- +1-408-508-6255
- 830 Hillview Court, Suite 100 Milpitas, CA 95035
- info@prophetstor.com
Milpitas, CA, Feb. 26, 2026 — ProphetStor Data Services, Inc. today announced that the United States Patent and Trademark Office (USPTO) has granted a U.S. patent (Application No. 18/238,569), titled “Method and System for Optimizing GPU Utilization.” The ProphetStor patent covers a predictive approach to GPU resource management that forecasts future GPU demand using correlation analysis and time-series modeling, instead of waiting for usage spikes to happen and then scrambling to respond.
The problem with reactive GPU management
Most GPU resource management today is reactive. Usage goes up, the system notices, and then it scales. The delay costs money: GPUs sit idle when they shouldn’t, or capacity arrives too late to prevent slowdowns. Unlike conventional auto-scalers that respond after the fact, ProphetStor’s patented technology works the other way around. It watches workload behavior across multiple nodes, identifies which resource metrics are most closely tied to GPU consumption, and uses time-series models to predict what’s coming next. GPU capacity gets adjusted before demand actually shifts.
“If you’ve run GPU workloads at any real scale, you know how fast things move. A training job spins up, inference traffic doubles overnight, and suddenly you’re either overpaying for idle GPUs or scrambling to allocate more,” said Eric Chen, CEO of ProphetStor. “What this patent covers is a way to get ahead of that. We use correlation analysis and time-series forecasting to predict demand and allocate accordingly, so GPU clusters stay utilized without the constant fire drills.
How ProphetStor’s GPU workload forecasting works
The patent, invented by Eric Chen, CEO of ProphetStor, and filed on August 28, 2023, covers 19 allowed claims for both method and system implementations. The technology works in several stages: it deploys across multiple nodes, collects workload and resource usage data over time, calculates correlation values between application behavior and GPU consumption, then applies time-series models to forecast demand at future time points. Resources that show high correlation above a set threshold get prioritized, and GPU capacity is allocated or released based on the predicted usage increments. The same predictive framework extends beyond IT workloads to operational technology (OT) resources such as liquid cooling. Through ProphetStor’s Federator.ai platform, organizations can manage both GPU compute allocation and cooling flow rates from a single predictive system. The whole system runs on VM and Kubernetes platforms, managing multiple GPU clusters.
Why predictive GPU management matters more with next-generation hardware
The need for predictive optimization becomes even more pressing with the latest generation of data center GPUs. Systems built on NVIDIA GB200, GB300, and the upcoming Vera Rubin architecture pack significantly more compute power per node, but they also generate far more heat. Liquid cooling is no longer optional for these GPUs — it is mandatory.
That creates a new operational problem. Dynamic liquid cooling is more energy-efficient than traditional air cooling, but adjusting coolant flow rates takes minutes to take effect. Meanwhile, GPU workloads can shift in milliseconds. A reactive system that waits to observe a workload spike before adjusting cooling is, by definition, too slow. By the time the coolant flow catches up, the hardware has already been running outside its optimal thermal envelope.
This is exactly the kind of gap that predictive workload forecasting is built to close. Because ProphetStor’s patented technology predicts GPU demand before it arrives, cooling infrastructure can be adjusted in advance, matching flow rates to anticipated workloads rather than chasing them after the fact. As liquid-cooled GPU data centers become the industry standard, the ability to forecast workload behavior ahead of time is no longer just a cost optimization — it is a thermal management requirement.
This patent is one of several that the USPTO has granted to ProphetStor. Together, they comprehensively cover a temporal, spatial, and holistic view of both IT resources (GPU compute, storage, networking) and OT resources (liquid cooling, power distribution, airflow) within AI-defined data centers (ADDC). Where most optimization tools focus on a single layer, ProphetStor’s patent portfolio spans the full stack — predicting how workloads will behave over time, where resources are needed across the physical infrastructure, and how IT and OT systems interact with each other. This is the technology foundation behind ProphetStor’s vision for the next generation of AI factories: data centers where compute, cooling, and power are managed as a single predictive system rather than in separate silos.
Organizations interested in ProphetStor’s predictive GPU optimization technology can contact the company through prophetstor.com.
ProphetStor Data Services, Inc. is based in Milpitas, California. The company builds AI-native solutions for operating and optimizing AI factories.
For more information, visit www.prophetstor.com .
ProphetStor’s U.S. patent (Application No. 18/238,569) covers a method and system for optimizing GPU utilization using correlation-based analysis and time-series forecasting to predict GPU demand and dynamically allocate resources in advance.
Reactive scaling adjusts GPU capacity after usage changes are detected, causing delays and wasted resources. Predictive GPU resource management forecasts demand using time-series models and historical workload data, adjusting capacity before demand shifts.
Yes. The patented system runs on VM and Kubernetes platforms and is designed to manage multiple GPU clusters in cloud-native and multi-tenant AI environments.
GPU workloads shift in milliseconds, but adjusting liquid coolant flow rates takes minutes. Reactive cooling systems cannot keep up. Predictive workload forecasting allows cooling adjustments to be made in advance, keeping hardware within its optimal thermal envelope.
Most solutions optimize a single layer — either GPU compute or cooling or power — in isolation, and they react to changes after they happen. ProphetStor’s patented technology takes a holistic approach, covering both IT resources (GPU, storage, networking) and OT resources (liquid cooling, power distribution, airflow) within a single predictive system. It uses time-series forecasting and correlation analysis to anticipate demand across all these layers simultaneously, giving AI-defined data centers (ADDC) a unified, forward-looking view instead of siloed, backward-looking dashboards.
An AI-defined data center (ADDC) is a data center where compute, cooling, and power are managed as a single predictive system using AI and machine learning, rather than in separate silos with manual or reactive controls. ProphetStor’s Federator.ai platform is designed to operate and optimize these next-generation AI factories.
Federator.ai is ProphetStor’s platform for predictive resource management. It manages both IT workloads (GPU compute allocation) and OT resources (such as liquid cooling flow rates) from a single system, using the patented time-series forecasting and correlation analysis technology.
Please select the software you would like a demo of: