文 | 半导体产业纵横
数年前，氮化镓以快充头为载体走进大众视野，依靠高开关频率、高功率密度的特性，完成功率半导体赛道的一次大众化普及。如今赛道格局早已改写，产能铺开之后价格战持续压缩利润，消费市场的增长天花板肉眼可见。行业开始向外寻找增量，车载OBC、48V辅助电源、AI服务器HVDC电源成为厂商重点攻坚方向。
从消费级产品转向车规、算力场景，验证标准、失效风险、交付逻辑全部重构。算力浪潮滚滚向前，车载验证周期漫长，氮化镓能不能接住这一波新需求，走出快充红利消退后的增长困局，答案仍在漫长的测试与定点落地之中。
快充红利见顶，产品线走到分化路口
近些年，氮化镓以高端快充为卖点，20W到140W充电器一场接一场地铺开，手机、笔电、配件厂商扎堆进来。从最早的65W GaN充电器被称为充电革命，到如今几十块钱就能买到一个支持多协议的快充头，这个市场只用了几年就完成了从爆款到标准品的转换。
根据行业调研数据，2025年国内消费GaN快充出货量突破2.2亿只，但是平均单价较2022年高点下跌超60%，毛利率从早期35%以上回落至12%至18%区间。产品形态集中在650V横向E-mode分立管和简易半桥，设计逻辑优先控制BOM成本，可靠性验证周期短，谁家能把一个管子的成本压下去，谁就能在整机报价里拿到优势。当产能放出来、方案越来越同质化，价格战不可避免，靠效率参数讲故事的时代就过去了。
快充市场红利消退，直接反映在厂家的账本上。一方面，氮化镓器件单价持续下探，量涨但利润空间收窄；另一方面，整机厂对元器件的议价能力越来越强，器件厂被迫在成本、性能和可靠性之间做更极限的取舍。厂商被迫做产品线取舍，一部分继续深耕消费赛道，迭代方向压在Qg优化、封装小型化、晶圆良率提升上，靠成本优势走量，把快充市场做成一条维持现金流的存量生意；另一部分则开始单独搭建车规和工业产品线，重新定义器件规格，建立批次管控规则，把验证周期拉长，用时间换更高的毛利和更稳的订单。车规级GaN器件的目标毛利率普遍维持在25%以上，是消费级产品的一倍左右，但研发、验证投入同样成倍增加。
聊氮化镓的应用前景，绕不开它和碳化硅的关系，而这两者常被混为一谈。很多行业外的观察会把宽禁带半导体当成一个整体，觉得氮化镓和碳化硅是同一类东西，谁性能强谁就能赢。但它们的应用广度和价值差异，根源在于电压等级和开关特性的分工，而不是谁性能更强。从市场规模来看，2025年全球SiC功率器件市场规模约38亿美元，同期GaN功率器件市场规模约8.7亿美元，体量差距明显，背后就是场景分工带来的天然差异。
碳化硅的本命在高电压。它是宽带隙材料，耐压、耐高温、导热出色，主战场在800V甚至1200V以上的车载主逆变器、大功率充电桩和光伏逆变器。这些场景功率大、母线电压高，碳化硅能直接顶替硅IGBT，换来更低的导通损耗和更高的开关频率，单个器件的价值量也高，是宽禁带功率器件里单价最贵、出货规模增长最猛的品类。所以市面上大量三代半报道扎堆在碳化硅衬底和主驱上，因为碳化硅的故事就是大功率、高价值、集中放量，讲起来最直接。
氮化镓则不同，它擅长的是低压高频。当前商用的GaN-on-Si横向器件，主力电压区间在100V到650V，比碳化硅低得多，但开关速度极快、效率高、体积小。它的主场在6.6kW到11kW的车载OBC、48V/12V DC-DC、工业伺服、通信电源，还有正在升温的数据中心板载电源。碳化硅做的是高功率、高压的粗重活，氮化镓做的是高频、高效、紧凑的精细活，二者在系统里常常是分工协作，而不是互为替代。这也解释了为什么氮化镓市场一直显得比碳化硅小，它的单价、单场景功率量和总体价值量，天然低于碳化硅。
同样是做氮化镓，英诺赛科、纳微、华润微、TI、EPC、VisIC的产品定位已经明显分化。有的押注产能和成本，有的押注系统级方案，有的押注集成度，有的干脆绕开红海去卡位细分场景。这种分化本身，就是行业进入下半场的信号，企业必须在快充带来的短期营收，和车规产品线漫长且高投入的验证周期之间做平衡，而这份平衡术，恰恰决定了各家未来几年能走多远。
三次技术跃迁，每一代都对应不同产品
真正决定这场转型走向的，是氮化镓器件技术自身的演进，而每一轮演进背后，都有各家企业不同的产品解法。把这二十年的产业史拉长看，氮化镓功率器件其实打了三场硬仗，每一场都对应一代可以追溯的产品。
第一次跃迁，是从耗尽型走向增强型。早期的D-mode耗尽型器件栅极要靠负压关断，必须外接级联硅MOSFET才能进系统，结构复杂，可靠性隐患也多，仅在少量工业场景小批量试用，整体占比不足GaN功率器件出货量的5%，根本谈不上装车。p-GaN栅极工程把器件做成E-mode增强型之后，单管直驱、不用负压电路，这才成为当前商用的底座。
但共性的痛点也随之而来，阈值电压Vth稳不稳定、高压动态应力下动态导通电阻Rds-on会不会退化，决定了这个管子敢不敢上车。围绕这个问题，各家动作不同。英诺赛科迭代第三代Inno GaN p-GaN平台，优化缓冲层结构去抑制界面电荷俘获，同时并行把多版栅工艺样品交给Tier1对比测试；华润微的车规横向氮化镓器件主打稳健的栅偏压管控方案，先面向工业电源和车载48V辅助电源，节奏偏保守；EPC则以低压增强型器件为根基，刻意避开650VOBC的红海，把精力放在工业高频变换和机器人伺服上。这第一场仗的核心，就是谁能在保持高开关速度的同时，把栅极做得更稳，谁的管子才敢被真正装进系统里长期运行。
第二次跃迁，是从横向走向垂直。横向GaN-on-Si是当下的量产主力，但在650V这个电压等级上已经接近耐压和热管理的天花板，再往上堆电压和电流，横向结构在导通电阻、散热路径上的代价会急剧上升。厂商的优化集中在缓冲层、外延缺陷管控，去改善动态特性，但横向结构的物理极限摆在那里。而要做到900V甚至1200V，就得押注垂直氮化镓，也就是GaN-on-GaN衬底、电流垂直流过的结构。截至2026年三季度，垂直GaN整体仍处于工程样片送样阶段，产业化进度落后横向GaN至少3到4年，同质衬底成本是同规格GaN-on-Si外延片的8至12倍，卡点集中在同质衬底成本、位错密度和大尺寸良率。
各家企业在这方面的打法分歧最大。VisIC选择直接切入垂直氮化镓，瞄准车载主驱逆变器，和AVL联合开发样车，路线相当激进；英诺赛科同步布局垂直氮化镓研发，但眼下把资源优先放在横向车规合封产品的落地，把垂直路线当作中长期技术储备；而国内多数厂商现阶段仍以横向平台迭代为主，垂直氮化镓还停留在实验室流片和学术合作层面，没有明确的送样计划。这第二场仗，比的是谁能在衬底和外延上先突破，谁就能把电压等级的天花板往上再顶一截。
第三次跃迁，是从分立器件走向集成方案，这也是当前路线分歧最大的赛道。分立器件灵活性高，适合自研能力强的Tier1，短板是环路寄生和EMI调试难度大，华润微、EPC主推这条线。集成方案则把GaN FET、栅极钳位、米勒钳位、短路保护封装在一起，降低系统开发门槛，是当下OBC落地的主流方向。英诺赛科推出的Solid氮化镓车规合封器件ISG6127ATPG-Q、ISG6129BTQ-Q，650V，集成米勒钳位，专门解决高频dv/dt带来的栅极误导通，已经通过AEC-Q100认证，面向6.6kW OBC和车载DC-DC。纳微的氮化镓Safe车规功率IC走得更极致，单片集成FET、驱动、短路保护和ESD防护，内置GaN Sense检测，主打数字入功率出，还配套整车厂OBC平台开发，甚至用二十年有限质保来打消车厂的可靠性顾虑。TI的做法又不一样，它不单独主推氮化镓器件，而是依托完整电源生态，把氮化镓功率管和自研驱动、控制芯片打包成参考设计交付，降低传统Tier1的移植难度。这第三场仗，本质上是商业模式之争，卖一个管子和卖一套电源方案，对客户能力的要求完全不同，也决定了各家的客户结构和议价能力。
三次跃迁有一条暗线贯穿始终，每次演进的真正驱动力，都不是开关速度又提高了多少，而是可靠性和系统可用性往前走了一步。理解了这条线，才能看懂为什么氮化镓明明在消费端很能打，一到车规就要重新过一遍鬼门关。
车规验证那道看不见的鸿沟
车规一直是是检验氮化镓成色的最严考场。消费产品考核的是短期台架指标，常规快充器件可靠性测试周期一般不超过1000小时；车规GaN要持续承受高低温循环、长时间高压应力和负载突降这些动态工况，AEC-Q100体系下基础可靠性测试时长就达到数千小时，叠加动态应力增补测试，整体验证时长大幅拉长。GaN身上有几类特有的失效问题，电荷俘获会让器件在持续高压动态应力下出现电流崩塌，动态Rds漂移幅度最高可达初始值的20%以上，栅极在高温下退化，温度循环还会带来封装疲劳。大量样品在脉冲短时间测试里指标很漂亮，一放到长时间连续应力测试，性能就开始偏移。
更麻烦的是，行业现有的AEC-Q101本身就是围绕硅功率器件建立的，对氮化镓动态失效的覆盖并不充分。硅器件的测试体系里，很多关键失效模式在氮化镓身上根本不适用，而氮化镓特有的电荷俘获、动态漂移又没有对应的标准测试项，各家器件厂和Tier1只能自行增补测试项，各测各的，结果很难横向对比。这就是为什么车规氮化镓的验证周期动辄以年计，因为很多测试规范还在摸着石头过河。
而且氮化镓的高频特性，对驱动IC、磁性元件、PCB布局、EMI抑制都提出全新要求，单换一个氮化镓器件根本不够，整套电源系统方案都要同步迭代。一个管子性能再好，如果驱动电路没跟上、磁性元件不匹配、PCB布局不合理，整机一样跑不出效果，甚至可能因为环路问题出可靠性事故。这也解释了为什么英诺赛科、纳微都要做集成产品，本质上就是厂商把一部分系统层面的抗扰和保护能力内置进器件，把调试压力从Tier1往器件端转移，用更高的集成度去换更低的系统落地门槛。
数据中心里，氮化镓正悄悄进场
如果说车载是氮化镓眼下最热的主叙事，那AI数据中心正在成为它最被低估的增量。AI浪潮推高了算力集群的功耗，主流 AI 服务器机柜功率从早年5kW提升至40kW甚至60kW级别，服务器的板载电源和机架电源从传统的12V架构，逐步转向48V甚至更高电压的HVDC高压直流架构。电压一抬上去，硅MOS在效率、功率密度和散热上的短板就暴露出来，测试数据显示，在48V转12V板载拓扑中，氮化镓方案相比传统硅MOS整机效率可提升1.5到3个百分点，功率密度提升40%以上，正好适配AI服务器对电源效率和功率密度的苛刻要求。
这也是为什么氮化镓厂商都在往这条线使劲。纳微的GaN Safe和配套方案，一大块就打在AI服务器电源上，把高频氮化镓和高效拓扑结合，去拼数据中心每瓦成本的账。英诺赛科也在布局数据中心场景，把横向650V和低压产品线同时往服务器电源侧延伸。工业侧，通信电源、储能变流器支路、AI板载HVDC，被普遍认为是比车载更早放量的市场，因为数据中心客户不像车企那样需要18到24个月的车规验证周期，决策更快、迭代更急，一款电源方案的验证和起量节奏要快得多。机构预测，2026年全球数据中心GaN电源相关市场增速将超过45%，是所有细分赛道里增速最高的板块。
从这个角度看，氮化镓能不能跟上AI浪潮，答案不只是能不能，而是已经在路上。AI越往后发展，算力密度越高，单位面积的电源转换压力越大，对高频、高效、高功率密度的氮化镓需求就越明确。只是氮化镓吃到的是数据中心电源效率这一块，而不是算力芯片那种聚光灯下的位置，它不在聚光灯下，但几乎每一块GPU服务器的电源板上，都开始有它的身影。随着800V、48V这类高压母线架构在数据中心里铺开，氮化镓在AI生态里的存在感，大概率还会继续往上走。
把时间轴拉长看，快充赛道的红利正在降温，车载、数据中心等新赛道的验证长跑才刚刚开始。氮化镓很难复刻快充时代的爆发路径，它更像一场慢而稳的长周期功率器件生意。未来一两年市场主线还是定点验证和小批量试产，收入贡献有限，放量载体集中在6.6kW车载OBC、通信电源、储能辅助支路和数据中心板载电源。中期看点在于横向650V集成产品持续优化稳定性，垂直氮化镓继续打磨衬底和外延良率，等经济性拐点。风险同样具体，新一代硅MOS还在持续迭代降价，会挤压氮化镓的性价比窗口；车规验证周期漫长，持续的产品迭代和验证投入考验企业的现金流，不是每一家都扛得住。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
The semiconductor industry
A few years ago, nitrogenized aluminum entered the public view with a fast-capture vehicle, using high switch frequency and high power density to complete a popularization of the success rate semiconductor track. Today, the pattern of the track has been rewritten, with price wars continuing to reduce profits following the opening of the production capacity, and the growth of the ceiling in the consumer market is visible. The industry began to look for incremental amounts, and the vehicle-borne OBC, 48V Auxiliary Power, and the AI server HVDC power supply became the main focus of the plant.
The shift from consumer-grade products to car regulations, power scenarios, certification standards, risk of failure, and delivery logic are fully re-engineered. The wave of calculus is moving forward, the vehicle-mounted certification cycle is long, and whether the argon can catch up with this new wave of demand and emerge from the growth trap of a fast-filled dividend, the answer is still in the process of long testing and setting.
Quick to the top. The product line to the split.
In recent years, nitrous aluminum has been filled with high-end speed, and 20W to 140W chargers have been laid in one by one, and mobile phones, pens, and spare parts have come in. From the earliest 65W GaN charger, known as the charger revolution, to the point where a few dozen dollars can now be bought to support a multi-negotiated capital, a market that has only taken a few years to complete the conversion from explosive to standard.
According to industry research data, domestic consumption of GaN was more than 220 million in 2000 but average unit prices fell by over 60 per cent over 2022, with the Māori rate falling from over 35 per cent to between 12 and 18 per cent in the early stages. The product form is concentrated on the 650V horizontal E-mode split tube and simple semi-bridge, with design logic giving priority to BOM costs, a short reliability verification cycle, and who can press down the cost of a tube and who can gain advantage in the overall machine offer. When production capacity is released and programmes are increasingly homogenized, price wars are inevitable and the time for storytelling is over with efficiency parameters.
The collapse of the fast-charging market dividend is directly reflected in the manufacturer ' s books. On the one hand, the unit price of nitrogenized aluminum units continues to increase, but the profit space is shrinking; on the other hand, the whole plant has become more and more bargaining power over metaware, and the plant has been forced to make more extreme trade-offs between cost, performance and reliability. The manufacturer was forced to make product-line trade-offs, in part to continue deep-farming consumption lanes, with the iterative direction pressping on Qg optimization, encapsulation miniaturization, and a higher crystal circle rate, with a cost advantage to make the fast-market a cash-flowing stock business, and in the other part to begin to build a separate car set and industrial product line, redefine the specifications of the device, establish batch control rules, lengthen the certification cycle, and trade time for higher Maori and more stable orders. The target Māori rate for car-graded GaN devices has generally remained at more than 25 per cent, or about double that of consumer-grade products, but R & D and validation inputs have also multiplied.
Talk about the future of the use of nitrogenized beryllium and its relationship to silicon carbide, which is often confused. A lot of observations outside industry see a wide-banded semiconductor as a whole, and feel that nitrogenized aluminum and silicon carbide are the same kind of thing that can be won by anyone who can perform better. However, the differences in the breadth and value of their application are rooted in the division of the voltage grade and switch characteristics rather than in who is more capable. In terms of market size, the global market for SiC power devices in 2025 was about $3.8 billion, while the market for GaN power devices in the same period was about $870 million, with significant volume differences behind the natural differences in the division of scenes.
Silicon carbide is in high voltage. It is a wide-banded, stress-resistant, high-temperature-resistant, heat-guided, main-mounted reverser, high-power charger and light-volatilizer in the main battlefield, at 800 V and above 1200 V. The high power of the scenes, high main-line voltage, the ability of silicon carbide to replace Silicon IGBT directly, in exchange for a lower conductive loss and higher switch frequency, as well as the high value of individual items, are the most expensive and increasing in size of the unit in a wide-banded device. So there's a lot of three-and-a-half-generation stories on the base and main drive of carbide silica, because the story of silica is the most direct one.
It's not the same for nitrogen, it's good at low-pressure HF. Current commercial GaN-on-Si horizontal units, with main voltage ranges of 100 to 650 V, are much lower than carbide silicon, but are very fast, efficient and small. Its main site is on board OBC, 48V/12V DC-DC, industrial server, communications power, and the hotning data centre plate. Silicon carbide is used for high-power, high-pressure heavy work, and nitrous oxide for high-frequency, efficient and compact fine work, which is often a division of labour rather than a substitute in the system. This explains why the market for nitrogenized aluminum has always appeared to be smaller than that of silicon carbide, with unit price, single-situation power and total value, and naturally lower than that of silicon carbide.
Also for nitrogenized nitrousium, the product positioning of Innosico, Nawe, Hua Yunmi, TI, EPC, and VisIC has become significantly different. There are bets on capacity and cost, some on a system-level approach, some on the integration of bets, and some simply bypass the Red Sea to break down the scene. This division in itself is a sign of the second half of the industry, and it is important for enterprises to balance the short-term revenue that comes from fast-filling and the long and high-input certification cycle that determines how far they can go in the coming years.
Three technological leaps. Each generation corresponds to different products.
What really determines the course of this transition is the evolution of the nitrogen-based equipment technology itself, and there are different product solutions for companies behind each rotation. The 20-year-old industry Slade has actually fought three hard battles with a nitrogenized argon power, each of which corresponds to a retroactive generation of products.
The first leap was from depletion to enhancement. The early D-mode-depleted device fence is highly dependent on a negative pressure break, which requires an external cascade of silicon MOSFET to enter the system, with complex structures and high reliability risks, with only a small number of industrial scenes tested in small batches, and an overall share of less than 5 per cent of the output of the GaN power unit, which is not loaded at all. The p-GaN fence works to make the device an E-mode enhancement, with one-barrel direct and no negative voltage circuit, which is the base of the current commercial.
But the common pain point followed, and the fact that the threshold voltage Vth was unstable and that the dynamic conductivity of Rds-on under high-pressure dynamic stress would be degraded determined that the tube was afraid to get in. There are different movements around this issue. The Inno GaN p-GaN platform for the third generation of the Inno GaN p-GaN, which optimizes the buffer layer structure to contain interfacing charge captures, is accompanied by the handing over of multi-barrel process samples to the Tier1 comparison test; a robust fence bias control programme for the main horizontal nitrogen oxides of the valorized car, which first targets industrial and vehicle-borne 48V auxiliary power, with a conservative rhythm; and the EPC uses low-pressure boosters as the basis for deliberately avoiding the red sea of 650VOBC and placing its energy on industrial high-frequency conversion and robotics. At the heart of this first battle is who, while maintaining high switch speed, will be able to make the fence more stable and whose pipes will be put into the system for a long time.
The second leap was from horizontal to vertical. The horizontal GaN-on-Si is the current mass power, but it is close to the pressure-resisting ceiling at the level of 650 V, which then stacks up the voltage and currents, and the cost of the horizontal structure on the conductive resistance, dispersing thermal path increases dramatically. The plant ' s optimization is focused on buffer, outer fault control to improve dynamic properties, but the physical limits of horizontal structures are there. For 900 Vs or even 1200 Vs, you have to bet on the argon nitrous oxide vertically, i.e. the structure of the GaN-on-GaN lined, which runs vertically. As of the third quarter of 2026, the vertical GaN as a whole was still in the engineering sample delivery phase, and the industrialization was lagging behind the horizontal GaN for at least three to four years. The cost of the homogeneity liners was 8 to 12 times the same size of the GaN-on-Si extension, with the card points concentrated on the homogeneity liner costs, misdensity and high-size good rates.
Businesses have the most divergent approaches in this regard. VisIC chooses to cut directly into the vertical argon, to target the vehicle's main retrovertor, and to develop a prototype jointly with AVL, which is quite radical; the Innosyco Synchronized Vertical aluminum Research and Development, which now prioritizes resources on the land of horizontal car-covered products, using the vertical route as a medium- and long-term technology reserve; and the majority of domestic manufacturers at this stage are still dominated by horizontal platform overlaps, with vertical aluminum remaining at the level of laboratory flow and academic cooperation, with no clear delivery plans. This second battle is better than who can break through the bottom and out, and who can top the voltage-grade ceiling.
The third leap, from a stand-alone device to an integration programme, is the most divisive track on the current route. The separation device is highly flexible and suitable for a self-researched Tier1, with the shortboard being that ring-carrying and EMI debugging is very difficult, and Zhu Xiaowei and EPC push this line. The integration programme has combined the GaN FET, fence pliers, Miller pliers and short-circuit protection, lowering the threshold for system development, and is the current mainstream direction of the OBC landing. ISG6127ATPG-Q, ISG6129BTQ-Q, 650V, an integrated Miller Clip, which was launched by Innosyco, and which specifically addresses the highly misleading flaring of the high frequency dv/dt, has been certified through AEC-Q100, addressed to 6.6 kW OBC and the DC-DC vehicle. Nawe ' s nitrogen-nitrous safe car traffic-power IC travels more extreme, with single-film integration FET, drive, short-circuit protection and ESD protection, built-in GaN Sense testing, primary digitization of power, as well as the development of the whole plant ' s OBC platform, and even 20 years of limited quality security to address the plant ' s reliability concerns. TTI has a different approach, not to single-handedly push nitrogen oxides, but to rely on a complete power source ecology to package nitrogenized argon power tubes and self-researched, self-propelled, control chips into reference design delivery and to reduce the difficulty of transplanting traditional Tier1. This third battle is essentially a business-model struggle, with the sale of a pipe and a power supply programme, with completely different requirements for customer capacity and determining the client structure and bargaining power of each household.
The three leaps have a dark line running through, and the real driving force of each evolution is not how much more switch speed, but rather a step forward in reliability and system availability. To understand this line, it's possible to understand why the nitride is really good at the consumption end, and once in the car code, it's going to go through the door again.
Car regulations confirm the invisible gap.
The rules of the car have always been the strictest test for the coloring of the nitrogen. The consumer product is evaluated for short-term stand-up indicators, with the reliability test cycle for conventional fast-filled devices generally not exceeding 1,000 hours; and the dynamic work of the vehicle code GaN, which sustains the high-temperature cycle, long-term high-pressure stress and load drops, with the basic reliability test under the AEC-Q100 system lasting thousands of hours, superseding dynamic stress additional tests, and significantly lengthening the overall validation period. GaN has several specific types of failure problems, where charge capture gives the device a current collapse under constant high-pressure dynamic stress, where the dynamic Rds drifts up to 20 per cent of the initial value, where the fence deteriorates at high temperatures, and where the temperature cycle leads to containment fatigue. A large number of samples have good indicators in pulse short-duration tests, and once they have been tested for a long period of continuous stress, the performance begins to shift.
Even more problematic is the fact that the AEC-Q101 that exists in the industry is itself built around silicon power units, and the coverage of nitrogenized aluminum dynamic failure is inadequate. Many of the key failure models in the testing system for silicone devices do not apply at all to nitrousium, while the nitroglycerin-specific charge capture, dynamic drift and no equivalent standard test are available, and individual plants and Tier1 can only update their own tests, which are measured in different ways and the results are difficult to compare horizontally. That's why the verification cycle for car azidium is year-round, because many test norms are still touching rocks across the river.
Also, the high-frequency characteristics of nitrogenized aluminum, with completely new requirements for driving IC, magnetic components, PCB layouts, and EMI inhibition, are not enough for a single nitrogenized aluminum device, and the entire power supply system programme needs to be synchronized. The performance of a pipe is better, and if the motor circuit is not up to speed, magnetic components are not matched, the PCB layout is not reasonable, the whole machine as it is, and may even have a reliability accident due to the loop problem. This explains why Innosico and Navi are all going to be integrated products, which essentially mean that the manufacturer is putting some of the system-level anti-turbation and protection capabilities into the device, shifting debugging pressure from Tier1 to the end of the device, and replacing the lower system landing threshold with a higher integration.
In the data centre, the nitro is sneaking in.
If the vehicle is the hottest major narrative of the azimuthine at the moment, the AI data centre is becoming the most undervalued increase. The power of the main AI server has increased from 5 kW in the early years to 40kW or even 60kW in the first days, and the server's panel and rack power has moved from the traditional 12V structure to the high-pressure direct flow structure of the HVDC at 48V or higher voltage. As soon as the voltage is lifted, the silica MOS is exposed to a short plate on efficiency, power density and heat dispersion, and test data show that in the 48V turn 12V board lift, the acetate programme can increase by 1.5 to 3 percentage points compared to the conventional silica MOS unit efficiency and power density by more than 40 per cent, meeting the stringent requirements of the AI server for power efficiency and power density.
And that's why the nitride manufacturers are pushing this line. GaN Safe in Nawe and the package, a large piece of which is placed on the AI server power, combines high-frequency nitrogen oxides with high-efficiency uplifts, to consolidate the data centre per watt. Innosaic is also setting the data centre scene, extending horizontal 650V and low-pressure product lines to the server side. On the industrial side, communications power, storage transformer feeder, and AI Board HVDC are generally considered to be markets with an earlier discharge than a vehicle, as data centre clients do not require a 18 to 24-month car-testing cycle, as in the case of car companies, faster decision-making, more iterative, and much faster validation and start-up of a power programme. The agency predicts that the GAN-related market growth of the global data centre will exceed 45 per cent in 2026, the highest rate of increase in all broken tracks.
From that point of view, the answer is not just whether or not to keep up with the AI wave, but is already on the road. The more AI develops, the higher the calculator density, the greater the pressure of power conversion in a unit area, the clearer the demand for high frequency, high efficiency and high power density for nitrogen oxides. It's only the data centre power efficiency section, not the location under the spotlight like the calculator chip, which is not under the spotlight, but it's starting to appear on the power panel of almost every GPU server. As high-pressure matrix structures such as 800 V, 48 V are spread in data centres, the sense of presence of nitrous oxide in the AI ecology is likely to continue upwards.
Looking at the length of the time axis, the dividends of the fast-filling track are cooling, and the validation of new tracks, such as the vehicle, data centres and so forth, is just beginning. It is difficult to reset the path of a fast-growing era, more like a slow and steady long-cycle power machine business. The main market line over the next year or two is also site validation and small-volume trial production, with limited revenue contribution, with a concentration of discharge carriers in 6.6 kW-borne OBC, communications power, storage support feeder and data centre panel power. The medium-term focus is on the continued optimization of stability in horizontal 650V integrated products, and vertical nitrogenization continues to grind down and spread rates, such as economic turning points. The risks are also specific, as a new generation of silica MOSs continue to be priced down in an iterative manner, squeezing the price-for-money window for nitrous sulfide; and the long cycle of certification of cars, continuous product-overs and validation of the cash flow of the firm, not all of which can handle it.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.