The Puck Won: Why XR Compute Moved Off Your Face

The headset stopped being the computer

A month ago this site argued that the interesting XR design question of 2026 is not how to fit more compute onto a face, but what you refuse to put there. That piece is now backed by three product decisions and one corporate retreat, all of which landed inside a single week.

Valve told Steam Frame reservation holders that their order email arrives either by December or in 2027. ByteDance’s Pico had its next flagship headset, Space Pro, leaked from Japanese regulatory filings down to the teardown photos, and the photos show an external tethered main unit alongside a fan inside the visor. Meta’s VR Glasses, which ship in Spring 2027 at $1,300, put roughly 100 grams on your head and move everything else to a compute puck. Sony, meanwhile, quietly transferred dozens of XR patent properties to Meta.

Read together, they say one thing: in the premium tier, the headset is a display and a tracking rig. The computer is something you own separately, or something you stream from. And that has consequences no spec sheet lists, starting with what a “launch review” even means anymore.

The queue into 2027 is a verdict on the architecture

Valve’s own news post is unusually precise about the mechanics. Signups for Steam Frame opened September 14, the list was randomized once on September 17, and order fulfilment began September 18. Reservation holders can now see on the product page whether their order email comes by December or in 2027, and the 2027 group gets firmer timing in January. A purchase email comes with a 72 hour window before Valve moves to the next person in line.

The detail worth reading twice is regional. Valve runs a separate reservation queue for each region and only dips into a region’s waitlist when that region’s reservations are exhausted, with no leapfrogging between groups. That is the queue design of a company that knows it is shipping in constrained batches over a long horizon, not one sitting on a warehouse of units.

Here is the architectural read. Steam Frame renders locally on a Snapdragon 8 Gen 3 when you want standalone play, at visibly lower quality, and it renders on your PC’s GPU when you stream. The most expensive component in the product is optional hardware you may already own. That is why the queue can run deep into next year without the product looking broken: the buyers who most want fidelity are the ones already holding a gaming PC, and they are the same buyers Valve is selling to. Background on what the two approaches cost you: Quest 3 vs Steam Frame: Specs, Streaming, and Which to Buy.

SteamOS 0.4 proves the fidelity now lives in the link

The second half of Valve’s post is more revealing than the queue, because it is a changelog of things that were wrong.

SteamOS 0.4 raises power efficiency and battery lifetime by up to 12% in both streaming and local modes. For context on the size of that fix, mixed use previously landed around 1 to 1.5 hours. It improves color clarity during VR streaming, fixes a set of “Error 17” protocol failures that occur between headset and host PC (which requires opting into SteamVR 2.18.2 or later on the beta branch), fixes wireless adapter connectivity problems that hit users in some countries, corrects Wi-Fi channel lists in manual configuration, improves audio reliability while streaming, makes downmix and binaural rendering match volume across formats, and unfreezes the keyboard while streaming.

Notice the shape of that list. Almost none of it is about the renderer on the headset. It is about the network path, the codec, the host client and the compositor. When compute sits on the other end of a link, your product quality is set by packet loss, protocol negotiation and color reproduction through a video pipeline, not by the mobile SoC you shipped with. A headset review written in September is partly invalidated by an October software update, and that is now normal.

Two smaller items in the same post point at where third parties fit. Valve published Steam Frame CAD files and electrical interface information for non-commercial use, with a commercial path on request. It also added a “Track Dominant Eye Only” accessibility mode and a dominant eye setting that decides which eye feeds spectator view and screenshots. That is the eye tracking pipeline surfacing as a user-facing control rather than a hidden optimization, which matters if you build foveated rendering. For background: how VR eye tracking works.

Pico’s leaked teardown is the same architecture, plus a fan

Pico’s Space Pro story is the counterexample that proves the pattern. The device has been public in fragments for a year: micro-OLED panels at 4,000 pixels per inch, a dual-chip architecture with a self-developed coprocessor handling computer vision and image processing, and a new main processor rated at twice the CPU and GPU performance of the Snapdragon XR2 Gen 2 inside Quest 3 and Pico 4 Ultra. Pico OS 6 and the RealityKit-style Pico Spatial Engine were detailed alongside it, and Pico ran a closed “Global Early Access Program” asking for rigorous feedback on hardware and software before release.

Then the schedule slipped. A launch event announced in August was cancelled days later, and the device moved to Q4 2026. Filings in the United States listed two components: a “Mixed Reality Headset” and a “Mixed Reality Main Unit,” with the puck as the model B3110 main unit. Japanese filings under certification number 202-JWM032 now show the same B3110 hardware from the outside, and teardown pages show the front camera and sensor array, the optical modules, and a central cooling fan.

The fan is the tell. Even with a tethered compute unit in the box, Pico keeps a thermal budget on the face for the camera and vision pipeline, because that loop cannot leave the head without breaking motion-to-photon latency. The part that can leave is the general purpose GPU, and it does. The delay itself is the other half of the argument: the silicon was clearly not the blocker, the software was.

That puts Pico in the same design space Valve and Meta occupy, with one differentiator left, which is whether Pico OS 6 arrives in a state worth shipping. Rivals are converging on the same hardware: XR Daily: Pimax Takes On Steam Frame.

Meta bought 100 grams by spending 40 degrees of FOV

Meta’s VR Glasses are the cleanest expression of the trade because its numbers are public.

The device puts about 100 grams on your head. It reaches that by moving compute and battery into a puck and by shrinking the optical stack to small pancake lenses over micro-OLED panels. The price is field of view: 70 degrees horizontal by 66 degrees vertical, roughly 88 degrees diagonal, against Quest 3’s 110 by 96 degrees, or about 119 degrees diagonal, with a 515 gram on-head package. You give up about 40 degrees of horizontal vision and $1,300 buys you a device that Meta positions around productivity, communication and casual media rather than games.

Continuity is the compensating move. VR Glasses run the same operating system as Quest, so the Quest library carries over, and although the device is hand-tracking first, it pairs with Touch Plus controllers to reach the back catalogue. Meta’s framing is explicit: VR Glasses sit on a continuum that runs from audio glasses to full immersion, and Meta’s VP of VR, Ryan Cairns, described them as a different product for a different market. The reading that matters for developers is that Meta is buying software continuity with an optical compromise, which only works if the puck is doing serious work.

Sony did the math and left

The week’s clearest signal about who can afford this structure came from a company that decided it could not. According to reporting relayed by Road to VR, 43 U.S. patent properties moved from Sony to Meta between June and September 2026, covering XR information processing, head-mounted displays and XR glasses types filed between 2010 and 2022. The underlying confirmatory assignment, signed in December 2025, covers 419 patents worldwide: 180 in the United States and 239 in Japan, China, Korea, Europe and elsewhere. Neither company has commented. Our earlier coverage of the transfer: XR Daily: Meta Buys Sony’s XR Patents.

Sony’s trajectory explains the decision. PSVR 2 arrived in February 2023 at $550, $150 above the original PSVR, with a thin launch lineup, and Sony later sold a PC VR adapter that let the headset run SteamVR content on a gaming PC, which is the opposite of a walled garden strategy. Building an on-head computer means owning a silicon roadmap, a display supply chain and a platform that has to earn back its cost over years. Selling components and licensing is cheaper. The patent transfer does not change Meta’s roadmap much, but it removes a rival that was already walking away.

The catch: three performance classes and a waiting room

None of this is free, and the bill arrives in three places.

First, immersion. Moving compute off the face buys grams and battery, and the currency you pay with is field of view, cabling and cost. Meta’s 100 gram, $1,300 device gives up roughly 40 degrees of horizontal FOV against a $500 class standalone, which is a large amount of peripheral vision to trade for comfort.

Second, fragmentation. A developer targeting 2027 hardware now picks among a standalone mobile SoC, a tethered puck with a dedicated vision coprocessor, and a streamed PC render. Those are three different performance envelopes, three different failure modes, and in the streaming case a fourth variable that is entirely outside the developer’s control: the player’s network. Valve’s changelog is effectively a public list of those failure modes, and version 1.0 of any streaming stack reads like that.

Third, supply. A reservation queue that runs into 2027 is not a victory lap for anyone who wants the device this year. Valve is shipping in batches, region by region, and the buyers at the front are the ones who already own the GPU. That is a healthy niche, and a niche is exactly what it sounds like when the mass market tier disappears into glasses.

Where this leaves the buyer and the developer

The practical consequence is that the specs that used to settle arguments no longer settle much. Resolution, refresh rate and chipset naming are all secondary to questions nobody prints on a box: how good is the link, does the tracking loop close in single digit milliseconds, will this vendor still be shipping meaningful updates in eighteen months, and can I get the thing at all.

That last question has an answer for Valve’s customers, and it is January. For developers, the more useful takeaway from this week is that the hybrid architecture, which we sketched as a design argument in Get the Computer Off Your Face: XR’s Real Design Argument, is now the default in the premium tier rather than a bet. Steam Frame’s update notes, Pico’s teardown fan and Meta’s 100 gram puck all describe the same pipeline split in different proportions.

The interesting engineering work in 2027 is not on the visor. It is in the codec, the latency budget, the compositor, and the coprocessor that keeps computer vision on your face while everything else leaves.

---

Sources

• Valve: Steam Frame updates: Reservation timing, software updates, and more
• Road to VR: Steam Frame Backorder Queue Stretches into 2027
• Road to VR: Steam Frame Update Boosts Battery Life, Improves Streaming Quality & More
• UploadVR: Pico Space Pro Headset Spotted In Japanese Regulatory Filings
• Road to VR: Pico Space Pro Leaks Ahead of Planned 2026 Launch
• Road to VR: What Meta VR Glasses Mean for the Future of Quest
• Road to VR: This App Lets You Actually Experience Meta VR Glasses’ Field of View
• Road to VR: Sony Sells off Dozens of XR Patents to Meta