Supercomputing power that doesn't require ripping out existing infrastructure
At Computex 2026, HP unveiled the ZGX Fury GB300 — a desktop workstation capable of running trillion-parameter AI models entirely on local hardware, without cloud dependency. Built around Nvidia's GB300 GPU and packed with 784 gigabytes of unified memory, the machine collapses what was once the exclusive domain of server farms into something that can sit beside a monitor. Priced between $94,000 and $200,000, it speaks to a quiet but significant shift: enterprises are beginning to reclaim ownership of their most powerful AI, pulling it back from the cloud and into their own walls.
- The gap between datacenter and desktop is closing fast — HP's new machine delivers 20 petaflops of AI compute to a single workstation, a scale that would have required a dedicated server farm just years ago.
- The pressure is competitive and immediate: Nvidia's own DGX Station announcement on May 31st set the stage, and Dell, MSI, ASUS, and Supermicro are all racing to field rival systems on the same GPU architecture.
- For enterprises handling sensitive data, the stakes are high — cloud-dependent AI inference means proprietary information leaves the building, and these machines offer a way to keep billion-parameter models entirely in-house.
- The price tag — starting near $94,000 and climbing past $200,000 — narrows the field to organizations already committed to scaling AI internally, making this a capital decision as much as a technical one.
- With Q4 2026 shipments expected, the market is moving from asking whether to own AI infrastructure to figuring out how to procure, deploy, and integrate it into existing enterprise systems.
At Computex 2026, HP announced the ZGX Fury GB300, a workstation built around Nvidia's top-tier GB300 GPU that can run trillion-parameter AI models entirely on local hardware. With 784 gigabytes of unified memory and 20 petaflops of FP4 compute, it brings datacenter-scale processing to a machine that fits beside a monitor — a threshold that, until now, required distributed server infrastructure.
The announcement is a direct response to Nvidia's DGX Station reveal at GTC Taipei, and HP has tailored its offering for Windows environments, where more than 70 percent of enterprise computing already runs. The company's customers, according to HP's senior vice president for advanced compute solutions Jim Nottingham, have been asking for exactly this: supercomputing power that doesn't require dismantling existing infrastructure.
The implications are most significant for organizations managing sensitive data. Rather than routing proprietary information through third-party cloud services, enterprises can now fine-tune and run massive models entirely on their own hardware. The tradeoff is cost — reseller listings for comparable systems start around $94,000 and can exceed $200,000 — making these machines capital investments for organizations already committed to scaling AI internally.
HP is not alone. Dell, MSI, ASUS, and Supermicro are all building systems on the same GPU architecture, with shipments expected in Q4 2026. The broader moment is telling: Computex 2026 ran under the theme 'AI Together,' and the show floor reflected a market that has moved past debating AI's relevance and into the harder work of owning it. The ZGX Fury GB300 is both a product and a signal — that the era of centralized, cloud-dependent AI is beginning to fracture, and that for those with the budget and the need, a new option is arriving.
At Computex 2026, HP unveiled a machine that collapses the distance between the datacenter and the desk. The ZGX Fury GB300 is a workstation built around Nvidia's top-tier GB300 GPU, configured with 784 gigabytes of unified memory and capable of running trillion-parameter AI models entirely on local hardware. For the first time, the kind of computational power that once required a dedicated server farm can now sit beside your monitor.
The timing is deliberate. Nvidia announced its DGX Station for Windows on May 31st at GTC Taipei, speaking to an audience of more than 30,000 people from nearly 190 countries. The DGX Station itself is a deskside supercomputer—a machine that performs at scales previously reserved for enterprise datacenters. HP's ZGX Fury GB300 is the company's answer to that announcement, built on the same foundational technology but tailored for Windows environments where most enterprise computing still happens. According to Jim Nottingham, HP's senior vice president for advanced compute solutions, over 70 percent of enterprise PCs run Windows. The company's customers, he said, have been asking for this: supercomputing power that doesn't require ripping out existing infrastructure.
What makes this possible is the sheer amount of memory and processing capacity packed into a single unit. The ZGX Fury GB300 delivers up to 20 petaflops of FP4 compute—a measurement of floating-point operations per second in a specific precision format optimized for AI work. That memory footprint of 784 gigabytes means enterprises can load and run models that would otherwise demand cloud access or distributed computing clusters. For organizations handling sensitive data, this changes the equation entirely. They can fine-tune massive models on their own hardware, keeping proprietary information local instead of sending it to third-party cloud services.
The machines are not cheap. HP has not released a price, but the underlying DGX Station has appeared on reseller lists starting around $94,000 for mid-range configurations, with fully loaded systems approaching $200,000. These are not impulse purchases. They are capital expenditures aimed at enterprises that have already committed to scaling AI across their operations—companies that need to run AI agents internally, that require the ability to customize billion-parameter models for their specific workflows, or that simply cannot afford the latency and security implications of cloud-dependent inference.
The broader context matters here. Since ChatGPT launched in November 2022, the conversation around AI has shifted from theoretical capability to practical deployment. Computex 2026 ran under the theme "AI Together," and the show floor reflected a market that has moved past asking whether AI will matter to asking how to build it into existing business systems. Nvidia's parallel announcements at GTC 2026 underscored the same point: the company is not just building chips anymore. It is building the entire stack—from the silicon up through the software and enterprise integrations—that allows organizations to own their AI infrastructure.
HP is not alone in this space. Dell is preparing its own Pro Max GB300 workstation. MSI, ASUS, and Supermicro are also building systems around the same GPU architecture. The expectation is that these machines will begin shipping in the fourth quarter of 2026, likely arriving in waves as partners finalize their configurations and enterprise customers work through procurement cycles. For now, the ZGX Fury GB300 remains a promise—a signal that the era of centralized, cloud-dependent AI is beginning to fracture, and that organizations with the budget and the need are about to gain a new option: keeping their most powerful models at home.
Notable Quotes
As enterprises scale AI agents across their organizations, they need AI infrastructure that can connect directly to the applications and workflows that power their business.— Chris Marriott, VP of Enterprise Platforms at Nvidia
Over 70% of enterprise PCs run Windows, and our customers have asked for AI supercomputing power that can seamlessly integrate into their existing environments.— Jim Nottingham, SVP and Division President of Advanced Compute and Solutions at HP
The Hearth Conversation Another angle on the story
Why does a company like HP need to build this machine at all? Couldn't enterprises just rent compute time from cloud providers?
They could, and many do. But there's a hard limit to that approach when you're dealing with proprietary data or models you don't want leaving your building. A trillion-parameter model running on your own hardware means you're not sending customer data, trade secrets, or training data anywhere. That's not a luxury for some enterprises—it's a requirement.
So this is really about data sovereignty and control?
Partly, yes. But it's also about latency and cost at scale. If you're running inference thousands of times a day, cloud costs add up fast. And every call to a remote API introduces delay. A local machine eliminates both problems. You get instant response times and predictable expenses.
At $94,000 to $200,000, who actually buys this?
Large enterprises that have already committed to AI as a core part of their business. Financial services firms doing risk analysis. Pharmaceutical companies training models on proprietary research. Tech companies building AI agents for internal use. These are organizations where the cost of the hardware is trivial compared to the value of keeping their data private and their models responsive.
Is this the beginning of the end for cloud AI?
Not the end. But it's a fracture. Cloud will remain dominant for most use cases. But for organizations with the budget and the sensitivity around their data, this creates a real alternative. The fact that multiple manufacturers are building these machines suggests the market sees genuine demand.
What happens next?
We'll see how many of these actually ship and how enterprises use them. If they become standard in certain industries, you'll start seeing a two-tier AI infrastructure: cloud for commodity tasks, local supercomputers for the work that matters most.