Power Integrations’ 2,200 V GaN Claim Needs Context

Power Integrations appeared in Google News with a striking claim: 2,200 V gallium nitride technology aimed at AI data centers. That number would place GaN much deeper inside territory long associated with silicon carbide devices.

However, the headline needs technical context. Power Integrations currently presents its commercial PowiGaN platform around 1,250 V and 1,700 V device ratings. Its published reliability work uses stress conditions reaching 2,200 V, but that does not automatically create a 2,200 V product rating.

The distinction matters because AI infrastructure companies are redesigning how electricity moves through increasingly dense racks. NVIDIA’s proposed 800 VDC architecture creates an opening for higher-voltage GaN, while placing it against established silicon carbide options.

What the 2,200 V GaN Headline Actually Represents

The available evidence supports a 2,200 V reliability test condition, not a clearly documented 2,200 V commercial device rating.

Power Integrations develops PowiGaN, its proprietary gallium nitride power semiconductor technology. GaN is a wide-bandgap material that can switch electricity quickly while producing relatively low switching losses.

The company’s public PowiGaN roadmap describes three major voltage advances. Power Integrations moved from 750 V to 900 V, then introduced 1,250 V and 1,700 V technology.

Those ratings identify the voltage classes in which the company positions its devices. They are not the same as every voltage applied during laboratory qualification.

Power Integrations’ reliability material for its 1,250 V technology describes accelerated testing between 2,100 V and 2,200 V. Engineers use elevated electrical and thermal stress to estimate how devices behave over longer operating periods.

The company says its model predicts a cumulative failure rate of one part per million after more than 15,000 years at 1,000 V and 100 degrees Celsius. That prediction comes from extrapolating accelerated tests, rather than operating one device for thousands of years.

This is an important reliability claim. It suggests the 1,250 V switch has substantial voltage margin under the company’s test model. It still does not make 2,200 V the device’s normal operating rating.

A rated voltage defines the operating boundary that a supplier supports under specified conditions. A breakdown measurement identifies the point where a device stops blocking voltage as intended. An accelerated stress voltage is selected to produce failures sooner for statistical modeling.

Those three numbers can differ substantially. Removing their labels makes a technically meaningful test look like a different product announcement.

Power Integrations has previously said its lower-voltage GaN families also break down well above their published ratings. That practice gives designers safety margin against manufacturing variation, switching transients, and long-term degradation.

The underlying engineering result remains notable. High-voltage lateral GaN has traditionally faced material and manufacturing constraints, especially when built on silicon substrates.

Power Integrations uses GaN-on-sapphire construction. Sapphire provides electrical isolation that helps the company extend lateral GaN devices toward higher blocking voltages without the same conductive-substrate limitations found in common GaN-on-silicon processes.

Its architecture also uses a cascode arrangement for the 1,250 V switch. A cascode combines a high-voltage GaN transistor with a low-voltage silicon device to produce normally off behavior and a familiar control interface.

That approach lets designers gain GaN’s switching characteristics without treating the device like an exposed, normally on transistor. Integration can also shorten critical electrical paths and coordinate protection with the switch.

The Google News phrasing therefore compresses several layers into one number. Power Integrations has high-voltage GaN, its 1,250 V technology was stressed at 2,200 V, and the company targets future AI power systems.

Each statement is relevant. They should not be treated as interchangeable.

Why AI Data Centers Are Moving Toward 800 VDC

The voltage race is happening because rack power is rising faster than conventional low-voltage distribution can comfortably support.

A data center does not deliver utility electricity directly to a GPU. Power passes through switchgear, rectifiers, distribution buses, power shelves, board converters, voltage regulators, and protective systems.

Every conversion stage occupies space, creates heat, and loses some energy. The penalty becomes larger when thousands of accelerators operate continuously.

Traditional server racks commonly distribute power through relatively low-voltage buses. That design worked when rack demand was measured in tens of kilowatts and current stayed within manageable limits.

AI racks are moving past 100 kW, while planned systems approach much higher levels. Power Integrations describes megawatt-scale racks as a design target for emerging infrastructure.

Electrical power equals voltage multiplied by current. Raising distribution voltage allows a system to transmit the same power with less current.

Lower current reduces resistive losses, which rise with the square of current. It can also shrink the copper conductors needed to carry electricity through the rack.

That is the appeal of NVIDIA’s 800 VDC proposal. It shifts more power distribution to a high-voltage direct-current bus before converting electricity closer to the computing load.

Power Integrations has positioned 1,250 V and 1,700 V PowiGaN devices for this transition. Its 800 VDC paper argues that these voltage classes can support several conversion stages inside future racks.

A switch cannot be rated at exactly the bus voltage. Normal operation includes overshoot, ringing, load changes, and fault conditions that can temporarily raise device stress.

An 800 V bus therefore needs switches with meaningful voltage headroom. The appropriate margin depends on topology, control behavior, insulation requirements, and expected transient conditions.

Power Integrations argues that a single 1,250 V PowiGaN switch can replace arrangements using stacked 650 V GaN devices. Stacking lets multiple transistors share voltage, but it complicates timing, balancing, protection, and control.

A single higher-voltage switch simplifies that part of the circuit. It can reduce component count and remove the risk that voltage divides unevenly between stacked devices.

The company also compares its high-voltage GaN with 1,200 V silicon carbide. This is the more consequential contest because silicon carbide already serves high-voltage industrial, automotive, and infrastructure applications.

Power Integrations says its GaN technology offers faster switching, lower switching losses, and greater integration. These characteristics can help reduce magnetics, cooling hardware, and enclosure volume.

Yet those benefits depend on the complete converter. A fast transistor does not guarantee that an assembled power supply will achieve better efficiency, density, cost, or reliability.

Magnetic components, capacitors, control algorithms, thermal interfaces, circuit layout, electromagnetic interference, and packaging all affect the final result. A successful device must work within that larger system.

For data center operators, even a small conversion improvement can matter when multiplied across a large facility. However, operators also demand predictable maintenance, proven failure behavior, and a secure supply chain.

This combination explains why Power Integrations is emphasizing reliability data alongside switching performance. AI racks need compact power conversion, but they cannot trade away service continuity to obtain it.

Google News Focuses on Voltage, but Integration Is the Bigger Bet

Power Integrations is not simply trying to sell a higher-voltage transistor; it is betting that integrated GaN can simplify complete power stages.

Discrete power transistors give engineers flexibility. Designers can choose a separate gate driver, controller, current sensor, protection circuit, and switching device for each application.

That flexibility also creates engineering work. Fast-switching GaN devices are sensitive to parasitic inductance, gate-loop layout, timing, and protection response.

Power Integrations builds the power switch into controller families such as InnoSwitch and InnoMux. The company can coordinate switching behavior, sensing, isolation, protection, and control around its own transistor.

This integration is central to its competitive argument. It can shorten development cycles and reduce the number of interactions that a power-supply designer must validate independently.

For AI infrastructure, the company has discussed both main power conversion and auxiliary supplies. Main power moves large amounts of energy toward processors, while auxiliary supplies run management, control, cooling, and standby functions.

Power Integrations announced ultra-slim auxiliary power-supply reference designs for NVIDIA’s Kyber 800 VDC architecture in June 2026. Those designs use 1,700 V PowiGaN technology and target the limited space available inside dense racks.

Reference designs matter because they turn device specifications into testable circuit arrangements. Customers can examine efficiency, temperature, protection behavior, board area, and electromagnetic performance before developing production hardware.

They are not equivalent to mass adoption. A reference design proves that a circuit can be assembled and evaluated, while a production deployment must satisfy a customer’s qualification and sourcing requirements.

Power Integrations identifies itself as an NVIDIA ecosystem partner for the 800 VDC transition. That connection gives its technology a defined target architecture rather than a hypothetical data center application.

NVIDIA is not the only company involved. The broader supplier roster includes firms working on rectification, solid-state protection, intermediate conversion, point-of-load power, connectors, and cooling.

A direct path from 800 V to low processor voltages creates an extreme conversion ratio. Power systems must transform hundreds of volts into single-digit voltages while responding rapidly to changes in accelerator demand.

Some architectures retain an intermediate 48 V stage. Others seek to remove stages through direct conversion or modular transformers.

Each route creates different opportunities for GaN and silicon carbide. High-voltage devices can serve the rack input and primary conversion stages, while lower-voltage GaN can operate closer to processors.

Power Integrations’ strongest advantage is not necessarily the largest number in a headline. Its bet rests on controlling the switch, driver behavior, protection, and controller as one product architecture.

That strategy can be valuable when data center power engineers face aggressive schedules. A more integrated design reduces some component-selection and layout decisions.

Integration also creates dependence on a supplier’s architecture. Customers give up part of the flexibility available with discrete components, and redesigns become harder if an integrated part becomes constrained.

This tension will influence adoption. Hyperscale buyers often use multiple suppliers to reduce operational and geopolitical risk.

Power Integrations must therefore show that integration creates enough system value to justify deeper architectural commitment. A claimed voltage margin alone will not settle that decision.

GaN Is Entering Silicon Carbide Territory

The primary competitive pressure falls on silicon carbide because higher-voltage GaN challenges one of its clearest market boundaries.

GaN gained broad commercial visibility through compact chargers and adapters. Many of those products use 600 V or 650 V devices that switch faster than traditional silicon alternatives.

Silicon carbide became the usual wide-bandgap choice at higher voltages. It is widely used in electric-vehicle traction systems, renewable-energy equipment, industrial drives, and high-power infrastructure.

The boundary was partly technical. Common lateral GaN devices struggled to reach higher voltage classes while maintaining yield, cost, reliability, and acceptable on-resistance.

Power Integrations has methodically moved beyond that boundary. It introduced 900 V and 1,250 V products in 2023, followed by a 1,700 V GaN switch in 2024.

The company’s GaN technology history shows how that progression moved GaN beyond compact consumer power supplies. The current data center campaign extends the argument into infrastructure.

Silicon carbide retains meaningful strengths. It has established high-voltage product families, multiple device architectures, mature modules, and experience in harsh power applications.

Its thermal conductivity is also higher than GaN’s. That property helps move heat away from active regions, although packaging and cooling design determine the practical system result.

GaN switches faster and can reduce switching energy. Higher frequency can shrink transformers, inductors, and filters, increasing power density.

Fast edges also create electromagnetic interference and voltage-overshoot challenges. Designers must control layout, insulation, and switching behavior carefully.

The contest is therefore not a simple ranking between two materials. Different stages can use different semiconductors according to voltage, frequency, thermal load, fault requirements, and cost.

Infineon, onsemi, Navitas, Innoscience, Texas Instruments, and other suppliers are positioning products around future AI power chains. Some emphasize GaN, some combine GaN with silicon carbide, and others support multiple materials.

Onsemi, for example, presents silicon carbide for higher-voltage conversion while expanding GaN into AI power shelves and downstream stages. Its data center portfolio illustrates how suppliers can divide the power path between technologies.

Navitas has promoted GaN power ICs and silicon carbide products for data center conversion. Infineon combines an extensive silicon carbide business with growing GaN manufacturing and intellectual property.

Innoscience brings high-volume GaN-on-silicon production. Its scale adds price pressure, even though its product mix and technology path differ from Power Integrations’ sapphire-based strategy.

These competitors prevent Power Integrations from defining the market alone. AI infrastructure buyers can evaluate several materials and vendors at each conversion stage.

Power Integrations’ 1,250 V and 1,700 V products still shift the competitive frame. They force silicon carbide suppliers to defend applications that once had fewer credible GaN alternatives.

They also pressure other GaN vendors to explain their high-voltage roadmaps. A supplier concentrated around 650 V risks being confined to downstream conversion while rivals pursue the rack’s higher-voltage stages.

The 2,200 V test result adds evidence of device margin, but it is not the main competitive product specification. Commercial success will depend on qualified systems using the rated devices.

That is why efficiency measurements, thermal results, customer validation, and production availability deserve more attention than an isolated voltage number.

What the Company’s Reliability Model Does Not Prove

Accelerated stress testing strengthens the engineering case, but it cannot answer every question about field deployment.

Power semiconductor failures are rare under normal conditions. Waiting for enough failures at the intended operating voltage would make reliability studies impractically slow.

Engineers instead raise voltage and temperature to accelerate degradation. They record time-to-failure data, fit a statistical distribution, and model behavior under less stressful operating conditions.

Power Integrations says it tested devices at off-state voltages from 2,100 V to 2,200 V and temperatures from 80 to 120 degrees Celsius. It then used voltage and temperature acceleration models to estimate long-term reliability.

This is standard engineering practice. The usefulness of the prediction still depends on whether the selected stress activates the same failure mechanisms expected during service.

If excessive stress creates a different physical failure mode, the extrapolation can misrepresent normal operation. Engineers must inspect failures and validate the model across several conditions.

The published result also addresses intrinsic device reliability. A deployed converter includes solder joints, substrates, magnetic components, capacitors, cooling interfaces, isolation barriers, and control circuitry.

Any of those elements can limit system lifetime. A highly reliable transistor cannot compensate for a poorly cooled capacitor or an insulation defect elsewhere in the power supply.

Dynamic operation adds another layer. Devices inside AI power systems experience repetitive switching, changing load, temperature cycling, startup events, and occasional transients.

An off-state stress test examines one important dimension. It does not replace short-circuit testing, switching endurance, thermal cycling, surge qualification, or system-level fault evaluation.

The company’s model should therefore be read as evidence of design margin. It is not a guarantee that every converter using the device will reach the predicted lifetime.

Commercial readiness presents another uncertainty. Power Integrations has published technology material and reference designs, but widespread 800 VDC rack deployments remain ahead of the current market.

Data center operators qualify equipment conservatively because a power failure can interrupt costly computing clusters. New architectures must integrate with facility power, batteries, cooling systems, service procedures, and safety rules.

High-voltage direct current also changes protection requirements. Direct current does not cross zero every cycle like alternating current, making arcs harder to interrupt.

Solid-state circuit breakers, fuses, connectors, insulation monitoring, and maintenance procedures must mature alongside the semiconductor switches.

Supply diversity matters as well. Hyperscalers rarely want a critical power architecture to depend on one proprietary technology without alternatives.

Power Integrations controls its PowiGaN process and integrated designs. That control supports optimization, but customers must evaluate manufacturing capacity, second-source options, and recovery plans.

Cost remains difficult to judge from public technical material. A device can produce lower system cost by shrinking cooling and magnetic components, even when the semiconductor itself costs more.

The reverse can also happen. Specialized packaging, qualification, and redesign work can consume expected savings.

The fairest interpretation is that Power Integrations has produced credible evidence for high-voltage GaN reliability. The evidence supports continued evaluation inside 800 VDC systems.

It does not establish a universally available 2,200 V product, prove superiority over every silicon carbide design, or confirm large-scale customer adoption.

That gap is not unusual for an emerging power architecture. It is precisely what the next round of engineering disclosures must address.

What to Watch After the Google News Headline

Three signals will show whether high-voltage PowiGaN becomes an AI infrastructure platform or remains an impressive qualification result.

The first signal is a production product with an explicit 2,200 V rating. Power Integrations should provide a part number, datasheet, operating limits, package details, and qualification data if it intends to commercialize that voltage class.

Without those materials, readers should describe 2,200 V as an accelerated stress condition. Continued use of the number as a product rating would weaken confidence in the headline.

The second signal is customer-qualified 800 VDC hardware. Reference designs are useful, but production power shelves or auxiliary supplies reveal whether efficiency, temperature, protection, and density survive customer requirements.

NVIDIA’s architecture is the clearest near-term reference point. Watch for named manufacturers adopting Power Integrations parts, shipping schedules, and evidence that the designs enter real racks.

The third signal is comparative system data against silicon carbide. Useful results should cover efficiency across the load range, power density, thermal performance, fault behavior, component count, and total system cost.

A single peak-efficiency figure would offer limited guidance. AI accelerators change load rapidly, so performance across realistic operating conditions matters.

The broader GaN research supports using different device types at different points in the power chain. That makes converter-level comparisons more valuable than declaring one semiconductor material the universal winner.

For engineers, the immediate action is straightforward. Separate rated voltage, breakdown behavior, and accelerated stress voltage whenever a headline merges them.

Then examine the system around the switch. Ask which conversion stage it serves, what protections accompany it, how heat leaves the package, and whether the supplier has production-qualified hardware.

For enterprise buyers, track named deployments and qualification evidence before treating 800 VDC as settled infrastructure. For investors and industry watchers, follow design wins rather than laboratory superlatives.

Google News surfaced a compelling number, but Power Integrations’ real story is more substantial and more demanding. Can integrated 1,250 V and 1,700 V GaN earn a durable place inside 800 VDC AI racks? The next product datasheets, customer systems, and comparative tests should provide the answer.