Power GaN Expands Into AI Data Centers and EV Drivetrains

Google News has surfaced a telling shift in power semiconductors: gallium nitride is moving beyond chargers into megawatt AI racks and core EV systems.

The important change is not a new consumer adapter. Power GaN suppliers are targeting infrastructure where small efficiency gains affect cooling capacity, copper requirements, vehicle range, and operating costs. Those applications also impose stricter reliability standards than phone chargers.

NVIDIA’s planned 800 VDC AI architecture gives this transition a firm deadline. The company expects megawatt-class racks to adopt high-voltage direct-current distribution beginning in 2027. That plan pressures power suppliers to qualify GaN, silicon carbide, and conventional silicon for distinct stages of the same electrical chain.

GaN will not replace every incumbent technology. Silicon carbide still suits many higher-voltage and higher-power stages, while silicon remains inexpensive and deeply established. The contest is about which material controls each conversion step, not which material wins everywhere.

What the Google News Headline Really Signals

Power GaN is leaving its first mass-market niche and entering systems where electrical efficiency determines how much computing or driving performance fits inside a fixed space.

Gallium nitride, or GaN, is a wide-bandgap semiconductor that switches electricity faster than conventional silicon while losing less energy as heat. Its early commercial visibility came from compact USB chargers. Higher switching frequencies allowed manufacturers to shrink transformers and other passive components.

That success established a manufacturing base, but chargers were only an opening market. A consumer adapter handles tens or hundreds of watts. An AI server power supply can handle several kilowatts, while an entire rack can demand hundreds of kilowatts.

The shift changes what customers value. A smaller charger offers portability. A denser server power system creates room for additional accelerators, memory, networking, or cooling equipment inside a constrained rack.

Texas Instruments says GaN devices can operate above 500 kHz and enable magnetics up to 60 percent smaller than comparable components. Its GaN technology also integrates drivers and protection functions that previously required separate parts.

Integration matters because faster switching creates its own engineering problems. Poor board layouts can produce voltage overshoot, electromagnetic interference, and unwanted heat. A fast transistor alone does not guarantee an efficient or dependable system.

Data-center suppliers must therefore deliver complete designs around the device. Those designs include controllers, gate drivers, sensing, isolation, protection, packaging, and thermal management. Customers are buying a qualified power architecture, not merely a better switch.

The same principle applies inside electric vehicles. GaN first appeared as an option for onboard chargers and DC-to-DC converters. Suppliers now want to extend it toward traction inverters, the systems that turn battery power into controlled current for electric motors.

That expansion carries greater consequences. A charger operates during a limited part of the vehicle’s life. A traction inverter must respond continuously to acceleration, regenerative braking, temperature changes, vibration, and fault conditions.

This is why the Google News framing matters. It connects two markets that share a basic constraint: more power must pass through less space without creating unacceptable heat. However, their qualification cycles and failure risks remain very different.

The transition is also broader than GaN. Silicon carbide, or SiC, has gained ground in high-voltage EV inverters and industrial power systems. Conventional silicon remains competitive where switching speed or size does not justify a more expensive material.

Power GaN must earn specific positions within this mixed-material system. Its strongest argument is not novelty. It is the ability to raise switching frequency, reduce conversion losses, and shrink surrounding components where space carries economic value.

AI Data Centers Turn Efficiency Into Compute Capacity

In an AI data center, better power conversion does more than lower the electricity bill. It determines how many processors can operate within the rack’s physical and thermal limits.

Traditional cloud racks were commonly designed around power levels measured in tens of kilowatts. Current AI systems have moved beyond 100 kW, and future designs are heading toward one megawatt per rack.

That increase exposes the limitations of low-voltage distribution. Delivering a fixed amount of power at a lower voltage requires more current. Higher current needs thicker conductors and produces more resistive loss.

NVIDIA says a one-megawatt rack using 54 VDC distribution would require up to 200 kilograms of copper busbar. It also says power shelves could consume most available rack space in a comparable design.

The company’s proposed 800 VDC architecture raises the distribution voltage and lowers the required current. Power would travel as high-voltage direct current before being converted near the processors.

NVIDIA plans to support one-megawatt racks and higher beginning in 2027. Its partner list spans semiconductor vendors, power-system manufacturers, and infrastructure suppliers. Infineon, Navitas, Texas Instruments, Delta, Eaton, Schneider Electric, and Vertiv are among the named participants.

This ecosystem approach is essential. An 800 VDC system needs more than efficient conversion. It requires safe connectors, circuit protection, isolation, battery systems, monitoring, compatible racks, and maintenance procedures.

GaN fits particularly well in high-frequency conversion stages close to the computing load. Faster switching can shrink transformers, inductors, and capacitors. Lower losses can also reduce the burden placed on liquid-cooling systems.

Infineon introduced two CoolGaN reference designs in March 2026 for converting high-voltage DC inside AI infrastructure. One converts 800 VDC to 50 V, while the second converts 800 VDC directly to 12 V.

The company reports that its 800 VDC design exceeds 98 percent efficiency at full load. That figure comes from the supplier and requires validation in customer deployments across different operating conditions.

Still, the architecture reveals where GaN vendors see an opening. GaN can handle the high-frequency intermediate stage, while SiC may serve grid-facing or higher-voltage conversion. Silicon remains useful near processors and in cost-sensitive auxiliary circuits.

Navitas has demonstrated another route. Its 12 kW reference power supply combines SiC devices in the input stage with GaN components in the isolated conversion stage. The company reports 97.8 percent peak efficiency.

That 12 kW platform targets 120 kW server racks and follows Open Rack v3 requirements. It illustrates how suppliers are dividing jobs between materials instead of betting on a single universal switch.

The economics become meaningful at facility scale. A fraction of a percentage point saved at one conversion stage looks modest. Multiplied across thousands of power supplies operating continuously, it reduces wasted energy and associated cooling demand.

Texas Instruments has estimated that a 0.8 percentage-point efficiency gain in a 100 MW data center could save substantial energy costs over ten years. The exact result depends on utilization, local electricity rates, cooling design, and system lifetime.

Efficiency also affects deployment speed. Grid access has become a limiting factor for new AI infrastructure. Using available electricity more effectively lets an operator install more computing capacity before reaching its contracted power ceiling.

The International Energy Agency reported that data-center electricity demand grew 17 percent during 2025. Its energy outlook expects total consumption to double by 2030, while AI-focused facilities could triple their power use.

Those projections strengthen the case for better conversion. They do not mean GaN can solve the wider energy problem. Efficient electronics cannot create generation capacity, shorten every grid queue, or eliminate the environmental effects of additional demand.

GaN’s role is narrower but commercially important. It can reduce losses and equipment volume between the facility connection and the processor. In a constrained AI campus, that improvement can translate directly into usable compute capacity.

GaN Versus Silicon Carbide Is the Wrong Contest

The central competition is not GaN against SiC across an entire system. It is a contest for the conversion stages where each material creates the most value.

GaN and SiC are both wide-bandgap semiconductors, but their characteristics differ. GaN is attractive at high switching frequencies and medium-to-high voltages. SiC performs well at higher voltages, high temperatures, and heavy power loads.

AI power architectures contain several conversion stages. Electricity may arrive from a medium-voltage grid, pass through facility-level rectification, enter an 800 VDC bus, and step down near the computing equipment.

No single device optimizes every stage. A grid-facing converter prioritizes voltage handling and fault tolerance. A converter beside a GPU prioritizes density, switching speed, and minimal heat.

The emerging division places SiC near high-power input stages and GaN closer to high-frequency intermediate conversion. Conventional silicon can remain at lower voltages or in control functions where its cost and maturity dominate.

That division is visible in commercial reference designs. Navitas combines SiC and GaN in its 12 kW platform. Infineon sells all three materials and presents them as complementary parts of a grid-to-chip portfolio.

This mixed approach pressures pure-play vendors. A customer building a complete power shelf may prefer a supplier that can recommend devices across several stages. A specialist must prove that its technology advantage offsets the convenience of a broader portfolio.

It also complicates market forecasts. A headline about the growth of power GaN does not show how much semiconductor content GaN will capture per vehicle or server. Revenue depends on device count, voltage class, packaging, integration, and customer pricing.

Design wins offer better evidence than broad market estimates. A reference board shows technical intent, but a qualified production contract shows that an operator accepts the cost, reliability, and supply chain.

The same distinction applies to efficiency claims. Peak efficiency is usually measured under selected laboratory conditions. Real systems operate across changing loads, temperatures, input voltages, cooling states, and transient events.

AI workloads can produce sudden changes in power demand. Accelerators shift between idle, communication, and computation phases. Power equipment must maintain stable output during those transitions without wasting excessive energy at partial load.

GaN’s switching speed can improve response and density, but it raises design sensitivity. Parasitic inductance from packaging or circuit traces can create overshoot. Engineers must control switching behavior without surrendering the expected efficiency gain.

Protection is another challenge. High-voltage DC does not naturally cross zero as alternating current does. Interrupting faults therefore requires suitable breakers, solid-state protection, sensing, and carefully coordinated controls.

These surrounding systems can determine whether an 800 VDC design reaches production. A converter with impressive laboratory efficiency will not matter if operators cannot isolate faults safely or service equipment without extended downtime.

This is why the next two years matter. NVIDIA has provided a target architecture and a 2027 adoption window. Power suppliers now need to turn demonstrations into interoperable, serviceable equipment.

Google News may frame the story around the rise of GaN, but customers will evaluate the complete electrical path. The winning material at one stage can depend on the packaging, controller, and protection choices around it.

EV Adoption Must Move Beyond Onboard Chargers

GaN’s automotive opportunity becomes much larger only when it advances from charging hardware into propulsion, but that move sharply raises the reliability threshold.

An onboard charger converts grid electricity into regulated DC power for the vehicle battery. A DC-to-DC converter supplies lower-voltage systems from the high-voltage battery. Both applications reward smaller components and efficient switching.

GaN can reduce the weight and volume of those systems. Smaller magnetics and improved efficiency can free packaging space, lower cooling needs, or support faster charging designs.

Traction inverters present a more demanding prize. They control the electricity flowing between the battery and motor. Better inverter efficiency can preserve energy during driving and reduce heat inside the vehicle.

However, traction hardware operates under frequent load changes. It must survive high temperatures, vibration, moisture exposure, rapid acceleration, regenerative braking, and abnormal electrical conditions.

Automakers also expect long service lives and extremely low failure rates. Qualification extends beyond a successful prototype. Suppliers need manufacturing consistency, traceability, automotive-grade packaging, and stable output across millions of components.

SiC already holds a strong position in high-voltage EV traction systems. Its voltage handling and thermal performance suit 800 V vehicle platforms. GaN must therefore prove a system advantage, not merely comparable transistor specifications.

The likely path is incremental. GaN can expand through onboard chargers and auxiliary conversion, where suppliers already understand its behavior. It can then enter selected inverter designs as packaging and high-voltage devices mature.

Vehicle architecture will influence that pace. Lower-power vehicles may value GaN’s compactness and switching frequency. Larger vehicles may favor SiC for higher current, voltage, and thermal margins.

Cost remains central. GaN may shrink passive components and cooling hardware, potentially reducing the total system bill. Yet the semiconductor itself must be available at automotive scale with predictable yields.

Manufacturing programs are moving in that direction. Imec launched a 300 mm GaN program with partners including GlobalFoundries, KLA, Synopsys, AIXTRON, and Veeco.

Larger wafers can produce more devices per manufacturing cycle and use established semiconductor equipment. Imec says the program targets lower manufacturing costs and more advanced low-voltage and high-voltage devices.

A larger wafer does not automatically make GaN inexpensive. Yield, wafer bow, epitaxial quality, defect control, packaging, and testing still determine production economics. Automotive qualification adds more time and expense.

Supply concentration also matters. Automakers resist dependence on a single source for critical propulsion components. Device compatibility and second-source options can influence adoption even when one supplier has better performance.

For that reason, the expansion beyond chargers will be measured in named vehicle platforms and production volumes. Technical presentations indicate direction, but sustained shipments reveal whether GaN has crossed the automotive threshold.

The AI market can indirectly help. Data-center demand supports investment in fabrication, packaging, controllers, and high-voltage device development. Those capabilities can strengthen the wider GaN supply chain.

Automotive demand can return the favor by creating larger and steadier production volumes. The two markets have different qualification requirements, but both reward efficient power conversion within limited space.

This shared manufacturing base explains why suppliers discuss AI data centers and EVs together. The applications are not identical, yet progress in wafer scale, packaging, reliability, and integration can benefit both.

What Power GaN Claims Still Need to Prove

The case for GaN is technically credible, but reference designs and supplier efficiency figures do not yet establish broad production adoption.

The first uncertainty is reliability at scale. Consumer chargers helped suppliers accumulate operating experience, but AI racks and traction inverters create different electrical and thermal stresses.

Data-center equipment runs continuously and must support strict service targets. A small failure rate multiplied across thousands of power modules can create maintenance costs that outweigh an efficiency benefit.

Automotive systems face even harsher physical conditions. Suppliers must demonstrate stable performance over temperature cycling, vibration, electrical surges, humidity, and years of real driving.

The second uncertainty is economics. GaN can reduce losses and shrink surrounding components, but buyers evaluate the complete system cost. That includes devices, controllers, cooling, protection, manufacturing, and field support.

A design can deliver greater power density without producing a lower total cost. In some deployments, mature silicon remains sufficient. In others, SiC offers a better balance of voltage capability and qualification history.

The third uncertainty is standardization. NVIDIA’s 800 VDC plan gives the industry a common direction, but equipment from different vendors must work together. Protection methods, connectors, controls, and maintenance procedures require coordination.

Open Compute Project specifications can help align suppliers. Even so, operators will test interoperability, fault behavior, and serviceability before committing critical facilities to a new distribution architecture.

The fourth uncertainty is timing. NVIDIA points to 2027 for its next-generation architecture. Facility development, procurement cycles, and component qualification can move more slowly than processor roadmaps.

Early systems may use sidecars, centralized power equipment, or hybrid arrangements. Those transitional designs could support GaN, but they might distribute semiconductor content differently from the final architecture.

The fifth uncertainty involves workload behavior. AI accelerators generate fast power swings, and rack-level storage may become part of the solution. GaN converters must operate efficiently during those changing loads.

Peak efficiency alone is insufficient. Operators need efficiency curves, thermal data, fault-response results, electromagnetic compatibility, and lifetime estimates across realistic conditions.

Supplier concentration creates another risk. The industry includes large diversified companies and smaller specialists. Customers need confidence that chosen components will remain available throughout lengthy facility and vehicle programs.

Manufacturing scale should therefore be treated as evidence, not a footnote. New 300 mm programs are promising, but yields and qualified capacity matter more than wafer diameter alone.

Environmental claims also require discipline. More efficient conversion reduces waste during operation. It does not erase the electricity demand, construction impact, water use, or material requirements associated with expanding AI infrastructure.

The IEA expects data-center electricity consumption to rise substantially even as hardware efficiency improves. Growing compute demand can exceed the savings created by better converters.

GaN should be judged against a practical standard: whether it reduces the energy, space, cooling, and material needed for each delivered unit of compute or mobility.

That is a meaningful contribution, but it is not a complete answer to either sector’s energy constraints.

Three Signals Will Show Whether the Shift Is Real

The next phase will be decided by production deployments, measured operating data, and manufacturing depth rather than additional laboratory records.

The first signal is customer adoption of 800 VDC equipment. Watch for named hyperscalers, server manufacturers, and power suppliers qualifying interoperable systems for 2027 deployments.

A production announcement should specify where GaN appears in the power chain. It should also identify rack power, conversion stages, efficiency ranges, protection methods, and applicable standards.

Broad partner lists show alignment, but purchase commitments reveal commercial confidence. Repeated orders would strengthen the argument that GaN has become necessary infrastructure for dense AI computing.

The second signal is a production EV traction platform. GaN already has a reasonable path through onboard chargers and DC-to-DC converters. A high-volume traction inverter would represent a more consequential step.

The strongest evidence would include a named vehicle, qualification status, expected production timing, and the system benefit. Claims about size or efficiency should cover normal driving conditions rather than a single optimal point.

A traction design would not mean GaN has displaced SiC. It would show that automakers see room for both materials across different vehicle classes and voltage levels.

The third signal is qualified manufacturing capacity. Watch whether 200 mm and 300 mm GaN programs produce automotive and infrastructure devices with consistent yields.

Capacity announcements should be accompanied by customer sampling, reliability data, and production schedules. Multiple qualified sources would reduce procurement risk for vehicle and data-center buyers.

These signals also provide a better filter for future Google News coverage. A new transistor or reference design is a development milestone. A deployed rack, shipping vehicle, or qualified fabrication line marks adoption.

Readers following this market should track the electrical architecture around AI chips, not only the processors themselves. Computing performance increasingly depends on power delivery, cooling, storage, networking, and grid access.

The same systems view applies to EVs. Battery chemistry attracts attention, but conversion efficiency influences charging hardware, auxiliary systems, propulsion, heat, and vehicle packaging.

Teams evaluating these developments can preserve source material, architecture notes, and supplier claims inside a searchable AI knowledge base. That makes it easier to compare announcements with later production evidence.

The central question is now concrete: will GaN remain an efficient component inside selected converters, or become a standard layer across AI and automotive power systems?

Watch the first shipping 800 VDC racks, the first scaled GaN traction programs, and the first qualified large-wafer supply agreements. Those events will show whether the Google News headline captured a durable market transition.