HP has confirmed a mobile workstation built to run large language models entirely offline, and the price tag makes clear this is not a mainstream laptop. The ZBook Ultra G3a, detailed by TechRadar on September 25, 2026, pairs AMD’s newest flagship mobile chip with 192GB of unified memory and a claimed ability to load AI models as large as 300 billion parameters without touching the cloud. HP has since launched the 16-inch ZBook Ultra G3a in October 2026 with prices from $3,149, and the Japanese launch that month is priced at ¥639,100, the first real retail figures to accompany the earlier $7,449 estimate the company’s own Local AI Value Calculator had generated for the top US configuration.
The announcement lands in the middle of two collisions the PC industry can’t avoid discussing right now: a scramble among laptop makers to build machines that keep AI inference on the device instead of a data center, and a memory market where DDR5 prices have risen so fast that laptop vendors are raising sticker prices mid-cycle. HP’s new machine sits at the intersection of both stories.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What HP Actually Announced
According to TechRadar’s report, the ZBook Ultra G3a’s top-tier configuration pairs AMD’s Ryzen AI Max+ PRO 495 processor with 192GB of memory and a 512GB SSD, priced at an estimated $7,449. HP also lists a 128GB version of the same chip at an estimated $5,999, and a step-down configuration built around the Ryzen AI Max PRO 490 with 64GB of memory at roughly $3,899. TechRadar was careful to note these are calculator-generated estimates rather than confirmed retail prices, and HP has not released a separate technical brief covering the laptop’s display, battery capacity, port selection, or weight.
That gap in the public spec sheet is starting to close. HP’s Japanese product listing, published alongside the October 2026 launch, puts the 16-inch ZBook Ultra G3a at roughly 1.9kg and 18.4mm thick, giving buyers their first concrete sense of the chassis carrying that 192GB memory pool. Screen resolution, battery life, and full thermal design details still haven’t surfaced for the US market, though, so buyers considering a $7,449 mobile workstation are left piecing together the full picture from regional listings rather than a single global spec sheet.
Inside the Ryzen AI Max+ PRO 495
The processor driving HP’s pitch is AMD’s newest member of the Ryzen AI Max PRO 400 family, built on the Zen 5 architecture. Per AMD’s own published specifications (cited in ACEMAGIC’s September 3, 2026 announcement of a rival mini workstation using the same silicon), the Max+ PRO 495 carries 16 Zen 5 cores and 32 threads with boost clocks up to 5.2GHz, integrated Radeon 8065S graphics with 40 RDNA 3.5 compute units, and an XDNA 2 neural processing unit rated at up to 55 TOPS. HP’s own October 2026 Japanese product listing for the ZBook Ultra G3a corroborates those same 16-core, 32-thread, 5.2GHz-boost figures for the laptop itself, not just for AMD’s reference silicon. AMD lists the total platform AI throughput at up to 132 TOPS when the CPU, GPU, and NPU are combined. The chip supports up to 192GB of LPDDR5X-8533 memory across a 256-bit bus, PCIe 4.0, and USB4.
That 55 TOPS NPU figure is a meaningful, if incremental, step over the 50 TOPS rating AMD assigned to earlier Ryzen AI Max PRO 400-series parts, a jump previously covered when the Ryzen AI Max PRO 400 platform first shipped with the same 192GB memory ceiling. The “+” in the Max+ PRO 495 name doesn’t signal a different core count or a second CPU die: the practical differences between 400-series variants come down to graphics compute unit count, NPU clock, and how aggressively a given OEM configures sustained power and cooling.
| Processor | CPU Cores/Threads | Integrated GPU | NPU (TOPS) | Max Memory |
|---|---|---|---|---|
| Ryzen AI Max PRO 490 | 16 / 32 | Radeon 8060S-class | Up to 50 | Up to 192GB |
| Ryzen AI Max PRO 395 | 16 / 32 | Radeon 8060S-class | Up to 50 | Up to 192GB |
| Ryzen AI Max+ PRO 395 | 16 / 32 | Higher-config Radeon (platform-dependent) | Up to 50 | Up to 192GB |
| Ryzen AI Max+ PRO 495 | 16 / 32 | Radeon 8065S, 40 CUs, up to 3.0GHz | Up to 55 | Up to 192GB |
Source: AMD official specifications, as reported by PRNewswire and TechRadar, September 2026.
The 300B-Parameter Claim, Checked Against the Math
TechRadar’s headline figure, that the ZBook Ultra G3a “can handle up to 300B parameters,” is the kind of number that needs a second look before anyone buys a machine expecting to run a dense 300-billion-parameter model at full precision. A model of that size requires roughly 600GB of memory just to hold 16-bit weights, around 300GB at 8-bit quantization, and approximately 150GB at aggressive 4-bit quantization, before accounting for context windows, activations, and operating system overhead.
A 192GB system therefore sits right at the edge of what a heavily quantized 300B-class model needs, with little headroom left for a large context window or multiple running applications. What the ZBook Ultra G3a genuinely enables is comfortable local inference on 70B-to-140B models with room to spare, and the theoretical ability to load some 300B-class models if a buyer is willing to accept aggressive quantization and a tight memory budget. AMD’s own footnote frames it the same way: as of 5/11/2026, the Ryzen AI Max+ 495 PRO supports up to 160GB of dedicated graphics memory, which it describes as capable of running 300 billion+ parameters at 4-bit quantization, a figure that refers to memory allocated for graphics within the shared pool rather than 160GB of discrete VRAM. That’s still a real capability few laptops on the market can match, but it is a narrower claim than the marketing framing suggests.
How It Stacks Up Against Apple’s MacBook Pro Line
The most direct rival for a portable, large-memory AI machine is Apple’s MacBook Pro line. Apple’s M4 Max configuration topped out at 128GB of unified memory, meaning HP’s 192GB ceiling gives it a real capacity advantage for anyone trying to load the largest possible open-weight models locally. Apple’s counterargument has always been software maturity: Metal, Apple’s MLX framework, and a large ecosystem of Mac-tuned inference tools give the MacBook Pro a smoother experience running models that fit inside its memory budget, even if that budget is smaller.
AMD’s side of that equation depends on ROCm, Vulkan, DirectML, and community-maintained llama.cpp backends, all of which have narrower and less consistent optimization coverage than Apple’s first-party stack. Raw TOPS figures also don’t translate cleanly into real-world tokens-per-second: memory bandwidth, quantization format, and how well a given inference engine is tuned for a chip matter more than the peak number printed on a spec sheet. The honest framing is that HP wins on capacity, Apple wins on polish, and neither claim is proven without independent, apples-to-apples benchmarks using identical models and quantization settings on both machines.
Where Nvidia’s DGX Spark and RTX Mini PCs Fit
Nvidia’s DGX Spark occupies a different niche: a compact desktop system built around CUDA, TensorRT-LLM, and a mature PyTorch deployment pipeline. For developers already locked into Nvidia’s software stack, DGX Spark is generally the lower-friction path, even though it requires external power, a monitor, and peripherals rather than folding into a battery-powered laptop.
Discrete RTX-based mini PCs and gaming laptops split by VRAM tier. Cards with 8GB or 16GB of dedicated memory handle small-to-medium models well but struggle with 30B-and-up models without offloading to system RAM, which HP’s recent 24GB VRAM push was partly designed to address. A 24GB RTX 5090-class laptop GPU will generally out-throughput the ZBook Ultra G3a on any model that fits inside its VRAM, because dedicated GDDR memory is simply faster than shared LPDDR5X. The ZBook’s edge only appears once a model is too large for that VRAM ceiling and would otherwise require slow CPU offloading. In short: RTX systems win on speed for models that fit, HP wins on capacity for models that don’t.
An OEM Race Nobody Announced Officially
HP isn’t alone in building around this exact chip. ACEMAGIC announced what it called the first Ryzen AI Max+ PRO 495 mini workstation at IFA 2026 on September 3, 2026, also spec’d with 192GB of LPDDR5X-8533 memory and the same 55-TOPS NPU rating. Three weeks later, Windows-focused outlets reported that Thunderobot was teasing a “Station” workstation cluster built on the same silicon, following an earlier Thunderobot launch of the AI Master M7000, a system pairing the smaller Ryzen AI Max+ 395 chip with 64GB of memory for 120B mixture-of-experts models, according to TechRadar’s September 25 report.
Three vendors building around one chip family inside a single month is a signal worth taking seriously, even without an analyst forecast to attach a market-size number to it. It suggests AMD’s 400-series platform has become the default reference design for OEMs chasing the “AI PC that doesn’t need the cloud” pitch, the same way a handful of reference chipsets tend to define a hardware category before it matures into dozens of differentiated products.
The Memory Price Backdrop HP Can’t Avoid
HP’s timing collides with an increasingly ugly memory market. Tom’s Hardware reported on September 25, 2026 that laptop maker Schenker said DDR5 SO-DIMM memory had become roughly six times more expensive than a year earlier, and announced its own price increases: roughly $113 (€100) more on many gaming laptop models, $228 (€200) more on others, and a $342 (€300) jump on one professional laptop line.
That wasn’t an isolated data point. A separate Tom’s Hardware report from August 17, 2026 found memory prices had climbed roughly 500% over 12 months, with some 128GB DDR5 desktop kits reaching about $3,399, up to 10 times the lowest price ever tracked for that capacity. A month earlier, on July 28, 2026, Tom’s Hardware listed the going rate for a 16GB DDR5 kit at $209 and a 32GB kit at $379, prices that now look almost quaint next to the 128GB figure.
| Memory Product | Reported Price | Date Reported | Source |
|---|---|---|---|
| 16GB DDR5 kit | $209 | July 28, 2026 | Tom’s Hardware |
| 32GB DDR5 kit | $379 | July 28, 2026 | Tom’s Hardware |
| 128GB DDR5 kit | $3,399 (up to 10x lowest-ever price) | August 17, 2026 | Tom’s Hardware |
| DDR5 laptop SO-DIMMs | ~6x pricier than a year earlier | September 25, 2026 | Tom’s Hardware / Schenker |
| HP ZBook Ultra G3a, 192GB config | ~$7,449 (estimated) | September 25, 2026 | TechRadar / HP calculator |
This is why the ZBook Ultra G3a’s memory architecture is a double-edged detail. LPDDR5X is soldered to the board rather than sold as replaceable SO-DIMMs, which is standard for this class of unified-memory chip and part of what makes 192GB possible in a laptop chassis at all. But it also means buyers commit to a memory configuration at purchase time, with no path to upgrade later if a heavier model comes along, right as the underlying memory market is in the middle of its steepest price run in years. The same dynamic has already pushed handheld gaming PC prices higher, as covered in Valve’s own DRAM-driven Steam Deck price increase and the broader memory crisis Acer flagged across its consumer laptop lineup.
Why Companies Want AI Off the Cloud Meter
The economic argument for local inference is straightforward even without a hard industry-wide cost-per-token figure to cite. Running inference on owned hardware avoids per-token cloud billing, keeps sensitive data inside the organization’s own walls, removes network latency from the response loop, and keeps work running during an outage. None of that is free: electricity, cooling, IT support, and hardware depreciation are real costs that a $7,449 machine has to earn back over its service life.
The calculation tends to favor local hardware for high-volume, repetitive, or privacy-sensitive workloads, such as a legal team running document review against a fixed model all day, and to favor cloud access for bursty or rapidly changing workloads where renting scales better than owning. HP pricing the ZBook Ultra G3a like a professional workstation rather than a premium consumer laptop suggests the company is chasing that first category of buyer, not casual enthusiasts experimenting with open-weight models on the side.
Historical Context: From Copilot Branding to Bare Memory Numbers
It’s worth noting how quickly the PC industry’s AI marketing has shifted register. Two years ago, the pitch was a branding exercise: “Copilot+ PC” stickers on laptops that met a minimum NPU threshold, regardless of whether buyers ran any AI workload at all. Tom’s Hardware’s September archive notes Microsoft has since quietly dropped Copilot+ branding from some of its newest Surface laptops, even as those devices continue to meet the same underlying hardware requirements.
HP’s approach with the ZBook Ultra G3a is the opposite instinct: instead of a badge, the pitch is a specific, checkable number, 192GB of memory and a defined maximum parameter count. That’s a more falsifiable claim than a branding sticker, which is partly why the caveats around what “300B parameters” actually means in practice matter so much here. The industry moved from vague AI branding to concrete memory specs in roughly two years, and the ZBook Ultra G3a is a useful marker of how far that shift has gone.
Market Impact: A Premium Niche, Not a Mainstream Shift
At an estimated $7,449 for the flagship configuration, the ZBook Ultra G3a is priced well outside what most individual buyers or small teams will spend on a laptop, AI capability or not. Its immediate market is enterprise IT departments and AI development teams evaluating whether owning hardware beats renting cloud GPU time for specific, well-understood workloads. The fact that AMD’s 400-series platform is showing up simultaneously in HP’s professional workstation, ACEMAGIC’s mini PC, and Thunderobot’s cluster-oriented system suggests vendors see a genuine commercial opening in that enterprise segment, even without third-party analyst figures yet to size it.
For the broader PC market, the more immediate and more painful story remains the memory shortage. Tom’s Hardware’s September 4, 2026 report on Jon Peddie Research data found discrete GPU shipments hit a four-year record of 13.24 million units despite high prices, with desktop GPU shipments down 4% quarter-over-quarter while notebook GPU shipments surged 16.8%, evidence that demand for AI-capable and gaming hardware is holding up even as component costs climb. That tension, rising demand meeting rising memory costs, is the real backdrop against which HP’s high-memory pitch has to be judged. It’s the same memory-capacity-versus-cost trade-off already playing out at data-center scale, as seen in Majestic Labs’ 128TB Prometheus server design, just shrunk down to a single laptop chassis.
What to Watch Next
- HP has now confirmed an official Japan launch for the ZBook Ultra G3a in October 2026 at ¥639,100 and US launch prices from $3,149, though the $7,449 figure for the top configuration remains the calculator-based estimate TechRadar reported rather than a confirmed US list price; availability still varies by configuration and country.
- DDR5 and LPDDR5X prices show no sign of stabilizing based on Tom’s Hardware’s tracking through September 2026, which likely means the $7,449 estimate could shift as more configurations reach more markets.
- ACEMAGIC and Thunderobot’s parallel launches around the same Ryzen AI Max+ PRO 495 platform suggest more OEMs will announce similarly specced machines before the end of 2026, turning this into a genuine product category rather than a one-off HP experiment.
- Independent, apples-to-apples benchmarks comparing the ZBook Ultra G3a against an M4 Max or M5 MacBook Pro on identical quantized models haven’t been published yet; that comparison will be the real test of HP’s capacity claims.
- Whether enterprise IT buyers actually choose owned hardware over scaling cloud GPU rental will depend more on utilization rates and internal cost accounting than on any single spec sheet, and that verdict will take quarters, not weeks, to become clear.
The Bottom Line
The ZBook Ultra G3a is a real, specific product built around a real, specific chip, not vaporware or a rebranded reference design. Its 192GB memory ceiling genuinely sets it apart from most laptops on the market, including Apple’s MacBook Pro line, and its price reflects that it’s aimed at enterprise buyers rather than consumers. But the “300B parameters” framing needs the context this article laid out: it describes a theoretical ceiling under aggressive quantization, not a smooth, general-purpose experience running the largest open models at full precision. Buyers evaluating this machine, or any of the AMD 400-series-based competitors now following it to market, should treat the memory number as the headline feature and the parameter-count marketing as the asterisk underneath it.
Frequently Asked Questions
What is the HP ZBook Ultra G3a?
It’s a mobile workstation laptop from HP, reported by TechRadar on September 25, 2026, built around AMD’s Ryzen AI Max+ PRO 495 processor with up to 192GB of unified memory, aimed at running large AI models locally rather than through a cloud service.
How much does the HP ZBook Ultra G3a cost?
HP has published an official retail price for Japan, where the 16-inch ZBook Ultra G3a launches in October 2026 at ¥639,100. In the US, HP’s launch pricing starts from $3,149 for a base configuration with the Ryzen AI Max Pro 480, 16GB of RAM and 512GB of storage; TechRadar separately reported calculator-generated estimates of $7,449 for the 192GB/512GB configuration, $5,999 for a 128GB version, and roughly $3,899 for a Ryzen AI Max PRO 490 model with 64GB, none of which HP has confirmed as final.
Can the ZBook Ultra G3a really run a 300-billion-parameter AI model?
Only under aggressive quantization. A 300B-parameter model needs roughly 150GB of memory at 4-bit precision, which leaves little headroom in a 192GB system once context and overhead are included. The laptop is much better suited to comfortably running models in the 70B-to-140B range.
How does the Ryzen AI Max+ PRO 495 compare to earlier Ryzen AI Max chips?
It keeps the same 16-core, 32-thread Zen 5 CPU design as earlier 400-series chips but steps up to a 40-compute-unit Radeon 8065S GPU and a 55-TOPS NPU, versus roughly 50 TOPS on prior Ryzen AI Max PRO parts, according to AMD’s published specifications.
Is the HP ZBook Ultra G3a better than a MacBook Pro for AI work?
It depends on the workload. HP’s 192GB memory ceiling exceeds Apple’s M4 Max maximum of 128GB, giving it more headroom for very large models. Apple generally offers a more mature and better-optimized software stack for the models that do fit inside its memory budget.
Why are AI laptops getting so expensive right now?
Memory prices have surged industry-wide. Tom’s Hardware reported DDR5 prices climbing roughly 500% over 12 months as of August 2026, and laptop maker Schenker said DDR5 SO-DIMMs had become about six times pricier than a year earlier as of its September 25, 2026 statement. Laptops like the ZBook Ultra G3a that rely on large amounts of soldered LPDDR5X memory are especially exposed to that price pressure.
Are other companies building laptops or mini PCs with the same chip?
Yes. ACEMAGIC announced a mini workstation using the Ryzen AI Max+ PRO 495 at IFA 2026 on September 3, 2026, and Thunderobot was reported teasing a cluster-oriented “Station” system on the same platform in late September 2026, alongside its own AI Master M7000 built on the related Ryzen AI Max+ 395 chip.
When can I actually buy the ZBook Ultra G3a?
HP has announced a confirmed retail availability date for Japan: the 16-inch ZBook Ultra G3a launches there in October 2026 at ¥639,100. HP has also launched the laptop in the US in October 2026 with prices from $3,149, though availability varies by configuration and country; the $7,449 figure for the top configuration comes from TechRadar’s September 25, 2026 coverage of HP’s announcement and remains a calculator-based estimate.
