ExplorXR

Why XR's Next Headsets Render at a Quarter of Their Resolution

By Toni Tan · · 13 min read

The number that describes the product is not the one on the box

Valve’s kernel source code leaked a second Steam Frame this week. Codenamed Plano, it carries an OLED panel at 3,552 × 3,840 per eye, up from the shipping model’s 2,160 × 2,160 LCD, with 90Hz native refresh and a 120Hz experimental mode. That is the number that will end up in a headline, and it is roughly three times the pixel count of the headset Valve is selling right now. It is also the second time this month the same source tree has told us something about the hardware before Valve did, as the first OLED leak showed.

Then comes the number that describes what the device actually does. According to the same source dig, Plano’s default render resolution is 1,776 × 1,920 per eye, exactly half along each axis. One quarter of the panel’s pixels. Rendered at 90Hz across both eyes, that is about 0.61 gigapixels per second of shading work, which is less than the current Steam Frame pushes at the same refresh rate (roughly 0.84 Gpix/s at its native 2,160 × 2,160).

Read that again. The next-generation panel comes with a lighter workload than the current one.

That inversion is the story of this week in XR hardware, and it is not really about Valve. In the same seven days: Meta’s VR Glasses were confirmed to ship with 128GB of internal storage and an expandable microSD slot in the compute puck, with Sandisk as the exclusive flash supplier. A credible hardware leaker put PlayStation 6 RAM and NAND at over $680 on their own, with a bill of materials near $1,000. Valve confirmed a deluxe headstrap with hot-swappable batteries for a headset that runs about 1 to 1.5 hours on a charge. And XREAL shipped a $599 refresh that keeps its panel, resolution, field of view and chip identical, and spends the entire upgrade on brightness.

The thesis is uncomfortable but hard to avoid: in 2026 the specs that decide what an XR device can do are no longer the display specs. They are the memory budget and the energy budget. Everything on the marketing sheet is downstream of those two line items.

What half resolution buys, and what it costs

Start with the arithmetic, because it explains the design.

At 3,552 × 3,840 per eye, Plano’s panel holds 13.64 million pixels per eye, 27.3 million across both. At 90Hz that is 2.46 Gpix/s of shading if you light every one of them, plus the bandwidth to write and read the surfaces behind them. A naive RGBA8 color target alone would be 109MB per stereo frame and roughly 9.8 GB/s of write traffic at 90Hz before depth, motion vectors, or any intermediate buffer. Mobile memory systems in a standalone headset do not have that headroom, and the Snapdragon 8 Gen 3 in Plano is the same part as in the shipping Steam Frame.

So the default is a quarter of the pixels: 3.41 million per eye, 6.82 million stereo. Even that, at 90Hz with an RGBA8 target, is 27MB per frame and about 2.5 GB/s of write traffic, and that is a lower bound that ignores overdraw entirely.

This is not a Valve-specific trick. Every standalone headset does it, and the honest ones publish the panel number and stay quiet about the framebuffer. What makes Plano interesting is that the gap is unusually large and unusually visible in the code, because the panel jump was so aggressive while the compute stayed fixed. It is the clearest public illustration of a rule developers have lived with for years: you render at the resolution the memory bus can feed, then reconstruct the rest with upscaling and, where eye tracking exists, foveation. The full comparison between Quest 3 and Steam Frame is the background here if you are weighing the shipping hardware against a refresh that has not been announced.

The trade is real. A quarter-resolution render upscaled to a 4K-per-eye OLED is a soft image with reconstructed detail, and on a panel that sharp you will see it. The 90Hz native cap, against the LCD model’s 72 to 120Hz with a 144Hz experimental mode, is the other half of the compromise: OLED motion handling and brightness are better, refresh headroom is worse. That is what it looks like when a product is designed around a budget instead of a benchmark.

Memory is the line item that decides the product

Now the reason the budget exists.

Meta’s VR Glasses, launching in spring 2027, ship with 128GB of internal storage and, for the first time in a Meta standalone, an expandable microSD slot. The slot is not in the headset. It is in the compute puck, the box that already holds the battery and the Snapdragon Reality Elite. Sandisk announced this week that it is the exclusive provider of next-generation flash for the device, selling microSD Express cards and USB-C flash drives under the Made for Meta program, with no capacities or prices disclosed yet.

UploadVR calls it “a move likely intended to ease rising memory costs.” A company does not add a card slot to a sealed wearable unless the arithmetic forces it. NAND at 128GB is a fixed cost per unit that scales with every unit shipped, and it is being priced in a market where every AI cluster is bidding for the same fab capacity. A slot converts a per-unit cost into a customer purchase, and it lets Meta ship a $1,300 device without eating the difference on 512GB.

The architecture matters too. A puck has volume and thermal headroom that a 100g headset does not, which is the whole argument behind why XR compute moved off the face. Once the electronics live in a box on your belt or in your pocket, storage stops being a design constraint and becomes a consumable.

The software is catching up: Horizon OS now lets you browse and manage external drives in the Files app, opens compatible files in 2D apps through an “Open With” menu, and the 2.9 PTC lists external storage locations next to internal storage in Settings. That is a quiet but significant admission that the library is bigger than the device.

For XR specifically, storage is not a nice-to-have. A scene with dense photogrammetry or Gaussian splat captures is gigabytes before it is interactive, and a headset cannot fall back on streaming assets from a disk the way a console can, because the user is wearing the display and latency is the product. Expandable storage is also slower storage, and microSD Express is the only variant worth the name here, since it moves data over a PCIe lane rather than the legacy UHS bus. Whether the certification program keeps card prices sane is a separate question, and one Meta and Sandisk have not answered.

The shortage is upstream of every roadmap

The console leak explains why the slot exists.

Kepler_L2, posting on NeoGAF, put PS6 RAM and NAND at over $680 combined, with a full bill of materials near $1,000, up from a March estimate of around $760. The context he was replying to assumed a 20 to 30 percent error could be absorbed by platform subsidies. His answer was that memory is the subsidy problem. Sony has reportedly pushed the console to 2028 or 2029 over the shortage, which is usually traced to Micron and others prioritizing RAM supply for AI customers; Nikkei reported in April that it would persist into next year.

XR sits at the worst end of that queue. Headset volumes are a rounding error next to phones and a small fraction of consoles, so an XR program gets allocation last and pays spot pricing first. Valve’s own engineer Yazan Aldehayyat has said the memory situation was worse than the team expected and that Steam Machine pricing took a heavy hit from it, which is the same company that priced Steam Frame at $1,059 for 256GB and $1,299 for 1TB.

Put the three facts together and the design logic writes itself. If DRAM and NAND are the expensive, scarce inputs, then the cheapest way to ship a better headset is to leave the compute alone and change the panel. That is exactly what Plano appears to be: new display stack, same Snapdragon 8 Gen 3, default framebuffer tuned so the memory system never has to feed the panel it is attached to. It is a refresh shaped by supply, not by ambition.

The other scarce resource is watt hours

Power has the same structure, and Valve’s answer this week was an accessory.

The Steam Frame runs about 1 to 1.5 hours of mixed use. Valve has confirmed to PC Gamer that it is building a deluxe headstrap with hot-swappable batteries and Balanced Mode Radiator drivers meant to match the off-ear speakers on Index, still the audio reference seven years on. The upgrade is described as swap-in, and it is likely to exploit the fact that the display assembly detaches completely from the strap and facial interface. No price, no date, and backorders on the headset have already stretched into 2027.

Two things are worth noticing. First, the shipping headset has no 3.5mm jack, so the audio fix and the power fix arrive in the same part. Second, a hot-swappable battery is a statement about the energy budget: it says the internal cell is not going to get bigger, so the design moves the problem to the user’s pocket. Whether that works depends on details Valve has not published, including whether the headset can stay alive across a swap or has to sleep through it, and how the added mass at the rear changes balance. Comfort is the part of this that does not show up in a spec table, and it is where Valve has historically competed, which is the argument made here when the headset was first detailed.

The accessory layer is turning into the product. Virtuix is now an official Meta reseller, selling Omni One treadmills bundled with a Quest 3 or 3S for $599 or $349 on top of a $2,595 base, which is a strange sentence to write about a headset that costs less than the accessory. When the core device is squeezed on BOM, the margin and the differentiation migrate outward, into straps, batteries, treadmills and certified cards. None of those are bad products. All of them are admissions about the base unit’s budget.

XREAL shows where the logic ends

The cleanest version of the argument is a pair of glasses with no battery worth mentioning.

XREAL’s 1 Pro+ costs $599, the same as the One Pro it replaces. It keeps the Sony 0.55 inch micro-OLED at 1,920 × 1,080 per eye, the 57 degree field of view, the X1 spatial chip, 3DOF (6DOF with the Eye accessory), 120Hz and the two IPD variants. The changes are peak brightness from 700 to 1,200 nits, AI-driven HDR, a redesigned nose pad, and 2D-to-3D conversion moving from 30 to 60fps. It is tethered over USB-C with DP Alt Mode, because the compute and the power live in the laptop or phone at the other end of the cable.

That is what a device looks like when its energy budget is zero. The only levers left are optical efficiency and panel brightness, so brightness is what the refresh buys. Twelve hundred nits matters outdoors in a way that resolution does not, because the glasses are fighting daylight through a dimming layer rather than controlling the room. Meanwhile the 2D-to-3D conversion at 60fps is a compute-cheap feature that improves perceived quality without touching the display stack, which is the same instinct as upscaling a quarter-resolution framebuffer: spend where the budget allows, and reconstruct the rest.

Three devices, one pattern. Plano spends on the panel and claws it back in the framebuffer. Meta’s glasses spend on the puck and claw it back with a card slot. XREAL spends on nits and claws it back by leaving the computer in your pocket, which is the same bet on the latency budget that decides what can leave the headset.

The catch

Memory cycles turn. Every shortage in the last thirty years has ended with capacity coming online and prices collapsing, and when that happens the constraint moves back to silicon and optics. A headset designed around 2026 pricing may look badly under-specced in 2029, which is roughly when a Plano-class device would actually ship if Valve follows the Steam Deck playbook of LCD first, OLED eighteen months later.

There is also a real risk that cost-driven design becomes an excuse. A 1 to 1.5 hour battery is a bad number no matter how good the strap is, and selling the fix as an accessory turns a design flaw into an upsell. A quarter-resolution default on a 4K panel is defensible engineering, but if the upscaler is not good the user experiences it as a blurry headset, and “it renders at half resolution” is not a sentence anyone wants in a review. Expandable storage has the same shape: certified microSD Express cards are expensive, and a slow card in a fast slot produces stutter that gets blamed on the device.

The honest version of the argument is not that scarcity produces bad hardware. It is that scarcity produces hardware with a different centre of gravity, and reviewers who keep grading on panel specs will keep misreading it.

What to watch

Three numbers are worth more than the display spec sheet from here on.

The first is default render resolution, not panel resolution. If a vendor quotes per-eye pixels and not a framebuffer, the gap is where the compromise lives. The second is memory: gigabytes of RAM, and whether storage is soldered. The third is watt hours, and whether the battery is swappable. A headset that scores well on those three will feel better in a year than one that wins the resolution column.

For developers, the implication is a pipeline decision rather than a hardware one. Plan for dynamic resolution and foveated rendering as defaults, not options, since the panel you are targeting will almost certainly be fed fewer pixels than it holds. Budget your assets for storage that might be a card, and design scene loads that survive a slow read. The hardware will meet you at a fraction of the resolution it advertises, and the teams that build for that reality will ship the sharper-looking games anyway.

The display was never the product. It was the promise. This year the industry started showing its arithmetic.


Sources

Related reading