WiCi Puts a GPU on Wi-Fi 7 to Run 800B Models at Home

What WiCi is actually building

WiCi describes itself as a systems research lab, not a gadget startup. Its focus is wireless GPU computing and local inference for frontier AI, with an emphasis on keeping models and data on your own network.

The core claim is simple to state and hard to pull off: a GPU sitting nearby should feel local to the application using it. WiCi says its runtime cuts network overhead through caching, data reuse and asynchronous execution, so software keeps running on the device while the heavy math happens on a GPU across the room.

That matters because GPU software is chatty. It makes thousands of small calls, and over Wi-Fi each one carries latency that adds up fast. WiCi lists network latency, limited memory, competing workloads and real-time response as the four constraints it is attacking.

The lab also says it is working on scheduling so that independent applications stop loading duplicate model weights into GPU memory, sharing weights while keeping each app's state separate.

WiCi One: price, storage and what it runs

InferSSD: 800B parameters on consumer hardware

The most concrete piece of WiCi's research is InferSSD, an open-source engine that keeps expert weights on an SSD and streams the working set to the GPU.

Large models can outgrow both GPU memory and system RAM. InferSSD sidesteps that by treating the SSD as the storage tier for expert weights and moving only what is needed into the GPU. WiCi says it has results running 800B-parameter LLMs on Apple Silicon and NVIDIA systems.

The WiCi One ships with a 4TB PCIe 5.0 SSD built into the enclosure, which is the detail that makes the streaming approach plausible on a desktop rather than a data center rack. Fast local storage is doing work that used to require a bigger memory budget.

There is a familiar trade-off hiding here: streaming weights from an SSD is slower than holding them in VRAM. WiCi frames it as the difference between running a model at all and not running it, which is a fair pitch for people who want frontier-scale models on hardware they own.

The WiCi One and the voice latency problem

WiCi's first product, the WiCi One, combines the wireless GPU runtime and SSD inference in one local system. Pricing starts at US$2,599.

The company positions it for homes, studios and workplaces, with a variety of devices connecting to a dedicated AI inference box on the local network. Marketing quotes on the site describe it as offering high-end graphics power over Wi-Fi 7 to any compatible device, and as pushing wireless external graphics beyond conventional eGPU designs.

WiCi's voice research is the other half of the story. Instead of waiting for a user to finish speaking, the system starts forming candidate answers while speech is still arriving, then uses a learned timing policy to decide when to respond. The lab's own framing, 'answering is not buzzing,' is a dig at assistants that acknowledge you instantly and then make you wait.

Developers are told they can reach local GPU compute through familiar APIs, SDKs and supported driver integrations, which is the line that will decide whether this stays a research demo or becomes infrastructure.

The catch: Wi-Fi is still Wi-Fi

Nothing in WiCi's material claims wireless matches a GPU on a PCIe slot, and that gap is the whole question.

A local runtime can hide latency for bursty, cacheable work. It cannot hide physics. Wi-Fi 7 helps, but interference, walls and competing traffic are part of every home network, and a shared GPU serving several devices has to divide its attention.

The $2,599 starting price also puts the WiCi One in competition with a well-specced workstation that keeps its GPU inside the case. The pitch only lands if sharing one GPU across laptops, tablets and assistants is worth more to you than raw local speed.

WiCi has published research write-ups on its runtime, InferSSD and voice timing, but independent benchmarks are not in yet. Treat the 800B-parameter claim as the lab's own result until outside testing appears.

Availability and pricing for US, UK and EU buyers

WiCi lists the WiCi One at US$2,599 but does not state shipping regions, UK or EU pricing, or a release date on the material available. Buyers outside the US should expect import duties and VAT to be added on top of the dollar figure, and should confirm local availability before planning a purchase.

The software side is easier to try: InferSSD is described as open source, so the SSD-streaming approach can be tested on hardware you already own, including Apple Silicon and NVIDIA systems.

WiCi is worth watching for the research rather than the price tag. If the runtime and InferSSD hold up outside the lab, the idea of one GPU serving every device in a home stops sounding strange, and the PC you already own gets a longer, more useful life.