WiCi has unveiled the WiCi One, an external GPU box designed specifically for running large language models locally. It connects over Wi-Fi 7 rather than Thunderbolt or PCIe, and includes a 4TB PCIe 5.0 SSD inside the enclosure. According to Tom's Hardware, the bundled driver makes the remote card appear as a normal GPU to client machines, while the SSD stores model weights to reduce cross-network transfers.
The wireless approach requires careful handling of GPU traffic. Standard host-to-card communication is extremely chatty, so the company says its driver performs data caching, deduplication, and compression, and decides which hardware requests can be handled locally versus sent over the air. WiCi also advertises multi-client support: Python libraries can pass the WiCi One to LLM frameworks as a regular GPU, and a standard HTTP API allows network access without special per-client configuration.
The base configuration ships with an RTX 5060 Ti 16GB and is priced at $1,999 for early signups or $2,599 otherwise. WiCi plans a preview in Q4 2026 and has mentioned a variant with an RTX 5090, but has given no release timing or price for that model. Gaming is mentioned as possible, but the documentation only cites Vulkan and Direct3D without offering performance expectations, and Wi-Fi's latency makes it a questionable fit for real-time rendering compared with conventional eGPUs.