The local-AI hardware conversation tends to start with which graphics card to buy. A new IT Chosun report by Kwon Yong-man makes a different case: for an individual or a small team running a model in the 30-billion-parameter class, the more practical default may be a mini PC with 64GB of shared memory, not a desktop tower with a discrete GPU.
The appliance in the middle
The report places the personal AI appliance between an AI PC and an AI server. Its hardware looks like a mini PC or workstation, but its job is closer to a server: agents and APIs connect to it, not a monitor, keyboard and mouse. That sounds like a small distinction. It changes the purchase.
One appliance can serve several devices and users, cut cloud token bills, reduce the burden of sending sensitive data off-site, and avoid putting a GPU, memory and SSD in every laptop and desktop. The report's argument is consolidation, not raw speed.
The 30B line, and why 64GB clears it
For local use, the report puts the practical model size at roughly 30B parameters. Smaller models can disappoint; larger ones are hard to fit. Two workarounds matter: mixture-of-experts models, which activate only part of the network, and offloading part of a model to system memory, which lets a machine run a model larger than its GPU memory.
Those workarounds do not erase the memory ceiling. A mini PC with 64GB of shared memory can run a 30B-class model at a practical level, according to the report. It may be slower than a discrete-GPU desktop, but the report judges it sufficient for individuals and small teams. Its total power draw can be lower than a single high-end desktop graphics card.
The discrete-GPU desktop still has more peak compute. The catch is capacity. Many consumer graphics cards carry only about 16GB of memory, so a full 30B model may not fit. Adding cards helps, but a PC platform is generally limited to about two GPUs and a workstation to about four. More cards also mean more power, noise and heat.
The distinction to keep straight is capacity versus speed. Shared memory addresses the fit problem; it does not make an integrated GPU as fast as a big discrete card. The report is describing a machine that can hold the model, not a machine that wins token-per-second contests. For agent workflows that queue work in the background, that trade can be acceptable. For latency-sensitive, high-throughput serving, a workstation still has the advantage.
Who is selling what
Nvidia is covering both ends. DGX Spark is an appliance with its own DGX OS; it now offers a 64GB memory option alongside 128GB, and ConnectX-7 can link two to eight units, up from two at launch. RTX Spark targets high-end PCs and workstations and can run Windows. DGX Station, the workstation/server-class machine, now supports Windows too. Nvidia's pitch is not just hardware; it is a software stack and a link fabric that can grow.
AMD's Ryzen AI Max 400 series raises maximum supported memory from 128GB to 192GB, a 50% increase. More memory means larger models can fit. The report also says AMD offers more memory than Nvidia's comparable desktop workstation cards. If capacity is the buying criterion, that is a direct argument.
Intel does not yet have a dedicated product line for this market, according to the report. Its available answer is the Core Ultra Series 3 X-series laptop with a stronger GPU, plus the 32GB Arc Pro B70 workstation card. That is a component strategy, not a turnkey appliance.
Qualcomm is pushing Snapdragon X2 Elite mini PCs. The chip has a strong CPU and NPU, but weaker GPU performance than rivals; the report notes that software support can hide some of that gap. The fit is best where the workload leans on CPU or NPU, not where GPU throughput is the whole game.
PC makers are turning these parts into packaged categories. HP's AI Station spans 10B to 600B models. Dell's deskside agentic AI is presented as scaling from the desk to the data center and includes code, research and private assistant solutions. Asus has the Ascent QN10 with Snapdragon X2 Elite, the NUC 16 Pro with Core Ultra Series 3 X-series, and DGX/RTX Spark mini PCs. MSI plans RTX Spark mini PCs. Corsair has the AI300 with Ryzen AI Max.
How to choose without overbuying
The report's advice is to start with model size and the performance it needs, then weigh installation space and power. The real fork is between a high-performance desktop or workstation and a mini PC with an integrated GPU and large shared memory.
That is a useful correction to spec-sheet shopping. A 192GB machine is not automatically better than a 64GB machine if the work is a 30B model for one user; the extra memory buys headroom for larger models, not free speed. A 16GB graphics card with a fast GPU is not automatically better either if the model does not fit without offload.
Two checks matter before buying. First, software: the appliance category depends on runtimes and agents that actually support the chip and operating system. Second, measurement: the report gives a rule of thumb, not a benchmark matrix. It does not test specific 30B models on each machine, so 64GB should be treated as a practical threshold, not a guarantee for every model, quantization or serving setup.
The broader shift is easy to miss if you read the market as a GPU race. The personal AI appliance is winning attention because it changes the unit of purchase: one box serves many devices, keeps data closer, and is sized around the model rather than the monitor. For a 30B-class local workload, 64GB of shared memory is the number that makes that box real. The discrete-GPU desktop remains the faster tool. It is just no longer the obvious default.