Musashi Seimitsu might be the missing piece in power distribution for datacenters

Musashi Energy Solutions Technology might be the missing piece in power distribution for datacenters

One of the sectors with highest growth in the stock market since 2025, has been electrical power related stocks. Names like Bloom Energy, Infineon, GE Vernova, and Delta Electronics have posted triple-digit gains. The fact that developers no longer size projects by compute capacity. They measure them in electrical power. Power, not processing, has become the real constraint.

Traditional servers in data centers essentially execute several tasks running in parallel. While some applications are running, others remain idle. The effect translates into a nominal, stable and predictable power usage. AI factories are the current paradigm for delivering AI inference at scale. They pressure the electrical grid at unprecedented rates. The difference with traditional server racks is in the amount of electricity and the instantaneous power variation. 

Historically, the server racks use just CPUs for processing, and their power consumption varies between 15 kW and 80 kW. On the other hand, GPU-based Vera Rubin racks can reach up to 600 kW. Company reports such as Infineon’s expect the power needs in AI-focused server racks to surge up to 1 MW by 2030. 

Not only does power usage differ between traditional and AI focused server racks, the instantaneous power needs also change. In traditional servers, power draw stays mostly constant, and the operating system and server hardware absorb any changes in demand. In the AI factories, GPUs work in lockstep under the bulk synchronous paradigm. That translates into a unison execution of tasks in which power demand can go from 10 % to 100% in a couple of milliseconds. 

Consequences and risk mitigation 

‘This time is different’, is something we have all heard over the news regarding the stock market and the AI capex boom. At least for the electric grid, it seems like the change is real. As the energy demand changes from just predictable power usage, to a savage, explosive and volatile demand,  the industry is developing different strategies to mitigate the new problems.

In a recent study with the name ‘Power Stabilization for AI Training Datacenters‘, researchers from OpenAI, Microsoft and NVIDIA published the power requirements for the AI Factories.

The main risks identified by the researchers arise from the violent swings in the power demand. The sudden changes induced by the synchronized computing loads, force tens, if not hundreds, of megawatts of power to be either required or be ready to be used in the grid.  The problem is that the generation equipment is relatively slow and can’t react to the abrupt changes in power. When the load drops, the generation equipment must dissipate or absorb the excess power, which places mechanical stress on the machinery. On the other hand, if the grid can’t meet the power demand, the data center sits underused. To operate optimally and protect the data center’s equipment, operators must maximize instantaneous power efficiency.

The Software approach

The study advocates for cross-industry co-design through 3 strategies to maximize power efficiency and stabilize the power use. Software-based approaches are the easiest to deploy, using firmware and application software to smooth out violent power swings. During low-activity periods, the system injects controlled filler workloads to keep power draw steady. Examples include matrix multiplication sequences, GPU context sharing, and idle jobs. The approach is flexible and requires no hardware changes. It is the least efficient of the methods since it might impose performance limitations on the AI workloads, and the need for careful calibration of jobs, changes in the OS, and continuous monitoring.

Not only that, but the approach has limited effectiveness, as not all the power spikes can be filtered using this method. Consequences of this approach expand to the memory and compute resources needed just for monitoring and stabilization of the power usage. This translates to less compute resources for the business logic and overall energy burned. Finally, the coupling between application and infrastructure is far from ideal from a software architecture standpoint.  

The Hardware approach

Second, the paper introduces the hardware level changes at server compute tray level. At this level power optimization approaches rely heavily on NVIDIA’s firmware and changes in the hardware introduced in the server blades. The firmware based approach focuses on limitations on the ramping rate for power usage. The feature enforces minimum power thresholds and regulates power fluctuations. Finally, the stop delay approach forces a minimum power usage threshold before the GPU finally ramps down. Similar to the software based solution, the wasted energy is the main downside of the hardware based approaches. 

The Infrastructure approach

Third, the best-case solution to the training power-stabilization challenge is an energy-storage solution. It shall include enough capacitance to support the workload and handle the sudden rise/drop needs in power and can quickly switch between charge and discharge. It could charge during low-power phases and discharge during intense compute usage. The proximity with the rest of the compute modules makes rack-level storage the best option. The study highlights batteries and supercapacitors as the optimal approach to tackle the GPU induced power spikes.

Instantaneous power demand current industry approach

The industry is working in parallel to match the power requirements, both at hardware and software level of present and future data centers. To illustrate, NVIDIA recently proposed a completely new power distribution architecture. At datacenter level, NVIDIA’s 800 VDC Architecture development path is described as the natural path to keep pushing the compute boundaries with more power efficiency.

Not only that, At server rack level electrical changes were introduced in Blackwell, the last generation of NVIDIA’s compute modules. GB300 NVL72, a Blackwell based server rack, designed to integrate hardware and software to smooth power spikes and reduce the peak demand up to 30%. In addition, NVIDIA’s last generation,  VR200 NVL72 Vera Rubin based racks will increase the power consumption compared to GB300. This architectural change is expected to increase the number of passive capacitor components (MLCC) by 182% to respond to the high-frequency power transients and provide local energy storage around the computing units. In fact, Morgan Stanley analysis cited in recent coverage estimates the MLCC content in Vera Rubin VR200 at about $4320, compared to $1530 from Blackwell GB300 systems. 

Tackling Power Spikes and Industry leader’s approach

The focus of energy reliability extends to suppliers such as Vertiv ($VRT). The Battery Energy Storage System, DynaFlex BESS is used to set microgirds that use different power sources. The system guarantees Site-level energy assets are used to manage power demand, act as reliable backup during outages and constrain energy supply conditions and operational flexibility. The company also offers Battery based UPS systems such as Vertiv™ Liebert® EXL S1 UPS, designed to maintain uninterrupted operations of compute equipment with capacities up to 1200 kW. 

Although these solutions act to stabilize the power supply and reliability of the grid, both Dynaflex BESS and Liebert EXL S1 rely on electrochemical batteries as the core energy buffer. The EXL S1 platform, for instance, handles 125% overload for 2 minutes, or 150% for just 15 seconds. Vertiv built the UPS for sustained, near-nominal power draw.

As previously mentioned, GPU clusters can swing from idle to 100% of load in milliseconds, and this is where batteries struggle to handle the quick load changes. Moreover, every cycle accelerates the battery degradation. That is why Vertiv developed features like Input Power Smoothing. The system uses its batteries as a short-term buffer to smooth input and output power peaks. This reduces stress on utility sources and the wider grid. The method works, but it carries several disadvantages. Those drawbacks tend to limit the system’s performance.

The current approach: batteries

The most important is perhaps that the battery behaves asymmetric in load and discharge conditions. In addition, the rate at which batteries can charge and discharge is also limited to a certain threshold, which leads to oversizing the batteries for the system. To stabilize the load,  batteries must be properly sized and be capable of sustaining very high charge rates. High C-rate batteries let the UPS absorb and inject large amounts of power within a fraction of a second. This compensates for sudden AI workload swings and keeps input demand from the generator steady.

An undersized system limits the performance and compromises long-term reliability. From an energy storage perspective, when the battery recharge rate is too low, the UPS reduces its ability to sustain input power smoothing (IPS), allowing load fluctuations to propagate upstream and placing additional stress on generators and the electrical grid. 

Supercapacitors might be the missing piece to fight transients

Vertiv’s approach can stabilize load transients to a certain extent, but the batteries’ chemistry ultimately limits it. As previously mentioned, the alternatives rely on physical rather than chemical energy storage. Supercapacitors can fill the role in high power quick transient mitigation and several products in this category already exist. To illustrate, Flex’s ($FLEX) CESS is a capacitive based energy solution for AI datacenters, built around Musashi Energy Solutions’ Hybrid SuperCapacitor (HSC) Cells. Musashi Energy Solutions also offers their own equipment based on supercapacitors. For instance the  ESS400 Energy Storage System  can handle AI workloads with power capacities up to 400 KW per rack which can be combined to support operations of up to 2 MW.

Supercapacitors inherent advantage

Super capacitor cells can act on the same physical timescale as the load itself, thus reacting in milliseconds. Unlike battery chemistry, supercapacitors charge and discharge symmetrically. That symmetry removes the need to oversize the cells. In addition, Musashi’s cells reach peak discharge currents of up to 1,200 A, withstand up to 1 million charge-discharge cycles, and carry no thermal-runaway risk — unlike lithium-ion batteries.

Furthermore, Musashi’s hybrid supercapacitor cells combine high power density with a prismatic format. That design dissipates heat and packs more efficiently than competitors’ cylindrical cells. Compared to batteries, HSC cells tolerate one to two orders of magnitude more cycling than even high C-rate lithium-ion. They also charge and discharge symmetrically. Lithium’s asymmetry forces Vertiv’s IPS algorithm to oversize battery capacity just to keep recharge ahead of discharge.

Architecturally, the two approaches solve different problems: Vertiv’s UPS targets outages, while Musashi’s HSC-based buffering isolates load transients. UPS and HSC can be employed in synergy to reduce stress on the electric grid. 

The super capacitor business

Musashi’s advantage is clear: prismatic design of supercapacitors and the highest publicly verifiable power density for supercapacitors. This approach makes it ideal to match the industry needs in power transients. However, the company is not the only nor the largest supplier in the supercapacitor industry. Companies like Eaton ($ETN), Kyocera (6971.T), Panasonic (6752.T) and Murata (6981.T), to name a few,  also produce supercapacitors, are more diversified, and have larger market caps than Musashi. 

On top of that, Musashi appears to be in a weak financial position. The parent company, Musashi Seimitsu Industry Co. (7220.T), reported full-year revenue mostly flat at ¥347B ~$2.3B), while net income collapsed 84% YoY to just ¥1.26B, cutting net margin to 0.4%. The stock has also crashed abruptly, losing roughly 62% of its value from ¥10,550 to ¥4,035 in June 2026.

Musashi Energy Solutions current status and conclusions

Power generation remains a priority for data center developers, and Musashi’s technology might be a missing piece. Still, its financials suggest this thesis remains early-stage.

Musashi’s technology can definitely help in the transient power suppression, however what does not yet hold up is the financial investment thesis based on its financial performance. For now, Musashi is best understood as a speculative stock backed by real technology, but whose financial results have yet to materialize.

Investors can gain exposure to the technology without buying Musashi directly. One route is Flex, the integrator that already collaborates with Musashi on the CESS. Flex carries a far larger market cap. Its fiscal 2026 revenue reached roughly $28 billion, up 8% year-over-year, with operating margins near 6%.

Contact [email protected] for any questions or corrections.