Heterostacks: Japan’s power-semiconductor answer to AI’s electricity problem

AI is eating electricity. Someone has to make the meal efficient.

The loudest AI stories are written in software language: models, tokens, agents, prompts, benchmarks and very expensive demos. Heterostacks lives in the less glamorous basement of the AI economy, where electricity arrives as one thing and must be converted into another before GPUs can turn it into answers, images, code and occasionally very confident nonsense.

That basement matters. Generative AI has made data centers more power-hungry, more thermally stressed and more sensitive to every wasted watt. If the world wants larger AI clusters, denser racks, electric vehicles, renewable-energy systems and smarter factories, the bottleneck is not only chips that calculate. It is also chips that switch, convert and control power.

RealTech Fund’s investment in Heterostacks is therefore a small but telling business signal. Japan’s AI story does not have to be another chatbot. It can be a device architecture story: silicon, silicon carbide, losses, heat, reliability, cost and the brutal physics between the grid and the processor.

What Heterostacks is trying to build

According to UntroD Capital Japan, which operates RealTech Fund, Heterostacks is an Osaka Metropolitan University-origin startup developing heterojunction power semiconductor devices that combine silicon, or Si, with silicon carbide, or SiC. The company’s core idea is direct bonding of different semiconductor materials, using a heterojunction architecture to chase a combination that is difficult for single-material devices: lower loss, higher reliability, lower cost and easier implementation.

The company’s announced development targets include a Si/SiC junction transistor called SABFET and a Si/SiC junction diode called JGSD. These are aimed at medium- and high-voltage areas, including data-center power supplies, where reducing power-conversion losses can lower electricity use and carbon dioxide emissions.

That is the heart of the story. Heterostacks is not claiming that silicon alone is useless or that SiC alone is magic. It is asking whether a hybrid architecture can use silicon’s controllability and reliability with SiC’s high-voltage and low-resistance strengths. In a world increasingly built from mixed systems, that is a very modern semiconductor bet.

Why power semiconductors suddenly sound strategic

For decades, power semiconductors were the kind of components that only engineers and procurement teams loved openly. They were not cultural objects. Nobody lined up outside a store to buy a new diode. But they sit inside the systems that make modern industry work: inverters, chargers, servers, renewable-energy equipment, factory drives, rail systems, air conditioners, and every machine that must change electricity into the exact form it needs.

AI has made that old industrial layer feel urgent again. AI servers draw power in large, dynamic, heat-producing bursts. Modern data centers increasingly use 48-volt distribution and complex conversion stages to reduce losses and scale density. In those chains, wide-bandgap materials such as SiC and GaN can offer advantages in voltage, temperature and switching speed, while silicon remains cost-effective and mature.

Heterostacks fits into that broader movement by arguing for a material-combination route. Instead of worshipping one material, it asks where each material performs best and whether direct bonding can create a device with fewer trade-offs.

Osaka Metropolitan University, deep tech, and the Japan angle

The company’s university-origin identity matters. Japan has world-class semiconductor history, but the country’s challenge is turning research depth into venture speed. Osaka Metropolitan University-origin Heterostacks sits in that national tension: excellent science, hard manufacturing questions, patient capital, and the need to move from laboratory promise to industrial relevance.

RealTech Fund is a fitting investor because its strategy is deep tech rather than fashionable internet software. The fund describes itself as backing research-and-development startups solving global and human challenges, especially in fields where ordinary capital may hesitate. That is exactly the zone power semiconductor materials occupy. The upside is enormous. The path is slow, technical, expensive and unforgiving.

For Japan.co.jp, the interest is not merely that a startup received money. The interest is that Japan may be trying to re-enter the advanced semiconductor story from the side door: not only leading-edge logic fabs, but power devices, materials, packaging, bonding, reliability and AI electricity efficiency.

The products: SABFET and JGSD

Dealroom.co summarized the investment as funding to accelerate research and development of two key products: the Si/SiC junction transistor SABFET and the Si/SiC junction diode JGSD. The funding amount was not disclosed publicly. The key technical idea is direct bonding between silicon and silicon carbide, not merely packaging them near each other.

A transistor controls current. A diode allows current to flow in one direction. That sounds elementary until the voltage rises, the switching frequency increases, the temperature gets ugly, and the operator wants smaller equipment with less wasted energy. Then a “simple” component becomes a strategic part of the machine.

The promise is that Si/SiC junction devices could help reduce power-conversion loss in medium- and high-voltage applications. The risk is equally plain: bonding different semiconductor materials at scale is difficult. Deep-tech press releases often sound clean because the dirty work happens in process development, yield learning, thermal stress, interface defects and manufacturing repeatability.

The data-center power problem is not theoretical

Infineon, one of the major global players in power semiconductors, has described AI data centers as highly dynamic engines that create rapid current spikes, sustained high loads and extreme thermal stress. Its technical blog argues that SiC power supplies matter because traditional silicon-based approaches hit limits as rack power rises and thermal margins shrink.

This does not mean Heterostacks is competing directly with every large SiC vendor tomorrow. It means the problem Heterostacks is addressing is real. AI infrastructure is not only about who has the fastest accelerator. It is about who can feed those accelerators with less waste, less heat, less space and less failure.

In the end, a data center operator does not want a philosophical argument about semiconductor materials. It wants efficiency, reliability, availability, compactness, supply stability and cost. If Heterostacks can deliver a meaningful improvement in those trade-offs, a small Japanese startup could become relevant in a very large market.

The risk: from clever architecture to manufacturable device

The danger is that “promising architecture” is not the same as “manufacturable business.” Semiconductor history is full of elegant ideas that struggled in production. Direct bonding of different materials raises interface questions. Power devices must survive heat, voltage, cycling and harsh operating conditions. Reliability is not a nice extra; it is the product.

Heterostacks must prove high-voltage performance, stabilize its bonding technology, establish core device processes, and eventually persuade customers that the performance, price and reliability are worth changing designs. That is a long road. It is also the kind of road where Japanese process discipline may matter.

The company will also face giants. Established players in SiC, GaN and silicon power semiconductors are not asleep. They have fabs, customers, qualification history and manufacturing scale. A startup must find the opening where its architecture is not merely interesting, but necessary.

What to watch

A Japanese answer hiding below the chatbot layer

The Heterostacks story is refreshing because it refuses the obvious AI narrative. There is no digital assistant with a cute name. No viral demo. No promise that a model will replace your entire back office by Tuesday afternoon.

Instead, there is a harder and more useful question. As AI demands more electricity, who makes that electricity easier to manage? Who cuts loss between the grid and the chip? Who reduces heat before cooling becomes a tax on intelligence? Who makes electrification less wasteful?

Japan has spent decades living inside these questions, even when they did not sound fashionable. Heterostacks is one small company, very early, with a difficult technical road ahead. But the direction is right. The future of AI may depend not only on what models think, but on how efficiently the machines beneath them breathe.