A startup called WiCi is shipping a wireless eGPU box that tries to solve local AI inference's storage-versus-VRAM problem by putting a full graphics card, a network stack, and a 4TB PCIe 5.0 SSD in a single $1,999 enclosure. The WiCi One, as detailed by Tom's Hardware, uses Wi-Fi 7 instead of a PCIe cable to connect a remote GPU to a client machine, and the pitch is squarely at people who want to run LLMs locally without buying a full workstation around the card.
The mechanism is the interesting part. A client machine runs a GPU driver that makes its OS believe the card is installed locally, when in fact every instruction is going out over Wi-Fi 7 to the WiCi One box. Standard GPU-to-host communication is, in the report's words, "extremely chatty" — fine over PCIe 5.0's sub-microsecond latency, but according to Tom's Hardware's description of WiCi's own explanation, the same chattiness over a wireless link would turn basic status queries into multi-second affairs and could stretch large computations to hours if handled naively. WiCi's answer, per the outlet, is a driver layer that does heavy caching, deduplication and compression, and that decides which hardware requests can be resolved locally versus which actually need to cross the network.
That software layer is also what makes the multi-client story work: because the GPU is abstracted behind a network protocol rather than a physical slot, WiCi says the same card could serve more than one machine. The company offers Python libraries that expose the WiCi One as a normal GPU to LLM frameworks, plus a standard HTTP API for network access — both signals, per Tom's Hardware's read of the documentation, that WiCi expects several clients to share one card without special setup.
The onboard SSD is aimed at a specific pattern in current model serving: Mixture-of-Experts architectures that fall back to storage when VRAM runs out. WiCi built a 4TB PCIe 5.0 SSD directly into the enclosure so model weights can sit close to the GPU, letting larger models run while cutting down on repeated transfers over the wireless link. The base configuration ships with an RTX 5060 Ti with 16GB of VRAM; WiCi has said it plans an RTX 5090 variant with 32GB, though Tom's Hardware reports no price or date has been set for that version, and 5090 cards alone are currently listed well above $7,000.
On price, Tom's Hardware's own component breakdown puts the box in believable territory: an RTX 5060 Ti 16GB runs around $800 on Newegg, a 4TB PCIe 5.0 NVMe SSD around $850, plus an estimated $100 for a power supply and another $100 for the case, Wi-Fi card and other parts — landing near $1,850 before markup. WiCi is charging $1,999 for early signups and $2,599 at general availability, which the outlet calls reasonable if the packaging and software stack actually deliver.
What's conspicuously absent is any inference latency number. Tom's Hardware's account never states a Wi-Fi 7 throughput figure, a tokens-per-second result, or a measured round-trip delay for GPU commands under the new driver — the entire chattiness argument is described qualitatively, not benchmarked. The same gap shows up around gaming: WiCi says the card can run games unmodified, but the documentation reportedly only mentions Vulkan and Direct3D support with the caveat that "[not] every API must be complete on day one," and Tom's Hardware notes Wi-Fi is a poor fit for pushing frames at 16.67ms intervals for 60 FPS. There's also no driver compatibility list, no benchmark against a wired eGPU baseline, and no shipping date — WiCi has only committed to a Q4 2026 preview.
For a team evaluating this as a dev-rig building block: treat every WiCi latency and caching claim as unverified marketing until someone publishes tokens-per-second or round-trip numbers against a wired PCIe eGPU, because the vendor's own account admits the naive version of this design is unusably slow.