After the fast-charging dividend is exhausted, can GaN keep up with the AI wave?

Fast-charging dividends exhausted, can GaN keep up with the AI wave? Original source: Semiconductor Industry Observer, published 2026-10-30 19:39, author: Zihao. GaN AI-generated poster. The road to transformation faces multiple tests at once: technology, reliability, and commercialization.

A few years ago, GaN entered the public eye in the form of fast-charging heads. Relying on high switching frequency and high power density, it completed a popularization of the power semiconductor track. Today the track's landscape has long been rewritten: after capacity was rolled out, price wars have continued to squeeze profits, and the growth ceiling of the consumer market is visible to the naked eye. The industry has begun looking outward for incremental growth, with automotive OBCs, 48V auxiliary power supplies, and AI server HVDC power supplies becoming key targets for manufacturers.

Moving from consumer-grade products to automotive-grade and computing-power scenarios means validation standards, failure risks, and delivery logic are all restructured. The computing-power wave rolls on, while automotive validation cycles are long. Whether GaN can catch this new wave of demand and escape the growth dilemma after the fast-charging dividend fades—the answer still lies in lengthy testing and design-win implementation.

01 Fast-charging dividends peak, product lines reach a fork in the road

In recent years, GaN has used high-end fast charging as its selling point, with 20W to 140W chargers rolled out one after another, and phone, laptop, and accessory manufacturers piling in. From the earliest 65W GaN chargers being called a charging revolution to today, when a multi-protocol fast-charging head can be bought for a few dozen yuan, this market took only a few years to complete the transition from hit product to standard commodity.

According to industry survey data, in 2025 domestic consumer GaN fast-charging shipments exceeded 220 million units, but the average unit price fell more than 60% from its 2022 peak, and gross margins dropped from above 35% in the early days to the 12% to 18% range. Product forms are concentrated in 650V lateral E-mode discrete transistors and simple half-bridges; design logic prioritizes controlling BOM cost, and reliability validation cycles are short. Whoever can push down the cost of a single transistor gains an advantage in the complete-system quote. Once capacity is released and solutions become increasingly homogeneous, price wars are inevitable, and the era of telling stories through efficiency parameters is over.

The fading fast-charging market dividend is directly reflected in manufacturers' books. On one hand, GaN device unit prices keep falling, with volumes rising but profit space narrowing; on the other hand, complete-system makers have increasingly strong bargaining power over components, forcing device makers to make more extreme trade-offs among cost, performance, and reliability. Manufacturers are forced to make product-line choices: some continue to deepen the consumer track, with iteration focused on Qg optimization, package miniaturization, and wafer yield improvement, pursuing volume through cost advantages and turning the fast-charging market into an existing business that sustains cash flow; others begin building separate automotive-grade and industrial product lines, redefining device specifications, establishing batch control rules, and lengthening validation cycles to trade time for higher gross margins and steadier orders. Target gross margins for automotive-grade GaN devices are generally maintained above 25%, roughly double those of consumer-grade products, but R&D and validation investment also increases exponentially.

When discussing GaN's application prospects, one cannot avoid its relationship with silicon carbide, and the two are often conflated. Many observers outside the industry treat wide-bandgap semiconductors as a whole, thinking GaN and silicon carbide are the same kind of thing and that whichever performs better wins. But the root of their difference in application breadth and value lies in the division of labor by voltage class and switching characteristics, not in who performs better. In terms of market size, the global SiC power device market was about $3.8 billion in 2025, while the GaN power device market over the same period was about $870 million—a clear gap in scale, behind which lies the natural difference brought by scenario division of labor.

Silicon carbide's calling is high voltage. It is a wide-bandgap material with excellent voltage withstand, high-temperature tolerance, and thermal conductivity, and its main battlefield is automotive main inverters, high-power charging piles, and photovoltaic inverters at 800V or even above 1200V. These scenarios have high power and high bus voltage; silicon carbide can directly replace silicon IGBTs in exchange for lower conduction losses and higher switching frequency, and the value per device is also high, making it the most expensive and fastest-growing category by shipment scale among wide-bandgap power devices. That is why a large number of third-generation semiconductor reports cluster around silicon carbide substrates and main drives: the silicon carbide story is high power, high value, and concentrated volume ramp, the most straightforward to tell.

GaN is different; it excels at low voltage and high frequency. Current commercial GaN-on-Si lateral devices have a main voltage range of 100V to 650V, much lower than silicon carbide, but with extremely fast switching speed, high efficiency, and small size. Its home turf is 6.6kW to 11kW automotive OBCs, 48V/12V DC-DC, industrial servos, communications power supplies, and the warming data center board-mounted power supply segment. Silicon carbide does the heavy, high-power, high-voltage work; GaN does the fine, high-frequency, high-efficiency, compact work. In systems the two often divide labor and cooperate rather than substitute for each other. This also explains why the GaN market has always appeared smaller than silicon carbide: its unit price, per-scenario power volume, and total value are naturally lower than silicon carbide's.

Even among GaN players, the product positioning of Innoscience, Navitas, China Resources Microelectronics, TI, EPC, and VisIC has already clearly diverged. Some bet on capacity and cost, some on system-level solutions, some on integration level, and some simply bypass the red ocean to stake out niche scenarios. This divergence itself is a signal that the industry has entered its second half: companies must balance the short-term revenue brought by fast charging against the long, high-investment validation cycle of automotive product lines, and this balancing act precisely determines how far each company can go in the coming years.

02 Three technology leaps, each corresponding to different products

What truly determines the direction of this transformation is the evolution of gallium nitride device technology itself, and behind each round of evolution lie different product approaches from different companies. Looking at the industry history of these two decades in full, GaN power devices have actually fought three hard battles, each corresponding to a generation of traceable products.

The first leap was from depletion mode to enhancement mode. Early D-mode depletion devices required a negative voltage to turn off the gate, and had to be paired with an external cascode silicon MOSFET to enter a system; the structure was complex, with many reliability risks, and they were only trialed in small batches in a few industrial scenarios, accounting for less than 5% of overall GaN power device shipments—far from being ready for automotive use. After p-GaN gate engineering turned the device into E-mode enhancement mode, enabling direct single-transistor drive without a negative voltage circuit, this became the current commercial foundation.

But common pain points followed: whether the threshold voltage Vth is stable, and whether the dynamic on-resistance Rds-on degrades under high-voltage dynamic stress, determine whether this transistor dares to go into a vehicle. Around this issue, companies acted differently. Innoscience iterated its third-generation Inno GaN p-GaN platform, optimizing the buffer layer structure to suppress interface charge trapping, while in parallel delivering multiple versions of gate process samples to Tier1s for comparative testing; China Resources Microelectronics' automotive-grade lateral GaN devices focus on a robust gate bias control scheme, initially targeting industrial power supplies and automotive 48V auxiliary power supplies, with a rather conservative pace; EPC, building on low-voltage enhancement-mode devices, deliberately avoids the crowded 650V OBC market, focusing its efforts on industrial high-frequency conversion and robotic servo drives. The core of this first battle is who can make the gate more stable while maintaining high switching speed—only then does their transistor dare to be truly installed in systems for long-term operation.

The second leap was from lateral to vertical. Lateral GaN-on-Si is the current mass-production mainstay, but at the 650V voltage class it is already approaching the ceiling of voltage withstand and thermal management; pushing voltage and current higher would sharply increase the cost of lateral structures in on-resistance and heat dissipation paths. Manufacturers' optimization efforts focus on buffer layers and epitaxial defect control to improve dynamic characteristics, but the physical limits of lateral structures are there. To reach 900V or even 1200V, one must bet on vertical GaN, i.e., GaN-on-GaN substrates with current flowing vertically. As of the third quarter of 2026, vertical GaN as a whole is still in the engineering sample delivery stage, with industrialization progress lagging lateral GaN by at least 3 to 4 years; homogeneous substrate cost is 8 to 12 times that of GaN-on-Si epitaxial wafers of the same specification, with bottlenecks concentrated in homogeneous substrate cost, dislocation density, and large-size yield.

Companies' approaches in this area diverge the most. VisIC chose to go directly into vertical GaN, targeting automotive main-drive inverters, and jointly developed a demonstration vehicle with AVL—a fairly aggressive route; Innoscience is simultaneously laying out vertical GaN R&D, but currently prioritizes resources on landing lateral automotive-grade co-packaged products, treating the vertical route as a medium- to long-term technology reserve; while most domestic manufacturers at this stage still focus on lateral platform iteration, with vertical GaN remaining at the laboratory tape-out and academic collaboration level, without clear sample delivery plans. This second battle tests who can break through first in substrates and epitaxy, and whoever does can push the ceiling of voltage class up another notch.

The third leap was from discrete devices to integrated solutions, and this is also the track with the greatest divergence in routes today. Discrete devices offer high flexibility and suit Tier1s with strong in-house capabilities, with the shortcoming of loop parasitics and difficult EMI debugging; China Resources Microelectronics and EPC mainly promote this line. Integrated solutions package the GaN FET, gate clamp, Miller clamp, and short-circuit protection together, lowering the system development threshold, and are the mainstream direction for current OBC deployment. Innoscience's Solid GaN automotive-grade co-packaged devices ISG6127ATPG-Q and ISG6129BTQ-Q, 650V, integrate a Miller clamp, specifically addressing gate false turn-on caused by high-frequency dv/dt, have passed AEC-Q100 certification, and target 6.6kW OBCs and automotive DC-DC. Navitas' GaN Safe automotive-grade power IC goes even further, monolithically integrating the FET, driver, short-circuit protection, and ESD protection, with built-in GaN Sense detection, featuring digital input and power output, and also supporting OEM OBC platform development, even using a twenty-year limited warranty to dispel automakers' reliability concerns. TI's approach is different again: it does not separately push GaN devices, but relies on its complete power ecosystem to package GaN power transistors with its own drivers and control chips into reference designs for delivery, lowering the migration difficulty for traditional Tier1s. This third battle is essentially a battle of business models: selling a transistor versus selling a power solution places completely different demands on customer capabilities, and also determines each company's customer structure and bargaining power.

A hidden thread runs through all three leaps: the real driver of each evolution is not how much switching speed has improved, but that reliability and system availability have moved forward one step. Only by understanding this thread can one see why GaN, despite being highly competitive in consumer applications, must pass through the gates of hell all over again once it reaches automotive grade.

03 The invisible chasm of automotive-grade validation

Automotive-grade qualification has always been the strictest testing ground for proving GaN's true quality. Consumer products are judged on short-term bench metrics, and reliability testing cycles for conventional fast-charging devices generally do not exceed 1,000 hours; automotive-grade GaN must continuously withstand dynamic conditions such as high-low temperature cycling, prolonged high-voltage stress, and load dump, and under the AEC-Q100 framework alone, baseline reliability testing runs to several thousand hours, which, when supplemented with additional dynamic stress tests, greatly lengthens the overall validation period. GaN has several unique failure mechanisms: charge trapping causes current collapse in devices under sustained high-voltage dynamic stress, with dynamic Rds drift reaching as much as 20% or more above the initial value; the gate degrades at high temperatures; and temperature cycling also brings package fatigue. Many samples look excellent in short pulsed tests, but once subjected to prolonged continuous stress testing, their performance begins to shift.

What is more troublesome is that the industry's existing AEC-Q101 was itself built around silicon power devices and does not adequately cover GaN's dynamic failures. In the silicon device test framework, many key failure modes simply do not apply to GaN, while GaN-specific charge trapping and dynamic drift have no corresponding standard test items, so device makers and Tier 1 suppliers can only add test items on their own, each testing in its own way, making results hard to compare across the board. This is why automotive-grade GaN validation cycles routinely run for years: many test specifications are still being figured out by trial and error.

Moreover, GaN's high-frequency characteristics impose entirely new requirements on driver ICs, magnetic components, PCB layout, and EMI suppression; simply swapping in a GaN device is not enough—the entire power system solution must be iterated in tandem. No matter how good a single transistor performs, if the driver circuit does not keep up, the magnetic components do not match, and the PCB layout is unreasonable, the whole unit still will not deliver results, and reliability incidents may even occur due to loop issues. This also explains why Innoscience and Navitas both pursue integrated products: essentially, manufacturers build part of the system-level immunity and protection capability into the device, shifting debugging pressure from Tier 1 suppliers to the device side, trading higher integration for a lower barrier to system deployment.

04 In data centers, GaN is quietly entering

If automotive is currently the hottest main narrative for GaN, then AI data centers are becoming its most underestimated growth area. The AI wave has driven up the power consumption of computing clusters; mainstream AI server racks have risen from 5 kW in earlier years to 40 kW or even 60 kW, and server onboard power and rack power are gradually shifting from the traditional 12V architecture to 48V or even higher-voltage HVDC architectures. Once the voltage goes up, the shortcomings of silicon MOSFETs in efficiency, power density, and heat dissipation are exposed. Test data show that in a 48V-to-12V onboard topology, a GaN solution can improve overall unit efficiency by 1.5 to 3 percentage points compared with traditional silicon MOSFETs and increase power density by more than 40%, precisely matching AI servers' stringent requirements for power efficiency and power density.

This is also why GaN manufacturers are all pushing hard along this line. A major part of Navitas's GaN Safe and accompanying solutions targets AI server power supplies, combining high-frequency GaN with efficient topologies to compete on data center cost per watt. Innoscience is also laying out data center scenarios, extending both its lateral 650V and low-voltage product lines toward server power supplies. On the industrial side, telecom power supplies, energy storage converter branches, and AI onboard HVDC are widely regarded as markets that will scale earlier than automotive, because data center customers, unlike automakers, do not require an 18-to-24-month automotive-grade validation cycle; they decide faster and iterate more urgently, and the validation and ramp-up pace for a power solution is much quicker. Institutions predict that in 2026 the global data center GaN power-related market will grow by more than 45%, the highest growth rate among all sub-segments.

From this perspective, whether GaN can keep up with the AI wave—the answer is not just that it can, but that it is already on the way. The further AI develops, the higher the computing density, the greater the power conversion pressure per unit area, and the clearer the demand for high-frequency, high-efficiency, high-power-density GaN. It is just that GaN captures the data center power efficiency segment rather than the spotlight position of computing chips; it is not in the spotlight, but on almost every GPU server's power board, its presence is beginning to appear. As high-voltage bus architectures such as 800V and 48V spread in data centers, GaN's presence in the AI ecosystem will most likely continue to rise.

Taking a longer view, the dividends of the fast-charging segment are cooling, and the long validation marathon for new segments such as automotive and data centers has only just begun. GaN will find it hard to replicate the explosive path of the fast-charging era; it is more like a slow and steady long-cycle power device business. Over the next year or two, the market's main thread will still be design-in validation and small-batch trial production, with limited revenue contribution, and volume ramp-up concentrated in 6.6kW automotive OBCs, telecom power supplies, energy storage auxiliary branches, and data center onboard power supplies. The medium-term focus is on continuously optimizing the stability of lateral 650V integrated products and continuing to improve substrate and epitaxy yields for vertical GaN while waiting for the economic inflection point. The risks are equally concrete: a new generation of silicon MOSFETs continues to iterate and fall in price, which will squeeze GaN's cost-performance window; automotive-grade validation cycles are long, and continuous product iteration and validation investment test companies' cash flow—not every company can withstand it.

This content is original to the platform, and the copyright belongs to the platform. Reproduction without permission is prohibited; to reprint, please contact the platform. Comments: No user comments yet. Latest News More Related Recommendations

Winter has arrived, and AI is abandoning phones and PCs 2026-10-30 19:36

Computing Gathers in Guangzhou, Chips Create the Future—Yuxian Microelectronics Successfully Hosts Guangzhou Computing Chip Industry Ecosystem Conference 2026-10-30 19:36

From 3nm to 2nm, what exactly makes Apple's A20 Pro strong? 2026-10-30 19:34