Novosense’s GaN and 1,500V Isolation Push Faces a Hard Qualification Test

Novosense entered Google News with two connected claims: its GaN products target AI server power, while wider isolators address high-voltage systems reaching 1,500 volts. The combination sounds timely because AI racks are forcing power designers to reconsider voltage, conversion stages, component density, and electrical safety. Yet the products occupy different parts of that transition, and neither guarantees a place inside a deployed AI data center.

The strongest evidence comes from Novosense itself and its regulatory disclosures. The company says high-voltage GaN drivers for AI server power have reached batch shipments. It also sells digital isolators with package options offering 15-millimeter creepage, the surface distance separating conductive regions. Those are concrete developments, but shipment scale, named customers, qualification status, and data-center deployment volumes remain undisclosed.

That distinction matters as Nvidia builds an 800 VDC power ecosystem for future megawatt-class racks. Novosense is entering a supplier field that already includes Analog Devices, Infineon, Innoscience, Navitas, onsemi, Renesas, Rohm, STMicroelectronics, and Texas Instruments. Its contest is therefore not GaN against isolation. It is component availability against the far harder standard of system-level qualification.

What the Google News Headline Actually Represents

Novosense has assembled relevant components, but the evidence does not establish a new, unified 1,500V AI data-center platform.

The headline joins two product directions under one strategic story. The first covers GaN drivers and integrated power stages. Gallium nitride, or GaN, is a wide-bandgap semiconductor material that supports fast switching in compact power converters. The second direction covers galvanic isolation, which transfers signals across a protective electrical barrier without a conductive connection.

Novosense’s 2026 interim reporting provides the clearest commercial update. According to its regulatory filing, the company’s high-voltage GaN driver for AI server power had achieved batch shipments. The filing also said medium-voltage and low-voltage GaN integrated products had completed sample testing with target customers in AI power applications.

Those statements establish progress beyond a laboratory announcement. Batch shipment means devices have moved into a repeatable delivery process, although it does not reveal volume or whether the chips power production data centers. Sample testing represents an earlier stage because customers can still reject, redesign, or delay the component.

One relevant product is the NSD2622N, a high-voltage half-bridge driver designed for enhancement-mode GaN transistors. A half-bridge driver controls two power switches that alternate current through a conversion stage. Novosense says the device supports common-mode voltage from negative 700 volts to positive 700 volts and switching slopes reaching 200 volts per nanosecond.

The GaN driver specifications also list 2-amp source and negative 4-amp sink capability. Integrated regulators generate an adjustable positive gate voltage and a fixed negative 2.5-volt bias. That negative bias helps hold the GaN transistor off during fast voltage transitions, reducing false turn-on risk.

This integration addresses a real engineering problem. GaN switches can reduce converter size by operating at higher frequencies, but their gate requirements leave less tolerance for noise and layout errors. Removing external bias circuitry can reduce the bill of materials and shorten sensitive electrical paths.

Novosense’s isolation story has a longer history. The company introduced ultra-wide-body NSI82xx isolators before the present AI infrastructure cycle. Their connection to 1,500V systems originally centered on photovoltaic equipment, energy storage, charging equipment, and industrial power.

The current headline therefore combines a newer AI server opportunity with an established high-voltage isolation portfolio. That is strategically coherent because both depend on faster, denser power conversion. It is not proof that the same 1,500V isolator has entered an AI rack or that AI facilities have standardized around 1,500V distribution.

Google News readers should treat the development as a supplier-positioning story. Novosense has products aligned with high-voltage conversion, and it reports early commercial progress in AI power. The missing evidence concerns adoption depth, customer identity, and deployment architecture.

Why AI Data Centers Need a Different Power Chain

The opportunity exists because rack power is rising faster than traditional low-voltage distribution can scale economically.

A power system must deliver the same wattage through some combination of voltage and current. When voltage remains low, increasing power requires more current. Higher current demands thicker conductors, creates more resistive loss, complicates connectors, and consumes valuable rack space.

That problem becomes severe as AI systems concentrate accelerators, networking, memory, and cooling infrastructure inside tightly integrated racks. Traditional 48-volt or 54-volt rack distribution was developed for a much lower power range. Extending it toward hundreds of kilowatts forces designers to manage currents measured in thousands of amps.

Nvidia has proposed 800-volt direct-current distribution as its answer. The company says the architecture will support one-megawatt IT racks and beyond, beginning in 2027. Its design converts facility power to 800 VDC outside the compute rack, then distributes that higher voltage closer to the load.

The 800 VDC architecture reduces current for a given power level. It also seeks to eliminate intermediate conversions, reduce copper requirements, and move bulky equipment away from compute space. Energy storage remains necessary because GPU workloads can produce sharp, subsecond power changes.

This shift creates demand at several component layers. High-voltage switches control power conversion. Gate drivers command those switches. Isolators protect low-voltage controllers from dangerous power domains. Current sensors and isolated amplifiers provide measurements needed for regulation and fault detection.

GaN competes for the fastest and most space-constrained conversion stages. Silicon carbide, or SiC, often serves higher-power stages where voltage tolerance and thermal performance carry more weight. Conventional silicon remains attractive where cost, manufacturing maturity, and familiar design practices dominate.

Novosense does not manufacture every layer of this chain. Its relevant role includes gate drivers, integrated GaN power stages, digital isolators, isolated drivers, signal-chain devices, and current sensors. That breadth can help customers source compatible control and protection components, but it does not remove the need for external power transistors, magnetic components, firmware, cooling, and system certification.

The company’s NSD2622N illustrates the role of a driver rather than a complete power converter. It controls GaN switches and supplies suitable gate voltages. The customer must still select the transistors, design the power stage, manage electromagnetic interference, validate thermal behavior, and prove reliability under workload transients.

Novosense also offers the NSI66x2x family of isolated dual-channel gate drivers. The company lists switching support up to 2 megahertz, 6-amp source current, 8-amp sink current, and typical common-mode transient immunity of 150 kilovolts per microsecond. Common-mode transient immunity measures whether the isolation channel continues operating during rapid voltage changes.

Those characteristics matter because faster switching improves power density only when the surrounding control circuit remains stable. A false pulse can damage a power switch, interrupt a rack, or trigger protective shutdown. Isolation is therefore part of converter performance, not simply a compliance accessory.

This is why the dual push makes technical sense. GaN increases switching speed and raises stress on gate control, sensing, and communication. Better isolation helps the control system survive that environment. The two product lines reinforce each other, even when customers purchase them separately.

Novosense Faces an Ecosystem Qualification Contest

The decisive competition is not the specification sheet. It is whether system vendors trust these components through qualification, production, and long operating lifetimes.

Nvidia’s published partner list shows how crowded the power transition has become. It names established analog suppliers, wide-bandgap specialists, power-system manufacturers, and infrastructure vendors. These companies bring different advantages, including fabrication capacity, packaging knowledge, reference designs, field history, and direct relationships with hyperscale operators.

For Novosense, matching headline specifications is only the entrance requirement. AI data-center equipment must operate continuously under demanding electrical and thermal conditions. Component failures can affect expensive accelerators and interrupt workloads spanning many interconnected systems.

Qualification can include accelerated aging, temperature cycling, surge testing, electromagnetic compatibility, partial-discharge evaluation, and validation across manufacturing lots. Customers also examine process controls, supply continuity, failure analysis, and the vendor’s ability to support redesigns.

The NSD2622N integrates functions that would otherwise require external parts. That can reduce board area and simplify the negative gate-bias circuit. Integration also concentrates responsibility inside one device, however. Customers must verify that its regulator behavior, propagation delay, protection response, and transient immunity remain consistent across operating conditions.

Established competitors can answer those questions with deeper field records. Texas Instruments and Analog Devices have extensive isolation portfolios. Infineon, onsemi, Rohm, and STMicroelectronics combine power semiconductors with drivers and control products. Innoscience and Navitas have focused their strategies more directly on GaN adoption.

Novosense’s position is broader than that of a pure GaN supplier. Its isolation portfolio can give it access to designs that use GaN, SiC, or silicon switches. This reduces its dependence on one transistor technology and lets it participate wherever a controller must cross a high-voltage barrier.

The company also competes through packaging. Physical distance remains one of the clearest tools for controlling flashover along a component’s surface. Wider packages provide more creepage, but they occupy more board area. Designers must balance safety distance against density, cost, and routing complexity.

Novosense says its ultra-wide-body package reaches 15 millimeters of creepage. The NSI822x family includes one-channel and two-channel configurations, several data rates, and multiple package widths. Its NSI822x datasheet lists 15-millimeter clearance and creepage for the SOWW16 package.

The same datasheet lists a maximum working isolation voltage of 1,500 volts RMS for wide package variants and 2,121 volts for DC operation. Those are insulation ratings under defined standards and test conditions. They should not be confused with a universal claim that every 1,500V system automatically satisfies reinforced-insulation requirements.

System voltage is only one input to an insulation design. Engineers must also consider pollution degree, overvoltage category, altitude, material group, expected transients, equipment standard, and whether the barrier provides basic or reinforced insulation. A suitable component rating does not certify the finished system.

This qualification burden favors suppliers that provide clear documentation and application support. It also gives Novosense a path to differentiate. If the company can pair competitive isolation specifications with dependable supply and responsive engineering, it can win sockets without displacing an incumbent across the entire power chain.

That is a more realistic reading than a winner-takes-all contest. AI power systems contain many converters, sensing points, protection boundaries, and communication links. Multiple semiconductor vendors can participate in the same rack, while customers maintain alternative sources for critical components.

The 1,500V Label Needs Careful Interpretation

The ultra-wide-body isolator supports high-voltage design, but 1,500V does not describe the announced AI rack distribution standard.

Nvidia’s public architecture centers on 800 VDC distribution. Novosense’s 1,500V isolation material comes from products designed for several high-voltage markets. Photovoltaic installations are especially relevant because utility-scale solar systems already operate with 1,500V DC buses.

Novosense previously explained that photovoltaic equipment can require creepage exceeding 14 millimeters under certain 1,500V reinforced-insulation conditions. Its ultra-wide NSI82xx package provides 15 millimeters. The company also identifies energy storage, charging equipment, industrial controls, and high-voltage power supplies as target applications.

That history matters because solar and electric-vehicle systems provide useful engineering precedents. They have forced suppliers to develop wide-bandgap power devices, high-voltage packaging, isolated sensing, and fault protection. Nvidia explicitly presents these adjacent markets as a source of mature components and design experience for data centers.

Yet migration is not automatic. A solar inverter experiences different load patterns, service conditions, safety requirements, and lifetime stresses from an AI rack. Automotive components face demanding temperatures and vibration, but data centers place unusual value on continuous availability and coordinated fault containment.

The voltage numbers also refer to different things. An 800 VDC bus describes nominal power distribution. A 1,500-volt RMS isolation rating describes the permitted continuous voltage across an insulation barrier under specified certification conditions. Surge withstand voltage addresses short transient events rather than normal operation.

Mixing those categories can exaggerate product relevance. A component rated for a higher isolation voltage can provide design margin, but engineers do not select it from the headline number alone. Excessive package width can even become a disadvantage when the required clearance is lower and board density is critical.

The NSI822x datasheet provides another useful distinction. Its widest package has 15-millimeter external clearance and creepage. Clearance is the shortest distance through air, while creepage follows the insulating surface. Both influence high-voltage safety, but environmental conditions can affect them differently.

The company’s older product announcement linked the package to IEC 62109, a safety standard used for photovoltaic power converters. AI data-center equipment can fall under other standards depending on placement and function. Designers must map the device’s certifications to the exact equipment boundary.

An independent isolation analysis published by EE Times describes Novosense’s use of a dual-chip architecture, a silicon-dioxide dielectric, and on-off keying modulation. On-off keying encodes data by switching a carrier between present and absent states across the barrier.

The article also connects 15-millimeter packages with electric vehicles, solar systems, energy storage, and AI server power. However, it does not identify a hyperscale operator deploying Novosense’s widest isolator in a production 800 VDC facility. That commercial gap remains central.

The accurate conclusion is narrower than the Google News framing. Novosense owns isolation technology with ratings relevant to high-voltage conversion. That portfolio can support equipment around future AI power systems. Public evidence does not yet show 1,500V becoming the standard distribution voltage inside AI data centers.

GaN Integration Solves One Problem and Creates Others

GaN can shrink conversion hardware, but faster switching transfers difficulty into control, layout, thermal design, and electromagnetic compatibility.

GaN transistors can switch faster than traditional silicon devices and avoid the reverse-recovery behavior associated with a conventional body diode. Those properties can reduce switching loss and allow smaller magnetic components. The result can be a more compact power supply.

Novosense has worked on this approach for several years. Its earlier NSG65N15K power stage combined a half-bridge driver with two 650-volt GaN transistors. The device used transistors with 150 milliohms of on-resistance and supported a working current up to 20 amps, according to the company’s GaN product release.

Integration reduces the distance between the driver and switches. Shorter paths can lower parasitic inductance, which is unwanted inductance created by package leads, traces, and interconnections. Lower parasitic inductance helps control voltage overshoot during rapid switching.

The approach also restricts customer choice. A discrete design lets engineers combine a preferred driver with transistors selected for voltage, resistance, package, and sourcing. An integrated stage fixes more of that combination inside one product.

Thermal design presents another tradeoff. Higher frequency can shrink passive components, but losses still produce heat in switches, drivers, magnetics, and interconnects. A dense package must move that heat into a board or cooling structure without exceeding junction-temperature limits.

Electromagnetic interference creates further work. Rapid voltage edges can couple noise into nearby circuits, cables, and isolation channels. Designers may need slower gate transitions, filters, shielding, or different layouts. Each mitigation can surrender part of GaN’s theoretical density advantage.

The NSD2622N addresses one risk through negative turn-off bias. During a high-speed transition, capacitance inside a GaN transistor can inject current into its gate. A negative off-state voltage gives the circuit more margin before that disturbance accidentally turns the device on.

Novosense also specifies high immunity to rapid switch-node movement. These are relevant characteristics, but they remain company claims until customers validate them within their converter topologies. A test board cannot reproduce every interaction inside a dense rack power module.

AI workloads add another complication because their demand can change quickly. A converter must maintain voltage while accelerators move between computational phases. System designers coordinate power supplies, busbar capacitance, battery backup, control firmware, and workload management to prevent instability.

A gate driver contributes to that response but does not determine it alone. The overall transient behavior depends on the complete control loop, energy storage, converter topology, and distribution network. This is why component marketing should not be mistaken for an end-to-end AI power solution.

GaN’s competitive position is also stage-dependent. It can be attractive in high-frequency power supplies near the load. SiC can remain preferable for higher-voltage, higher-power conversion closer to facility input. Silicon MOSFETs can still win at lower voltages where their cost and mature supply base outweigh frequency advantages.

Novosense can participate across those choices because it offers drivers for several switch technologies. That flexibility strengthens its strategic position. It also makes customer qualification, not allegiance to GaN, the clearest measure of progress.

Three Signals Will Show Whether the Push Is Working

The next evidence must connect product specifications to named architectures, repeat orders, and independently validated system performance.

The first signal is customer progression from samples to sustained production. Novosense already reports batch shipments for high-voltage GaN drivers and sample testing for integrated GaN products. Future filings should reveal whether those programs expand, generate repeat orders, or enter more customer platforms.

Revenue alone will not answer every question because Novosense sells into automotive, industrial, energy, and consumer markets. Investors and buyers need segment detail linking growth to AI server power. Named customers or design wins would provide stronger evidence, although confidentiality can limit disclosure.

A move from sample testing to volume production would strengthen the dual-push thesis. Repeated language about testing without shipment growth would weaken it. Delays can arise from product performance, customer schedules, architecture changes, or broader data-center construction constraints.

The second signal is inclusion in public 800 VDC reference designs or partner programs. Nvidia’s ecosystem already identifies several semiconductor suppliers. Novosense is not named on the company’s published partner list, although absence does not prove exclusion from downstream equipment.

A public collaboration with a power-supply maker, rack integrator, or infrastructure vendor would close part of that gap. More valuable evidence would identify the component, conversion stage, qualification status, and deployment timetable. A general memorandum would carry less weight than a released reference design.

The third signal is independent system validation. Buyers should watch for efficiency curves, thermal results, electromagnetic compatibility testing, fault behavior, and long-duration reliability data from complete converters. Component maxima provide context, but production designs operate across load, temperature, and input-voltage ranges.

Certifications will also matter for the isolation portfolio. Novosense must show that each package and device variant satisfies the standards attached to its intended equipment. Engineers should verify the exact working voltage, insulation class, creepage, clearance, and environmental assumptions.

These signals are more informative than another Google News headline. They separate a technically plausible market position from verified commercial adoption. Novosense has already crossed one threshold by reporting batch shipments, but the available disclosures do not quantify its AI exposure.

Developers and knowledge workers may never select an isolator, yet the outcome still affects them. Power-delivery constraints influence where AI systems can be installed, how densely accelerators can operate, and how infrastructure costs evolve. Better components can help, but architecture and qualification decide whether those benefits reach deployed compute.

Enterprise buyers should ask vendors which power architecture supports a planned system, what redundancy model protects it, and how component changes affect serviceability. Hardware teams should compare full operating curves instead of isolated peak specifications. Investors should distinguish design activity from production revenue.

The central question now is measurable: can Novosense turn a broad high-voltage portfolio into qualified positions inside the emerging 800 VDC supply chain? Track the filings, partner designs, and independent test data. Those records, not aggregation momentum, will determine whether this Google News appearance marked commercial traction or only timely positioning.