Get the Computer Off Your Face: XR's Real Design Argument

Get the computer off your face

Four XR products moved this week, and none of them competes on resolution, field of view or refresh rate. Valve’s Steam Frame went on sale and got reviewed. Unseen Reality closed a $1,288,712 Kickstarter for a 93 gram headset with no battery. Snap priced its first consumer AR glasses at $2,195 and put the cellular modem in the carrying case. Meta got reported to be shipping a pair of smart glasses with no camera at all, days before its Connect keynote on September 23 and 24.

Read those four together and they are one argument, stated four ways. A device you wear on your face has a fixed budget of grams and milliwatts. The interesting engineering question in 2026 is no longer how to fit more into that budget. It is what you refuse to put there, and how you get it back.

The tracking loop is the part that cannot leave

Start with the cheapest way to move mass off a head: don’t put it there in the first place.

Unseen Reality’s URXR One raised $1,288,712 from Kickstarter backers after blowing past a $50,000 goal in minutes when it opened on August 4. It weighs 93 grams. That number is possible for exactly two reasons: thin mini pancake optics, and no compute and no battery on the device. It tethers over USB-C DisplayPort Alt Mode to a laptop, phone, tablet or handheld, and its Power Hub can even charge the phone it is plugged into.

Look at what Unseen Reality kept on the glasses anyway: a spatial processing unit plus a dedicated NPU, dual-camera SLAM with a 1000 Hz IMU, four forward-facing cameras, and passthrough video at 2,328 x 1,728 running at 90 fps with under 10 ms latency. The display stack is dual 1.03-inch micro-OLED panels at 2,448 x 2,064 per eye, 36 pixels per degree, 90 Hz, 90 degrees diagonal. Background: how passthrough works.

That is the whole thesis in a spec sheet. Tracking and display stayed. Power and the general purpose GPU left. Nothing that closes the tracking loop was allowed to cross the cable, because a back-and-forth over USB has to fit inside a motion-to-photon budget measured in single-digit milliseconds, and the IMU runs at 1000 Hz. The SoC that renders a spreadsheet can tolerate a round trip. Your vestibular system cannot.

The dongle is a design decision, not an accessory

Valve’s approach to the same problem is more interesting than the spec comparison suggests.

Road to VR’s Steam Frame review lands on a $1,060 total against the Quest 3’s $600, a $460 premium for a headset it describes as Quest 3-class in almost every visual respect: similar size, similar field of view, with Quest 3 still holding an edge on clarity. Frame wins on one axis that matters, weight, at 435 g against 515 g, and it wins for a structural reason. The battery sits at the back of the strap, so the mass is behind the head rather than hanging off the front. Background: Quest 3 vs Steam Frame.

Then the reviewers say something worth slowing down for. With the included wireless dongle, PC streaming was good enough that the tester did not think the average user could tell by latency or fidelity whether they were looking at a compressed stream or a dedicated tether. That dongle is the load-bearing decision in the product. Valve did not hand the link to the user’s Wi-Fi and hope. It shipped its own radio so that the latency budget for offloading rendering was something it controlled, and then it sold 130-plus titles on the strength of both paths: the “Great on Frame” page started in July with about five or six entries and now lists over 130 games and apps, roughly half of them flatscreen titles like Hollow Knight: Silksong, Portal 2 and Hades.

Valve is also honest about where the local path breaks. In standalone mode, demanding or poorly optimized games are described as largely unplayable, and battery life came in at 1 to 1.5 hours of mixed use against roughly 2 hours on Quest 3. Offloading is not free either. It just moves the cost from your neck to your network.

The rumored next Meta headset makes the same bet with a different connector. Project Phoenix is reported to be a slim, light mixed reality headset that tethers to a pocketable compute puck, deliberately moving a large share of the weight off the head. The Information also reports Meta is preparing a demo of hologram-like video calls using an advanced version of Codec Avatars, the photorealistic avatar system Meta has been developing since 2015, with Mark Zuckerberg and Andrew Bosworth expected to present it on stage at Connect.

Consider what Codec Avatars actually are as a workload. Photoreal neural avatars are a per-frame inference problem: heavy, parallel, tolerant of a few milliseconds, and exactly the sort of thing you would move to a puck rather than render on your face. Phoenix’s puck is not only a weight fix. It is a compute tier that makes a rendering technique possible.

The camera is now a legal object

The fourth device takes the logic to its conclusion.

The Information reported that Meta is preparing a camera-less line of smart glasses, internally codenamed Luna, for this fall. Instead of cameras, Luna reportedly carries six onboard microphones for talking to Meta AI and its Muse agent, a quick action button on the side, and speakers in slimmer-than-usual temple arms. Road to VR frames it as a direct reaction to public perception of Meta’s existing glasses lineup, and the surrounding context is not subtle: new legislation in US states and cities including New York state and Philadelphia, a class-action suit alleging captured footage was sent to a Kenya-based subcontractor for AI training, and a v26 update that disables the camera entirely if the capture LED has been physically removed or damaged, closing a loophole where users drilled out the privacy light.

Now put Snap next to that. Snap’s Specs arrive in the US, UK and France later this fall starting at $2,195, or $2,395 bundled with a charging case that adds four charges plus cellular connectivity, with Verizon data plans from $10 a month for customers and $20 for non-customers, alongside Orange in France and EE in the UK. Snap states plainly that the glasses do not use facial recognition and do not continuously record audio, video or images, with an external LED whenever recording is active.

Two companies, opposite sensor strategies, same finish line: keep the camera from becoming the story. Meta removes the sensor. Snap keeps the sensors but writes the restrictions into the product description and offloads the heavy context work to “Specs Intelligence,” an anticipatory AI service that runs across Specs, iPhone and Mac, uses on-device processing plus secure cloud infrastructure, and promises that personal content is not used to train models or serve ads.

This is where the hardware conversation collides with the software one. An anticipatory assistant needs sensors. Context about your goals, routines and relationships is built by watching the world. Luna with six microphones hears but does not see, which caps what any agent on it can ever infer about a room. Snap’s premise requires outward sensing, so it buys permission with explicit, checkable limits instead. Regulation and public opinion are now setting the sensor complement on consumer glasses, and the sensor complement sets the ceiling on what the AI layer can do. That is a hardware roadmap being written by city councils. See also Snap’s agent-first launch.

What stays on the head, and why Apple keeps betting on it

If the split line is the latency budget, Apple’s visionOS 27 shows what sits on the far side of it.

Vision Pro uses always-on, eye-tracked foveated rendering in the Shared Space and in many Full Space apps, rendering only the region you are looking at at full resolution. UploadVR notes this is essential to how the headset makes good use of its 4K micro-OLED panels, and that the tradeoff shows up in screen recordings, where everything you were not looking at is a low-resolution blur. Previous versions could only capture unfoveated footage through a connected Mac; visionOS 27 adds it on the headset.

Foveation is a hard local constraint. The gaze signal has to be sampled, fused and applied inside the frame you are about to display. You cannot ship eye tracking to a puck and get the frame back in time. The same applies to two other additions in visionOS 27: a high frame rate object tracking mode for moving and held objects, which developers estimate runs around 30 Hz, and new APIs that let any company build infrared LED tracked accessories, with optical tracking at 90 Hz fused with inertial data, the same approach Sony’s PSVR 2 Sense controllers use. DFRobot and MikroE already have reference hardware.

Note the shape of Apple’s argument. It moved nothing off the head and paid for that choice in grams, and in exchange it owns a rendering pipeline that a thin client cannot reproduce over a radio link.

The numbers on where the wall is

Gaussian splat viewing on Quest gives the split a measurable edge, because it is one workload tested on both sides.

UploadVR’s survey of Quest splat viewers found GSVision recommends a ceiling of about 700,000 splats for local playback, with lag becoming noticeable around 1.2 million. SuperSplat, the browser option with the largest library, needed a performance mode and still produced lag bad enough at times to cause motion sickness. The same reviewer then streamed a 5 million splat scene over CloudXR and got 30 to 40 frames per second off an Nvidia RTX 4070 Super, describing the result as comparable to cloud content with a visible speed boost.

Read that ratio. A three-year-old standalone headset handles roughly 0.7 million splats comfortably. A discrete desktop GPU handles 5 million at frame rate. That is close to an order of magnitude in scene complexity, and it is the difference between a splat being a technical demo and a splat being a place. It also explains why splat viewers keep shipping PC streaming modes and why cloud storage plus a subscription, which is how AirVis structures its free 5 GB tier, is starting to look like the normal distribution model for volumetric media. Background: real-time Gaussian splatting.

The catch: thin clients rent their capabilities

Every device that offloads compute inherits a dependency, and this week’s reporting is unusually clear about the bill.

Steam Frame only reaches its full capability with a PC present, or with a title that happens to be well optimized for standalone, and its battery life is the weakest number in the review. URXR One’s 93 grams are only 93 grams until you need the laptop in your bag, which is a reasonable trade for a display replacement and a poor one for a walk in the park. Snap’s cellular case means connectivity is a subscription, priced per month, from a carrier, and the glasses that rely on it are useless as an always-connected device without it. Splat streaming swaps local frame rate for cloud storage tiers and, in the CloudXR case, a desktop GPU on the other end of the link.

Apple’s counterexample is worth keeping in view precisely because it is expensive. Everything on the head, no tethers, and the weight is what users complain about.

The honest version of the thesis is not that compute is leaving the head. It is that the XR stack is being cut along its latency boundary, and the cut is different for every product. Tracking, foveation and the display chain stay on the head because physics will not negotiate. General rendering, AI inference over network-scale models, splat workloads and connectivity move out to a puck, a laptop, a dongle, a case or a datacenter, and each of those has an owner, a monthly fee, or a battery you now have to charge separately. See also Meta’s off-head compute puck.

Where this lands

Connect opens on Wednesday with Phoenix, Codec Avatars and Luna all expected on stage, which means the industry is about to make this argument in public, on one stage, in one keynote.

The thing to watch for is not how thin the headset is. It is how the compute puck is positioned: as an accessory you accept, or as the first tier of a platform that Meta controls end to end. Valve has already shown that owning the radio link is what makes streaming good enough to be the primary path rather than a fallback. The company that owns the puck, the link and the marketplace in one stack gets to decide which half of the experience is on your head and which half is on a rental plan. Over the next two days, Meta gets to answer whether it wants to be that company, or just the company that ships the glasses.

Get the Computer Off Your Face: XR's Real Design Argument