Infinity and HP Double AI Inference Performance on NVIDIA GB300-Powered HP ZGX Fury
Infinity, also known as the Infinity Artificial Intelligence Institute, has announced a collaboration with HP Inc. to improve artificial intelligence (AI) inference performance on HP’s ZGX Fury workstation powered by the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip. The collaboration uses Infinity’s automated AI research technology to optimize the NVIDIA Nemotron 3 Super model, demonstrating the potential to increase AI processing capacity without upgrading the underlying hardware.
During testing, the companies achieved approximately twice the concurrent workload capacity and doubled overall token throughput on the evaluated GB300 configuration while maintaining a defined responsiveness threshold. The initial performance improvement was achieved in less than 24 hours of engineering work, with further throughput optimization continuing after the initial results.
The announcement highlights the growing importance of software optimization as enterprises expand their use of AI applications. Rather than relying exclusively on additional computing hardware to meet increasing demand, businesses can improve the performance of existing infrastructure by optimizing how AI models use available processing resources.
Infinity develops software designed to make AI chips ready for inference workloads, while HP provides computing systems intended to support demanding professional and enterprise applications. Their collaboration demonstrates how software-level improvements can help organizations serve more AI requests using the infrastructure they already have.
Improving AI Performance Through Software Optimization
AI inference refers to the process of running a trained AI model to generate responses, predictions, or other outputs. As organizations deploy AI assistants, enterprise applications, and autonomous agents, inference performance has become an increasingly important consideration.
Businesses must balance processing capacity, response times, operating costs, and the number of users who can access AI systems simultaneously. When inference software cannot efficiently utilize available hardware, organizations may experience reduced throughput or limitations on concurrent workloads.
Infinity’s optimization technology aims to address these challenges by improving the software components responsible for executing AI models on computing hardware.
The company’s work with HP focused on optimizing the NVIDIA Nemotron 3 Super model running on the HP ZGX Fury system. According to the companies, the optimized configuration supported approximately twice as many concurrent requests while maintaining the specified responsiveness requirements.
This improvement increases the amount of AI work that can be processed on the same hardware configuration, potentially helping organizations improve resource utilization and accommodate additional users without immediately investing in more computing capacity.
Jeremy Nixon, founder and chief executive officer of Infinity, emphasized the importance of extracting greater performance from existing AI infrastructure.
“The next frontier in AI is about extracting the greatest performance from every chip in use, and HP is moving swiftly to ensure their customers lead the way,” Nixon said.
The collaboration reflects Infinity’s approach of using automated research and optimization to identify performance improvements that conventional software development processes may take considerably longer to discover.
HP ZGX Fury Supports Shared AI Computing
The collaboration centers on HP ZGX Fury, HP’s AI workstation based on the NVIDIA DGX Station platform and powered by the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip.
The system is designed to provide substantial AI computing capabilities in an office-deployable, rack-mountable configuration. Its architecture supports shared computing resources, enabling multiple users to run inference workloads simultaneously or development teams to experiment with models and conduct fine-tuning activities in parallel.
This capability can be valuable for organizations that need access to advanced AI computing without depending entirely on remote cloud infrastructure.
For enterprise developers and research teams, shared AI workstations can support model development, application testing, experimentation, and deployment preparation. However, maximizing the value of these systems depends on efficiently managing computing resources and ensuring that individual workloads do not unnecessarily limit overall system capacity.
Jim Nottingham, senior vice president and division president of Advanced Compute Solutions at HP Inc., said the company is focused on helping customers deploy AI at scale through solutions that combine powerful local computing with practical productivity improvements.
According to Nottingham, the collaboration with Infinity demonstrates how HP ZGX Fury can improve efficiency, accommodate more concurrent AI workloads, and help organizations maximize returns on their AI investments.
By combining HP’s computing platform with Infinity’s optimization technology, the companies aim to help users obtain greater performance from advanced AI hardware.
Automated Optimization Across the Inference Stack
Infinity’s engineering team developed optimizations across multiple components of the AI inference stack. The work incorporated techniques associated with established inference frameworks, including vLLM, NVIDIA TensorRT-LLM, and SGLang, alongside Infinity’s proprietary optimization methods.
The objective was to increase the number of concurrent requests the system could handle without compromising the defined responsiveness requirements for individual users.
Traditional inference optimization can require substantial engineering effort because developers must understand hardware architecture, model characteristics, memory management, and the software operations involved in executing AI workloads efficiently.
Infinity seeks to automate significant portions of this process through its AI research technology. Its system can generate, test, and optimize low-level computing components that influence how efficiently AI models operate on different chips.
The company says its automated approach can shorten inference software development timelines that have historically required months or years of specialized engineering work.
The HP collaboration builds on Infinity’s previously published research involving the Qwen3-8B model. In that work, the company’s optimization system generated an inference engine from scratch and delivered up to 34.3% more tokens per second than vLLM on identical hardware and under the reported test conditions.
Infinity has also published a case study involving chip partner d-Matrix. According to the company, it achieved 92% of a new chip’s theoretical peak performance within 10 hours of first accessing the hardware.
These results illustrate Infinity’s broader objective of making AI inference software more adaptable to different hardware platforms.
Growing Demand for AI Inference Infrastructure
The collaboration comes as organizations increase investment in AI infrastructure and deploy more applications that require continuous model execution.
According to a Gartner forecast cited by Infinity, worldwide spending on AI-optimized cloud infrastructure was projected to increase by 96% in 2026, reaching $42 billion.
The cited forecast also anticipated that global spending on AI inference would exceed spending on AI training for the first time in 2026, with projected inference spending of $23.3 billion compared with $19 billion for training.
This shift reflects the growing importance of operating AI models after their initial training. As businesses introduce AI assistants and autonomous systems into everyday workflows, they need infrastructure capable of handling repeated requests and complex sequences of operations.
Agentic AI can intensify these requirements because autonomous systems may perform multiple steps, retrieve information, invoke tools, and execute actions before completing a task.
For infrastructure providers and enterprise customers, improving inference efficiency can therefore become an important part of controlling costs and expanding AI deployment.
Software optimization offers one potential way to increase capacity without relying exclusively on additional hardware purchases.
Infinity’s Automated AI Research Technology
Infinity’s technology is built around an AI research agent called Ignition, which autonomously writes, tests, and optimizes low-level computing kernels used to execute AI models.
These kernels determine how efficiently computing hardware performs the operations required by AI workloads. By evaluating performance and repeatedly testing potential improvements, Ignition aims to generate optimized inference software for different chip architectures.
Infinity argues that hardware performance alone does not determine how effectively a chip can run modern AI models. Software ecosystems, optimized kernels, and supporting infrastructure also play important roles in real-world performance.
The company believes automation can help chip developers address software optimization challenges more quickly, potentially reducing the time required to make new processors suitable for production AI inference.
Founded in August 2025 and headquartered in San Francisco, Infinity is a privately held company backed by Touring Capital and angel investors, including technology executives and researchers associated with major AI organizations.
The company is also working with d-Matrix through the Infinity d-Matrix Cloud and has reported discussions about potential partnerships with other chip companies.
Testing Conditions and Performance Considerations
The reported HP results were based on testing conducted by Infinity and HP using the NVIDIA Nemotron 3 Super model, described in the company’s disclosure as a 120.6-billion-parameter workload, on the GB300 platform.
The evaluation used HP’s existing vLLM benchmark configuration and compared performance against a baseline of 82 concurrent users while maintaining specified throughput requirements.
Validation included NVIDIA SPEED-Bench, MMLU, and WikiText quality benchmarks.
The companies noted that actual results may vary depending on the AI model, workload, hardware configuration, software version, and other operating conditions. Infinity’s previously published performance comparisons are likewise subject to the conditions described in the relevant research.
Advancing Efficient Enterprise AI Deployment
The collaboration between Infinity and HP demonstrates how automated software optimization can help improve the performance of advanced AI computing systems.
By doubling concurrent workload capacity and overall token throughput in the tested configuration, the companies have shown the potential to increase AI inference capacity without changing the underlying hardware.
As enterprise adoption of AI continues to expand, the ability to maximize existing computing resources may become increasingly important for businesses seeking to balance performance, infrastructure costs, and growing demand.
Infinity and HP are positioning their combined technologies to help organizations make more effective use of AI workstations and accelerate the deployment of AI-powered applications.
Source Link: https://www.businesswire.com/