Tech • AI • Robotics • Game
Local AI's Biggest Problem... They Solved It?
7/10TL;DR
WiCi is introducing a wireless external compute box built around an NVIDIA RTX 5090 that lets laptops and Macs use desktop-class GPU power over Wi‑Fi, while also enabling surprisingly large local AI workloads through aggressive data, memory and scheduling optimizations.
KEY POINTS
A GPU that appears local over Wi‑Fi
WiCi packages an RTX 5090, a mini PC and OpenWRT networking into a single box that can present the GPU to another device as if it were installed locally. In testing, a laptop recognized the card in its task manager and used it to run Cyberpunk 2077 and Blender, while a Mac simultaneously accessed the same hardware for AI and creative workloads. The concept is not standard remote desktop or cloud gaming: the front end remains on the user’s device while graphics compute is offloaded wirelessly.
Why wireless GPU use is normally impractical
A conventional GPU depends on PCIe, which offers roughly 512 gigabits per second and sub-microsecond response times. By contrast, a Wi‑Fi link may deliver only about 1 gigabit per second with round-trip latency around 1.5 milliseconds under good conditions. That gap is crippling because games and graphics apps exchange thousands of small, time-sensitive messages per second; if every command had to wait on Wi‑Fi, frame rates would collapse.
How WiCi cuts the network bottleneck
To make the setup usable, WiCi installed what amounts to a virtual GPU layer and built its own transport protocol. The system bundles and compresses commands, sends only changed data, and predicts some responses so they can be delivered before they are explicitly requested. It then sends compressed video output back to the client device. Those changes reportedly reduce traffic that would have required about 8 gigabits per second to roughly 400 megabits per second, enough to sustain about 92 frames per second in gaming tests.
Local AI is the bigger target
The gaming demo is eye-catching, but the company’s larger ambition is local AI. Consumer GPUs with 32GB of VRAM typically cannot host frontier-scale open models at full quality. One example highlighted was DeepSeek 4.1 Flash, described as a 750 billion parameter model that would normally require more than 500GB of memory, far beyond what a single desktop GPU can provide.
Using SSD storage as a substitute for missing memory
WiCi addresses that limit by leaning on a 4TB SSD as an extension of the memory hierarchy. On paper, that sounds unworkable: GPU VRAM can move data at around 1.8TB per second, while an SSD may top out near 14GB per second in ideal large transfers and much less for tiny reads. For large language models, where data movement often determines speed, naive SSD offloading would make generation effectively unusable.
Mixture-of-experts makes the problem solvable
The company focuses on mixture-of-experts models, where only part of the model is active for each token. In the DeepSeek 4.1 Flash example, only 6 experts out of 384 are selected per token, reducing the amount of data needed each step from roughly 508GB to around 15GB. That still leaves storage throughput as a major challenge, but it creates a path that dense models do not.
Four optimizations drive usable token speeds
WiCi says it rewrites model files so expert blocks are equal-sized and aligned for fast SSD access, allowing reads of 64 chunks at a time and lifting effective throughput from about 1GB/s to around 11.5GB/s. It also caches popular experts in RAM, meaning only about 1GB of experts may need to come from SSD for many tokens. Finally, it splits the math between CPU and GPU depending on where the weights are loaded from. In demonstrations, that produced roughly 8 to 15 tokens per second in favorable runs, with some prompts closer to 3 to 4 tokens per second.
A scheduler aims to turn one GPU into a shared household resource
Another obstacle is that GPUs are poor at multitasking across large AI models because each model can occupy most or all of available VRAM. WiCi’s answer is a scheduler that “parks” inactive models on SSD so they can be restored faster than a full reload. In side-by-side tests, switching an image model without the scheduler took about 18 seconds; with parked copies, the same task dropped to about 8.6 seconds. Returning from image generation to chat fell from about 17 seconds to around 4 seconds.
Early hardware with commercial ambitions
The device shown is an early engineering model, delivered in rugged packaging and described as already available to order. WiCi is offering versions based on the RTX 5090 and a cheaper RTX 5060 Ti, with shipping estimated at about eight weeks. The company’s pitch is a “family GPU” that can serve laptops, tablets, phones and Macs as a centralized local compute resource.
CONCLUSION
WiCi is attempting to turn a high-end GPU into shared wireless infrastructure for gaming, creative apps and local AI. If its networking, storage and scheduling techniques hold up beyond early demos, it could broaden access to desktop-class compute without putting a large GPU in every machine.
Ask a question