Niv-AI Raises $12M to Address the Hidden Power Bottleneck in AI Infrastructure

A new startup is entering the increasingly crowded AI infrastructure space with a focus on a constraint that rarely makes headlines but is quickly becoming one of the industry’s most pressing challenges: power.

Niv-AI has emerged from stealth with $12 million in funding from Glilot Capital, Grove Ventures, Arc VC, Encoded VC, Leap Forward, and Aurora Capital Partners. The Tel Aviv–based company is positioning itself at the intersection of energy systems and high-performance computing, targeting what it describes as an “instantaneous power capacity” crisis inside modern data centers.

The Problem: AI’s Growing Appetite for Power

As AI workloads scale, particularly with the adoption of increasingly power-dense GPUs, data centers are encountering a physical limitation that software optimization alone cannot solve. While compute capacity continues to grow, the ability to deliver stable power at the millisecond level is lagging behind.

Traditional monitoring systems were not designed for the rapid, spiky power consumption patterns of modern AI workloads. To avoid equipment damage or grid instability, operators often overcompensate by limiting usage. The result is significant underutilization of existing infrastructure, with as much as 30% of contracted power capacity effectively sitting idle.

This inefficiency hasect financial implications. Data center operators are paying for capacity they cannot fully use, while AI companies face constraints that slow deployment and increase costs.

A New Layer Between Power and Compute

Niv-AI’s approach introduces a new control layer that sits between energy delivery and computational workloads. At the core of its platform is what the company calls an “electrical fingerprint” — a high-resolution view of how AI workloads consume power in real time.

Using specialized sensors, the system captures detailed power signals that traditional meters miss. These signals are then processed by AI models designed to predict short-term fluctuations in demand. Instead of reacting after a spike occurs, the platform proactively adjusts workload timing, subtly staggering compute operations to smooth out power usage.

In practice, this functions like a traffic management system for electricity inside the data center, allowing operators to push infrastructure closer to its true limits without triggering instability.

Moving Beyond Hardware Fixes

Most current attempts to address power constraints rely on physical solutions such as batteries, capacitors, or conservative throttling of workloads. While effective to a degree, these approaches add cost, complexity, or reduce performance.

Niv-AI is betting that a software-driven orchestration layer can unlock similar or greater gains without requiring additional hardware. By improving visibility and control at a granular level, the company aims to enable operators to extract more value from existing infrastructure.

This shift mirrors broader trends in data center optimization, where software-defined approaches are increasingly used to manage physical constraints.

The Broader Implications for AI Infrastructure

If this category of technology proves effective, it could reshape how data centers are designed and operated over the next decade. Rather than treating power limits as fixed constraints, operators may begin to view them as dynamic variables that can be actively managed in real time.

This has implications beyond efficiency. It could delay or reduce the need for costly grid upgrades and new facility construction, particularly in regions where energy availability is already a bottleneck. It may also influence how AI workloads are scheduled, priced, and prioritized, introducing a new dimension of optimization that blends compute orchestration with energy management.

At a systems level, the convergence of power and compute control suggests a future where infrastructure is increasingly coordinated across layers that were historically siloed. As AI continues to scale, the ability to harmonize these layers could become as important as advances in model architecture or chip design.