HP’s Jim Nottingham sees a growing problem with how companies are paying for AI as rising usage makes the economics harder to predict. Employees are turning to models more often, while agents can keep working after the person who assigned them a task has moved on, consuming compute along the way. Inside HP, some developer teams have already run into token constraints, exposing the strain that heavier AI usage can put on the infrastructure underneath it.
Nottingham, SVP and Division President of Advanced Compute Solutions at HP, expects enterprises to use both cloud and local compute as AI workloads expand.
“That shift is already happening, but I wouldn’t describe it as cloud versus local. The future is hybrid,” he tells me. “Consumption is growing incredibly quickly as more people use AI, use it more often and start deploying agents that can be working continuously. At some point, companies naturally start asking whether every one of those interactions really needs to go back to the cloud.”
But the economics are only part of the equation. Nottingham also points to data sovereignty and privacy, the need for real-time responsiveness, and what he calls “data gravity/physics” when large volumes of data are costly or impractical to move. “If you’re processing a huge amount of computer-vision data in a factory or retail environment, for example, it often makes no sense to move all of that data somewhere else just to make an inference and then throw the data away,” he says.
The result is a more distributed approach: compute-intensive AI workloads stay in the cloud, while data-heavy or latency-sensitive work runs closer to where the data is generated.
AI Model Boom Is Changing What Enterprise PCs Can Do
The next important enterprise AI decision, in Nottingham’s view, is not “cloud or PC?” It is which model, on which infrastructure, for which task. “What’s changing quickly is the quality of smaller models,” he says. “Models that fit on local devices today can do things that would have required much larger models not very long ago. At the same time, the local hardware is becoming dramatically more capable.”
HP says its AI PCs can run models of roughly 30 billion parameters locally for personal productivity use cases, while workstations and dedicated AI systems can move into much larger local agentic AI workloads.
“Instead of asking, ‘What’s the biggest model I can use?’ The better question is, ‘What level of model gives me the quality, governance, security and responsiveness this task actually requires?’ Then put that workload in the place that makes the most sense,” he says. “I think companies are still learning how to make that decision.”
As companies start making those choices workload by workload, the role of the corporate PC begins to change as well. Instead of serving primarily as a standardized endpoint built around predictable workloads and centralized management, the machine can increasingly take on some of the AI work itself. It can run models locally, use context from the user and their work, support workflows, and handle parts of tasks through AI agents.
“We’re experiencing this ourselves at HP,” he says. “As internal AI adoption increased, we found there were workloads that made a lot more sense to move from metered cloud resources onto local workstation infrastructure. As agents increase consumption further, we think that economic and architectural conversation only gets more important.”
HP Z Boost Lets Enterprises Reuse Idle GPU Capacity
Cloud providers can keep expensive GPUs working across thousands of workloads, while a company’s own workstation may leave that same capacity unused for hours when its primary user has no demanding AI job running.
HP Z Boost is designed to solve that utilization problem by turning a workstation’s unused GPU into a shared resource. The software allows data scientists and AI developers to remotely access available GPU capacity from other HP Z workstations. That gives companies a way to get more out of GPUs they have already purchased rather than adding dedicated hardware whenever demand spikes. It helps create a distributed pool of compute built from machines enterprises already own.
HP claims Z Boost can pool up to four GPUs from a host workstation and make them available to multiple users and workloads. “Instead of treating every workstation as an isolated island, you can make more of the compute you already own accessible across the team,” Nottingham says.
He rejects the idea that HP is simply recreating a miniature cloud inside the enterprise. “It isn’t utilization for utilization’s sake or just trying to squeeze a few dollars out of the hardware. The real value is getting work done faster,” he says. “If an engineer, data scientist or creator can tap available compute instead of waiting for a resource to free up, that can shorten development cycles and accelerate time to value.”
The race to bring AI beyond the cloud
Dell, Lenovo, Nvidia and other vendors are pursuing similar local and hybrid AI strategies, with many relying on the same Nvidia chips to power their systems. Nottingham acknowledges that the underlying hardware could quickly become commoditized if every vendor offered essentially the same configuration, which is why HP is looking to differentiate through the software, services and broader experience built around those components.
“We spend a lot of time understanding how our customers actually work, where they’re getting stuck and what would meaningfully improve their workflows,” he says. “Then we work closely with our technology partners to build the right solution around those needs. We don’t want to hand somebody a very powerful piece of hardware and say, ‘Good luck figuring out what to do with it.'”
The company is building a new line of AI-focused workstations designed to bring high-performance compute closer to where developers, data scientists and other users work. Its ZGX Nano is a compact AI workstation built around Nvidia’s GB10 platform, while the larger ZGX Fury is designed for more demanding AI workloads and uses Nvidia’s GB300 platform. Both give enterprises a way to run AI workloads outside traditional cloud infrastructure, but the core compute platforms come from Nvidia and are available to other hardware vendors.
As the underlying hardware becomes more standardized, HP needs to show why its systems offer more than the Nvidia processors and GPUs inside them. That becomes more important as AI workloads spread across PCs, workstations and cloud infrastructure, creating more systems for IT teams to monitor and control while agents gain access to company data and begin taking actions on employees’ behalf.
Nottingham says the security challenge exists regardless of where the AI runs. “This isn’t a security problem unique to local AI,” he says. “Whether deploying AI in the cloud, a private cloud or locally, you need observability around what agentic systems can access and what they’re allowed to do. Likewise, governance and security have to evolve alongside the AI itself.”
The Control Layer Verdict
HP is positioning the PC as enterprise AI infrastructure just as the broader PC market faces a projected high-teens decline in unit demand in the second half of 2026, while AI demand is also driving up memory costs. Microsoft owns the operating system layer, Nvidia controls key accelerator and AI software platforms, and Dell and Lenovo can connect edge systems to broader infrastructure portfolios.
With HP no longer owning a server business after the HPE separation, its opportunity is to differentiate through the broader experience around its AI systems. The company wants PCs to contribute compute, process sensitive data locally, support AI agents and connect across private and public clouds. The open question is whether HP can prove its AI systems offer something beyond the Nvidia hardware inside them before every competitor chasing the same shift makes the same promise.
