HP ZBook Ultra G3a: A 192GB Laptop Built to Run 300B-Parameter AI Locally
HP announced the HP ZBook Ultra G3a on Sept. 15, 2026, a 16-inch mobile workstation built around an AMD Ryzen AI Max+ PRO 495 processor and up to 192GB of unified memory. HP says up to 160GB of that pool can be assigned to the integrated Radeon GPU, the figure behind its claim that the machine can run large language models of up to 300 billion parameters on the device itself. U.S. availability is expected in October.
The memory is LPDDR5X-8533, soldered around the CPU package, according to HP. CPU, GPU and NPU sit on one die and draw from the same pool, which is how a laptop with no discrete graphics card reaches HP's 300B-parameter ceiling. The top configuration runs 16 cores and 32 threads with a 5.2GHz boost clock, adds a 55 TOPS NPU for lighter on-device work, and supports up to 8TB of PCIe 5.0 NVMe storage.
HP frames the system as a self-contained AI workstation for professionals. It ships with Perplexity Computer for Windows and integration with Perplexity's agentic platform, and HP lists Autodesk Revit among the applications the system is tuned for. The chassis measures 17.9mm thick and starts at 4.2 lbs. Pricing has not been disclosed.
The 300-billion-parameter figure comes from HP and AMD, not from independent testing, and it is a claim about memory rather than about speed. A model of that size fits inside 160GB only at reduced numeric precision; the same weights at 16-bit precision would need roughly 600GB. HP has not published token-generation rates, benchmark results or the quantization settings its number assumes, so the ceiling is best read as a configuration limit rather than a measured performance figure.
Why Memory Sets the Ceiling
On a laptop, memory capacity settles which models run locally before chip speed enters the conversation. A 300-billion-parameter model has to fit somewhere, and the 160GB the GPU can address is the number that decides whether it runs with a usable context window beside it.
Capacity also buys context. A model's key-value cache grows with the length of the document or conversation it processes, so a machine that only just fits a 300B model at short context has little room for the long inputs agentic tools generate. AMD says the extra memory is headroom for context as much as for parameter count.
HP's pitch rests on concurrency as much as raw size. HP says the configuration is meant to run multi-agent workflows and professional applications at the same time, a harder test than loading one model in isolation. Keeping a large model resident while a CAD package, an agent framework and a browser session stay open is where a 192GB pool earns its place.
AMD says the 192GB configuration carries 50% more memory than its previous Ryzen AI Max PRO 300 series, the generation behind HP's 14-inch ZBook Ultra G1a. The move to a 16-inch chassis widens the thermal budget too, with the platform rated for up to 100W.
| Specification | HP ZBook Ultra G3a 16 |
|---|---|
| Processor | Up to AMD Ryzen AI Max+ PRO 495, 16 cores / 32 threads, 5.2GHz boost |
| Unified memory | Up to 192GB LPDDR5X-8533, soldered |
| GPU-addressable memory | Up to 160GB |
| NPU | 55 TOPS |
| Storage | Up to 8TB PCIe 5.0 x4 NVMe |
| Local model ceiling | Up to 300B parameters (manufacturer claim) |
| Chassis | 4.2 lbs, 17.9mm |
| Availability | U.S. in October 2026; pricing not disclosed |
The HP ZBook Ultra G3a Against the Cloud
Professionals who need a 300B-class model today choose between cloud inference and a desktop fitted with one or more professional GPUs. Cloud inference carries no hardware cost, bills per token, and sends prompts and data off-site. A desktop keeps everything local, at the price of a machine that does not travel.
The G3a compresses the second option into a 4.2 lb package. For architecture, engineering and construction teams working in Revit on client drawings, or developers handling proprietary code, the practical difference is that sensitive material never has to leave the device. That argument holds whether or not the 300B ceiling survives independent testing, because data residency applies to smaller models just as well.
Shipping Perplexity Computer for Windows alongside the hardware is the other half of the strategy. Preloading an agentic platform ties the workstation's value to software that assumes local inference, in the same pattern PC makers have used to bundle assistant software with AI-branded hardware. Whether buyers want that stack preinstalled is a separate question from whether the silicon can run it.
Where the Trade-Offs Sit
The clearest limit is upgradeability. Because the LPDDR5X sits soldered around the CPU, buyers commit to a memory configuration at purchase. A shared pool of this size also favours capacity over the raw bandwidth dedicated graphics memory provides, which shapes how quickly a 300B model produces tokens once loaded.
The NPU covers the always-on features Windows now expects, while the GPU and its 160GB allocation carry the large models. Splitting the work that way keeps background tasks off the big pool, though it also means AI performance depends on which engine an application chooses.
The processor is not exclusive to HP. AMD's Ryzen AI Max PRO 400 series, built on Zen 5 and codenamed Gorgon Halo, will appear in other vendors' machines, so HP's differentiation rests on chassis design, software bundling and price rather than on the chip alone.
Anyone weighing this against a desktop should compare two numbers: how much video memory the alternative offers, and whether it can be expanded. Consumer graphics cards remain far below 160GB of addressable memory, and workstation cards that reach that range bring their own power and cost requirements.
Battery life is the other unknown. A 100W platform with 192GB of memory running a 300B model will not behave like a thin-and-light ultrabook under load, and HP has not published endurance figures for the configuration.
Price is the missing variable. HP has not published one, and that number decides whether the G3a competes with a cloud subscription or with a desk-bound workstation. Until it appears, the machine's case rests on data residency, a memory configuration no competing laptop matches, and a model ceiling that only HP and AMD have stated.
Why This Matters
The G3a moves an AI workload that previously required server hardware into a laptop bag, and it does so on memory capacity rather than compute. For buyers weighing local against cloud, the decision turns on how much they value keeping data on the machine and what they will pay upfront for it. Memory capacity, not processor speed, is the number to measure against any alternative, and October is when HP is expected to put a U.S. price on it.
Sources
AMD Brings Local AI to Autodesk University 2026
Photo by Jayasahan Hansana on Unsplash
Related Articles
- HP Korea Launches HP IQ On-Device AI Platform and High-Performance Workstations
- AWS Debuts Amazon EC2 P6-B300 Instances Featuring NVIDIA Blackwell Ultra GPUs
- Microsoft Debuts Surface Laptop Ultra and New Reasoning-Focused AI Models
✔Human Verified
Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.