HP ZBook Ultra G3a Economics: Can a $7,449 AMD Laptop Replace Cloud AI?
The ZBook Ultra G3a can load a model with 300 billion parameters. Its memory runs too slowly to use one heavily, and the same open models cost less than $1,000 a year to rent in the cloud. Regulated firms that cannot send data off the device should buy it. Companies trying to cut token bills should not.
HP is selling its new ZBook Ultra G3a as a way to stop paying for cloud AI. The top configuration pairs AMD's Ryzen AI Max+ PRO 495 with 192 GB of unified memory. HP and AMD say it can run language models with up to 300 billion parameters on the laptop. It ships in October at $7,449, according to HP's own cost calculator. The capacity claim holds up. The savings claim does not. The laptop is worth buying for legal teams, defence suppliers and engineering firms that are not permitted to send data to a cloud. For a company trying to reduce its token bill, it is poor value.
Speed depends on memory bandwidth, and here the specification works against the machine. A language model has to read its active weights from memory for every token it produces. The Gorgon Halo chip uses a 256-bit memory bus. HP's own footnotes say the 128 GB and 192 GB configurations run memory at 8,000 MT/s, not the 8,533 MT/s available on smaller configurations. That gives about 256 GB/s. A dense 300-billion-parameter model compressed to four bits takes up about 150 GB, so the laptop could produce fewer than two tokens a second. Only mixture-of-experts models, which activate about 20 billion parameters per token, run at a usable speed. A 235-billion-parameter model with 22 billion active parameters reads about 12 GB per token, which puts the ceiling near 20 tokens a second and real-world output, once overheads bite, at roughly 10 to 12. That is fast enough for one person using a chat window, and too slow for an agent that works through tokens nonstop.
HP's savings case overlooks that limit. According to TechRadar, HP's Local AI Value Calculator assumes a heavy user processes about 6.3 million tokens a day. That averages 73 tokens a second around the clock. Generating output at 10 to 12 tokens a second, the laptop cannot reach that volume with a large model, even if it runs day and night. The heavy-use scenario that makes the payback look attractive requires more work than the hardware can do.
The relevant comparison is with cloud prices. A local user is running open-weight models, so the fair comparison is renting those same models, not paying for a closed frontier model. On Fireworks' serverless price list, DeepSeek V4 Flash costs $0.22 per million input tokens and $0.66 per million output tokens. MiniMax M3 costs $0.30 and $1.20. Apply HP's 6.3 million tokens a day with a typical four-to-one split between input and output, and the bill is $2 to $3 a day, or about $700 to $1,100 a year. The 192 GB laptop costs $3,550 more than the 64 GB configuration, which HP prices at $3,899 with the Ryzen AI Max PRO 490. The extra memory alone takes three to five years to pay back at cloud prices, and the laptop will be out of date in less time than that.
The comparison with frontier models changes the answer, and it is the strongest argument for HP. Moonshot's Kimi K3 costs $3 per million input tokens and $15 per million output tokens on the same list. At HP's usage profile, that comes to about $34 a day, or more than $12,000 a year, so the laptop would pay for itself in months. Chinese open-weight models are also close enough to be useful. Larridin, which measures enterprise AI use, told Fortune that they handle about three-quarters of engineering tasks reasonably well. But the model that fits in 192 GB and runs quickly enough is a mixture-of-experts model smaller than Kimi K3, served far more slowly than a cloud provider would serve it. Anyone who needs frontier quality will still pay for frontier quality. HP and Perplexity have designed for this: the preinstalled Perplexity software starts tasks locally and sends them to a cloud model when needed. That is a hybrid product, and in a hybrid product the local hardware reduces the cloud bill without removing it.
The competition shows where the product fits. Apple's new Mac Studio with an M5 Ultra and 256 GB of memory has bandwidth of up to 1.2 TB/s, almost five times the ZBook's, but costs $11,299 with 2 TB of storage and stays on a desk. Nvidia's DGX Spark has 128 GB of memory and similar bandwidth for $4,699, and runs Linux on Arm. HP's advantage is that it is an x86 Windows laptop that IT departments already know how to buy, secure and service, with ISV certifications for Revit and 3ds Max. It is a workstation that can carry a large model with it.
That points to the buyers: law firms that are not allowed to upload discovery documents, defence engineers working on air-gapped networks, and architects whose clients prohibit cloud processing of drawings. They value data control more than tokens per second. For them, $7,449 is cheap compared with a compliance failure. AMD benefits whatever happens: each of these machines raises the average selling price of a PC chip and moves AI spending that would have gone to data-centre GPUs into its client business.
The CIO's question is simpler than HP's calculator makes it. If the data can go to a cloud, rent the open model there. It is faster and cheaper, and it gets better without anyone buying new hardware. If the data cannot leave the device, the ZBook is the most portable option available, and the price is secondary.
The desk expects independent reviews after the October ship date to confirm usable speeds for mixture-of-experts models between 100 billion and 235 billion parameters. We expect them to show that no dense model close to 300 billion parameters runs at a speed anyone would use for work. We also expect the 128 GB configuration, not the 192 GB one, to account for most corporate orders through the first half of 2027. That will show buyers treating local AI as a matter of governance rather than of cost.