Toshiba International Corporation (TIC), Toshiba's U.S. subsidiary, and DIMAAG-AI announced on September 9, 2026 that they will collaborate on developing energy storage systems for AI data centers. Toshiba's lithium titanate oxide (LTO) battery, "SCiB," will be built into DIMAAG's power delivery platform, the "ZettaWatt Power Platform." The plan is to support continuous charging and discharging at 10C, assuming adequate cooling, and to target a service life of more than 10 years. The aim is to absorb, as a matter of routine, the sharp swings in power demand that come with computing workloads. Achieving that will require cooling and power control that can sustain the battery's high-output performance.
AI power demand is a problem of both scale and fluctuation
In large-scale AI training, many GPUs repeatedly synchronize computation and data exchange. According to NVIDIA's technical explanation, this behavior produces sudden load swings across an entire facility. Unlike a group of servers running separate jobs, the machines draw power all at once and cut usage all at once. Beyond average power consumption, how quickly demand rises and falls shapes the design of the power supply.
An energy storage system fills this timing gap. When compute load surges, the batteries discharge to cover the shortfall; when load drops, power flows back into charging. If the batteries move power in and out in step with demand inside the facility, the fluctuations seen from the external grid can be softened. This is load smoothing.
What DIMAAG and Toshiba are combining is a system that spans everything from battery cells to control. Toshiba supplies high-output SCiB cells, while DIMAAG designs and manufactures the battery system and integrates it into ZettaWatt. The announcement of the collaboration lists demand adjustment and grid stabilization as applications, in addition to backup power.
The more jobs a battery takes on, the harder it becomes to judge performance by stored energy alone. What matters is how much power it can deliver or absorb at a given moment, and how many times it can repeat that.
SCiB and cooling for repeated 10C operation
SCiB is a lithium-ion battery that uses lithium titanate in its negative electrode. Replacing the carbon-based anode material with LTO gives it high input/output performance and long life. This collaboration does not involve inventing a new battery material; it is better described as applying the characteristics of a battery already used in industrial settings to AI infrastructure.
Toshiba's SCiB overview cites features such as no metallic lithium deposition even when charging at low temperatures and little degradation even with repeated rapid charging. It also indicates a cycle life of more than 20,000 cycles, but characteristics vary with cell type and operating conditions. A general description of a cell cannot simply be read as a guaranteed value for this system.
The "C" in 10C expresses the magnitude of charge/discharge current relative to battery capacity. In an idealized calculation that holds a 10C current constant, the time to move an amount of charge equal to the rated capacity is one-tenth of an hour, or six minutes. Actual full-charge time also depends on control conditions, so this conversion alone does not determine it. The time a facility can be supported during an outage and how fast the system reacts to load changes are separate metrics.
In its explanation of the Zenius module, DIMAAG also positions continuous 10C charging and discharging for applications with frequent operating cycles, such as construction machinery and ports. For the AI collaboration, it will use Zenius modules, which employ immersion cooling, a method that circulates coolant to cool the batteries. Being able to pass a large current once is different from being able to pass it repeatedly while managing the heat, and the two call for different designs.
This is why the companies make a "properly designed cooling system" a condition. The module's cooling design is what links the electrical capability of 10C to the targeted lifespan.
What 800V and energy storage each solve
ZettaWatt uses a modular power architecture that connects grid power to 800V DC. It combines power conversion equipment, energy storage and real-time control, and is designed so that multiple units can be added in parallel as a data center expands.
The move to 800V is meant to deliver large amounts of power efficiently. For the same power, a higher voltage reduces the current required and makes it easier to cut resistive losses in wiring. NVIDIA also presents a concept of connecting from the facility to the rack with DC to reduce redundant power conversion.
Raising the voltage, however, does not remove the rises and falls in GPU workload. NVIDIA presents a design that combines 800V delivery with energy storage and divides roles by the timescale of the fluctuation. The idea is to use capacitors and similar components near the racks for fluctuations from milliseconds to seconds, and facility-side batteries for changes from seconds to minutes. It combines a mechanism for carrying large amounts of power with one for absorbing timing differences in demand.
This explanation describes the design background of AI power systems, not the internal configuration of ZettaWatt. The collaboration announcement does not say which component handles each timescale. Still, it explains why the SCiB cells' high-output performance alone does not determine the stability of the whole facility, and why a combination of power conversion equipment and control is needed.
How to read the figures in the announcement
The companies explain that ZettaWatt holds cyclical AI load fluctuations to under 1% as seen from the grid side. However, the announcement does not give the base against which the percentage is measured or the input waveform used in testing. To compare performance, one needs to know how large the fluctuation was and over what time span it was suppressed.
10C is a charge/discharge rate, the sub-1% figure is a company explanation about grid-side fluctuation, and 10+ years is a lifespan target. The announcement contains no test conditions for the completed system and no deployment information.
| Figure in the announcement | What it represents | Conditions and caveats needed to judge it |
|---|---|---|
| Continuous 10C charge/discharge | Magnitude of charge/discharge current relative to capacity | Assumes a properly designed cooling system. The announcement gives no details such as temperature conditions |
| Under 1% on the grid side | The companies' explanation that cyclical AI load fluctuations are smoothed | The base, input waveform and time span evaluated are not given |
| 10+ years | Targeted service life | Depth of charge/discharge and the capacity retention that defines end of life are not disclosed |
| Over 100 million units since 2010 | Shipment record for SCiB cells overall | Not a shipment count or operating record for this AI system |
Source: TIC's collaboration announcement dated September 9, 2026. Items marked as not disclosed are those that cannot be confirmed in the body of that announcement.
Separating what each number means shows that a cell's mass-production record does not substitute for long-term validation of a completed system. In particular, the sub-1% figure does not mean power consumption is cut by 99%. A battery stores and releases power; it does not eliminate the energy that computing requires.
The announcement also says the system complies with ERCOT's low voltage ride-through requirements. This relates to the ability to keep operating when grid voltage temporarily drops, but no third-party certification results are provided. The end of the announcement also states that the capabilities, performance targets and applications described depend on future design, testing and validation. At this stage, the figures should be read as company explanations that come with these caveats.
Balancing smoothing with reserve power in real facilities
NVIDIA's explanation of designing battery energy storage systems points out that absorbing load fluctuations, holding reserve power for grid disturbances and participating in demand response can compete for the same battery headroom. It says the battery's state of charge must be managed and priorities among uses decided.
If the battery is drained by everyday load smoothing, there will not be enough power to draw on in an emergency. Conversely, if there is no room to charge, it is hard to absorb power when load drops. Beyond choosing cells that can withstand frequent charging and discharging, designers must decide what state of charge to keep the batteries in and when to use them for what.
The announcement does not give a commercial launch date, customers, or the system's power output and storage capacity. Price and power losses, including for cooling, are also undisclosed. As a result, comparisons such as cheaper than other approaches, smaller footprint or higher facility-wide efficiency cannot yet be made.
Whether SCiB's high-output performance can be turned into a countermeasure for AI power swings depends on integration testing with real loads. If smoothing performance can be maintained with cooling running and with provision for outages still in place, operators will find it easier to balance everyday power control with securing emergency power.