In one volatile week, GaN power electronics went from “great for phone chargers” to “running the backbone of AI factories.” On August 1, 2025, Innoscience said it’s in NVIDIA’s 800 VDC data-center power program with a “full-link” GaN stack—from the 800-volt feed all the way down to the GPU supply rails. The same day, a Munich court issued a first-instance injunction siding with Infineon that restricts Innoscience from making, selling, or marketing certain GaN products in Germany, pending appeal. Velocity meets moats. And that clash tells you where the power game is heading.
Decode the tech
What 800 VDC actually means (and why it’s a big deal)
Most modern data center racks distribute power at 48/54 V DC. That works until you try to feed ~1 MW per rack systems for next-gen AI. At those levels, the copper required to carry the current balloons, and every step where electricity is changed from one form to another, consumes efficiency and space. 800 VDC pushes high-voltage DC closer to the load, slashing current, copper mass, and the number of times the electricity needs to be converted. You then step down inside the rack and server to the GPU’s low-voltage rails, which are the current supply lines that power the chips with rapid increases in demand during intense processing. To manage these demands, power supply components must work precisely, especially when the current surges during workload spikes, typical of AI training, which is why the efficiency of the final connection to the power distribution network is so crucial.
Texas Instruments calculates that a 1 MW rack at 48 V would need ~204 kg of copper busbars and cabling. Moving to 800 V dramatically reduces the copper mass and associated resistive losses. That’s a design win you can feel in your back and your operating costs.
Why GaN (Gallium nitride semiconductors) matters
GaN switches are quicker and have lower energy losses compared to silicon when used at high frequencies. This quicker switching means that the components like magnets and capacitors can be smaller, which in turn leads to smaller circuit boards and reduced sizes for power supplies. Additionally, GaN can handle higher temperatures and voltages better than regular silicon.
The market is no longer hypothetical. Yole Group expects the GaN power market to top $2 billion by 2029; Infineon quotes a GaN-power CAGR of ~36% toward ~$2.5 billion by 2030. Either way, we’ve crossed from “promising” to “industrial.”
What changed in 2025
- AI is transforming energy systems. As we design powerful computer systems that use a lot of electricity, traditional voltage levels are becoming less effective. Companies like NVIDIA are introducing new blueprints that utilize higher voltage levels, and many leading manufacturers are ready to provide the necessary components. New technologies are helping improve the efficiency of how power is converted and used in these systems, making them more effective as we move forward.
- The GaN supply map is changing. TSMC will stop providing GaN wafer foundry services by July 31, 2027. This means companies that rely on TSMC for GaN production will need to find new manufacturers. Navitas, which has used TSMC’s services, has already formed a new partnership with PSMC for production, showing a proactive approach. We can expect X-FAB and other manufacturers to take on the extra workload.
- Scale meets IP. Infineon is focusing on intellectual property rights in Europe and the U.S. A German court recently ruled in favor of Infineon, impacting certain products from their competitor. In the U.S., Infineon is involved in active legal cases that could influence the market. These developments are likely to shape the industry both in the courtroom and in manufacturing.
Innoscience’s moment validated (and contested)
NVIDIA has officially listed Innoscience as one of the chipmakers helping supply the special power electronics for its future data centers. Innoscience says its GaN (gallium nitride) chips can now handle the entire “power chain” inside those systems—from the very high 800-volt input all the way down to the tiny voltages that feed NVIDIA’s GPUs. That’s like passing a final exam: it proves GaN is no longer just for small gadgets like phone chargers, but ready for the world’s most advanced computing.
From a market perspective, this is also about positioning. Innoscience is the only Chinese company on NVIDIA’s 800-volt supplier list, holding about ~30% of the GaN power-device market worldwide, which puts it in the leadership tier. That said, NVIDIA is still only testing the technology, and it’s not yet in mass production. But even being on the shortlist is a significant stamp of credibility.
Of course, the celebration comes with challenges. Innoscience is currently fighting legal battles with Infineon over patents. This matters because big customers like NVIDIA care as much about legal certainty as they do about technical performance. If a supplier is tied up in lawsuits, buyers worry that future products could be blocked or disrupted. That’s why strong intellectual property isn’t just a legal shield, it’s a competitive weapon.
A Third Way: How ST and Innoscience Are Hedging the Supply Chain
Instead of fighting in court, STMicroelectronics (ST) and Innoscience decided to work together. STMicroelectronics is one of Europe’s biggest semiconductor companies, headquartered in France and Italy. They make a wide range of chips, from car electronics and industrial sensors to power devices, and are a key player in Europe’s tech independence push. Recently, they signed an agreement to share technology and manufacturing capacity. The deal means Innoscience can use ST’s factories outside of China, while ST gets access to Innoscience’s large GaN production lines inside China.
This partnership isn’t just about making chips; it’s about supply chain security. With trade tensions and export controls shaping the semiconductor world, both companies are reducing their risk. By linking their factories across regions, they’re ensuring they can continue producing GaN chips even if politics or regulations disrupt one part of the world.
The wafer-size race and why it matters to costs
Innoscience made an early decision to focus on a specific technology that uses 8-inch (200 mm) wafers made from a material called Gallium Nitride (GaN) on Silicon (Si). This choice enabled them to produce their products at a lower cost than others who are still trying to catch up. Now, they’re moving into a new phase where they will use larger 300 mm wafers. Infineon, a major player in the industry, announced that they are on schedule to send out initial samples of its new 300 mm products by the fourth quarter of 2025. Using these larger wafers means that they can produce approximately 2.3 times more chips from each wafer, which helps reduce costs if they can maintain good production quality. It’s crucial to monitor developments closely during this period, including key product releases and the number of high-quality chips they can produce, as these factors will significantly influence market pricing.
When the giants step aside, new players rise
The world’s biggest chipmaker, TSMC, has decided it will stop producing gallium nitride (GaN) power chips by 2027. That single move is shaking up the industry. Companies that relied on TSMC now need to find new partners fast. One clear example: Navitas, a major GaN designer, has already shifted production to Taiwan’s PSMC. Others are lining up deals with specialty foundries across Europe and Asia. When the biggest factory walks away, the smaller ones thrive. Foundries that specialize in niche technologies suddenly find themselves at the center of the action.
Why power electronics is stealing the spotlight
For the past two years, U.S. export controls have made it harder for China to access the most advanced logic chips—the kind used in cutting-edge AI processors. If you can’t compete head-on in that race, you double down where you still can: power electronics.
And that’s where wide-bandgap materials like GaN (gallium nitride) and SiC (silicon carbide) come in. These aren’t household names, but they are the backbone of the energy systems that feed AI. By 2025, the shift is crystal clear: the real bottleneck isn’t just who makes the fastest AI chip—it’s who can deliver enough clean, efficient power to keep those chips running. The race for AI dominance isn’t only fought on GPUs, it’s fought on the power rails that feed them.
What this means for the ecosystem
The real race in AI isn’t just about who builds the fastest chips. It’s about who can keep them powered. As data centers shift from today’s 48-volt systems to 800-volt architectures, the game changes completely. Suddenly, the winners aren’t only NVIDIA or AMD; they’re the companies that can deliver the power chain behind them. Gallium nitride (GaN) and silicon carbide (SiC) are no longer niche materials but the backbone of this transition. The factories that make them, the suppliers that refine them, and the startups that solve cooling, connectors, and reliability will quietly become the most strategic players in the AI boom.
TSMC’s retreat from GaN has blown the market wide open. Specialty foundries are rushing to capture that business, and the companies that move fast will lock in customers for the decade ahead. For anyone building or investing in AI infrastructure, this isn’t optional; testing 800-volt systems now is the only way to future-proof your clusters. And for governments, the message is just as stark: without domestic strength in power electronics, you’re not really in the AI race, no matter how many GPUs you buy.
The bottom line is that AI’s next bottleneck isn’t chips, it’s power. Whoever controls GaN and SiC controls the flow of electricity that fuels the entire industry.