WICI ONE — A Research Dossier

SEC.01

The spec sheet, translated.

Plate W1 — box, brains, and the gap between shows and shipments

In the brochure		What it actually says
“GPU OVER WI-FI”	→	The slogan hides the load-bearing wall: what stays by the GPU, what crosses the wire, and how the software hides the split from the app. That software layer is the product. The radio is just the radio.
OWN CPU · STORAGE · OS	→	Physically a standalone compute node — not a graphics card with a Wi-Fi module bolted on. Novelty lives in how capabilities are presented to applications, not in the box existing.
ONE BOX, MANY DEVICES	→	Different devices sharing one accelerator is the declared capability. Claimed, not confirmed — no independent multi-client evidence yet. The Steam-Headless family route you already run is a working alternative, not a worse one.
$2,599 / $9,999	→	RTX 5060 Ti 16 GB or RTX 5090 32 GB. At the top price, the software must prove value beyond the silicon — you're not buying a GPU, you're buying the integration tax back.
“SHIPS IN 8 WEEKS”	→	A promised window, not a witnessed delivery. Same filing rule as every dossier: promises go in pencil. NYC shows Oct 5–6 are a chance to see claims outside one promo video — still not independent results.
DEV PREVIEW · Q4 2026	→	Examples and interfaces on the technical pages are marked preliminary. The honest reading: architecture described, product unproven. That's not an insult — it's the correct epistemic shelf.

SEC.02

Games at the boundary.

Plate M2 — memory as territory: what stays, what crosses, what returns

The technical articles are better than the slogan. Where LUCEBOX distributed compute between two pools in one box, WiCi distributes it across a wire — and the research variable is new: the device boundary, and how rarely it truly needs crossing.

Keep the bytes by the GPU

Active data in VRAM, saved state in the box's RAM, model weights in an SSD cache next to the GPU. When one operation's output feeds the next, don't ship a fat tensor to the laptop just to send it back — the client references it through a logical identifier. The SSD path covers weights today; arbitrary results are named an open research task. Honest scope.

Cross rarely, not faster

The question isn't "can Wi-Fi carry PCIe's bytes?" — it's "what fraction of the exchange can be skipped entirely?" Cache immutable answers. Send independent commands asynchronously. Cut spurious synchronisations. The boundary stays slow; the number of crossings is the engineering variable.

Preserve true dependencies

Optimisation has a conscience: real data dependencies must survive the reordering. Their own test protocol says the proof must include cold start, competing clients, state recovery and latency spikes — not one flattering demo run. A vendor publishing the exam questions deserves the index's respect.

Our check, not theirs	What would count as proof
Compatibility	A named app, a named OS, a named mode (API / SDK / Driver). “Works” without those three words is marketing, not evidence.
Everydayness	What changes after the first launch, through a wall, with a second user attached. The demo is the first minute; the product is month one.
Correctness	Results and data separation preserved across memory reuse and reconnection. For the compliance line: can user context and authority control stay on the client while compute is delegated with the minimum necessary data? — our research question, not their advertised feature.

“Don't add gigabytes — change where data lives and how execution is structured, so the system's limit stops being the point.” — THE THROUGH-LINE FROM LUCEBOX'S POOLS TO WICI'S WIRE

SEC.03

Capability map, not datasheet.

Plate W3 — three modes, three confidence levels, one wire

CAPABILITY MAP STATUS: hypothesis / claim translation. Use the evidence label on each row before reading the adjective.

On WiCi One · 5060 Ti 16 GB or 5090 32 GB · own CPU, RAM, SSD, OS

API MODE · LLM SERVICE

Vendor
MOST PLAUSIBLE · ASK-RECEIVE FITS THE WIRE

SDK MODE · wici.gpu

Vendor
REAL BUT WRITTEN · PER-APP INTEGRATION

DRIVER MODE · EXISTING APPS

Vendor
THE AMBITION · PRELIMINARY, UNSIGNED

GAMES OVER WI-FI

Vendor
LATENCY'S COURT · THEIR OWN CAVEAT

SHARED · MULTI-DEVICE

Vendor
CLAIMED · NO INDEPENDENT PROOF

OFFLINE / NO NETWORK

Inferred
NO · NO WI-FI, NO GPU

CLIENT DATA STAYS CLIENT

Unverified
OUR QUESTION · NOT THEIR PROMISE

SEC.04

Eight tests. Wire-shaped.

Plate O4 — the Edinburgh test fails by design; that's information

T-01KettleTBDAPI mode should be minutes. Driver mode is the real kettle question: does an untouched app see the virtual GPU without a weekend? Dev preview Q4 2026 will say.

T-02ReadingHYPOTHESISFor LLM-shaped traffic (weights cached, state local), the wire can look wider than it is. For latency-chained workloads, physics keeps the invoice. Two sports, one scorecard — keep them apart.

T-03LibraryDOUBLE-EDGEDThe box can live in a cupboard — silent desk achieved by relocation. The laptop is quiet too. Whether the library's network is your network is the library's policy, not WiCi's.

T-04BackpackNOA stationary shared appliance — the opposite assignment. The backpack class remains the 300-gram node's job.

T-05Edinburgh trainNO (BY DESIGN)No network, no accelerator — the boundary isn't wirelessness, it's the network itself. A fair fail: the device never promised the train. File under "knows what it is", pending proof of what it is.

T-06Meeting roomHYPOTHESISFascinating case: the GPU can sit in your office while your laptop visits the client — the first machine in the series that decouples compute from presence. Data separation across that link is precisely the unsolved check.

T-07Electricity billTBDA 5090-class node idling as a shared appliance could be a £200+/yr background character. Idle watts unpublished; the always-on-shared-GPU idea lives or dies on them.

T-08Next doorIS THE PRODUCTThe upgrade path isn't parts — it's modes. API today, SDK as developers adopt, Driver if it matures. You're not buying a configuration; you're buying a software roadmap with a GPU attached.

SEC.05

The honest edges, up front.

Plate K5 — their caveats are better than their marketing

prosumer@index — wici-one.log

# filed before the NYC shows, updated after — one way or the other. $ ./wici-one --list-honest-edges [MODES] api ≠ sdk ≠ driver. a working chat over the api proves api. nothing else. the index keeps three files, not one folder. [WIFI] "wi-fi is not pcie" — their own article, their own words. some tasks are too latency-, data- or driver-sensitive. respect. [BYTES] vram for active, ram for state, ssd cache for weights. don't send the tensor home just to send it back. the real trick is skipping crossings, not accelerating them. [SHARE] one box, many devices — declared, not demonstrated. your steam-headless route already works. different way, not worse. [PRICE] $9,999. the card inside is not the product. the software around it has to be — or it's a very fancy dock. [STATUS] dev preview q4 2026. examples preliminary. shows on oct 5–6 are a chance to look — not to conclude. $ so is it the future? > it's a bet on the boundary: compute over there, life over here, > and as few crossings as honesty allows. > the index doesn't grade bets. it grades what ships. $

“The AORUS dream, one abstraction up: not my laptop gets a GPU — my devices get a shared heart, without stopping being my devices.” — WHY THIS ENTRY SITS NEXT TO A WORKING FAMILY SETUP, NOT ABOVE IT

SEC.06

The verdict, in prosumer.

Plate Y6 — filed with interest, promoted on shipments and third-party runs

Watch it if

You're building the shared-household-GPU thesis: several screens, one accelerator, zero admin theatre. Follow the Developer Preview, the NYC show outcomes, and the first owner reports — especially Driver mode and two-client contention.

Skip it if

You need it now (Q4 preview), offline (no network, no GPU), portable (stationary appliance), or proven (nothing independent ships yet). Your working Olares/Steam-Headless route already delivers the family scenario — don't pay $9,999 to re-buy what you own.

The one test

Does it help me do more, or give me more to maintain? WiCi's architecture says the right things: skip crossings, preserve dependencies, publish your own exam questions. If Driver mode matures, this is the first box that upgrades existing work instead of relocating it. Until then: the most honest unproven idea in the index.

GPU OPTIONSRTX 5060 Ti 16 GB · RTX 5090 32 GB

NODEOWN CPU · RAM · SSD · SOFTWARE STACK

MODESAPI · SDK · DRIVER (VIRTUAL GPU)

PRICE / PROMISE$2,599 / $9,999 · SHIPS ~8 WKS (STATED)

STATUSDEV PREVIEW Q4 2026 · NYC SHOWS OCT 5–6

CLOSING — BOUNDARY RULE №1 · PLATE F

AORUS gave a laptop a GPU.
WiCi asks: what if the GPU belonged to the room?

The competition here is not between beautiful boxes. It's between ways of making local compute available to a human: relocate the work, relocate the screen, or — the WiCi bet — leave everything where it is and make the boundary itself cheap to cross. Rarely. Logically. Honestly.

The index doesn't grade dreams. It files them, marks what's proven, and waits for the wire to show its receipts. Crossings skipped, dependencies kept — that's the whole religion.