Chinese silicon carbide (SiC) has been mass-produced and supplied for AI server power supplies, and has even entered the manufacturing processes of overseas major manufacturers. However, if we dig one level deeper: "Which voltage rating of devices is used at which stage of the 800V power supply chain?" In public information, no Chinese company can clearly answer this.
Therefore, this article does not focus on whether they have entered or not, but rather on two things: first, why their exact position cannot be clearly stated; second, under the premise of relying solely on public information, what can and cannot be deduced from this blind spot.
Yangjie Technology disclosed an investor relations activity record on July 1, 2026, in which it mentioned two things.
When discussing automotive-grade SiC, it is precise down to the factory level: one SiC automotive-grade wafer fab, three SiC automotive product packaging and testing factories, strict separate line management for automotive and non-automotive grades, equipped with a CNAS-certified laboratory. When discussing a 1200V SiC MOSFET, it is precise down to the project level: introduced into the core Tier 1 supply chain of mainstream domestic OEMs, used in 6.6kW/11kW OBC (On-Board Charger) systems on 400V/800V platforms, and also supporting an 11kW/22kW high-power OBC project for a Japanese brand's 800V high-voltage platform.
When it comes to AI server power supplies, it only has five words, sandwiched in a sentence stating it "can be widely applied in multiple downstream fields such as NEVs (New Energy Vehicles), photovoltaic energy storage, industrial control, AI server power supplies, and consumer power supplies."
The same company, the same document, the same day. It can specify the exact stages for automotive-grade applications, but cannot do so for data centers.
This gap itself is exactly what this article intends to discuss.
Figure 1 | Three original text excerpts from the same company and the same document: automotive-grade applications are precise down to factories and projects, while AI server power supplies only reach the application field level. Source: Yangjie Technology Investor Relations Activity Record Form 2026-006
I. Chinese Companies Have Already Entered: It's Just That Public Information Doesn't Show the Exact Position
Let's first set aside a popular claim: Chinese SiC hasn't entered AI data centers. This statement no longer holds true in 2026.
Sanan Optoelectronics stated very clearly in its interactive platform reply on April 23, 2026: In the field of data centers and AI server power supplies, the SiC products of Hunan Sanan have achieved mass production for leading customers such as Delta, Lite-On, Vertiv, Great Wall, and Flex. It named five customers and confirmed mass production. The company also disclosed its SiC diode and MOSFET product portfolio ranging from 650V to 2000V in its 2025 semi-annual report.
Yangjie Technology's SiC MOSFETs cover three voltage ratings: 650V, 1200V, and 1700V. The company claims that various products have been widely used in AI data centers.
Silan Microelectronics'8-inch SiC production line was commissioned in January 2026. According to the company, this line will focus on serving application scenarios such as NEVs, photovoltaic energy storage, charging piles, AI server power supplies, and industrial power supplies.
Here, the two pieces of evidence must be discussed separately. On the one hand, Sanan has achieved mass production for AI server power supply customers; on the other hand, STMicroelectronics has included its 200mm SiC joint manufacturing capability with Sanan in its Cloud & AI supply system. This indicates that Chinese manufacturing has entered the capacity layout of overseas IDMs (Integrated Device Manufacturers), but it cannot be asserted based on this that the joint venture's products have been supplied to a specific 800V project.
Turning the lens to the other side, the picture is completely different.
As of August 2026, based on public searches, the list of 14 chip suppliers announced in October 2025 has not yet been updated. The only mainland Chinese chip company on the list is Innoscience, which mainly focuses on gallium nitride (GaN), while the number of mainland Chinese SiC brands remains zero.
As for the position of high-voltage SiC, the corresponding relationships that can be found in public technical documents are as follows:
STMicroelectronics' 800V hot-swap protection uses 1200V SiC.
STMicroelectronics' 800V to 50V conversion uses stacked 650V GaN.
ROHM's BBU (Battery Backup Unit) backup power, with the battery side at about 560V, uses 750V SiC.
Infineon has included 1200V to 3300V SiC MOSFETs in its power module solutions for solid-state transformers; Navitas's GeneSiC has incorporated ultra-high voltage 3300V and high voltage 1200V SiC modules into a solid-state transformer solution for AI data centers, jointly demonstrated with EPFL; Wolfspeed also has ready-to-use mass-produced 2300V half-bridge modules, targeting the exact same medium-voltage scenario.
Note the difference between these two sets of information. On the Chinese side, they say, "Our products have entered these customers"; on the overseas side, they say, "Which voltage rating of our devices is used at which circuit stage, how the reference design is built, and how it coordinates with the platform."
The former is a supply relationship; the latter is a solution position.
A degree of restraint is needed here. The ecosystem list can only prove cooperation, not mass volume. The fact that overseas manufacturers publicly disclose solutions with clear stages and platform coordination does not mean they have shipped in large volumes. Conversely, the fact that Chinese manufacturers have not publicly disclosed the stages does not mean they haven't participated. The only thing that can be confirmed is this: at the level of public information, the granularity of the two sides is not on the same order of magnitude.
II. 800V Bus Does Not Equal 800V Devices
Voltage ratings are determined by the topology; voltage rating alone cannot explain it, nor can it clarify the position.
Let's first dismantle a misunderstanding that almost everyone falls into.
800V is not equivalent to 800V devices. It defines the bus voltage, not the device specifications.
Looking back at the previous set of corresponding relationships makes it clear: hot-swap protection uses 1200V, 800V to 50V conversion uses stacked 650V, and BBU uses 750V. The voltages 650V, 750V, and 1200V appear repeatedly in this chain; Chinese manufacturers have already covered multiple mainstream ratings. Yangjie's publicly disclosed product portfolio includes 650V, 1200V, and 1700V, while Sanan's product layout extends to 2000V.
Figure 2 | In the same 800V chain, the voltage ratings used at each stage are not the same; blue indicates the ratings already covered by mass production from Chinese manufacturers. Sources: Public technical documents from ST, ROHM, Infineon, Navitas, and Wolfspeed; periodic reports of each company.
The fact that voltage ratings are determined by the topology is also clearly seen in other scenarios. A three-level rectifier reference design by STMicroelectronics for 400V AC to 800V DC conversion—the original application is marked as EV charging, not data centers—uses both 650V SiC MOSFETs and 1200V SiC diodes. With the same 800V output, changing the topology leads to a completely different choice of device voltage ratings.
Then does voltage rating pose no problem at all? Not exactly.
What truly pushes up the voltage rating requirements is the medium-voltage solid-state transformer (SST) layer. But there are two facts to state simultaneously about this layer. First, it does not inherently require single 2300V or 3300V devices—multi-level and cascaded topologies can achieve the same voltage capability by connecting lower-voltage devices in series. Second, this layer is still in the stage of joint development and validation overseas—Infineon is jointly designing, optimizing, and validating a 2 to 5 MW SST module with SolarEdge, while Navitas and EPFL's 250 kW solution is an exhibition demonstration.
The situation on the Chinese side deserves a separate look. CRRC Times Electric has formed SiC device capabilities ranging from 750V to 3300V, and its 3300V products have been validated in high-voltage scenarios such as rail transit. According to the company's 2025 semi-annual report, its 6-inch dedicated SiC platform has been built, and the 8-inch platform has completed design and is about to be commissioned. In other words, China does not lack the device capability for the 3300V rating—it is developed along the rail transit line.
Therefore, the first explanation can be ruled out: The inability of Chinese manufacturers to clarify their position on the 800V power supply chain has nothing to do with "being unable to manufacture these voltage ratings."
Voltage rating is not that hurdle.
What needs to be clarified is that what is ruled out is this rough explanation, not the entire gap. Being able to make high-voltage devices for rail transit does not mean having completed the validation of 2300V/3300V SiC devices for data centers, solid-state transformer packaging, and the corresponding mission profiles. Device capability and scenario qualification are two different things.
As for the engineering cost of higher voltage ratings themselves, it mainly falls on materials and device structures: the higher the voltage rating, the thicker the drift layer needs to be and the lower the doping, making the trade-off among epitaxial uniformity, defect control, and on-resistance more stringent. There is one point worth noting here: stacking faults are precisely one of the most critical types of such defects, and the national standard for the "Test Method for Stacking Faults in Silicon Carbide Single Crystal Polished Wafers" was led and formulated by SICC (Tianyue Advanced). At the level of defect characterization, China has already started writing the rules.
The question thus becomes more pointed: Materials are writing the standards, devices are in mass production, and customers have already received the goods—why is there silence when it comes to specifying the stages?
III. It's Not That One Reliability is Harder, But That the Grading Rubrics for the Two Exams are Different
A common saying is that data centers are stricter than automobiles, so passing automotive qualifications is useless. The direction of this saying is correct, but the details are completely wrong.
AEC-Q101 includes items such as high-temperature bias, temperature cycling, and intermittent operating life; its appendix also uses a set of typical 15-year life assumptions to illustrate how mission profiles enter the judgment of reliability suitability. It is not a lenient exam. Infineon's public materials on SiC reliability simultaneously discuss gate oxide FIT rates, threshold voltage shift, short-circuit robustness, bipolar degradation, power cycling, and long-term application testing—these mechanisms are common to both automotive and data center applications, and there are no two completely different failure mechanisms.
The real difference lies in the weights and the passing thresholds.
Automotive applications focus on assessing wide temperature ranges, humidity, vibration, frequent thermal cycling, and extreme transients. A car must cold-start in sub-zero mornings and also endure bumpy roads for over a decade in summer.
AI data centers require a different set of things: long-term bias under high utilization, sustained efficiency, FIT rates, redundant switching, hot-swap capabilities, and the cost of downtime. It doesn't bump or freeze, but it never goes offline.
So the difference is not just about how long it runs, but how it runs. Automotive mission profiles convert a set of typical life assumptions into accelerated tests; data center customers, however, need to redefine a set of mission profiles according to their own utilization rates, load curves, and redundancy strategies.
Let's cite an example that is most easily written incorrectly. Bipolar degradation is often described as "data centers run continuously, so it degrades more easily." This is inaccurate. Its seed is basal plane dislocations, and the real mechanism is the expansion of stacking faults in the body diode under bipolar injection—it is related to whether the body diode is continuously conducting, that is, related to the freewheeling path and the dead-time design of synchronous rectification, rather than directly linked to long full-load times. The same device placed in different topologies can have vastly different degradation risks.
The mission profile is the process that translates usage patterns into assessment items. And the right to define this process currently lies in the hands of customers and platform providers.
IV. Track Records Are Not Invalidated, But Cannot Be Redeemed As-Is
Automotive reliability track records are not invalidated when applied to data centers, but they cannot be directly redeemed as-is.
Another misunderstanding needs to be clarified here: PPM (Parts Per Million) and FIT (Failures In Time) are not empirical values accumulated by "how many cars have been run." They come from both accelerated testing and statistical models, and also require continuous calibration of field failure rates. Chinese manufacturers do not lack data; what they lack is the caliber: that set of data was collected and converted according to a different mission profile.
It is like a driver's license obtained in another country. The skills are not invalidated, and the road is still the same road, but to drive here, one must retake the test according to local rules.
There are three items to retake.
First, whether the customer accepts this mission profile. Who defines the mission profile determines whose data counts.
Second, whether a complete set of validations can be provided. Not just devices, but also drivers, isolation, sensing, protection, and control. Data center power supplies are not procured by individual devices, but by solutions.
Third, and the hardest one: after a problem occurs, who is responsible for the entire power supply system. Who is accountable if the data center goes down for an hour, who closes the loop on failure analysis, and who signs the FIT commitment.
All three questions point to responsibility.
There is a visible connection between these three items and the granularity gap mentioned at the beginning. The premise for a manufacturer to clearly state "my 1200V device is used at the 800V hot-swap stage" is usually that it has actually performed validation at this stage; the premise for co-signing a reference design with a platform provider is usually that it has jointly assumed responsibility for this design. Being able to specify the stage often means having assumed responsibility for that stage.
Conversely, one must be careful when inferring backwards: being unable to specify the stage does not necessarily mean no participation. This point will be expanded in the next section.
What can currently be confirmed as missing is not the partners themselves, but their public, verifiable existence: there has been no joint signing of 800V reference designs by Chinese SiC manufacturers with platform providers or power supply manufacturers.
V. From Substrates to Solutions: The Three Layers of Chinese SiC Each Have Their Own Positions
Looking at the substrate and epitaxy layer, SICC (Tianyue Advanced) is at the forefront. According to the company's 2025 annual report, citing a report released by Fuji Keizai (Japan) in March 2026, in the 2025 global conductive SiC substrate market, SICC's share was 27.6%, of which 8-inch reached 51.3%; the aforementioned national standard for stacking fault testing was also led and formulated by it. However, by the first quarter of 2026, the company's gross profit margin dropped to 19.12%, down nearly 5 percentage points from 23.95% in the same period last year. Operating revenue was 366 million RMB, a year-on-year decrease of 10.41%, and net profit attributable to shareholders turned from a profit of 8.52 million RMB in the same period last year to a loss of 60.51 million RMB. Market share is rising, but prices and profits are falling.
Looking at the device and module layer, the most advanced public evidence comes from Sanan Optoelectronics. It named five customers—Delta, Lite-On, Vertiv, Great Wall, and Flex—and claimed mass production, making it the only one among the seven to reach the step of "naming customers." According to the company's first-quarter 2026 report, total operating revenue was 2.907 billion RMB, a year-on-year decrease of 32.59%; net profit attributable to shareholders was 67.49 million RMB, a year-on-year decrease of 68.15%; and the gross profit margin was 18.48%—the weight of SiC in the company's overall portfolio is not yet enough to change the direction of the financial statements.
The most intriguing is CRRC Times Electric. Its high-voltage capability rating falls exactly in the range required by medium-voltage solid-state transformers, but the publicly verifiable 3300V application validation is in rail transit, not data centers. According to the company's first-quarter 2026 report, operating revenue was 5.102 billion RMB, a year-on-year increase of 12.45%, and net profit attributable to shareholders was 643 million RMB, a year-on-year increase of 1.91%. The capability is there, but the public correspondence for data centers is not yet established.
Yangjie Technology's ratings cover 650V, 1200V, and 1700V, which are precisely the three most densely used ratings on the 800V chain. According to the company's first-quarter 2026 report, operating revenue was 2.13 billion RMB, a year-on-year increase of 34.87%, and net profit attributable to shareholders was 385 million RMB, a year-on-year increase of 41.01%, making it the fastest-growing among the seven. However, its public statement on AI server power supplies still stops at the opening phrase "can be widely applied in."
For Silan Microelectronics, StarPower Semiconductor, and NCE Power, their public disclosures in the direction of AI data centers all remain at the level of listing application fields; see Table 1 for data.
Looking at the system solution layer, this layer is empty. In public information, none of the seven Chinese companies has disclosed an 800V reference design jointly released with power supply manufacturers or platform providers, nor has any company specified which voltage rating of its devices corresponds to which stage of the chain.
A Ruler That Can Be Verified Year by Year: Not Asking Whether They Have Entered, But Asking to Which Level They Have Publicly Disclosed
Dividing the degree of disclosure into four levels is much more useful than the binary judgment of "whether they have entered or not."
L1, has a general statement of "used for AI server power supplies."
L2, names customers and states mass production.
L3, specifies which voltage rating corresponds to which circuit stage.
L4, reference designs are public, and jointly validated and signed with platforms.
Currently, those standing at L3 and L4 are STMicroelectronics, Infineon, ROHM, and Navitas. The highest level on the Chinese side is L2.
Figure 3 | The four-level staircase of public disclosure: it measures the granularity of disclosure, not technical capability. Sources: Periodic reports of each company, investor relations records, and interactive platform replies; public technical documents from overseas manufacturers.
Seven Chinese SiC Companies: Stages, Disclosure Levels, and Latest Financials
Table 1 | Financial data is taken from the first-quarter 2026 reports of each company. "—" indicates that no relevant disclosures in the direction of AI data centers were found in public information; SICC (Tianyue Advanced) is in the substrate and epitaxy layer and is not included in the L1–L4 classification.
The Boundary of This Ruler: It Measures Disclosure, Not Capability
This staircase measures the level of public disclosure, not the level of technical capability, and certainly does not directly equate to industrial roles.
Through its joint venture with STMicroelectronics, Sanan's degree of participation in the 200mm SiC manufacturing stage obviously exceeds what the L2 position can express. This itself illustrates that a low level does not mean shallow participation.
There are at least three other explanations for low disclosure levels that are unrelated to industrial position. First, customer non-disclosure agreements (NDAs)—power supply giants like Delta and Vertiv may well only allow suppliers to disclose "supplied" without disclosing the specific stages; moreover, confidentiality constraints are signed by customer and by project, so they may indeed only restrict one type of business and not another. Second, disclosure habits—the scale of technical details differs between A-share listed companies and those in Europe and the US. Third, business weight—which line a company expands on in its documents depends on which line is the narrative focus for the current period.
Therefore, the conclusion supported by this gap must be narrowed: based on public evidence, we can only identify these companies as device suppliers, and cannot yet prove that they participated in solution definition. The reverse is equally true—it cannot be proven that they did not participate.
Then what else is this ruler useful for?
Its usefulness lies in highlighting a blind spot: for industry observers, public information currently cannot distinguish between "entering the general material pool" and "participating in solution definition." And these two things are completely different in terms of industrial value. The former fluctuates with the procurement decisions of power supply manufacturers, while the latter, once embedded in a reference design, is very difficult to replace.
As for what lies behind this blind spot, the three items mentioned earlier—the right to define mission profiles, complete set of validations, and attribution of responsibility—represent the most explanatory inference currently available. But it is an inference, not a proven fact. To prove it, what is needed is precisely the kind of disclosure that is currently missing.
VI. When Will the Door Open: Three Signals That Can Be Verified One by One
Let's first state the premise for this judgment to hold, and the variables that might invalidate it.
The premise is that the 800 VDC architecture advances according to the current roadmap, and solid-state transformers indeed follow the SiC route. There are at least four variables: (1) domestic intelligent computing centers may build their own 800V ecosystem, bypassing the list system of overseas platforms; (2) the implementation pace of 800V may be delayed overall; (3) some stages may continue to use IGBT solutions as a transition; (4) the advancement of GaN in the intermediate conversion layer will continue to squeeze the boundaries of SiC—Innoscience's entry into that list of 14 is precisely following this path.
If the above judgment holds, what to watch next is not market share or capacity, but three signals of climbing up the disclosure levels.
First, whether 2300V and 3300V products move from sampling to mass production, and are explicitly targeted at data centers, not just flexible DC and rail transit.
Second, whether they jointly release verifiable reference designs with power supply manufacturers or solid-state transformer manufacturers; this is the threshold from L3 to L4.
Third, whether they disclose the correspondence between voltage ratings and circuit stages, as well as long-term reliability data under the corresponding mission profiles; this is the threshold from L2 to L3.
Among the three, the third is the most easily overlooked, and may also be the first to appear in public information. It does not necessarily mean building new production lines, but at least it means that Chinese manufacturers have truly participated in stage-level validation and have obtained permission for disclosure.
Entering the data center does not mean being able to clearly state which stage one is at. And on this chain, those who can clearly state their position are usually also those who can sign on the dotted line.
Disclaimer: All data and statements in this article are derived from public information and have been cited item by item in the reference sources. The classification of disclosure levels in the text is an observation of the granularity of public information and does not constitute a judgment on the technical capabilities, industrial positions, or business relationships of the relevant companies; "—" in the table indicates that no relevant disclosures were found in public information, and no speculative filling has been made. This article does not constitute any investment advice.
Reference Sources: (Slide up and down to view):
[1] Yangjie Technology (300373.SZ) | Investor Relations Activity Record Form 2026-006 (July 1, 2026)
[2] Sanan Optoelectronics (600703.SH) | Reply on Investor Interactive Platform (April 23, 2026)
[3] Sanan Optoelectronics (600703.SH) | 2025 Semi-Annual Report
[4] Silan Microelectronics | 8-inch SiC Production Line Commissioned, 12-inch High-End Analog Chip Production Line Starts Construction Simultaneously (2026-01-05)
[5] STMicroelectronics | STDES-30KWVRECT Three-Level Rectifier Reference Design (Original Application: EV Charging)
[6] Infineon | Reliability of SiC power semiconductors White Paper
[7] Infineon | EasyPACK power modules for solid-state transformers Product Page
[8] Infineon | Infineon and SolarEdge collaborate to advance high-efficiency power infrastructure for AI data centers (2025-11)
[9] AEC | AEC-Q101 Official Document Catalog (Rev E)
[10] CRRC Times Electric | SiC Product Page
[11] CRRC Times Electric (688187.SH) | 2025 Semi-Annual Report
[12] SICC (688234.SH) | 2026 First Quarter Report
[13] SICC (688234.SH) | 2025 Annual Report (citing Fuji Keizai Japan March 2026 report)
[14] CRRC Times Electric (688187.SH) | 2026 First Quarter Report
[15] StarPower Semiconductor (603290.SH) | 2026 First Quarter Report
[16] NCE Power (605111.SH) | 2026 First Quarter Report
[17] Sanan Optoelectronics (600703.SH) | 2026 First Quarter Report
[18] Yangjie Technology (300373.SZ) | 2026 First Quarter Report (Announcement No. 2026-017)
[19] Silan Microelectronics (600460.SH) | 2026 First Quarter Report
[20] STMicroelectronics | Cloud & AI Supply System Description
[21] STMicroelectronics | 800V Power Supply Solution
[22] ROHM | SiC MOSFET Adopted in BBU for AI Servers (2026-06-03)
[23] Navitas | Navitas, EPFL to Demonstrate Novel Solid-State Transformer Solution for AI Data Center (2026-03-04)
[24] Wolfspeed | Powering AI with Reliable Silicon Carbide-Based Solid-State Transformers
[25] NVIDIA | NVIDIA and Partners Drive Next-Gen Efficient Gigawatt AI Factories (2025-10-13)
[26] National Standard Information Public Service Platform | Test Method for Stacking Faults in Silicon Carbide Single Crystal Polished Wafers (GB/T 47082-2026)
#DeepChip Alliance #Bay Chip Expo #Semiconductor Industry Research #SiC #AI