Cloud VR

Cloud VR (cloud virtual reality) is the delivery of virtual reality (VR) content from remote servers in a cloud or network-edge data center to a headset over a network, with the server doing rendering or storage work that would otherwise run on a local PC, console or the headset itself. In an interactive cloud VR system the headset sends its tracking data upstream, a GPU server renders the scene from that viewpoint, and the result comes back as a compressed video stream. A 2020 book based on the work of Huawei's iLab describes the approach as introducing cloud computing and cloud rendering to VR services, with the display and audio output coded, compressed and transmitted to user terminals, and notes that it has strict bandwidth and latency requirements, which makes 5G and gigabit home broadband suitable networks for it. The term is also used for augmented reality (AR) and extended reality (XR) in the forms "cloud AR/VR" and "cloud XR".

The aim is to let light, inexpensive or untethered headsets show content that needs more graphics power than they have, and to let users play without owning a gaming PC. Its main obstacle is delay: every frame has to cross the network, and VR is far less tolerant of lag than a game shown on a television or monitor. Telecom operators and their suppliers promoted cloud VR from 2018 as a service for 5G and fiber networks, starting with China Mobile's Fujian branch, and NVIDIA CloudXR became available on the major public clouds for enterprise use. Consumer cloud streaming of PC VR games has had little official support: Meta does not allow streaming from cloud sources in Quest store apps without its written approval, and the service PlutoSphere shut down in 2024.

Reviewed 21 October 2026. Checked every claim against 3GPP TR 26.928 V16.0.0, the Huawei CG Cloud VR draft and 2018 press releases, the WinWin article, the GSMA 2019 booklet, Meta's VRC.Quest.Streaming.2 page, the cited Road to VR, UploadVR, MIXED, AEC Magazine, NVIDIA and Microsoft Learn pages, the full texts of Hou et al., Elbamby et al., Rossi et al., Warsinke et al. and Zhou and Popescu, and Crossref records for all cited papers. About review dates.

Definition and scope

Cloud VR is a form of remote rendering. When the device keeps part of the work, such as correcting each received frame to its latest head pose, the arrangement is usually called split rendering. That page covers the general technique, including streaming from a local PC (as in Air Link, Virtual Desktop, ALVR and Steam Link) and from a phone to AR glasses. "Cloud VR" refers specifically to rendering or content that comes from servers outside the user's home or office, usually in a commercial cloud or at the edge of a mobile or broadband network.

Two kinds of content are delivered this way. Huawei's draft CG Cloud VR Technical Specifications distinguish panoramic video with three degrees of freedom (3DoF), used for on-demand and live 360-degree streaming, from real-time computer graphics (CG) rendered by GPUs with six or more degrees of freedom, used for games, training and collaboration. The document calls CG renders "advanced forms of VR" and states Huawei's aim to "cloudify" GPU resources for them to reduce user investment.

Xueshi Hou, Yao Lu and Sujit Dey of the University of California, San Diego, set out the design space in a 2017 paper. They listed three things a server could send: the current field of view rendered remotely, several views rendered remotely and stitched into a 360-degree video, or compressed models and textures sent to a local edge device that renders them. They paired these with three places the stream could come from: a cloud server, a remote edge server in the mobile network, or a local edge device in the same room or building. Their analysis concluded that streaming a rendered field of view from a distant cloud server was "impossible due to the ultra-low latency requirement", which left 360-degree video as the cloud option, at the cost of limiting the experience to 3DoF.

How it works

3GPP Technical Report 26.928, Extended Reality (XR) in 5G, describes the simplest network architecture, "viewport rendering in network". The XR server generates the media on the fly from incoming tracking and sensor information, for example with a game engine, flattens the viewport into a 2D format, encodes it and sends it over the 5G network. The device decodes the stream and displays it "without using the viewport information". The report says this allows simple clients but poses significant compression and transport challenges, because tracking, pose delivery, rendering, encoding, delivery, decoding and display must together fit within the 20 ms motion-to-photon limit for immersive media.

For fully network-rendered XR, the report estimates that the network would have to provide significantly lower latency than 10 ms in each direction together with a stable bitrate in the range of 0.1 to 1 Gbps, because only lightweight, low-latency compression could be used. It contrasts this with cloud gaming on a flat screen, where the limit is the round-trip interaction delay, typically 50 ms for advanced games, rather than motion-to-photon latency. In the split-rendering variant, the device corrects each frame to its current pose with asynchronous timewarp (ATW), so the 20 ms motion-to-photon requirement is met inside the device, and the network requirements are set instead by the pose-to-render-to-photon time and the round-trip interaction delay, typically 50 to 60 ms. The report states that split rendering "requires the use of edge computing" because the pose-to-render-to-photon time is expected to be below 50 ms.

Placing the server close to the user is the other main technique. The GSMA wrote in 2019 that "using central clouds is not an option" for interactive AR and VR because of the motion-to-photon requirement, and that the delay budget would require in-network edge data centres equipped with GPUs and hardware video encoders. The same document described a "lightweight client" model in which either the whole application moves to the edge cloud, needing few changes to existing software, or only parts of it are offloaded in a hybrid design that requires the application to be adapted. Mohammed Elbamby and colleagues from the University of Oulu, the University of the Basque Country and Nokia Bell Labs similarly argued in 2018 that remote cloud servers suit low-resolution, non-interactive VR such as streamed 360-degree content, while interactive VR needs computing pushed to the network edge (mobile edge computing), which they wrote is expected to reduce communication delay to less than 1 ms in metropolitan areas.

Video coding adds its own delay. The GSMA listed several causes of high latency in the H.26x codec family: encoding, transmission and decoding happen one after another, so a frame can only be shown once it is fully decoded; a large buffer before decoding hides network jitter; a further buffer after decoding absorbs variation in decoding time; and software processing adds delay across the end-to-end system. It noted that emerging low-latency codecs reduce delay at the cost of more bandwidth but lacked standardization and hardware support.

Latency and bandwidth requirements

Published targets vary with the type of content and with how much work stays on the headset. The figures below are given as each source states them.

Source	Scope	Stated figures
3GPP TR 26.928 (2020)	Viewport rendered entirely in the network	Motion-to-photon at most 20 ms end to end, including uplink of the pose; below 10 ms in each direction and 0.1 to 1 Gbps for network-rendered XR
3GPP TR 26.928 (2020)	Raster-based split rendering with on-device ATW	Pose-to-render-to-photon and round-trip interaction delay typically 50 to 60 ms; new 5G QoS identifiers may be needed for latency of 10 to 20 ms with guaranteed bitrates of 50 to 100 Mbps
Huawei CG Cloud VR specifications (draft)	CG cloud VR over 4.5G	Assumed network capability: end-to-end RTT 10 to 15 ms; downlink 50 Mbps and uplink 2 Mbps (plus or minus 20%). Experience constraints: binocular resolution above 2K; above 60 fps; motion-to-photon below 20 ms
Huawei CG Cloud VR specifications (draft)	CG cloud VR over 5G	Assumed network capability: end-to-end RTT 5 to 8 ms; downlink 100 Mbps and uplink 20 Mbps (plus or minus 20%). Experience constraints: binocular resolution above 3K; above 90 fps; motion-to-photon below 20 ms; cloud rendering below 5 ms
China Mobile Fujian (Zhang Jinshui)	Cloud VR "fair-experience phase"	At least 200 Mbit/s and end-to-end latency under 20 ms; 5 GHz Wi-Fi required
GSMA (2019), China Mobile Fujian case	4K panoramic VR video at 40 Mbps; cloud VR gaming	60 Mbps for smooth playback; network round-trip delay including the air interface below 20 ms for gaming
Hou, Lu and Dey (2017)	Measured with a 1080p stream on an Oculus headset with head motion	28.2 Mbps (virtual classroom) or 39.7 Mbps (racing game) at 75 fps; acceptable total latency 22 ms
Elbamby et al. (2018)	Wireless VR in general	Motion-to-photon upper bound of 15 to 20 ms; about 14 ms left for computing and communication if display delay falls to 5 ms

History

Date	Event
July 2017	Hou, Lu and Dey present "Wireless VR/AR with Edge/Cloud Computing" at ICCCN 2017.
February 2018	Huawei shows an early version of its end-to-end Cloud VR solution at Mobile World Congress.
18 July 2018	China Mobile Fujian launches "AND Cloud VR" for home broadband users with Huawei, described by Huawei as the world's first operator cloud VR service and, in its September 2018 release, as launched as part of a trial.
14 September 2018	Huawei releases its Cloud VR Solution White Paper at the fifth Ultra-Broadband Forum (UBBF 2018).
2019	The GSMA describes cloud AR/VR streaming over 5G and edge cloud and sets up its 5G Cloud XR project.
16 April 2019	Huawei and its VR OpenLab partners announce the "Dual-G" Cloud VR initiative for "Gigaband and 5G" Cloud VR; China Mobile Fujian and China Telecom are named as running commercial trials.
22 October 2019	NVIDIA announces CloudXR for streaming cloud-rendered AR and VR, with SteamVR and OpenVR support.
March 2020	Blade announces a closed beta of a Quest app for its Shadow cloud PC service to stream owned SteamVR games.
May 2020	Facebook's Jason Rubin says cloud-rendered Oculus VR is further away than five years.
12 October 2021	CloudXR becomes publicly available on the Google Cloud Marketplace, following earlier availability on Amazon Web Services and Microsoft Azure.
1 February 2022	PlutoSphere launches in early access on SideQuest for Quest and Quest 2.
5 March 2024	PlutoSphere stops selling play tokens and begins shutting down.
18 March 2024	NVIDIA demonstrates streaming of Omniverse industrial scenes to Apple Vision Pro through its Graphics Delivery Network.
June 2024	Varjo stops developing its Reality Cloud streaming service.
30 September 2025	Microsoft retires Azure Remote Rendering.
17 March 2026	NVIDIA announces native CloudXR 6.0 support in visionOS, streaming from RTX PCs or the cloud.

Operator cloud VR services

China Mobile's Fujian branch launched its "AND Cloud VR" service for home users on 18 July 2018, together with Huawei and other VR partners. The launch was paired with an upgrade of optical broadband from 100 Mbit/s to 1000 Mbit/s. Huawei's September 2018 release described the launch as part of a trial. Its content covered VR live broadcasts, video on demand, "VR IMAX", games and education. Huawei said the service used cloud GPU rendering and a cloud server architecture, and that by moving computing off the head-mounted display it lowered VR hardware costs by 70% to 80%; that figure is the companies' own claim. In an article for Huawei's WinWin magazine, Zhang Jinshui of China Mobile Fujian compared a high-end PC VR setup, needing a PC of around US$2,000, with an all-in-one headset at about US$300. He gave the service's user count at the time as about 100 and set targets of 1,000 users in 2018, 10,000 in 2019 and more than 50,000 cloud VR users on 5G networks in 2020.

Huawei released a Cloud VR Solution White Paper on 14 September 2018 at UBBF 2018 in Geneva, describing a solution with four parts (a content platform, a service platform, a network solution and a terminal solution) for VR IMAX, live broadcast, 360-degree video, gaming and education. Huawei said the solution used 5 GHz Wi-Fi to remove the cable from the headset, and that aggregating and rendering content in the cloud protected content copyright. In April 2019 Huawei and its VR OpenLab partners, including NVIDIA, BOE, Skyworth, Pico, DPVR and iQIYI, launched a "Dual-G" Cloud VR initiative covering what the announcement called "Gigaband and 5G Cloud VR". The announcement said China Mobile Fujian and China Telecom had started commercial trials, that VR OpenLab had more than 60 cloud VR partners, and that Skyworth's and iQIYI's 8K headsets already supported the architecture. Huawei iLab authors later described the technology in the 2020 CRC Press book Cloud VR: Technology and Application.

The GSMA promoted the same model to mobile operators. Its 2019 booklet argued that operators could act as content aggregators, run cloud AR/VR game streaming or provide managed services such as VR training for enterprises. It described a joint deployment by Huawei, China Mobile and Deutsche Telekom of VR factory planning, fire-fighting training and school education applications, streamed with CloudXR from virtual machines at each company over a common edge platform.

Cloud streaming platforms

NVIDIA CloudXR

Main article: NVIDIA CloudXR

NVIDIA announced CloudXR on 22 October 2019 as a set of tools for other businesses rather than a consumer service. Road to VR reported that NVIDIA hoped carriers building 5G networks would offer CloudXR streaming to their customers. The early SDK had a server driver for the data center, a client library for streaming OpenVR applications to Android and Windows devices, and an SDK for portable clients. NVIDIA claimed "no detectable latency difference" compared with local rendering; Road to VR's Ben Lang was skeptical and argued that physical distance to the data center, and so the availability of nearby edge nodes, would matter as much as 5G. By October 2021 CloudXR was available on Amazon Web Services, Microsoft Azure and the Google Cloud Marketplace, where it was offered as a virtual machine image with NVIDIA RTX Virtual Workstations. In March 2026 NVIDIA announced that CloudXR 6.0 was natively integrated in visionOS, with dynamic foveated streaming from visionOS 26.4, and that content could stream from RTX workstations and GeForce RTX GPUs on PCs or in the cloud.

Enterprise visualization

Microsoft's Azure Remote Rendering rendered complex 3D models on cloud GPUs and streamed them as video to HoloLens 2 or Windows desktop PCs. Microsoft illustrated the need with an engine model of more than 18 million triangles next to a reduced version of around 200,000 triangles, the kind of simplification untethered devices otherwise need. For models too large for one GPU the service split the work across several GPUs and merged the result into one image. It was retired on 30 September 2025.

Varjo's Reality Cloud was a commercial XR streaming service for enterprise customers. In April 2024 Varjo announced support for Meta Quest 3 and Quest Pro, but the following month it said it would end development in June 2024 because enterprise requirements for XR cloud streaming were "more diverse and security-driven than originally anticipated", and that it would focus on third-party and local streaming instead. At GTC in March 2024 NVIDIA showed a car configurator built on NVIDIA Omniverse by the studio Katana streaming to Apple Vision Pro through Omniverse Cloud APIs and NVIDIA's Graphics Delivery Network. NVIDIA's Rev Lebaredian said that traditional spatial workflows require developers to decimate their datasets, which does not suit very large industrial engineering and simulation data.

Consumer PC VR in the cloud

Cloud PC services have been used to run PC VR games for standalone headsets. In March 2020 the French company Blade announced a closed beta of a Quest app for its Shadow cloud PC service, letting subscribers stream SteamVR games they already owned without a gaming PC. UploadVR's David Heaney, who described himself as especially sensitive to latency, found the delay very noticeable even with a high-end Wi-Fi router. MIXED reported in April 2025 that Shadow still supported VR cloud streaming but that the VR app had to be sideloaded.

Pluto VR launched PlutoSphere on SideQuest on 1 February 2022 as an early-access service for the Oculus Quest and Quest 2. It gave users a cloud PC with Steam preinstalled on which they could play games such as Half-Life: Alyx that they already owned, and asked for a 50 Mbps internet connection and a 5 GHz Wi-Fi 6 router. Road to VR listed the virtual rig as an Intel Xeon Platinum 8259CL processor at 2.50 GHz, an NVIDIA Tesla T4 GPU and 32 GB of RAM, and noted that all data was wiped after each session. Play time was sold as tokens at roughly US$1 to US$3 per hour. On 5 March 2024 UploadVR reported that the service had stopped selling tokens and was shutting down. In a statement, the company said that because it could not launch on the Meta store and grow its user base, it had been unable to honor its commitments with a third-party reseller of Amazon Web Services cloud services, and CEO Jared Cheshier said that being kept off the official Quest store had "hampered our marketing strategies, restricted our growth and prevented us from fulfilling our usage obligations".

Adoption and obstacles

Store policy has limited consumer cloud VR on the most popular headsets. Meta's developer requirement VRC.Quest.Streaming.2 lets apps stream stereoscopic, head-tracked or immersive content only from a local PC that the customer has physical access to, and states that "streaming from virtual devices or cloud sources is not allowed unless expressly approved by Meta in writing". As a result, PlutoSphere could not be distributed through the Quest Store or App Lab and had to be sideloaded. In May 2020 Facebook's Jason Rubin said of the time needed for cloud-rendered Oculus VR, "It's not a million. It's not five. It's somewhere between." MIXED reported in April 2025 that Meta had tested its own cloud VR streaming, codenamed Avalanche, for years, and that a test app had briefly appeared in the Horizon Store before being removed; the Xbox Cloud Gaming and GeForce Now apps available on Quest stream only flat 2D games. The outlet attributed the slow progress to latency and unstable image quality, which matter more in VR than on a TV or monitor, and to the small size of the VR market.

Operators and researchers point to network capability as the other limit. Huawei's draft specification notes that panoramic video had already been streamed over wireless wide-area networks, but that whether CG VR could use such connections "has not yet been proven in the industry". A 2024 study by Henrique Souza Rossi and colleagues at Luleå University of Technology and Infovista, using CloudXR to stream Serious Sam VR: The Last Hope, predicted mean opinion scores of 2.5 to 3.0 for VR cloud gaming over measured 4G and 5G mobile network traffic, against 4.2 to 4.4 over wired connections, and concluded that commercial 5G deployments needed to improve to deliver the service.

Research

A 2018 survey in ACM Transactions on Multimedia Computing, Communications, and Applications by Maha Abdallah, Carsten Griwodz, Kuan-Ta Chen, Gwendal Simon, Pin-Chun Wang and Cheng-Hsin Hsu grouped VR and AR with cloud gaming and telepresence as "conversational" video services. The authors wrote that these are hard to support from the cloud because most cloud servers are far from users while the services need high bandwidth, short delay and high heterogeneity, in contrast to presentational services such as YouTube and Netflix.

Hou, Lu and Dey measured how much bandwidth a single 1080p stream needed for a virtual classroom and a Unity racing game. With head motion on an Oculus headset at 75 fps, the streams needed 28.2 Mbps and 39.7 Mbps, and the acceptable total latency was 22 ms, against 100 to 200 ms for the classroom shown on a PC monitor. They proposed multi-user encoding to cut the bandwidth needed when many users share one virtual space, and noted that it is easier to apply to cloud rendering than to edge rendering. Elbamby and colleagues pointed out that the 4G loopback latency under ideal conditions is 25 ms, above the 15 to 20 ms motion-to-photon bound, and proposed predicting users' head movements so that frames could be computed in advance in remote cloud servers and cached at the network edge or in the headset. Their case study of an interactive VR gaming arcade combined millimeter-wave links, edge computing and proactive caching.

Later work measured real systems. Eugene Korneev, Mikhail Liubogoshchev, Dmitry Bankov and Evgeny Khorov analyzed the traffic of a cloud VR application on a Pico Neo 2 headset and reported that the 3GPP baseline underestimates the effective VR bitrate, the bandwidth needed to deliver the traffic, by 25%. They used their model to evaluate intra-refresh video coding and headset-side buffering tuned to motion-to-photon requirements. Rossi and colleagues ran subjective tests with 30 users and built regression models that predict quality of experience from round-trip time, jitter and packet loss; in their CloudXR setup a packet loss rate of 6% crashed the service. In a poster at ISMAR 2024, Maximilian Warsinke and colleagues at the Technische Universität Berlin simulated degraded networks (a fixed 40 ms delay, or 0.3% packet loss) while 16 participants played Beat Saber and Cubism streamed to a Meta Quest 3 over Air Link, as a model for VR cloud gaming. Packet loss affected the experience more than delay, increasing cybersickness in Beat Saber and reducing presence in Cubism.

Other research avoids sending finished frames at all. Yuqi Zhou and Voicu Popescu of Purdue University argued in 2024 that a pure cloud VR approach, with all rendering done on the server, was not yet tractable because networks could not yet meet VR's strict latency requirements. Their CloVR system has the server split the virtual environment into a near region sent as geometry and a far region rendered into an environment map, so that the headset renders its own frames with low latency; the near region starts small and grows progressively to shorten loading times. They tested it with groups of eight Quest 2 users supported by a laptop server running a collaborative application with a 25-million-triangle environment.

See also

• Split rendering
• NVIDIA CloudXR
• Edge computing
• 5G
• Motion-to-photon latency
• Asynchronous timewarp
• Foveated rendering
• Pluto VR
• Varjo Reality Cloud