EPFL’s 3.4 kV GaN Transistor Corrects a 4 kV Technology News Headline

EPFL researchers built gallium nitride devices that blocked more than 3,000 volts, but the transistor did not quite reach the widely reported 4,000-volt mark. This technology news matters because it exposes both a significant engineering advance and an important distinction. The 3.9 kV result belongs to a diode. The team’s two transistors sustained 3.5 kV and 3.4 kV.

Those figures still place the experimental devices far beyond most commercial GaN transistors, which commonly carry ratings around 650 volts. More importantly, the researchers reached those voltages on a silicon substrate while preserving relatively low resistance. They did so without field plates across the high-voltage region or intentional doping for charge balance.

The work puts pressure on the boundary separating gallium nitride from silicon carbide. GaN dominates compact, high-frequency power conversion at lower voltages. Silicon carbide remains the established choice for many higher-voltage systems. EPFL’s result suggests that GaN’s territory can expand, but it does not establish a production-ready replacement.

What This Technology News Actually Reports

The headline number combines three related devices, and separating them is essential to understanding the result.

Researchers at the Swiss Federal Institute of Technology Lausanne, commonly known as EPFL, reported a new power-device structure called an intrinsic polarization superjunction. The research appeared online in Nature Electronics on August 28, 2026. EPFL published its public explanation on August 31.

The peer-reviewed device study describes Schottky barrier diodes and two transistor configurations. A Schottky barrier diode controls current in one direction through a metal-semiconductor junction. A transistor provides active switching through a gate.

The highest reported breakdown voltage exceeded 3.9 kV in a diode with a 25-micrometer superjunction region. Breakdown voltage marks the point where an off-state device can no longer block rising voltage as intended.

The transistor results were slightly lower. A depletion-mode device sustained 3.5 kV, while an enhancement-mode device sustained 3.4 kV. Depletion-mode transistors normally conduct without a gate signal. Enhancement-mode transistors remain off by default, which generally makes them easier to use safely in power systems.

That difference explains why “nearly 4,000 volts” is directionally accurate for the research platform but imprecise for the transistor itself. The transistor reached at least 3,400 volts, not 4,000 volts. The distinction is not semantic because diodes and transistors perform different jobs inside a power converter.

The measurements also represent experimental breakdown limits, not commercial voltage ratings. Manufacturers normally rate components well below destructive breakdown to provide operating margin. A laboratory device that breaks down at 3.4 kV would not automatically become a 3.4 kV commercial switch.

The researchers fabricated the devices on a six-inch silicon wafer carrying a GaN, aluminum gallium nitride, and GaN layer stack. They then divided the wafer into two-centimeter-square chips for fabrication. This substrate choice matters because silicon wafers offer an established manufacturing base and lower expected costs than freestanding GaN substrates.

The diode combined its greater than 3.9 kV breakdown result with a specific on-resistance of 4.7 milliohm-square-centimeters. Specific on-resistance normalizes resistance by device area, allowing engineers to compare devices of different sizes.

The depletion-mode transistor recorded a specific on-resistance of 6.3 milliohm-square-centimeters. It also showed an on-to-off current ratio greater than five orders of magnitude. These measurements establish a functioning switch, not only an isolated high-voltage material sample.

The enhancement-mode transistor is especially relevant to practical systems because it defaults to the off state. However, the paper reports a 3.4 kV laboratory result rather than a qualified component, packaged module, or converter.

That boundary frames the entire story. EPFL has shown a promising device mechanism under controlled conditions. The work has not yet established manufacturing yield, lifetime, current capacity, packaging performance, or system-level economics.

Why Charge Balance Changes the GaN Voltage Limit

EPFL’s main advance is a new way to spread electric stress across a GaN device instead of allowing voltage to concentrate near one edge.

A power transistor behaves as a controlled electrical switch. When the transistor is off, it must block voltage without leaking excessive current or damaging itself. Higher voltage becomes difficult when the internal electric field gathers at a small location.

Conventional lateral GaN transistors conduct through a two-dimensional electron gas. This term describes a very thin, highly mobile sheet of electrons formed at the interface between semiconductor layers.

When a conventional device switches off, those mobile electrons leave the channel. Fixed positive charges can remain behind. The imbalance creates a steep electric-field gradient, often near the gate or another contact edge.

That concentrated field can trigger premature breakdown before the wider device region reaches GaN’s material limit. Engineers commonly add field plates, which are conductive extensions that reshape the electric field through a dielectric layer.

Field plates work, but their behavior depends on tightly controlled dielectric thickness and geometry. They can also add capacitance, complicate switching, and leave designers balancing voltage capability against dynamic resistance.

The new intrinsic polarization superjunction, abbreviated iPSJ, takes another route. It uses GaN’s spontaneous and piezoelectric polarization to create two mobile charge sheets on opposite sides of an aluminum gallium nitride barrier.

One sheet contains electrons, while the other contains holes, which behave as positive charge carriers. The team adjusted the undoped GaN cap thickness until the two carrier concentrations closely matched.

In the tested diode structure, the researchers measured an electron density of 1.03 times 10^13 per square centimeter. The measured hole density was 1.05 times 10^13 per square centimeter. That represents less than 2 percent charge mismatch.

When the device switches off, both mobile populations can deplete. The drift region therefore remains close to electrically neutral. The voltage spreads more evenly, reducing the local field peak that limits ordinary lateral GaN devices.

This mechanism resembles the superjunction concept used in silicon power components. Traditional superjunctions place alternating doped regions beside one another so their charges balance during blocking.

Applying that method to GaN has been difficult. Magnesium, which commonly creates p-type GaN, has poor activation efficiency. Selective implantation and regrowth also lack the maturity needed for precise, repeatable charge matching.

EPFL’s structure avoids intentional doping throughout the drift region. The device still uses a small magnesium-doped contact layer to reach the hole gas. However, that doped layer does not provide the superjunction’s balancing charge.

The distinction is important. Calling the entire device “doping-free” would overstate the paper. The charge-balancing region avoids intentional doping, while a doped contact remains part of the fabricated structure.

The team also fabricated comparison devices. A thin-cap version lacked the hole gas and therefore retained complete charge mismatch. Another version kept a p-doped cap across the drift region.

The p-doped comparison suffered catastrophic breakdown at 342 volts and produced seven times more leakage current. The finding supports the researchers’ mechanism, although it does not cover every competing GaN design.

Temperature measurements provided another useful test. The matched diodes maintained breakdown voltages above 3.3 kV from 25 to 125 degrees Celsius. Their limit came from the underlying GaN-on-silicon buffer rather than the lateral superjunction.

The researchers also performed 17 consecutive sweeps to 3.3 kV on the same diode. The leakage-current measurements showed less than 20 percent standard deviation. That repetition is encouraging, but it remains far removed from an industrial qualification campaign.

The Real Opponent Is Silicon Carbide’s High-Voltage Lead

The iPSJ result challenges the voltage boundary between GaN and silicon carbide, not the continued use of silicon in every power circuit.

Commercial GaN products already serve adapters, telecommunications equipment, photovoltaic systems, data-center power supplies, and selected vehicle applications. Their appeal comes from fast switching, small passive components, and high power density.

However, most widely available devices remain concentrated around the 650 V class. One current 650 V transistor, for example, carries a maximum drain-source rating of 650 volts and targets industrial power conversion.

That commercial rating should not be compared directly with EPFL’s destructive breakdown measurement. Product ratings incorporate margins, qualification results, packaging limits, and the manufacturer’s reliability model.

Even with that caution, a 3.4 kV experimental transistor represents a large movement in blocking capability. It suggests that lateral GaN-on-silicon devices can operate in territory previously associated with thicker vertical devices and other materials.

Silicon carbide currently holds a strong position in higher-voltage applications. Its commercial ecosystem includes 1,200 V and higher-rated switches, modules, gate drivers, packaging expertise, and automotive qualification experience.

The two materials also offer different operating strengths. GaN’s high electron mobility supports fast switching. Silicon carbide combines high-voltage capability with established high-current products and strong thermal performance.

The U.S. Department of Energy treats GaN and silicon carbide as the two leading wide-bandgap materials for electric-drive power electronics. Its vehicle research connects both materials with smaller converters, higher-temperature operation, and reduced cooling requirements.

EPFL’s result attacks one part of silicon carbide’s advantage: blocking voltage. It does not resolve every factor that makes silicon carbide attractive in traction inverters, industrial drives, charging equipment, or grid hardware.

Current handling remains one unanswered issue. A device can block thousands of volts yet provide too little current for a practical converter. The published figures emphasize breakdown voltage and resistance, but they do not describe a production-scale module delivering industrial power.

Packaging is another dividing line. Kilovolt switching creates demanding electric-field, insulation, thermal, and electromagnetic conditions outside the semiconductor die. A successful chip must survive those conditions after it enters a package and joins a circuit.

Short-circuit tolerance also matters in motor drives and grid equipment. Engineers need to know how the device behaves during overloads, faults, repetitive switching, and transient voltage events.

System designers therefore will not replace silicon carbide after reading one technology news report. They will compare efficiency across realistic operating points, including switching frequency, current, temperature, and gate-drive conditions.

Still, the result changes the research question. The issue is no longer whether lateral GaN can cross 3 kV on silicon. EPFL’s measurements indicate that it can.

The next question is whether the intrinsic polarization superjunction can preserve that advantage in large, normally-off devices. Those devices must then survive packaging, qualification, and years of field operation.

Low Resistance Matters as Much as 3.4 kV

A high breakdown figure becomes valuable only when the device can also conduct efficiently and switch repeatedly under realistic electrical stress.

Power-device design includes a persistent tension between blocking voltage and conduction loss. Increasing the distance across which a transistor holds voltage often raises resistance. Higher resistance converts more electrical energy into heat.

The EPFL team’s charge-balancing approach aims to weaken that tradeoff. A nearly uniform electric field lets more of the drift region contribute to voltage blocking. Designers therefore need less conservative spacing to avoid a damaging local peak.

The diode’s 4.7 milliohm-square-centimeter specific on-resistance is notable beside its greater than 3.9 kV breakdown measurement. The transistor’s reported 6.3 milliohm-square-centimeter value also supports the central mechanism.

However, those normalized values do not reveal a finished device’s total resistance. Contacts, access regions, interconnects, package connections, and thermal conditions all contribute to system performance.

Dynamic resistance is another central issue for GaN. Trapped electrical charge can make a transistor’s resistance rise after it blocks high voltage. A device may look efficient in a static measurement but lose that advantage during switching.

The researchers report less than 15 percent dynamic on-resistance degradation in diode testing up to 3 kV. They describe the behavior as near ideal for the tested structures.

The transistor measurement had a lower ceiling. The team measured negligible dynamic resistance degradation up to 650 volts using an on-wafer pulsed system. Parasitic turn-on from cables and probes prevented that setup from reaching the transistor’s full blocking range.

This limitation deserves attention. The transistors blocked more than 3.4 kV in static tests, but their reported dynamic-resistance evaluation stopped at 650 V on the first setup.

The paper also describes double-pulse testing for high-voltage dynamic operation. Such tests repeatedly switch current and voltage to approximate converter behavior more closely than static measurements.

Even so, laboratory pulses cannot replace lifetime testing. Commercial devices face billions of switching cycles, temperature changes, humidity, mechanical stress, and irregular grid or load conditions.

The use of silicon as a substrate adds both promise and risk. Silicon offers mature wafer infrastructure and potential cost advantages. Yet GaN-on-silicon stacks require buffer layers that accommodate lattice and thermal-expansion differences.

In the matched diode tests, the buffer became the breakdown limit. Improving the superjunction alone will not raise device voltage indefinitely if the vertical material stack fails first.

That observation redirects future work toward epitaxial quality, which describes how semiconductor crystal layers are grown. It also increases the importance of wafer uniformity and defect control.

Manufacturing tolerance presents another test. The iPSJ depends on closely matched electron and hole densities. The researchers tune that balance through layer thickness and polarization instead of conventional drift-region doping.

This approach may improve temperature stability, but factories must still reproduce the required layer dimensions and interfaces across an entire wafer. A small research chip does not reveal full-wafer yield.

A useful commercial process must tolerate normal variation without producing unstable leakage or sharply reduced breakdown. Device makers will also need nondestructive methods to identify weak structures before packaging.

The research paper offers credible evidence for a mechanism that handles voltage and resistance together. It does not yet provide the manufacturing statistics needed to price that mechanism.

Data Centers, EVs, and Solar Need Different Proof

The same high-voltage GaN transistor will not move directly from a probe station into every application named in the announcement.

EPFL identifies AI data centers, electric vehicles, and renewable-energy systems as possible beneficiaries. Each market needs efficient power conversion, but their voltage levels, duty cycles, safety rules, and qualification standards differ.

AI data centers contain several conversion stages. Electricity passes through grid interfaces, backup systems, rack distribution, and low-voltage regulators before reaching processors.

Faster GaN switching can reduce the size of magnetic components in some stages. Higher blocking voltage could also reduce the need to place multiple lower-voltage switches in series.

Series-connected switches add gate-control complexity and require careful voltage balancing. A single higher-voltage device could simplify parts of the architecture if it meets efficiency and reliability requirements.

Yet a 3.4 kV breakdown measurement does not establish suitability for an 800 V data-center bus. Designers need rated voltage with transient margin, appropriate current capacity, low switching energy, and proven fault behavior.

Electric vehicles create another set of demands. Their power electronics include traction inverters, onboard chargers, and DC-to-DC converters. The conversion functions differ, so they do not require identical switches.

A traction inverter must handle high current, repeated acceleration, regenerative braking, vibration, and extreme temperature cycles. Automotive qualification can take years and requires statistically significant device populations.

An onboard charger may value high switching frequency and compact size more heavily. GaN already participates in charging designs, although higher-voltage vehicle platforms continue to favor silicon carbide for major power stages.

Solar and grid equipment often prioritizes long service life. Inverters must tolerate daily thermal cycling, grid disturbances, moisture, and outdoor conditions over extended deployments.

All three markets care about efficiency, but a small improvement has different value in each system. A data center may monetize reduced cooling and rack space. A vehicle manufacturer may prioritize range, weight, and warranty risk.

Solar operators may focus on lifetime energy yield and maintenance costs. A new transistor must therefore prove more than an impressive voltage number.

The device’s lateral architecture creates another application question. Lateral devices place current flow mainly along the wafer surface, while vertical devices carry current through the semiconductor thickness.

Lateral GaN can support integration with drivers and additional components on one wafer. That possibility could reduce parasitic electrical paths and create compact power integrated circuits.

However, lateral device area generally increases as designers expand the high-voltage drift region. That scaling can affect die size and cost, especially when applications demand both high voltage and high current.

EPFL’s research overview describes a future combination of iPSJ voltage blocking with multiple conduction channels. The additional channels would distribute current and reduce resistance.

That proposed combination has not yet produced a commercial component. It represents the laboratory’s next technical direction, according to POWERlab head Elison Matioli.

Application claims should therefore remain conditional. The work identifies a plausible route toward smaller high-voltage converters. It has not demonstrated an AI data-center supply, vehicle inverter, or solar inverter using the new transistor.

What the 4 kV Claim Does Not Show

The strongest skeptical reading accepts the measurements while rejecting the assumption that breakdown voltage equals commercial readiness.

The most immediate issue is the headline itself. The diode exceeded 3.9 kV, while the transistor results reached 3.5 kV and 3.4 kV. Reporting every device as a nearly 4 kV transistor blurs separate results.

EPFL’s public summary uses “nearly 4 kilovolts” for the chip-sized device class. The underlying paper provides the precise values needed for a more careful interpretation.

The second issue concerns operating mode. The 3.5 kV transistor was depletion mode, meaning it naturally remains on without the correct gate bias. Many safety-focused power applications prefer normally-off switches.

The normally-off enhancement-mode device reached 3.4 kV. That remains the more relevant transistor benchmark for many engineers, although its gate structure introduces its own reliability questions.

Third, breakdown voltage alone is not a commercial rating. Engineers need derating margins because actual systems encounter overshoot, ringing, manufacturing variation, and temperature-dependent stress.

Fourth, the devices were tested on diced research chips. The paper does not establish high-volume yield across full six-inch wafers, much less production lots from multiple fabrication runs.

Fifth, long-term reliability remains unknown. The 17 repeated diode sweeps demonstrate short-term repeatability, not years of continuous field service.

The paper’s temperature testing reached 125 degrees Celsius for selected diode measurements. That is useful evidence, but qualification requires longer stress periods and broader operating conditions.

Sixth, the transistor’s full-voltage dynamic resistance needs further validation. Static blocking above 3.4 kV and dynamic testing at lower voltage do not automatically guarantee efficient high-voltage switching.

Seventh, the buffer already constrained the matched diode. Raising lateral breakdown will have limited value unless the vertical GaN-on-silicon stack can withstand comparable stress reliably.

The broader GaN literature identifies several breakdown mechanisms, including field concentration, buffer leakage, and gate leakage. A breakdown review also shows why one successful field-management technique cannot eliminate every failure route.

Finally, silicon carbide will continue improving. GaN does not compete against a stationary benchmark. Silicon carbide suppliers are expanding wafers, devices, modules, and system designs while GaN researchers pursue higher voltage.

These limitations do not invalidate the iPSJ work. They identify the evidence needed to turn a strong physics result into a credible product platform.

The right interpretation sits between two extremes. This is not merely another simulated device concept, because the team fabricated and measured working diodes and transistors.

It is also not a ready replacement for qualified silicon carbide modules. The experiment changes what GaN appears capable of achieving, while leaving commercialization unresolved.

Three Signals to Watch After the GaN Technology News

The next phase depends on dynamic high-voltage validation, scalable normally-off devices, and system-level demonstrations.

The first signal is full-voltage switching data from the enhancement-mode transistor. Researchers need to show dynamic on-resistance, switching loss, leakage, and stability near the intended operating range.

Repeated double-pulse tests should include realistic current, temperature, and switching speed. Strong results would reinforce the claim that iPSJ solves both voltage blocking and dynamic degradation.

Weak dynamic performance would narrow the achievement to static breakdown. That outcome would reduce its value for converters, where devices must transition between blocking and conducting states continuously.

The second signal is wafer-scale manufacturing evidence. Future publications or industry partnerships should report yield distributions, layer-thickness tolerances, and device variation across complete wafers.

A few excellent dies can demonstrate physics. A commercial process must produce large populations with predictable thresholds, resistance, leakage, and breakdown behavior.

Normally-off yield will matter most. Enhancement-mode operation provides the safer default state expected in many products, but its recessed gate and oxide introduce additional process controls.

Evidence across several fabrication lots would strengthen the case for silicon-based cost advantages. Poor uniformity would weaken that argument even if individual devices keep setting voltage records.

The third signal is an independently evaluated converter or power module. A useful demonstration would place packaged iPSJ transistors inside a data-center supply, vehicle converter, or renewable-energy inverter.

That system should report efficiency across the load range, thermal behavior, electromagnetic interference, fault response, and performance over extended cycling. Comparisons should use appropriately rated silicon carbide and conventional GaN devices.

A successful converter would show that the voltage result survives contact with packaging and circuit constraints. It would also reveal whether faster switching produces meaningful system benefits at multi-kilovolt stress.

Without that evidence, the iPSJ remains a research platform with credible potential. With it, GaN could begin contesting applications that currently default to silicon carbide.

This technology news deserves attention because it changes the demonstrated limit of lateral GaN-on-silicon devices. It also deserves precise language. The diode passed 3.9 kV, while the normally-off transistor sustained 3.4 kV.

Will the next result deliver full-voltage switching, repeatable wafer yield, or a packaged converter first? Watch those three milestones, not another isolated breakdown record. They will show whether intrinsic polarization becomes a manufacturing platform or remains an elegant laboratory solution.