In a sponsored content piece by HPE and NVIDIA, the two companies present the concept of the “AI factory”, an integrated infrastructure designed to repeatedly and efficiently transform data into intelligence. They propose three offerings, from turnkey solutions to sovereign architectures, and stress that the choice must start from use cases rather than hardware.
An infrastructure designed from use cases
An AI factory encompasses the entire AI lifecycle: data ingestion, model development, training, adaptation (fine-tuning), inference, meaning execution, as well as monitoring and continuous improvement. The text specifies that this integration should serve a doctor, an engineer, or a public service as well as a business. Choosing an accelerated processor is not enough: the compute must match the business model, the network and data flows must feed the accelerators, storage must keep up with data volume, and security, governance, and infrastructure sharing among multiple users must be planned upfront. Power supply and cooling must finally match the system’s density, today and tomorrow. Thierry Pienaar, HPE Fellow, Vice President and Chief Technology Officer (CTO) for HPC and AI Sales, states: “Without the right architecture, governance, and operational expertise, organizations cannot securely harness their data, integrate AI into core processes, or turn innovation into a sustainable competitive advantage.”
Three offerings based on needs
The HPE AI Factory with NVIDIA portfolio combines NVIDIA’s accelerators, networking, and AI software with HPE’s infrastructure, software, and services. The first offering, HPE Private Cloud AI, is a turnkey, on-premises solution designed for fine-tuning, RAG (retrieval-augmented generation, which connects a model to documents), and inference, with a need that can reach 256 GPUs. The second, HPE AI Factory at-scale, targets model creators, service providers, and large enterprises, with resources ranging from a few hundred to tens of thousands of GPUs, centralized control, and multi-tenant operation, where multiple users share the same infrastructure. The third, HPE Sovereign AI Factory, builds on this architecture by adding enhanced operational control, security and data residency, as well as “air-gapped” configurations, isolated from any external network.
Operating the whole as a single system
As AI expands within an organization, teams have different needs for resources, service levels, and data perimeters. According to the text, an artificial production deployment, built from separate components, requires skills that many IT departments have never had at this scale. A poorly designed system can work technically while being unprofitable: underutilized GPUs, bottlenecks in data flows, power or cooling constraints, or management systems that are difficult to operate. “The HPE AI Factory with NVIDIA portfolio provides enterprises with a range of AI solutions co-developed with NVIDIA, backed by HPE’s engineering expertise and technical capabilities, to design an AI factory suited to their needs and optimize it for performance at scale,” explains Thierry Pienaar.
Sovereignty and support
Cloud, private environments, and hybrid approaches can all play a role in an AI strategy. For organizations subject to sovereignty requirements, the choice depends mainly on the desired level of control: location for storing data and models, identity of administrators, applicable jurisdiction, governance rules, and degree of workload isolation. Kaushik Shirhatti, Vice President, AI Factory at NVIDIA, states: “HPE and NVIDIA’s sovereign AI tools give a company, or even a nation, complete control over how its AI systems are built, deployed, operated, and governed.” The text cites two recent deployments: the TELUS Sovereign AI Factory in Canada and a sovereign AI factory at the University of Utah in the United States. HPE AI Services supports customers from planning to operation, while HPE Financial Services offers, among other things, accelerated depreciation. The two partners refer to the HPE AI Factory page for more information.
