Prior to the AI & Big Data Expo held at the San Jose McEnery Convention Center from May 18–19, we sat down with Jerome Gabryszewski, HP’s AI & Data Science Business Development Manager, to discuss AI, strategies for making data suitable for AI training, and the choice between local and cloud computing.
Technology outlets often state that data represents “the new oil.” However, the practical reality is that even when organizations have access to abundant internal information, using it effectively for commercial gains remains challenging, particularly for large-scale enterprises.
Should businesses opt for cloud-hosted AI models or local hardware? How can a company organize its data infrastructure so that powerful algorithms can deliver actionable insights? As always, we encourage our guests to share their thoughts on the upcoming trends within the rapidly changing world of enterprise IT in the current AI-driven climate.
Artificial Intelligence News: Manually feeding data into systems was once the standard, and switching to automation is the goal. However, the process is known to be incredibly complex. Where are HP’s clients currently facing the biggest hurdles?
The most persistent issue we observe is that many organizations misjudge the organizational overhead and technical barriers surrounding their data. Before any automation is possible, they must address scattered data control across various teams, mismatched data structures within their tools, and older systems that were not built to work together. Implementing automation is generally less difficult than the necessary preliminary work of establishing governance protocols and system integration.
Artificial Intelligence News: When AI systems begin self-learning on their own, problems often arise. What is your strategy for helping clients manage risks such as data corruption or evolving data patterns?
An AI system that perpetually learns can shift from being an organizational asset to a full-blown liability if not properly supervised. Our guidance to clients is to manage model revisions just as rigorously as they handle updates to their computer software. No changes should be made to the live environment without passing a validation checkpoint. Addressing shifting data trends involves implementation of automated monitoring through MLOps systems, with human approval required before a model is retrained. Data corruption is equally a matter of security as it is of tracking origins; it is essential to understand exactly what information was used to train the models and who has permission to access it. The companies that succeed in this area aren’t necessarily those with the most advanced engineering, but those that weave AI governance into their core risk management protocols before expanding their operations.
Artificial Intelligence News: Let’s talk about the physical side of HP, specifically the hardware. What are the current requirements for a workstation or processing setup to cope with the heavy demands of a self-sufficient AI system?
HP’s background in the hardware sector is highly significant here. The Z series has been engineered specifically for intensive professional workloads for over a decade and a half; therefore, when outlining the requirements for a self-sufficient AI system, we are not speculating. We have been refining these solutions for longer than most of the competition.
There isn’t a single “magic box” to solve this; it is rather a spectrum of options. For individual engineers, you need local hardware powerful enough to conduct real-world tests without relying on the cloud for every step. The ZBook Ultra and Z2 Mini occupy the mobile and small-form-factor professional tiers, running local LLMs and intensive processes without issue.
The ZGX Nano is exceptionally promising for teams focused primarily on AI. It serves as a handheld AI supercomputer (measuring 15x15cm), equipped with the NVIDIA GB10 Grace Blackwell Superchip that offers 128GB of unified memory and 1,000 TOPS of FP4 AI performance. By itself, it handles models with up to 200 billion parameters locally. For models exceeding that size, you connect two units via a high-speed link to process up to 405 billion parameters utilizing zero cloud resources or data center assistance. It arrives ready for use out of the box with the NVIDIA DGX software stack and HP ZGX Toolkit, allowing teams to go from unboxing to their initial tasks in minutes instead of days.
At higher levels, the Z8 Fury provides power users with support for up to four NVIDIA RTX PRO 6000 Blackwell GPUs (384GB VRAM) within one unit, allowing for complete model development locally. On the bleeding edge, the ZGX Fury completely transforms what is possible. Featuring the NVIDIA GB300 Grace Blackwell Ultra Superchip and 748GB of coherent memory, it enables processing of massive trillion-parameter models directly from a desk rather than a data center. For teams constantly refining models and handling sensitive data, this hardware often recovers its initial cost in 8 to 12 months compared to equal cloud performance.
For those requiring further expansion, the entire Z series is designed for seamless rack-mounting into existing IT setups while maintaining strict data security.
The main takeaway here is that a self-sufficient AI workflow presents challenges in management and speed, not pure processing power. It is impractical to send sensitive data to the cloud each time a model needs an update. HP’s products give companies a progression that matches their work maturity, ranging from an individual’s desk to distributed local processing. The hardware is now finally capable of keeping up with the high ambitions of modern AI capabilities.
Artificial Intelligence News: Generative AI expenses are skyrocketing for many large companies. Is there a realistic solution to balance this massive expenditure with the efficiencies offered by the modern cloud?
The issue with costs is permanent, not temporary. Enterprise spending on Generative AI hit $37 billion during 2025, yet 80% of companies still failed to stay within 25% of their estimated budgets. The core issue is that while the cost for individual AI actions is dropping, total investment is expanding because adoption is outpacing the cost reductions. The current cloud API system was built primarily for small-scale experiments, not as the financial backbone of full-scale production AI.
The realistic fix requires discipline before infrastructure changes: establish a distinct boundary between testing projects and production outputs, and strictly avoid utilizing the same hardware setups for both. Preliminary work—such as building prototypes, fine-tuning, and testing algorithms—should be done locally on systems like the ZGX Nano or Z8 Fury. This way, you invest in equipment once, rather than wasting budget on tests without a clear business case. Companies that succeed here typically employ a three-tier approach: the cloud for unexpected peaks or running new massive models, local HP Z hardware for steady, high-volume processing, and edge devices for time-sensitive tasks. External benchmarks show that local setups can offer up to an 18 times cost benefit over a five-year span when compared to cloud options. The advice we give clients is straightforward: use cloud power for growth you have proven, not growth you are hoping to achieve.
Artificial Intelligence News: Most organizations aim to make their internal data ‘AI-ready.’ How can they accomplish this without compromising sensitive or isolated data sets?
The primary error many firms make is viewing ‘AI-ready data’ purely as a technical engineering challenge. Failing to look at…
What looks like a security issue is often really a data sovereignty issue, and each demands a distinct solution. Feeding confidential data into a cloud-based model for analysis isn’t merely a security risk; it’s a governance breakdown in the making, particularly in regulated sectors where the simple act of moving data outside the organization can itself breach compliance rules.
The architecture that addresses this challenge is Retrieval-Augmented Generation (RAG) deployed on local infrastructure. This approach allows a model to pull relevant information from your internal knowledge repository at the moment of each query, without ever being trained on that data or sending it beyond your perimeter. Your confidential data remains on-premises, housed within hardware under your direct control. For instance, a ZGX Nano or Z8 Fury running a self-hosted model can drive a complete RAG pipeline against sensitive internal documents, with no data ever leaving the premises and no usage fees flowing to an outside provider.
The access control layer is where this becomes operationally critical. A properly designed RAG system enforces role-based permissions at the point of retrieval, ensuring the AI surfaces only the information a particular employee is authorized to access, mirroring the way your existing document management system works. It is the combination of local compute, a locally hosted model, on-premises retrieval, and tightly governed access that truly makes confidential data AI-ready without introducing exposure.
The organizations getting this right aren’t shipping their most valuable assets off to the cloud for processing; they’re moving the intelligence to where the data lives, rather than the reverse.
Artificial Intelligence News: If we merge autonomous AI with today’s modern cloud platforms, what becomes of the everyday responsibilities of an enterprise IT team over the next few years?
I believe Jensen Huang articulated this idea most effectively. He argued that our role isn’t to wrestle with spreadsheets or hammer away at a keyboard, that our work carries deeper significance than that. He’s drawn a clear line between a job’s task and its purpose. In IT, for example, the task might involve provisioning servers or sorting through incidents, but the purpose is ensuring the business stays resilient and continues to advance. That very distinction is precisely what’s unfolding today.
Gartner forecasts that 40% of enterprise applications will feature embedded AI agents by the end of 2026, a dramatic jump from under 5% just a year earlier. This means the routine execution layer of IT is being rapidly absorbed, yet the governance and architecture layer is growing at an equal pace. What’s already taking shape in forward-thinking organizations is a shift from IT teams carrying out tasks to designing and overseeing the agents that carry out those tasks on their behalf.
The critical gap is that only one in five companies currently has a mature governance framework for this transition. This is precisely where local-first infrastructure becomes relevant once more. When your automation layer operates on hardware you fully control, you gain complete visibility into agent behavior, a level of transparency that simply isn’t available when those workloads are abstracted into the cloud. The IT team of the next two years won’t be the team focused on keeping systems running. It will be the teams that determine which agents earn the trust to make which decisions, and that ensure the infrastructure supporting those decisions is something the business can genuinely rely on.
(Image source: Pixabay, licence.)
Want to learn more about AI and big data from industry leaders? Check out the AI & Big Data Expo taking place in Amsterdam, California, and London. This comprehensive event is part of TechEx and co-located with other leading technology events. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
