HP's Ketan Patel on Turning the PC Into a "Personal Data Center" for AI Agents

Open on YouTube ↗
Overview

In this episode of 1on1 Tech, TBS CROSS DIG business editor Masahiro Nakagawa interviews Ketan Patel, president of HP's Personal Systems business. The conversation starts with HP's new machines built around NVIDIA's RTX Spark chip, which Nakagawa describes as a laptop that "becomes sort of a miniature data center." The central question is how much AI work will move from the cloud to devices people own. Patel's position is that the PC is becoming a "personal intelligence" device shared by one human and many agents. In his view, AI inference will become a hybrid of on-device, on-premises, and cloud computing, driven by token costs, security, and data sovereignty.

18 min read
1:59

From Tool to "Personal Intelligence": One User, Ten Agents

Patel argues that for many years the PC has been an input-output tool: you give it input and it returns output. He sees a new opportunity in what he calls a shift from personal computing to "personal intelligence."

The main driver, in his view, is agents. "Today one PC is getting used by a single user but tomorrow a PC will be used by one user and 10 agents onto it," he says. Agents will move through different applications and issue commands to reach outcomes for the user. Patel says this changes how compute is consumed. The machine has to support multiple actors working in the same environment and manage their tasks intelligently, so local hardware has to become much more powerful.

He presents the RTX Spark machines as HP's answer to this need. The flagship he shows is the OmniBook Ultra, which he says launched "last week." Its key feature, according to Patel, is a unified silicon architecture in which CPU, GPU, and memory share one pool. The machine has 128 GB of unified memory. Because of that, he says, workloads run locally with efficiency "like of a class of data center." He adds that the point is not agents for their own sake. What matters is how they change workflows, productivity, and experience for freelancers, developers, creators, small businesses, and enterprises.

4:55

"NVIDIA Inside"? Patel Says Multiple Architectures Will Coexist

Nakagawa asks whether the old "Intel Inside" PC is becoming an "NVIDIA Inside" PC, with a GPU in every machine. Patel does not accept that framing. He lists Intel, AMD, NVIDIA, MediaTek, and Qualcomm as silicon providers and says HP works with all of them. In his view, each architecture is more efficient for particular use cases and price points, so a range of options matters to users. "There are multiple architecture which will coexist bases the use cases," he says.

On GPUs specifically, Patel predicts that every PC will become more capable and will include strong graphics capability. Whether that comes from a unified memory architecture or a discrete graphics card is, in his words, "secondary."

6:29

Why Developers Are the First Audience

Asked what NVIDIA's entry means for a PC maker like HP, Patel calls RTX Spark "a great addition to the category" and names developers as its classic user. In the past, developers had to go to the cloud for many of their tools. Patel says a model of roughly 120 billion parameters can now be hosted locally on this machine. He lists the benefits he sees:

  • practically no token cost for tasks that can be done locally;
  • more security, adjusted to whether the user is a freelancer, gig worker, or enterprise;
  • no internet requirement, so a developer can keep running AI workloads while traveling with a weak connection.

Patel says this makes development cycles faster and more efficient and calls it "a breakthrough for developer segment." He also stresses NVIDIA's CUDA stack, which he says is very important to the developer community. Together, he says, it makes the product a fully integrated solution rather than hardware alone.

8:27

From 95/5 to 50/50: HP's Estimate of the Local–Cloud Split

Nakagawa asks how much data center computing can actually move to personal devices. Patel answers that the question is not about one device. HP's range, he says, will host models from about 4 billion to 20 billion, 50 billion, and 120 billion parameters across different products. He notes that even a mini desktop could run a 4-billion-parameter model for a specific task.

He then gives HP's assessment. Today, the majority of AI workloads run in the cloud, roughly 95% by his figure. Over "a couple of years," HP believes the split between local and cloud will be about 50/50. He describes this as HP's projection, not a measured result.

The main economic argument is token cost. Patel says token economics is becoming a big topic, especially in large organizations. He asks the listener to imagine giving AI access to all 10,000 employees and paying for it. Nakagawa adds that some companies spend a billion dollars, and Patel agrees. Organizations want every employee to have AI, Patel says, but they also care about cost and ROI. Balancing these, he concludes, the world "is going to be essentially hybrid," with inference spread across small local models, 120-billion-parameter local models, and cloud models.

10:48

On-Premises AI and Data Sovereignty: The Z Fury Workstation

Patel describes a third tier between the personal device and the cloud: on-premises AI. Many customers, he says, worry about security and data sovereignty. They do not want their company's intelligence to enter a public domain where it "becomes an intelligence for somebody else." But they still want the power of a strong model.

For that, he points to HP's Z Fury, part of its workstation line. He says it can host a model of up to one trillion parameters. Units can be stacked, and he gives the example that ten of them could host a 10-trillion-parameter model, all running locally. His summary of the future architecture is on-device, then on-prem, then cloud.

Later in the interview, Nakagawa returns to this workstation. He notes that it runs on NVIDIA's GB300 and can run a very large model on device, and Patel confirms the GB300. Nakagawa suggests these machines will appear in workplaces first but may eventually reach homes.

12:06

A Sales Professional's Day With Local AI

Nakagawa notes that much of the audience is general business professionals, not engineers, and asks for a concrete example. Patel describes a salesperson's day. The core job is to interact productively with clients, but much time goes elsewhere: more than 100 emails, at least about 30 minutes of preparation per meeting, note-taking during meetings, and then follow-up such as emails and task assignments.

In Patel's scenario, a local model already on the salesperson's machine accesses their applications, such as a CRM, before an 11:00 a.m. meeting and prepares the client briefing. He emphasizes that this is "all free all local." Confidential client data does not leave the network, and the salesperson skips the emails and preparation. During the meeting, a meeting agent powered by the local model captures the discussion. The salesperson can focus on the client instead of action items. Afterward, the agent sends emails and tasks to the relevant people. Patel says the same pattern applies to other roles in enterprises, small businesses, and consumer settings. He presents this as a vision of how the products could be used, not a deployment he reports.

14:51

Does Expensive Hardware Widen the AI Divide?

Nakagawa points out that high-end RTX Spark machines cost over 1 million yen, while cloud AI is also getting more expensive. He asks whether this could widen the gap between people who can afford AI and people who cannot.

Patel replies that the RTX Spark machine is not the only route to local AI. It is the highest-performing device at its price, aimed at users like developers and game creators. For them, he argues, the cost pays off through productivity and avoided cloud spending. A sales professional does not need that much capability. Patel says such a user could buy a $1,000 product and still get local AI for basic tasks. He names the OmniDesk Mini, a mini desktop, as HP's entry-level product with local AI capabilities. The range then goes up to RTX Spark and further to workstation-class on-prem devices. HP's position, he says, is to have solutions across price bands for different users.

16:42

Mac vs. Windows: The "Open Ecosystem" Argument

Nakagawa observes that many AI developers use MacBooks and keep the lids open so their AI workflows don't stop. He asks why someone should choose an HP Windows machine. Patel says every user can choose the ecosystem that suits them. HP's pitch is openness: users can pick any model or application and work across operating systems without being locked into one ecosystem.

This matters for agents, he argues, because someone may have a phone on one OS and a PC on another. In an agentic world, they may give a command from the phone and expect the PC to carry it out. Patel says HP focuses on interoperability and continuity across devices.

Nakagawa pushes back: Apple has smartphones, laptops, and PCs, while HP has no smartphone. Patel repeats that HP is "an open ecosystem player." Users can choose operating systems ranging from Android to Windows and are not locked in.

18:31

Memory Prices and the Total Cost Argument

Will RTX Spark-class machines get cheaper? Patel says scale tends to bring down technology costs over time. But he says, "I can't predict that this is going to be a cheaper cost after few years." He expects that wider adoption of unified memory architecture products and scale across silicon architectures will help users get better technology at better value over time.

Nakagawa then asks about the recent surge in RAM prices. Patel says hyperscalers' buildout has made memory the most constrained commodity and "the most important commodity in the world of AI." He notes that memory costs have risen significantly year on year over the past 12 months. HP's response, he says, is to design memory-efficient systems that deliver maximum output with a smaller memory footprint. HP is working on its own architecture and with Microsoft on optimization.

He also argues on total cost. According to HP's internal analysis, he says, using local AI on a PC saves about $650 per user for basic tasks and up to about $1,200–$1,300 per user for advanced tasks. He presents this as larger than the extra hardware cost, so the total cost still works out despite recent commodity price increases. These are HP's internal figures as reported by Patel.

21:24

Young Smartphone Users and the "Sleeping PC That Keeps Working"

Nakagawa says he has heard that many young people grow up on smartphones and start using PCs only when they join a company. He asks whether AI gives them a new reason to own a PC. Patel says HP believes so. He agrees the phone will remain everyone's primary device. But as more agents are deployed, he argues, they need "a little more powerful surface area" to run on, which makes the PC critical. He adds that the PC has always been a strong creation device and more local power improves that. He describes the relationship as complementary: the phone becomes a "command center" that gives instructions, and the PC executes them.

Patel predicts that time spent on the PC will rise, but not the user's own time. Agents will work on the user's behalf in the background. Today people use a PC for anywhere from 2 to 12 hours depending on who they are. In the future, he says, "when your PC is sleeping the agents are still working." That would use the machine's full capability around the clock, which he sees as a big advantage since PCs are not used 24/7 today. Asked whether "command on the phone, run locally on the PC" will become normal, Patel says device continuity will be part of normal life: commands given on one device, executed on another, and results checked on the phone.

24:44

Containing Unpredictable Agents: Wolf Security

Patel raises device-level security next. HP's Wolf Security stack, he says, includes work on a BIOS-level secure root of trust. This matters in the agentic world, he says, "because agents can operate in a way which can be unpredictable." HP is using its security stack to contain agents and apply governance. He frames HP's offering as more than a product at a price: it is a user experience, the ROI from the sales example, and security together.

Nakagawa mentions a recent rise in cyberattacks in Japan over the past few weeks and asks whether security is the reason to run AI locally. Patel says it is "one of the reasons," not the only one. He positions local AI plus HP's container technology as a reason to choose HP, calling security and trust a core problem that every customer in the agentic world will face. He notes that HP built an endpoint security controller, which he says was first in the industry, because it recognized that hacking can happen through the endpoint.

26:54

Managing 60,000 "Users" in a 10,000-Person Company

Nakagawa cites reports from AI labs suggesting agents could automatically take over connected devices and run models on them, and asks what cyber defense will mean in that world. Patel describes two layers. The first is on-device agent security. The second is HP's Workforce Experience Platform, a manageability platform HP is building. It monitors how many devices and agents exist in an organization's infrastructure. It also lets IT administrators define which agents may or may not be deployed on machines and what governance applies to them. Each agent might even have something like an Active Directory identity, the way human employees do.

Patel describes the scale with a number. Today a large customer might manage 10,000 devices with 10,000 users. Tomorrow, he says, it will manage 10,000 devices and 50,000 agents, which means effectively monitoring 60,000 users for a 10,000-employee organization. That, he argues, is why the manageability layer is critical.

28:51

Will Every Home Have a Data Center?

Nakagawa asks how workstations, not just laptops, could change how people interact with AI, including at home. Patel says a workstation like the Fury can host large models locally, which makes it a strong fit for enterprises and small businesses. People with that requirement could extend it to their homes. But a home could also have an OmniDesk Mini hosting agents, while a regular PC or phone handles other applications. He describes a range from basic "agentic PCs" HP is developing up to the Fury level. He again insists the world will not be "either local or cloud." Even at home, a local machine would handle most of the load but should be able to reach the cloud when needed. For a home user, he says, what matters most is that agents manage tasks such as the personal calendar and travel booking. Hardware should be sized to each customer's needs.

Nakagawa explains why he asked. A previous guest on his show, Silicon Valley angel investor and All-In podcast host Jason Calacanis, said that people running AI on powerful home workstations won't worry about token costs, which changes their behavior because they no longer watch a usage meter. Patel says, "I love what he said," and calls it exactly HP's thesis. He adds that homes will have different needs, so HP plans "good better best" options rather than one size fits all.

Nakagawa also mentions a Japanese startup he covered that is developing manipulation robots for homes and says each home would need a workstation. Patel says that for robotics, yes, workstation-class power would be needed. For basic use cases like the ones he described, a less powerful machine would be enough.

33:02

Beyond the Laptop: Form Factors, Robots, and Expanding Workstation Markets

Patel describes Personal Systems as more than laptops. It covers laptops, desktops, displays, peripherals such as headsets, keyboards, and mice, HyperX gaming peripherals, workstations, and related services and solutions. On what comes next, he returns to his opening theme. The PC will become "a digital twin of yourself." It will capture the user's personal knowledge graph and understand their context better than any other device, as an always-on, context-aware device for personalized experiences.

He says this will "absolutely" lead to new compute form factors beyond the classic laptop, but he declines to give concrete examples because he cannot share HP's roadmap. When Nakagawa asks whether HP is thinking about robots, Patel says HP's AI focus is currently on compute, print, workstations, and "AI stations." HP looks at both horizontal solutions and specific industry verticals. If a full solution required something like a robot, even one not made by HP, HP would consider partnering to deliver it. "But that's not something as of now we have thought of but you never know," he says.

On workstations, Patel says they have mainly served high-performance niches: CAD/CAM, architecture, engineering and construction, heavy graphics, and media and entertainment rendering. Local AI and agents, he argues, will expand their addressable market beyond those classic uses, because they combine compute power with the ability to host high-capability models on device. He places workstations at the center of HP's "edge AI" positioning.

On how personal AI should work across many devices, Patel says HP designs around a user's whole workflow, not single-device solutions. His examples are hardware-to-hardware: an HP PC working smoothly with another HP PC, with HP printers without the usual printing steps, or with HP headsets. He says the same principle will apply to full workflows in the agentic world.

38:35

Looking Five to Ten Years Ahead

Asked how they imagine HP five or ten years from now, Patel says the AI-at-the-edge opportunity will open up beyond where it is today. In conversations with customers, he says, HP sees that token cost, security, and data sovereignty are "all real issues" in real AI implementations. HP's aim over the next few years is to provide "outcome-based solutions" to these pain points by investing in its portfolio, solution stack, and services. He adds that technology is moving so fast that HP will keep innovating across its portfolio toward that vision. He offers this as the only "sneak peek" he can give, without a specific roadmap.