In the high-stakes race for AI compute, a fundamental bottleneck isn't just the chips—it's the electricity that powers them. Data centers, straining under the demands of training massive models, are often forced to throttle their expensive GPU clusters by up to 30% to avoid overloading the local power grid. This unused capacity represents a massive financial drain.
Enter Niv-AI. The Tel Aviv-based startup, founded in 2025 by Tomer Timor and Edward Kizis, emerged today with $12 million in seed funding to address this precise issue. The company is developing a system to measure and manage GPU power consumption with unprecedented, millisecond-level precision.
The core problem is one of erratic demand. When thousands of GPUs work in concert, sudden power spikes occur as they switch between computation and communication. These surges force data centers to either invest in costly temporary energy storage or, more commonly, cap their GPU usage. Both solutions cut into the return on multi-billion-dollar hardware investments.
Niv-AI's first move is deep measurement. The company is deploying custom sensors in partner data centers to build detailed power profiles for different AI workloads. The subsequent goal is to build predictive software—an operational layer that can forecast and synchronize power loads across a facility, effectively smoothing out the demand sent back to the electrical grid.
"The grid is apprehensive about data centers consuming too much power at once," Timor explained. "We're addressing both sides: helping data centers use more of their installed GPUs and creating a more stable, responsible interface with the power infrastructure."
With backing from Glilot Capital and Grove Ventures, among others, Niv-AI plans to have initial systems running in U.S. data centers within eight months. As expansion of new data centers faces permitting and supply chain delays, the promise of unlocking more capacity from existing facilities is a compelling proposition for an industry hungry for every watt.
Source: TechCrunch