Data Centers Can't Ignore Power Anymore

AI infrastructure is guzzling electricity at an unsustainable rate, and data center operators are hitting a wall. Nvidia CEO Jensen Huang put it bluntly at GTC 2026: "Every unused watt is revenue lost." Now, a new startup is tackling the biggest drain on those expensive GPUs.

Niv-AI, a Tel Aviv-based startup, emerged from stealth this week with $12 million in seed funding to solve a deceptively simple problem: GPU power surges.

The Power Problem

When training frontier AI models, GPUs constantly switch between computation and inter-GPU communication. These millisecond-scale power demand spikes force data centers into a tough choice: pay for temporary energy storage to handle the surges, or throttle GPU usage entirely. Both cost money. A lot of it.

Large frontier labs operating thousands of GPUs in concert can lose up to 30% of their throughput just managing power constraints. When you're running million-dollar monthly GPU bills, leaving 30% of performance on the table is catastrophic.

Niv's Approach: Measure First, Optimize Second

The startup's first move is data collection. Niv is deploying rack-level sensors that detect power usage at millisecond granularity across different deep learning workloads and GPU types. Once they understand the specific power profiles of training runs, code generation, and inference tasks, they can develop software to smooth out the demand spikes.

The long-term play is obvious: build an AI model trained on real power data to predict surges and automatically adjust workloads to prevent throttling. This isn't revolutionary—it's infrastructure optimization at scale. But it's also essential.

Backed by Glilot Capital, Grove Ventures, and Arc VC, Niv is betting that every data center operator running GPUs will soon have power management tools on their stack. Given the cost pressures and the frantic build-out of AI infrastructure, they're probably right.

Source: TechCrunch — Niv-AI exits stealth