I just saw NVIDIA's event at Computex 2026 where they unveiled their RTX Spark computers which houses a CPU-GPU combo on a single chip with up to 128 GB unified memory, primarily aimed at running local AI inferencing and may be some training as well.

Personally, I would like to run local AI models on my PC without compromising on speed or accuracy. Unfortunately, the biggest limiting factor in my case is VRAM on my RTX 5070 GPU. I understand that these GPUs are optimized for running video games and so it is not logical to expect these cards to have 48 GB VRAM. While APUs with unified memory can be a solution, but that would also mean that we would lose the flexibility to upgrade RAM and GPU on our system besides limiting performance due to thermal throttling.

Hence, I feel the only practical solution for desktop PC users who would want to run local AI also besides gaming is to add an AI accelerator card with some 32 to 48 GB VRAM on motherboard using PCIe slot. But for this, at first, we will need,

(a) Motherboards that would support 2 PCIe x16 slots and

(b) Affordable AI accelerator cards with sufficient RAMs

I would love to know how you people feel about it.