HP and the art of AI and data for the enterprise

Ahead of the AI & Big Data Expo at the San Jose McEnery Convention Center, May 18-19, we spoke to Jerome Gabryszewski, the company’s AI & Data Science Business Development Manager about AI, processing data for AI ingestion, and local versus cloud compute.

The technology media is fond of quoting that data is ‘the new oil’, but the reality on the ground is that, despite having access to plenty of first-party information, actually leveraging it to the business’s advantage can prove problematic, especially at enterprise scale.

Should you chose a cloud-hosted AI model, or local compute? How do you get your ‘data house’ in order, so the smart models can produce meaningful results? And as ever, we like to encourage our interviewees to help us predict the next chapter in the fast-moving story of business IT in this AI-dominated business landscape.

Artificial Intelligence News: Moving from manual to automated data ingestion sounds great in theory, but it’s notoriously difficult. Where is HP seeing companies get stuck right now?

One of the most consistent friction points we see is that organisations underestimate the organisational and architectural debt behind their data. Before automation can take hold, they have to reconcile fragmented data ownership across departments, inconsistent schemas in systems, and legacy infrastructure that was never designed for interoperability. The technical lift of automation is often smaller than the governance and integration work that has to precede it.

Artificial Intelligence News: When AI models start updating themselves continuously, things can easily go sideways. How are you advising clients to handle risks like concept drift and data poisoning?

Continuous learning is where AI goes from a project to a liability if it isn’t governed carefully. What we advise clients is to treat model updates the same way they treat code deployments. Nothing goes to production without a validation gate. For concept drift, that means MLOps pipelines with automated drift detection and human-in-the-loop triggers before retraining kicks in. For data poisoning, it’s a data provenance problem as much as a security problem. It’s critical to know exactly where your training data comes from and who can touch it. The clients who get this right aren’t necessarily the most technically sophisticated; It’s those who’ve embedded AI governance into their risk frameworks before they scaled.

Artificial Intelligence News: I want to touch on HP’s hardware roots. What does a modern workstation or compute setup actually need to look like today to handle the sheer weight of an autonomous AI lifecycle?

HP’s roots here actually matter. The Z series has been purpose-built for the most demanding professional compute for over 15 years so when we talk about what an autonomous AI lifecycle actually requires from hardware, we’re not guessing, we’ve been iterating on this problem longer than most!

The answer isn’t a single machine, it’s a spectrum. At the individual developer level, you need local compute powerful enough to run real experiments without being cloud-dependent for every iteration. The ZBook Ultra and Z2 Mini handle the mobile and compact deskside tier professional-grade machines capable of running local LLMs and heavy workflows simultaneously.

The ZGX Nano is where things get really interesting for AI-first teams. It’s an AI supercomputer that fits in the palm of your hand (15x15cm), but it’s powered by the NVIDIA GB10 Grace Blackwell Superchip with 128GB of unified memory and 1,000 TOPS of FP4 AI performance. A single unit handles models up to 200 billion parameters locally. And when a team needs to scale beyond that, you connect two units together via high-speed interconnect and you’re working with models up to 405 billion parameters… no cloud, no data centre, no queue. It comes pre-configured with the NVIDIA DGX software stack and the HP ZGX Toolkit, so teams go from setup to first workflow in minutes, not days.

Moving up, the Z8 Fury gives power-user teams up to four NVIDIA RTX PRO 6000 Blackwell GPUs in a single system (384GB VRAM): That’s the full model development cycle running on-premises. And at the frontier, the ZGX Fury changes the conversation entirely. Powered by the NVIDIA GB300 Grace Blackwell Ultra Superchip with 748GB of coherent memory, it delivers trillion-parameter inference at the deskside, not the data centre. For teams running continuous fine-tuning and inference on sensitive data, it typically pays for itself in 8 to 12 months versus equivalent cloud compute.

And for organisations that need to cluster and scale further, the entire Z portfolio is designed with rack-ready form factors that drop into managed IT environments without compromising security or data residency.

The larger point is this; the autonomous AI lifecycle creates a governance and latency problem, not a compute problem. Teams can’t keep sending sensitive training data to the cloud every time a model needs to update. HP’s portfolio gives organisations a hardware path that scales with their workflow maturity, from the developer’s desk all the way to distributed on-premises compute. The hardware finally matches the ambition of what these AI systems actually need to do.

Artificial Intelligence News: Gen AI compute costs are spiraling for a lot of enterprises. What is the practical fix for balancing that massive expense with modern cloud efficiency?

The cost problem is structural, not cyclical. Enterprise GenAI spend surged to $37 billion in 2025, and 80% of companies still missed their cost forecasts by more than 25%. The core tension is that unit inference costs are actually falling, but total spend keeps rising because use is growing faster than cost drops. The cloud API model was designed for experimental, low-volume workloads. It was never built to be the economic engine for production AI at scale.

The practical fix is a discipline problem before it’s an infrastructure problem: Draw a hard line between exploratory work and production workloads, and never use the same compute model for both. Early iterative work – prototyping, fine-tuning, model evaluation – should run on local hardware like the ZGX Nano or Z8 Fury, where you’re spending capital once instead of burning operational budget on experiments without a clear ROI path.

The organisations getting this right are running a three-tier model: Cloud for burst training and frontier model access you’ve genuinely earned, on-premises HP Z infrastructure for predictable high-volume inference, and edge compute where latency is critical. Independent analysis shows on-premises can deliver up to an 18x cost advantage per million tokens over a five-year lifecycle. The framing we use with clients is simple: cloud is for scale you’ve earned, not scale you’re hoping for.”

Artificial Intelligence News: Everyone wants their proprietary data to be ‘AI-ready.’ How do companies pull that off without exposing sensitive or siloed information?

The mistake most companies make is treating ‘AI-ready data’ as a data engineering problem when it’s really a data sovereignty problem, and those require different solutions. Sending proprietary data to a cloud model for processing isn’t just an exposure risk, it’s a governance failure waiting to happen, especially in regulated industries where even the act of transmitting data externally can trigger compliance violations.

The architecture that solves this is Retrieval-Augmented Generation (RAG) running on local infrastructure, which lets a model retrieve relevant context from your internal knowledge base at query time without ever training on it or exposing it externally. Your proprietary data stays on-premises, inside hardware you control. For example, a ZGX Nano or Z8 Fury running a locally hosted model can power a full RAG pipeline against sensitive internal documents with no data leaving the building and no token spend sent to a third party.

The access control layer is where this gets operationally serious; a well-architected RAG system enforces role-based permissions at the retrieval level, so the AI surfaces only what a given employee is entitled to see, the same way your document management system does. The combination of local compute, local model, local retrieval, and governed access is what actually makes proprietary data AI-ready without exposure.

The companies getting this right aren’t sending their crown jewels to the cloud to be processed; they’re bringing the intelligence to the data, not the other way around.

Artificial Intelligence News: If we combine autonomous AI with these modern cloud platforms, what happens to the day-to-day role of an enterprise IT team over the next couple of years?

I think Jensen Huang laid this concept out best. He said our job is not to wrangle a spreadsheet or type into a keyboard, that our work is generally more meaningful than that. And he’s drawn a sharp distinction between a job’s task and its purpose. In IT, for example, the task might be provisioning servers or triaging incidents, but the purpose is keeping the business resilient and moving forward. That distinction is exactly what’s playing out right now.

Gartner projects 40% of enterprise applications will have embedded AI agents by end of 2026, up from less than 5% just a year ago, which means the routine execution layer of IT is being absorbed fast but the governance and architecture layer is expanding just as quickly. What’s already happening in leading organisations is a change from IT teams executing tasks to designing and governing the agents that execute on their behalf.

The important gap is that only one in five companies has a mature governance model for that yet. This is where local-first infrastructure matters again. When your automation layer runs on hardware you control, you have full observability over agent behaviour that you simply don’t have when those workloads are abstracted into the cloud. The IT team of the next two years isn’t the team that keeps the lights on. It will be the teams that decide which agents get trusted with which decisions and makes sure the infrastructure underneath that judgement is something the business can actually stand behind.

(Image source: Pixabay, licence.)

Want to learn more about AI and big data from industry leaders? Check out the AI & Big Data Expo 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.