HP connects local AI development and serving through ZGX stations

HP’s ZGX AI stations pair local accelerated hardware with software for developing and serving models. Nano and Fury address different scales of work, while Z Toolkit and Z Runtime connect the developer workstation to that compute. The useful question is whether a department can operate the resulting service reliably. That requires more than fitting model weights into memory, especially while HP retires its older AI Studio service.

01 / ProductNano and Fury give HP two distinct local-compute starting points

HP’s AI stations page describes ZGX Nano around NVIDIA GB10, with 128 GB of unified memory and DGX OS. Fury uses the larger GB300 platform and Ubuntu. These are configurations within HP’s workstation business, rather than interchangeable names for a cloud subscription. The same page now presents Fury as available through a sales conversation; an actual delivered system still depends on location and configuration.

The memory architecture matters when choosing between them. HP lists Fury with 748 GB of coherent memory across CPU and GPU, while Nano’s 128 GB is shared unified memory. Those headline capacities do not establish how much a particular runtime can allocate to model weights, context and concurrent requests. Compare usable capacity under the intended software stack instead of dividing one advertised memory total by another.

The software overview separates two jobs. Z Toolkit helps developers discover and pair a station, then install development tools through Visual Studio Code. Z Runtime serves models through an OpenAI-compatible API and can retrieve models from supported repositories. Hardware selection and application deployment remain related but distinct decisions: a working API endpoint is not a complete departmental application.

02 / AudienceA department with a defined local workload has the clearest case

A product engineering team developing a document assistant is a plausible buyer. It may want a persistent local environment where developers can inspect retrieval behaviour, change model settings and evaluate confidential test material under its own access rules. The station is useful if that team can also assign responsibility for updates, storage, credentials and availability. Local ownership moves these jobs into the organisation.

A team whose demand arrives only in occasional bursts should calculate the practical cost of idle hardware and staff time. Conversely, a department with repeated development jobs may value a consistent environment even without using every compute cycle. Neither conclusion follows from the accelerator name. Estimate how often the same datasets and dependencies are reused, then compare the total development arrangement.

The Dell blueprint provides another workstation and enterprise infrastructure perspective. The HPE blueprint helps distinguish HP’s local workstation route from Hewlett Packard Enterprise’s separate systems offer. Compare the deployment and support responsibilities, rather than assuming the historical HP name makes these catalogues one product family.

03 / WorkflowProposed workflow: build a departmental document assistant on a controlled station

Begin with a permitted test collection and a short set of questions whose answers a human can verify. This is a proposed evaluation, not a tested HP deployment. Include questions requiring several documents and questions the collection cannot answer. The acceptance target is a useful, attributable answer with a clear refusal when evidence is missing; a locally running model alone does not establish that behaviour.

Choose the specific station and operating systems before installing software. The Toolkit repository documents device discovery, SSH pairing and tools including Ollama, Gradio, Streamlit and MLflow. It describes supported GB10 and GB300 devices and clients, but its platform list is broader than parts of the HP marketing page. A Windows 11 client and supported Fury configuration form a conservative starting proposal, subject to checking the release being installed.

Have the administrator establish the permitted network path and inspect how SSH keys are created and stored. Toolkit’s documented same-subnet discovery makes network design consequential: a station isolated from developers by policy may need a different approved connection arrangement. Avoid relaxing segmentation simply to make discovery convenient. Record which account performs installation and which account runs the application after development.

Next, choose a model whose licence permits the intended use and size its runtime against the full workload. Keep space for retrieval indexes, prompt context and concurrent requests. Establish a repeatable model download and record its revision. The software page describes Z Runtime integrations with Hugging Face, Azure ML and MLflow, but repository support does not transfer rights to every model stored there.

Build the retrieval and application layers around the approved document collection, then connect the application to the local serving endpoint. Keep the model endpoint behind the organisation’s intended access controls. An OpenAI-compatible interface can simplify client integration, but test the particular request fields, streaming behaviour and error handling the application needs. Interface compatibility is narrower than identical model behaviour.

Run the question set first with one user, then with the expected small departmental load. Measure completion time, unsuccessful requests and answer quality separately. Observe memory during long contexts and repeated sessions. A change that speeds generation but removes necessary evidence is not an improvement for this workload. Retain the settings that produced each result so the team can reproduce the comparison.

Finally, restart the serving process and restore the application using written instructions. Check that retrieval permissions still apply and that logs do not expose restricted document text to unauthorised operators. Test removing a developer’s access without stopping the shared service. These operational exercises determine whether the station can move from a convenient development desk into an owned departmental resource.

04 / PricingHardware quotes and software availability need separate treatment

The US Nano catalogue distinguishes 2 TB and 4 TB storage configurations, both described with 128 GB memory. That is useful configuration evidence, but it does not establish a current checkout price when the amount is absent from the rendered page. A quote should identify the storage variant and service package explicitly, particularly if a reseller supplies the system.

Keep the software budget separate from the chassis. Toolkit being free does not mean every model, enterprise support arrangement or external repository service is included. If the proposed assistant depends on a commercial model licence or managed identity service, include those costs in the departmental plan. Evaluate Fury’s preinstalled Runtime as a deployment convenience rather than assuming it removes application maintenance.

05 / DistinctionsThe development-to-serving connection is the practical distinction

HP’s useful combination is the connection between a familiar editor, a discoverable local station and a serving interface. Toolkit reduces several setup steps into a documented route, while Runtime provides a destination for the application being developed. The distinction is most valuable when the same team needs to iterate repeatedly and then hand over a stable service to colleagues.

That connection also makes version alignment important. Keep a record of the Toolkit release, station operating system and Runtime version used during acceptance. If the developer environment changes while the service stays fixed, the team needs to know which layer introduced a behavioural difference. The repository’s telemetry settings are another concrete administrative consideration; review them alongside the organisation’s VS Code configuration.

06 / QuestionsThe AI Studio sunset and conflicting platform lists affect adoption now

The AI Studio documentation explicitly states that the service ends on 29 October 2026. As of this review, that date is the following day. Existing users should resolve data and environment continuity through HP’s current guidance before assuming an older AI Studio workflow will remain available. This article does not treat an old waitlist or documentation page as a current onboarding offer.

Platform support needs similarly precise confirmation. The Toolkit repository lists macOS alongside Windows and Ubuntu clients, while the HP software page contains narrower client and host footnotes and a broader Nano/Fury FAQ. These are not a reliable basis for promising every combination. Ask HP which released Toolkit version supports the intended client, station and operating system, and retain that answer with the installation plan.

07 / DecisionChoose a station only after the development and operating roles are clear

HP fits teams that want to connect local AI experimentation to an owned service. Start with the workload, select a supported hardware and software combination, and prove the handover from developer access to routine operation. Treat the AI Studio transition as an immediate continuity question for existing users, separate from the attractiveness of the new ZGX hardware.