-Oxide-Semiconductor Field-Effect Transistor), promoting the expansion of silicon carbide power devices from diodes to power switching devices, and gradually entering a wider range of high-voltage, high-frequency and high-efficiency power electronics application fields. In the early stage, silicon carbide could already demonstrate performance advantages in industrial and high-end power electronics, but the cost of wafers, materials and manufacturing was very high, and the market size was always limited.

New energy vehicles provided this market. In 2009, China launched the demonstration of energy-saving and new energy vehicles in 13 cities; in 2010, subsidies for private car purchases were expanded; the industrial plan in 2012 further clarified the large-scale goal. By 2024, China's new energy vehicle sales reached 12.866 million units, accounting for 40.9% of new car sales.

The new energy vehicle policy does not require enterprises to choose silicon carbide. But because it has created a continuously growing demand for high voltage and high efficiency, it has obtained mass orders that were difficult to provide in the past industrial market, and has also undergone automotive-grade reliability verification and continuous cost reduction. The device, module, connector, testing and manufacturing capabilities formed around the high-voltage platform did not only stay in the automotive industry later.

This is exactly where the connection between automobiles and AI power supply takes place. When explaining the ±400V data center solution, Google explicitly mentioned that it can make use of the economies of scale, manufacturing capabilities and quality foundation already available in the electric vehicle supply chain. AI data centers do not need to re-establish a set of high-voltage DC industrial systems, as the automotive industry has completed a considerable part of the preliminary engineering work.

Silicon, an industry with uninterrupted progress

Silicon-based power electronics does not have a single turning point similar to the "Ten Cities, Ten Thousand Lamps" initiative or new energy vehicles. Its advantages come from decades of large-scale continuous use. From power Metal-Oxide-Semiconductor Field-Effect Transistors (MOSFETs) and Insulated-Gate Bipolar Transistors (IGBTs), to synchronous step-down, multi-phase power supply and Power Management Integrated Circuits (PMIC), silicon devices have successively entered home appliances, motors, factories, railways, photovoltaics, communications and servers. These markets have different requirements for cost, yield, packaging, control and reliability, but they have been jointly promoting the maturity of the same industrial system.

Servers themselves are also an important market for the long-term evolution of silicon power electronics. 20 years ago, the power supply of Central Processing Units (CPUs) already needed to handle hundreds of amperes of current. Today, the final stage power supply of advanced GPUs is moving towards the kiloampere level. After gallium nitride and silicon carbide enter the data center, silicon will not withdraw from the market. Silicon carbide will be more widely used in high-voltage and high-power links, gallium nitride leaves space for high-frequency and high-power density links, while control, protection, power management and a large number of low-voltage conversions will still rely on silicon for a long time. AI power supply is forming a hybrid material system.

This historical line of silicon also has an important significance: industrial capabilities are not necessarily shaped by a landmark policy. A long-term stable market will also form technological accumulation. Home appliances have pushed down costs, industrial frequency conversion and railways have improved power levels and reliability, communication equipment has strengthened high efficiency and high availability, and servers have promoted multi-phase power supply, digital control and higher current density. The superposition of decades of demand has finally turned silicon-based power supplies into the most mature technical system today.

"Hybrid evolution" is the underlying tone of technological development

Looking beyond China, you will find that "hybrid evolution" is not unique to China. Different economies choose different policy tools, but they are all dealing with the same problem: how to make an immature technology cross the breakpoint between scientific discovery and large-scale manufacturing.

The most representative path in China is to form large-scale demand through application scenarios. Semiconductor lighting created early orders through public lighting, and new energy vehicles gradually expanded from demonstration fleets to private consumption. Policies may not be able to pre-determine the final winning device route, but they can give new technologies early enough orders, a large enough engineering site, and a long enough cost reduction cycle.

The commercialization of blue LEDs in Japan is a typical "scientific research-enterprise" bridge. JST passed entrusted development to bring the gallium nitride breakthrough of the university team into the industrialization system of Toyoda Gosei; the New Energy and Industrial Technology Development Organization (NEDO) of Japan then promoted silicon carbide for a long time, from materials and devices to system applications such as railways.

The United States emphasizes continuous investment in technical bottlenecks. The Advanced Research Projects Agency-Energy (ARPA-E) of the US Department of Energy invested 27 million US dollars in 2013 to support low-cost wide-bandgap switches such as gallium nitride and silicon carbide, and later focused on solving manufacturing problems such as gallium nitride doping, and further expanded the scope to topology, packaging, control, thermal management and complete converters. PowerAmerica, the US power electronics manufacturing innovation center, further deals with manufacturing scale, cost, reliability, commercialization and talents. The policy focus changes with the maturity of technology, moving all the way from materials to systems.

The paths of China, Japan and the United States are different: some rely on large-scale markets, some connect scientific research and enterprises, some continuously overcome technical breakpoints, and some build joint pilot platforms. The results left behind are very similar: talents, equipment, processes, standards, supply chains, and the continuously reduced cost curve. These capabilities may eventually flow to new industries that were not originally written in the project documents.

Tracing all the way from lighting, automobiles and industrial electronics to AI, the most noteworthy thing is what these investments have left behind. Almost no one 20 years ago would have interpreted the LED policy as an investment in future AI power supplies, nor would anyone have written new energy vehicle subsidies as pre-engineering for megawatt-level data centers.

The return of a round of industrial investment should be evaluated by whether it can form transferable capabilities. Whether the equipment can serve the next type of device, whether engineering talents master reusable process knowledge, whether the supply chain can enter new voltage and power levels, and whether enterprises have established cross-material and cross-system development capabilities. These assets last longer than a certain generation of products.

This also explains the long-term value of international cooperation. The evolution of gallium nitride, silicon carbide and silicon-based power electronics has spanned universities, enterprises, government projects and industrial systems in multiple countries. If knowledge, talents and markets are prematurely locked into isolated "small yards", what we lose is not just a generation of products, but the opportunity for "hybrid evolution" between the next round of industries. Independent capabilities and international cooperation