Power delivery in AI data centers requires stepping down the voltage from the grid to the chip in four stages, with different power devices used at each stage. The high-voltage end is the battleground for SiC (Silicon Carbide), the intermediate bus is contested between GaN (Gallium Nitride) and silicon-based super-junction MOSFETs, and the board-level low-voltage end relies on silicon-based shielded gate MOSFETs and analog PMICs (Power Management ICs).
The voltage ratings across these four stages differ by orders of magnitude, making it difficult to transfer technological accumulation. A company's leadership at one stage does not automatically extend to adjacent stages. Therefore, the question of "how far domestic substitution has progressed" can only be answered when applied to specific voltage levels.
This article first explains the device division across the four voltage rails, then clarifies the boundary between GaN and SiC, and finally reviews the periodic reports of six domestic power semiconductor companies to examine the specific voltage level and progress of their disclosed products.
I. Four Voltage Rails, Four Types of Devices
From the grid to the GPU, the voltage undergoes four conversions.
The first stage is high-voltage power distribution from the grid to the data center. 10kV medium-voltage distribution and the main circuit of solid-state transformers require high-voltage SiC modules and flexible DC technology, with voltage levels ranging from 1700V to 6500V. The technological lineage of this stage originates from rail transit and UHV (Ultra-High Voltage) transmission, with almost no overlap with consumer electronics.
The second stage is the high-voltage DC bus of the rack. Under an 800V high-voltage DC architecture, if a ±400V bipolar system is adopted, the ground voltage of each pole is controlled at the 400V level, and the primary side of the intermediate bus converter only requires 650V-rated devices to cover it. This tier is exactly where GaN and silicon-based super-junction MOSFETs compete directly, and it is also the voltage level where 650V devices have achieved large-scale mass production and supply chain maturity in main drives and onboard charging for NEVs (New Energy Vehicles).
The third stage is from the intermediate bus to 48V. This stage has the highest requirements for switching frequency and power density, and the high-frequency characteristics of GaN yield the most direct benefits here. NVIDIA requires the integration of high-density supercapacitor modules inside the chassis and the introduction of GaN fast synchronous rectification on the secondary side to support transient current protection of up to 180% under transient overload conditions.
The fourth stage is from 48V to the chip core voltage. The GPU transitions from idle to full load within microseconds, with a current slew rate exceeding 1000A/μs. This stage is handled by multi-phase digital controllers, smart power stages, and silicon-based shielded gate MOSFETs, falling into the realm of analog design and specialized processes, which differs from the technological paths of the first three stages.
The key difference among the four stages lies in the voltage ratings. There is almost no transferable process and reliability accumulation between high-voltage SiC above 1700V and board-level devices below 48V. This explains a phenomenon: domestic manufacturers show distinct stratified clustering across the four stages, with few cross-tier cases, and the direction is mostly from low voltage to medium voltage.
Figure 1 | Four voltage rails and device division. The voltage levels are typical values stated in public technical materials, and actual selection varies with the topology.
II. The Boundary Between GaN and SiC Lies Not in Advancement Level
The two wide-bandgap materials are often discussed side by side, but they do not compete for the same position in AI power supplies.
GaN is a lateral device with high electron mobility, low gate charge, and no body diode reverse recovery, making it suitable for pushing up the switching frequency. As the frequency increases, the volume of magnetic components decreases accordingly, improving the power density of the power supply. The trade-off is voltage withstand capability; current mainstream mass production platforms are concentrated in the 650V and 700V tier, and scaling upward is relatively difficult.
SiC is a vertical device with higher critical breakdown field strength and thermal conductivity, making it suitable for high voltage withstand and high current. Therefore, it guards grid-side rectification, high-voltage buses, and the main circuits of solid-state transformers, with product voltage levels covering 650V to 6500V.
Silicon-based devices have not been replaced. In the 650V tier, silicon-based super-junction MOSFETs still compete directly relying on mature processes and cost advantages; below 48V, shielded gate MOSFETs remain the main force. Multiple manufacturers used terms like "symbiosis" and "coexistence" in their 2025 annual reports to describe this landscape.
Therefore, this round of replacement must be viewed at specific voltage levels. Currently, publicly evidenced replacements have only occurred in two places: the primary side of the 650V intermediate bus and the high-frequency synchronous rectification on the secondary side.
Figure 2 | The boundary between silicon, GaN, and SiC. The voltage withstand ranges are public calibers of current mainstream mass production platforms and do not represent the technical upper limit.
III. Three Criteria for Judging "How Close to Server Power Supplies"
Annual reports of power device manufacturers generally mention AI servers and data centers, but the certainty of the statements varies greatly. There are three verifiable criteria.
First, whether the product has been explicitly used in mass production for this scenario. "Can be widely applied in AI server power supplies" is a description of product capability, whereas "achieved mass production and shipment to core domestic data center customers" is the usage result; the certainty levels of the two differ by a tier.
Second, whether revenue from this application or the customer adoption stage is disclosed. Revenue caliber is the hardest, followed by "completed product adoption," "obtained certification and orders from leading customers," and then "entered reliability verification."
Third, the manufacturing model determines the upper limit of next-generation parameters. The three models—IDM with self-built production lines, virtual IDM, and asset-light manufacturing—have different levels of control over process iteration. This is particularly evident in wide-bandgap devices, as device performance highly depends on epitaxy and processes.
Based on these three criteria, the six companies can be placed in the same table for comparison.
IV. Positions of the Six Companies
The following content is extracted from the periodic reports of each company. Except for Innoscience, the 2026 semi-annual reports of the five A-share companies were disclosed sequentially from August 20 to 25. The financial and operational statements in this article are based on the semi-annual reports, and technical and product calibers cite the 2025 annual reports where the semi-annual reports do not cover. Non-disclosure does not mean non-achievement, but the externally verifiable parts are strictly limited to disclosed content.
Innoscience: The Only Company Disclosing Revenue from AI and Data Centers
Innoscience is the world's first pure IDM company to achieve large-scale mass production of 8-inch silicon-based GaN, following an asset-heavy route with self-built production lines.
Its 2025 annual report disclosed three developments directly related to this article: first, through deep cooperation with NVIDIA, the company became the only Chinese chip supplier in its 800V high-voltage DC power architecture, and more than half of the 800VDC power solutions constructed with its GaN devices as the core; second, 650V products achieved mass production and shipment to core domestic data center customers, and Gen3 650V high-power products completed product adoption at key AI high-voltage DC customers; third, the 100V series products with double-sided cooling En-FCLGA packaging have been adopted by multiple leading customers in AI 800VDC and 48VDC, with the company claiming a 20% increase in power density and a 1% increase in efficiency.
Notably, it disclosed revenue calibers: in 2025, GaN chip sales revenue in the AI and data center fields reached 63.19 million RMB, a year-on-year increase of 50.20%. During the same period, the company's total sales revenue was 1.213 billion RMB, and consumer electronics product revenue was 568 million RMB.
The reporting period needs to be clarified: Innoscience is listed on the Hong Kong Stock Exchange, and the statutory disclosure period for interim reports is within three months after the end of the reporting period. As of press time, its 2026 interim results have not been disclosed, so the above data remains the 2025 annual caliber, which is not in the same reporting period as the semi-annual data of the five A-share companies below.
Adoption has genuinely occurred, with clear customers and mass production, but the scale remains small. 63.19 million RMB accounts for a single-digit percentage of the company's revenue that year, with the main revenue still coming from charging equipment and consumer electronics. At this scale, the growth rate illustrates the progress of this route better than the absolute value.
CR Micro and Silan Micro: Dual-Line IDM Advancement, Revenue from This Scenario Not Yet Disclosed
The 2025 annual report of CR Micro states that the GaN epitaxy center was officially launched, and capacity expansion is advancing in an orderly manner; multiple products on the D-mode 650V G5 platform have entered the mass production stage. The company's newly launched fourth-generation D-mode GaN series products are stated to be "widely applicable in high-growth fields such as AI server power supplies, onboard chargers, LiDAR, robot joint drives, and high-end fast charging." On the low-voltage side, the E-mode 40V platform has passed ultra-high reliability verification, obtaining certification and orders from leading customers; meanwhile, 80V/100V/650V platforms are being developed for industrial control high-power power supplies and digital server power supplies. Regarding SiC, the new generation SiC MOS G5 and SiC Trench JFET G1 are expected to be mass-produced and launched in 2026.
The 2026 semi-annual report supplements the progress and growth rate. The company claims to have formed a full-link solution layout around cloud scenarios such as data centers and optical chip data transmission, achieving rapid volume ramp-up at leading customers in key fields; during the reporting period, MOSFET business revenue increased by over 20% year-on-year, of which shielded gate MOSFET revenue grew by 40% year-on-year, while revenue for advanced-generation super-junction MOSFETs (SJ MOS G4) corresponding to the high energy efficiency demands of AI data centers surged by 255% year-on-year.
Based on the criteria in Section III, CR Micro has a clear mass production platform and customer certification for GaN, and provided a growth rate caliber for silicon-based super-junctions, but still has not separately disclosed the revenue amount for the data center scenario.
Silan Micro is one of the few domestic IDM companies with both silicon-based and compound production lines. Its 2025 annual report shows that the 8-inch SiC power device chip manufacturing production line was operational in Q4 2025 and is expected to be officially put into production in the second half of 2026; the statement regarding GaN is "accelerating product R&D and technological maturity of 8-inch automotive-grade GaN power devices," pointing to automotive rather than server applications.
What is closer to server power supplies for Silan Micro is the chip side: the report states that DrMOS circuits, eFuse circuits, and multi-phase controller circuits applied to servers have been tested at customer sites or introduced into mass production. Its 2026 semi-annual report continues the same statement, claiming it will target automotive-grade analog circuits and power circuits related to computing servers, continuing to advance the R&D and mass production ramp-up of the 12-inch analog process platform.
Xinjien and Oriental Semiconductor: On the Silicon Side, Progress is Clearer
Xinjien's progress is concentrated on silicon-based shielded gate MOSFETs, which is the last stage of the four voltage rails. Its 2025 annual report states that the comprehensive performance figure of merit (FOM) of devices on the N150V third-generation SGT MOS platform is reduced by over 25% compared to the best competing products of the same specifications in the industry, and products on this platform have been used in large volumes in fields such as server power supplies; N40V third-generation SGT products have fully entered the mass production stage, involving industries including high-power data centers; the N25V third-generation platform is undergoing reliability verification, with target markets being AI computing and high-power data centers. The company also states that its products have passed strict certification by leading customers in tracks such as AI servers and data centers and achieved scaled mass production and sales.
Its 2026 semi-annual report lists AI servers and data centers as key application fields, and claims to have over 4,000 product models with voltage coverage across the full series from 12V to 1700V.
Based on the criteria in Section III, Xinjien's disclosure at the below-48V stage is the most specific, covering the three stages of mass production, certification, and under development, but this stage itself does not involve GaN.
Oriental Semiconductor's position is in high-voltage silicon and SiC. Its 2025 annual report states that SiC MOSFETs and SiC SBDs have achieved mass production, while simultaneously advancing the development of high-performance SiC JFETs and GaN HEMTs, simultaneously advancing the development of high-performance SiC JFETs and GaN HEMTs—GaN remains in the development stage in its disclosure. Regarding power modules, the company claims its product line forms a complete solution covering all voltage levels and is gradually being mass-produced in application fields such as computing server power supplies. Its 2026 semi-annual report further states that dedicated controllers have been launched for multiple advanced topologies common in server power supplies, with key technical indicators reaching levels comparable to international leading manufacturers, and a separate entry for Artificial Intelligence Data Center (AIDC) is listed in the glossary.
CRRC Times Electric: High-Voltage End, Not in the Same Tier as the Previous Five
The technological foundation of CRRC Times Electric comes from high-speed rail conversion, high-power IGBTs, and UHV flexible DC transmission. Its SiC MOSFET product line covers 650V to 6500V, and the 7th generation trench field-stop IGBTs have achieved mass production. Among the four voltage rails, it is one of the few domestic companies with engineering capabilities for the main circuit of 10kV-class solid-state transformers.
It is included in the comparison because if there is no domestic supply for the topmost tier among the four, no matter how well the lower three tiers are done, the entire chain remains incomplete.
The Six Companies Side by Side
First, look at their respective positions on the voltage rails and material routes.
Next, look at the latest financials. The five A-share companies are based on the 2026 semi-annual report caliber; Innoscience's interim report is not disclosed, so 2025 annual data is listed.
Net Profit Growth Rate Should Be Viewed After Deducting Non-recurring Items
After reading this financial table, it can be found that the growth rate of net profit attributable to shareholders of the parent company has limited reference value among this group of companies.
Silan Micro's net profit attributable to shareholders in the first half of the year increased by 94.84% year-on-year, but only grew by 0.67% after deducting non-recurring items; Oriental Semiconductor's net profit attributable to shareholders increased by 285.41% year-on-year, but decreased by 92.91% year-on-year after deducting non-recurring items, with the deducted amount being only 689,000 RMB. The profit growth of both companies mainly comes from non-recurring gains and losses, rather than main businesses. Xinjien is a different case: operating revenue increased by 25.57% year-on-year, net profit attributable to shareholders decreased by 0.75% year-on-year, and decreased by 6.49% after deducting non-recurring items, belonging to the category of increased revenue without increased profit. Among the four A-share power device companies, only CR Micro achieved simultaneous high growth in both attributable and deducted net profits, with a 51.48% year-on-year increase in deducted net profit.
This is not a financial flaw, but the current universal state of this segment: the time for product adoption in AI power supplies is not long, and the revenue scale is not yet large enough to change the overall profit structure. When citing the growth rates of this group of companies, the deducted caliber illustrates changes in the main business better than the attributable caliber.
Revenue Scale Does Not Correspond to Position
Placing the two tables side by side reveals a point contrary to common sense: Innoscience, which is at the forefront of the AI power supply chain and the only one disclosing revenue from this scenario, has the smallest revenue scale among the six and is not yet profitable; CRRC Times Electric, with the largest revenue scale, is positioned at the high-voltage end of the grid, still two tiers away from server power supplies, and its net profit attributable to shareholders in the first half of the year increased by 2.44% year-on-year, which is not synchronized with the high growth of AI demand.
The reason lies in the fact that the two measure different things. Current revenue reflects the scale of each company's existing main business, which is distributed in unrelated fields such as consumer electronics fast charging, automotive and white goods, industrial, as well as rail transit and the grid, with basically no intersection with data center power supply; whereas the position in AI power supplies reflects where the next round of increments may fall.
It needs to be clarified simultaneously that the absolute scale of this chain is still limited at present. The AI and data center GaN revenue disclosed by Innoscience is 63.19 million RMB, which is the only verifiable revenue data for this scenario among the six; CR Micro disclosed a 255% year-on-year increase in revenue for advanced-generation super-junction MOSFETs but did not provide the amount; the other four have not separately disclosed revenue or growth rates for this scenario. Under the condition of mutually different calibers, horizontal comparison at the scale level is still impossible for this segment.
Figure 3 | Positions of the six companies on the four voltage rails. Positioned according to the public disclosures in each company's 2026 semi-annual report and 2025 annual report, reflecting product positioning and disclosure degree, not a competitiveness ranking.
V. Conclusion
Based on current public disclosures, the replacement of GaN in AI data centers has indeed begun, but the scope is limited: verifiable mass production evidence has only been obtained in two places—the primary side of the 650V intermediate bus and the high-frequency synchronous rectification on the secondary side, which are exactly where its frequency advantages can be directly realized. The three materials coexist in tiers in AI power supplies, rather than replacing each other sequentially.
What needs more attention is the scale. Among the six companies, only Innoscience separately disclosed the revenue amount of 63.19 million RMB for this scenario; CR Micro disclosed the growth rate of advanced-generation super-junction MOSFETs but did not provide the amount; the other four did not separately disclose it. With only one data point, an industry caliber cannot be established, and the domestic progress in this segment can only be judged based on the product stages disclosed by each company, making horizontal comparison by scale impossible.
Therefore, the next signal worth paying attention to is very specific: whether a second company will disclose the revenue caliber for AI data center applications.
Disclaimer: This article is industrial research content and does not constitute any investment advice. The financial data in the text are all taken from the publicly disclosed periodic reports of the companies. The composition of the main business of each company is different and does not have a corresponding relationship with the data center power supply business. The voltage level positioning in the text is made based on the publicly disclosed calibers of each company, reflecting product positioning and disclosure degree, and does not constitute a ranking of market share or competitiveness; non-disclosure does not mean non-achievement. The product stage terms disclosed by each company (mass production, adoption, certification, reliability verification, R&D) are cited according to the original text of the annual reports and have not been treated as equivalent.
References: (Slide up and down to view):
[1] Innoscience (2577.HK) | 2025 Annual Report (P9, P14, P15) | HKEXnews (searched by stock code 2577) [2] CR Micro (688396.SH) | 2026 Semi-Annual Report (2026-08-25) [3] CR Micro (688396.SH) | 2025 Annual Report (P16, P17, P18) [4] Silan Micro (600460.SH) | 2026 Semi-Annual Report (2026-08-25) [5] Silan Micro (600460.SH) | 2025 Annual Report (P10, P15, P37, P38) [6] Xinjien (605111.SH) | 2026 Semi-Annual Report (2026-08-25) [7] Xinjien (605111.SH) | 2025 Annual Report (P19, P22) [8] Oriental Semiconductor (688261.SH) | 2026 Semi-Annual Report (2026-08-20) [9] Oriental Semiconductor (688261.SH) | 2025 Annual Report (P21, P24, P25) [10] CRRC Times Electric (688187.SH) | 2026 Semi-Annual Report (2026-08-20)
#DeepChipAlliance #GBAChipExpo #SemiconductorIndustryResearch #PowerSemiconductor #AI #ServerPowerSupply #Silicon #GaN