top of page

HP OmniBook Ultra 16 Review Part #2: I made an unplugged, local AI workstation

5 days ago
10 min read
Man holding an unplugged HP OmniBook Ultra 16 laptop showing the OLED touchscreen and a complex AI visual.
Photo provided by HP Inc.

NOTE: If you haven't read Part #1 of the review, I recommend you do so. The full specification and target user profile is included.


I need to say it again: the HP OmniBook Ultra 16 is a beast of a machine. But its power and performance capabilities mean that it is also somewhat daunting when starting to look at what it can be used for, and how to get the most out of it. Having used laptops and PCs since I was eight or nine years’ old, it’s not often that I am overwhelmed by a new piece of technology, but I was in this case. How does one start to get the most out of the capabilities of such a machine?


Having outlined the HP OmniBook Ultra 16’s design, its connectivity and the target market for this powerful laptop, I wanted to set it up as a functioning workstation using one of the use-cases HP provided to reviewers: the unplugged, local-AI workstation, functioning as a centralised resource for knowledge management, and AI and application development. In addition, I wanted to test how “sovereign” this build could be, assuming a scenario where extra-regional technology is limited, embargoed or even switched off. At an architectural level, the RTX N1X has native CUDA engine support, giving it the ability to operate completely independent of cloud services.


The premise of this setup was therefore to test the machine’s capabilities in non-cloud environments, where limited (or sporadic) charging is available, and/or where local AI is either a requirement, or a default due to a lack of connectivity. The objectives I set myself:


  • Use a variety of powerful, local AI tools and models for inferencing - no cloud-based models.

  • Where possible, use open-source and/or non-US/non-China applications, databases and operating systems.

  • Maintain performance using battery only, such that the laptop can be carried around while continuing to operate at maximum possible levels.


Sovereign Considerations: The Caveats

Given that I live in the UK, I took a UK/European perspective, meaning I first had to consider how close I could get to the objectives. I realised quite quickly that this would mean compromising on some of the choices. This included how to minimise some of the potential risks where sovereign considerations were concerned. I had to make sure that once built, nothing installed on the machine calls a cloud service. Finally, connecting to the internet would be rationed: only on switched on during planned maintenance windows or connecting with a local service.


Hardware & Operating System

HP’s partnership with NVIDIA and Microsoft defines the composition of the out-of-the-box experience with the HP OmniBook Ultra 16. Some choices are already made. Windows itself, Ollama and NVIDIA’s drivers are US software. They can run locally, but if full independence matters, the plan would be to move to native Linux OS once NVIDIA releases support for it on the RTX Spark N1X.


Software compatibility (and lack of)

The rule I worked to was to use a native Arm64 build of software tools if one existed. I could use an emulated build only when no native build exists and the software needs no drivers. Proprietary pieces (of which there are potentially many) had to stay only where the hardware gives no alternative. However, I found that compatibility status can change quickly and found myself bookmarking and checking web sites for updates regularly!


Sharing & Access

This configuration is also a build that works best as a single-user or small team workstation that also hosts services locally i.e. not as a production server or hosting platform for cloud-connected/remote users. For this review, I focused on my experience as a solo operator. With further review time, I could explore how this build functions as a mobile resource for multiple users - perhaps a long-term review is next?



Creating the Software Stack

Close-up of colorful JavaScript code on a dark screen, with diagonally blurred lines creating a busy, techy mood.

Without the requirements I imposed, the software install would have been quite easy. At setup time, the HP OmniBook Ultra 16 serves up a bunch of automation scripts and tools for creating your first AI agent, up to getting a full application development environment up-and-running using US manufactured tools and cloud-based services.

Sovereign considerations mean AI models and tools are limited to a relative handful. The Allen Institute for AI’s Open Language Model (aka OLMo) offers a range of open-source LLMs, and I considered this route, but found I lacked the skills to implement it (I can admit to failure). A second option was EuroLLM, another open-source suite but this one is “made in Europe”, having the additional benefit of supporting all 24 official European Union languages. While this seemed like the perfect fit for a sovereign AI build, I lacked familiarity with it. Perhaps this could be incorporated as an option in a future build.


The final choice was Mistral - the France headquartered AI LLM creator. I’ve used some older Mistral models in the past and was much more confident about choosing some of the larger models for this build. NVIDIA’s pre-installed RTX Spark app allows you to adjust the unified VRAM available via a simple slider. On this 128 GB system, I could dedicate up to 96 GB directly to GPU VRAM to accommodate massive LLM context windows and vector indexes. I set it to 85 GB to give myself a little more headroom.


Decision made, the next thing I did was to install Perplexity Portable Computer - the Windows version of Perplexity’s “local first” AI agent stack that can be operated without inferencing via the cloud. I’ve used Perplexity since it was first released, and for this build, it would assist with installing and configuring all the other components - a much-needed time saver. Connecting the various apps and local files would also help guide me through the trickier aspects of installing the correct versions of each tool, and with checking ongoing compatibility etc.


Tools, Tasks & Functions


  • Docker / Containers - I initially planned to use Ubuntu’s own docker.io build by default but ended up using the “early access” version of Docker Desktop for Windows on Arm.

  • Browser - Firefox, with a manual tweak to trust the local certificate authority.

  • Office-style documents - Collabora Online in the container stack, with LibreOffice (native Arm64) on the desktop.

  • AI Inference Engine - llama.cpp / llama-server in router mode. I kept Ollama as a fall-back option in case of compatibility issues.

  • App to Model API - OpenAI-compatible /v1 for every app, with engine-neutral model IDs.

  • AI Model(s) - It would have been great to run Mistral Large 4 (aka “Le Chonk”) on the HP OmniBook Ultra 16, but the RAM required for the French model maker firm’s latest-and-greatest was too much even for this laptop. Instead, I opted for Mistral Small 4 at 4-bit quantisation which requires 78.5 GB of VRAM.

  • NotebookLM-style research notebooks - Open Notebook is the best sovereign option here. Open source, it includes chat with citations, notes etc. Running in Docker it works fully offline with a local model.

  • Desktop Tools - (In addition to the above) there were a bunch of others including VS Codium, Git, Powershell 7, KeePassXC, WireGuard, and 7-Zip. Again, this is not an exhaustive list: others will be added over the coming weeks.


Person in a blue beanie and dark coat stands in a sparse pine forest, wearing a camouflage backpack, facing into the trees - represents going fully offline.

Going Fully Offline?


At this stage, I was deep into the build, testing the applications, tweaking settings etc. and had been reading and watching the US and UK government’s preparedness guidance during the annual Cybersecurity Awareness Month (October). With a heightened focus on resiliency, and the ongoing coverage of the threats to critical infrastructure, I wondered if this build could extend to having a completely offline digital knowledge base, as well as a “second brain” that could be queried if/when online services were unavailable. Lacking electricity, reading a book by torch or candlelight is one way, but keeping books for general knowledge has dwindled. The paper-based encyclopaedia is a rarity amongst households these days. Even if electricity is available, cloud services may not be, causing significant disruption to our online-dependent lives.

The HP OmniBook Ultra’s 2 TB of onboard storage was too tempting. Over a few days, I downloaded what I considered would be the essential resources to have on hand for temporary cloud outages:


  • Offline Encyclopaedia - Stored locally, Wikipedia is a roughly 100 GB download - which is quite modest given how much information it contains.

  • Offline Maps - Similarly, the OpenStreetMap project provides comprehensive data that can be downloaded, stored and viewed using various Windows-based OSM display/edit software.

  • Technical and/or Applied Knowledge - The Massachusetts’s Institute of Technology’s OpenCourseWare resource contains over 2,500 courses and materials that are free to download and use (MIT licence). I chose a selection of courses on AI, Analytics, Engineering and Programming.

  • The Internet - ok, maybe not the entire internet, but snapshots of favourite websites etc. There are many web archive tools applications out there, including newer, open-source ones such as Obelisk that work with Wayback Machine and other tools.


The obvious issue with all the above is that a laptop - even with the large battery that the HP OmniBook Ultra 16 has - will need charging at some point. Tweaking efficiency settings will help prolong its operation, but even so, its charge is not infinite. Keeping a few large, fully charged power banks to hand will help replenish its power, but better still is a portable charging station connected to solar panels. Recharging your laptop may not be a priority during a power cut, but intermittent top-ups are possible using easily available domestic equipment.


Security & Resilience

During the build and test process, there were a few additional things that I had not necessarily anticipated would be a challenge:


Running Hot

Like every laptop I have used, the HP OmniBook Ultra 16 runs hot under sustained load. While I lacked the equipment to reliably measure the machine’s temperature, the fans kicked-in as expected, and the exhaust port produced copious amounts of warm air along its entire length. Performance did not noticeably suffer - I couldn’t identify any thermal throttling - and the noise was a lot less that I’ve experienced with most other laptops. If in doubt (or you are already in a hot environment) you could use a cooling stand, keep the machine plugged in, and turn off sleep mode.


Single point of failure

From a security perspective, a portable workstation with a high level of local dependency that contains this much critical data and functionality represents significant risk. Once it leaves a secure physical environment, this risk increases. When mobile and disconnected, every hosted service that it runs goes down with it. If this build were a shared resource, it would be best to move the hosting layer to a server in a fixed location. Given that I am running this machine as a portable resource, hosted services would either be available sporadically, or not at all. For day-to-day use, a backup strategy would be needed: an external, encrypted 2 TB drive stored somewhere accessible but in a fireproof safe; a sovereign, cloud-based storage (only) service for failback situations.



Rusty chain and silver padlock secure a carved wooden door, with two brass bells above in a warm, weathered close-up. Represents strong encryption and security.

Encryption & Windows security

Again, as the only option currently is to run Windows 11 on Arm for the Nvidia RTX Spark chip set, the challenge then is to “harden” the Windows environment as much as possible. Running the Enterprise edition of Windows 11, the first step is to enable BitLocker for full-disk encryption, (assuming a TPM of v1.2 or later) but add a PIN. In addition, I followed the advice of making sure that recovery keys were stored offline and would be easily recoverable. Alternatively, as an open-source tool, VeraCrypt is independent of Window’s encryption architecture and would in theory have been perfect for the most extreme requirements for sovereign key management and control. However, VeraCrypt currently only supports encryption of non-system volumes in Arm64 environments. Therefore, the best option would be to run BitLocker on the system drive and use VeraCrypt for containers and other volumes.


Going Mobile

This build took me around two weeks to complete, but it could be done more quickly by someone with better technical skills. As a platform, the HP OmniBook Ultra 16 behaved extremely well: I didn’t experience any driver, startup or boot issues during the testing. I experienced incredibly smooth performance enhanced by a highly responsive keyboard and trackpad and the large bright screen (in the right lighting conditions). Couple with a seemingly never-ending battery, I didn’t feel silly carrying it around balanced on my arm.


What you can achieve with the hardware is albeit restricted by the current limitations of Windows 11 on Arm - which is not the fault of HP. Every laptop manufacturer using the Spark N1X SoC is facing the integration and compatibility challenges that any new chip set facing running on an unoptimised operating system. However, it’s not that the machines operations are restricted, rather that the range of choices open to customisation are fewer. If you are happy to install the recommended and/or suggested components, the setup and everyday operation would be much easier.


Did I achieve my objective? More-or-less. As an exercise in creating a portable, local AI workstation with sovereign capabilities, it worked. Sitting with it on my lap, unplugged, with airplane mode on running AI agents and querying sizeable LLMs seemed very odd. Response latency was low, while searching archives, wikis and databases was super-fast. The experience was like being hardwired to a server and mainframe that could sit comfortably on my knees.


The Gist

While dimensionally larger than the average laptop, the HP OmniBook Ultra 16 is nevertheless understated: a thin, grey slab, light enough to be truly portable, but with hints of its power if you look (and listen) closely. As a premium machine, its strength is that it can do everything, from deep AI computations and application development work, through to high-end gaming and pro-level creative projects. But as an all-rounder, this strength is also the HP OmniBook Ultra 16’s weakness: dedicated gaming laptops can maximise the performance of AAA titles for half the price; video editors, compositors and media professionals favour Mac-based systems; while AI and application developers are a tough bunch to please, favouring heavily customised environments and a choice of operating systems.


Where the HP OmniBook Ultra 16 shines is in matching the totality of its capabilities with a specific set of requirements - as opposed to attempting to be a jack of all trades. While still relatively niche, a portable, local AI-based, centralised resource for application development and knowledge management, that takes account (as much as possible) of sovereign requirements of a non-US/non-China location, is a use case that can realised without developer-level skills. In its Max configuration, the HP OmniBook Ultra 16 has enough storage, processing capability, and memory to undertake workstation-level tasks with some of the largest AI models. Powerful enough to maintain performance levels even when unplugged, yet discreet enough to be part of an EDC in a generic backpack.

Comments


bottom of page