local AI's biggest problem... they solved it?

Summary

NetworkChuck reviews WiCi One, a sponsored early-engineering product built around an NVIDIA RTX 5090, a mini PC running OpenWRT, and a 4 TB SSD. Its central idea is to make a powerful GPU available to multiple nearby devices over Wi-Fi, as if each device had its own GPU. The video tests gaming, creative applications, and local AI.

Wireless GPU and gaming

A laptop can see the remote RTX 5090 through a driver and use it for tasks such as playing Cyberpunk 2077 and running Blender. The host reports about 92 FPS in Cyberpunk. He emphasizes that this is not cloud gaming or remote desktop: the game runs on the laptop, while the GPU is accessed over Wi-Fi.

Wi-Fi has far less bandwidth and much higher latency than PCIe, so WiCi’s software protocol is designed to reduce communication. It bundles and compresses instructions, sends only changed data, avoids waiting for some replies, and returns compressed video. The host reports roughly 400 Mbps of Wi-Fi traffic during gameplay.

He also notes the system’s limitations: performance depends on the network, and the tested unit is an early engineering model.

Running a large local model from SSD

The main AI demonstration runs DeepSeek 4.1 Flash, described as a 750-billion-parameter model, on a system with 32 GB of GPU memory. The video explains why this is unusual: running the model at full quality would require more than 500 GB of memory, so much of its data must reside on SSD rather than in VRAM.

The video highlights several techniques WiCi uses to make this workable:

• Optimized model-file layout: Experts are stored in aligned chunks so the SSD can read them more efficiently. WiCi says this raises transfer speeds from around 1 GB/s to about 11.5 GB/s with its approach.
• Mixture-of-experts behavior: DeepSeek has 384 experts per layer and selects a small subset—described as six—for each token. Only some model weights are needed for each generated token, though the full model still has to be stored.
• Caching popular experts in system memory: Frequently used experts can be served from RAM instead of repeatedly fetched from SSD.
• Split processing: Work is divided between the CPU and GPU based on where the relevant weights are available.

The host’s live visualization shows experts being selected and data moving among SSD, RAM, and GPU memory. Generation varies across the demonstrations: one run reaches about 15 tokens per second, others are around 3–4 tokens per second, and the host characterizes the overall experience as roughly 8 tokens per second. These are demonstration results, not a guarantee for other workloads or setups.

Sharing the GPU between apps

WiCi also uses a scheduler to manage GPU sharing. When switching between AI applications or models, it can park a model on the SSD and restore it when needed rather than fully unloading and reloading it. In the host’s demonstrations, switching to an image model takes about 8.6 seconds with the scheduler versus roughly 18 seconds without it. Returning to a chat model takes about 4 seconds versus around 17 seconds.

The Mac demonstration also uses the GPU for image and video generation and DaVinci Resolve rendering. The host describes this as a promising way for laptops, phones, and other devices to share one powerful local GPU, while noting that the product and software are still early.

Overall assessment and guide content

The video is an explainer and product test, rather than a step-by-step setup guide. It walks through the bandwidth and memory bottlenecks behind remote GPUs and large local models, then demonstrates WiCi’s proposed solutions. The host is enthusiastic about the technology but presents it as an early product and a developing approach to consumer local AI. WiCi One is available in RTX 5090 and less expensive RTX 5060 Ti editions.

Main speakers and sources: NetworkChuck (host and demonstrations) and a WiCi team member, who briefly describes the idea’s origins in cloud gaming.