DeepAware AI Thinks Your Data Center Is Wasting 30% of Its Energy and They Have the Receipts

The Macro: The AI Boom Has an Energy Problem Nobody Wants to Talk About

I have been watching the data center energy conversation for the past year and it keeps getting worse. The International Energy Agency estimates that global data center electricity consumption could double by 2026, driven almost entirely by AI workloads. GPU clusters are power-hungry in a way that traditional compute never was. A single NVIDIA H100 draws 700 watts. Racks of them together can pull 100kW or more. Multiply that across a facility running thousands of GPUs and you get power bills that would make a utility executive blush.

The dirty secret of the industry is that most of this power is being used poorly. Studies from the Uptime Institute consistently show that average data center PUE (Power Usage Effectiveness) hovers around 1.58, meaning 37% of energy goes to overhead rather than compute. For GPU-heavy facilities the problem is even worse because GPU utilization rates are notoriously low. Run a training job and your GPUs might hit 80% utilization. Run inference at variable demand and you might see 30% utilization with 100% power draw.

The existing DCIM (Data Center Infrastructure Management) market is dominated by Schneider Electric’s EcoStruxure, Vertiv’s Trellis, and Nlyte. These products were built for traditional server environments. They monitor temperature, humidity, and power at the rack level. They are not designed to optimize GPU workload placement in real time based on energy costs and performance targets. The gap between what these legacy tools offer and what modern GPU data centers need is enormous.

The numbers make the case obvious. A 30MW data center spending $2 million per month on electricity that wastes 30% of it is lighting $600,000 on fire every month. Any product that can recover even a fraction of that waste has a straightforward ROI argument.

The Micro: Reinforcement Learning Meets the Server Room

DeepAware AI builds an automation layer for GPU data centers that optimizes workload placement using reinforcement learning. The system watches power, performance, and cost signals in real time and makes decisions about where to schedule GPU workloads to minimize energy waste.

Jerry Huang and Stela Tong are the founders. They are based in San Francisco at 989 Market Street and went through Y Combinator’s Summer 2025 batch. The company was also selected for the NVIDIA Inception Program, which provides access to NVIDIA’s technical resources and GPU credits.

The product has four main components. First is the RL scheduler, which optimizes GPU workload placement across power, performance, and cost dimensions. Rather than using static rules or manual scheduling, it learns the facility’s patterns and adapts. Second is real-time energy market integration, which shifts workloads to low-price windows and supports demand-response programs. If your utility offers lower rates at 2 AM, DeepAware automatically moves flexible workloads to take advantage of that pricing. Third is a unified dashboard for policy tuning, alerts, and scenario testing. Fourth is a robotics component for remote inspections and maintenance, which is still in development.

The results they are claiming are significant. Up to 30% energy waste reduction overall, with 15% energy savings demonstrated in simulation against a major 30MW-plus data center operator. They have a six-figure agreement with that operator, which suggests the simulated results were convincing enough to get a real contract.

The competitive landscape here is interesting. The legacy DCIM vendors like Schneider and Nlyte are not doing real-time AI optimization. Crusoe Energy takes a different approach by building their own data centers near stranded energy sources. Lancium does something similar with bitcoin mining and AI workloads. But DeepAware is not trying to build data centers. They are selling software to existing operators. That is a much lighter business model with faster deployment cycles.

The comparison I keep coming back to is what Samsara did for fleet management. Before Samsara, trucking companies had fragmented visibility into their operations. Samsara gave them a unified platform with real-time data and actionable insights. DeepAware is trying to do the same thing for data center operators who are currently managing GPU infrastructure with spreadsheets and gut instinct.

The Verdict

The timing is almost too perfect. Every major cloud provider is racing to build GPU capacity. Hyperscalers, colocators, and enterprise operators are all scrambling for power and cooling. Into that frenzy walks a product that says it can cut your energy waste by a third using software alone. The ROI pitch practically writes itself.

My concern is the sales cycle. Data center operators are conservative. They do not install new software on production infrastructure because a startup showed them simulation results. The path from pilot to production deployment in this market typically takes 12 to 18 months. DeepAware needs to survive that cycle with a small team and limited runway.

In thirty days, I want to know if the six-figure deal has moved from simulation to production deployment. Sixty days, the question is whether the 15% savings number holds up in a live environment with real workloads and real variability. Ninety days, I want to see a second customer. One contract is a proof of concept. Two contracts is the beginning of a business. The product addresses a real and growing problem. The RL approach is technically sound. If they can close the gap between simulation and production, this could be a very large company.