The Ceiling of Conventional CDN Streaming
For nearly twenty years, Content Delivery Networks (CDNs) have served as the foundation of all forms of digital media, providing a hierarchical distribution of geographically located servers which can store and provide video in close proximity to each viewing audience member. This model is simple and scalable for all types of non-interactive, pre-rendered content. However, CDNs were not built for the needs of immersive computing.
The issue is fundamentally architectural. All conventional CDN-based streaming models are optimised for maximum throughput rather than minimum response time. The HTTP-based adaptive bitrate protocols used in most CDN deployments today (HLS & MPEG-DASH) function by requesting multiple pre-encoded video segments (typically 2-6 seconds in length) from an edge cache. While both LL-HLS and LL-DASH may deliver content with latency times between one to three seconds, this represents latency that is still many orders of magnitude slower than the motion-to-photon (MTP) latencies required for interactive and spatially aware XR applications.
For example, extended reality applications require MTP latency times under twenty milliseconds to prevent the onset of cybersickness; conventional CDN technology does not come anywhere near meeting this threshold. According to IEEE researchers, “traditional edge-centric architectures find it difficult” to meet these very low MTP latency requirements due to the fact that per-frame rendering is highly computationally intensive and directly tied to user input issues that are completely outside of the capabilities of a static video cache (Wen et al., arXiv 2601.20625).
Edge computing, 5G/6G and real-time rendering pipelines as an alternative to CDNs
In order to solve the problem of CDNs, we have a shift from content delivery to compute delivery. Instead of delivering encoded content to the viewer, the future model delivers computation from a server located near the radio access point of a mobile device (i.e., edge computing) that responds to user requests in less than ten milliseconds.
The required technical foundation is provided by 5G low-latency networking capabilities such as low millisecond air interface latency and high-bandwidth connectivity. Telecommunications companies are therefore placing computational resources (GPU-based servers) directly at the base stations where they enable a motion-to-photon latency of about twenty milliseconds or lower for tethered head-mounted displays (HMDs) (Mordor Intelligence Spatial Computing Market Report, 2025).
The target latency for wireless networks in the 6G research community is expected to be around one millisecond. In addition, it is anticipated that these networks will provide terabit-level bandwidth, which should make it feasible to send the entire volumetric scene to a client over wireless links without needing to perform local rendering.
Simultaneously there are also new developments in real-time rendering pipelines. Techniques such as Neural Radiance Fields (NeRF), developed by NVIDIA, and Gaussian Splats enable us to render photorealistically complex 3D scenes based on sparse camera data. As a result, no longer do developers need to create all polygons by hand. In this regard, NVIDIA has recently published Slang as a shader programming language that automatically performs automatic differentiation in graphics pipelines so that neural rendering may take place within game engines in real time.
As a result, it becomes possible for the rendering intelligence to reside at the edge and deliver display streams of low weight to thin clients, headsets, smartphones, glasses, etc., that require merely decoding pixel information and no longer need to process geometric computations.
Edge Computing, 5G/6G, and Real-Time Rendering Pipelines
The response to CDN limitations is a structural shift from content delivery to compute delivery. Rather than moving pre-encoded assets to viewers, the emerging paradigm moves rendering workloads to the network edge – servers physically co-located with radio access nodes that can respond within single-digit milliseconds.
5G’s sub-millisecond air-interface latency and multi-gigabit throughput are essential enablers. Telecom operators are now embedding edge compute nodes – GPU-equipped servers – directly into base station infrastructure, trimming motion-to-photon latency below 20 milliseconds for tethered XR devices (Mordor Intelligence Spatial Computing Market Report, 2025). 6G research targets sub-1-millisecond latency and terabit-class throughput, promising wireless pipelines capable of streaming full-resolution volumetric scenes without local rendering hardware.
Parallel advances in real-time rendering pipelines are equally significant. NVIDIA’s neural radiance field (NeRF) and Gaussian splatting techniques allow photorealistic 3D scenes to be reconstructed from sparse camera data and rendered on the fly, rather than requiring artists to hand-author every polygon.
Khronos Group’s Slang shader language, launched in 2024, integrates machine learning directly into graphics pipelines through automatic differentiation, enabling neural rendering inside game engines in real time (Khronos Group, 2026). This means the rendering intelligence can live at the edge, serving lightweight display streams to thin clients, a headset, a phone, a pair of glasses – that need only decode pixels, not compute geometry.
Volumetric Video, Digital Twins, Sensor Fusion
Three technologies sit at the core of spatial presence: volumetric video, digital twins, and sensor fusion.
Volumetric video represents human subjects and objects as navigable 3D point clouds, rather than 2D arrays of pixels. Unlike conventional video, it provides viewers with six degrees of freedom – the ability to move around a subject, see it from any angle and naturally experience depth. The cost of the infrastructure is high; raw point cloud streams require orders of magnitude more bandwidth than H.265 video for the same visual quality.
Recent work from the University of Wisconsin-Madison and Microsoft Research Asia shows that lookup-table-based super-resolution techniques can reduce volumetric streaming bandwidth up to 70% and improve perceptual quality by 36.7% (VoLUT, arXiv 2502.12151), making consumer-grade delivery more plausible.
Digital twins generalise this principle to environments. They are driven by LiDAR scanning, photogrammetry and AI inference to create dynamic, millimetre-accurate replicas of physical spaces that update in real time as conditions change (INAIRSPACE, October 2025).
Gartner predicts that more than 50% of large industrial companies will be using digital twins for safety and decision-making by 2027 (Network Optix, 2024). The infrastructure to support a live digital twin – continuously ingesting sensor data, reconciling with a 3D model and serving updates to distributed clients – is orders of magnitude more complex than video streaming. It’s more like a distributed database than a CDN.”
Sensor fusion ties these threads together. Today, the perception stack of modern XR devices includes LiDAR, radar, high-resolution RGB cameras, inertial measurement units, and eye-tracking sensors. The multimodal stream uses on-device AI to maintain a semantic understanding of the environment, what surfaces there are, what objects there are, where people are looking (INAIRSPACE, November 2025). The result is a live spatial map, a prerequisite for anchoring virtual content to physical reality with sub-centimetre precision.
Infrastructure Challenges: Bandwidth, Synchronisation, and Distributed Orchestration
The infrastructure challenges of immersive media are not merely quantitative extensions of video streaming – they are qualitatively different.
Bandwidth. According to IEEE researchers, 4K multi-view 360° video with depthб a baseline for high-fidelity XR – requires significantly more bandwidth and processing resources than current mobile network norms (Dong & Lee, arXiv 2301.07740). Unlike flat video, where bit-rate adaptation can gradually degrade quality, XR quality degradation causes disorientation and motion sickness, making adaptive streaming more important.
Sync. In multi-user immersive environments, all users must experience the same state (object positions, physics interactions, avatar poses) within a small enough temporal window to feel realistic. This is a distributed-systems problem, at its core. Latency spikes from TCP’s congestion control mechanisms impact VR synchronisation, whereas UDP can lead to inconsistent states for different participants (Asim & Subramanian, arXiv 2507.20050). New transport protocols will have to balance these constraints, keeping reliability where consistency is important and sacrificing it where speed is critical.
Orchestration of compute. Serving a persistent shared 3D environment requires the coordination of the GPU rendering capacity, the radio resource allocation and the content delivery in a heterogeneous distributed system – in real time. O-RAN (Open Radio Access Network) architectures provide a promising framework: by exposing compute and radio resources through open interfaces, intelligent controllers can jointly optimise rendering quality, transmit power and bandwidth allocation. Experimental results on 5G O-RAN testbeds show that reinforcement learning agents can reduce median MTP latency by more than 11% compared to static allocation strategies (Wen et al., arXiv 2601.20625).
Interoperability Standards and Protocols
Fragmentation has always been a problem for immersive media. Each platform, from Meta’s Quest to Apple Vision Pro to Microsoft HoloLens, used its own proprietary APIs, spatial anchoring systems and asset formats, making cross-platform experiences impractical.
The standardisation landscape is settling down. The Khronos Group’s OpenXR has emerged as the de facto cross-platform API for XR application development, with support from Unity, Unreal Engine, and Godot, and adoption by Meta, Microsoft, Pico, and XREAL. OpenXR 1.1 was released in April 2024, bringing popular extensions into the core specification. More significantly, the OpenXR Spatial Entities Extensions, released in June 2025, set the first open standard for spatial computing proper: cross-platform plane detection, marker tracking, spatial anchors, and cross-session persistence (Khronos Group, June 2025).
OpenXR is complemented by the MPEG-I (ISO/IEC 23090) standard family, which covers immersive media compression and scene description, including spatial audio extensions for glTF, the 3D asset format that is rapidly becoming the ‘JPEG of 3D’. The Metaverse Standards Forum offers a way to coordinate these entities, connecting the work of Khronos, 3GPP and ISO (Metaverse Standards Forum, 2025).
WebRTC is still the dominant low-latency real-time communication technology at the network layer. WebRTC was designed for small-group video calls and does not scale well. New extensions like WHEP (WebRTC HTTP Egress Protocol) help in bridging WebRTC with CDN infrastructure. Emerging QUIC-based protocols offer better congestion control for spatial data streams (CDN Alliance Whitepaper, 2025).
Data Ownership, Privacy and Security in Spatial Ecosystems
Spatial computing engenders a new form of data sensitivity that has no equivalent in traditional media: the constant and granular mapping of private physical spaces. In everyday use, an XR headset doesn’t just know what content you are watching; it knows the geometry of your home, the layout of your workplace, the biometrics of your body, and the pattern of your gaze.
Meta’s own privacy policy says headsets collect “information about your environment, physical movements and dimensions” (echo3D / Medium, 2023). The aggregation of biometric, behavioural, and cognitive data in persistent XR environments presents acute risks of profiling, behavioural manipulation and identity theft (arXiv 2411.04508; Virtual Worlds, MDPI 2025).
The legal structures for ownership of spatial data are still largely under development. The detailed 3D scans created by devices have commercial value, are personally intimate and are not legally classified in most jurisdictions (INAIRSPACE, January 2026). A key industry response is the first principle of on-device processing, which means that environment maps are processed locally by neural processing units (NPUs), and only abstracted semantic data is transmitted to the cloud, ensuring the raw spatial record never leaves the device (INAIRSPACE, December 2025). Hardware kill switches for cameras and microphones are now a standard feature. Although regulatory clarity, particularly in the EU with the AI Act and future spatial data frameworks, is beginning to boost enterprise adoption confidence, all-encompassing legislation is still ahead of the technology (Mordor Intelligence, 2025).
A Forward View: Infrastructure as Experience
The trajectory is clear: immersive media infrastructure is converging to a real-time, distributed, multi-sensory compute fabric, a persistent layer of shared spatial intelligence mapped on top of the physical world.
Volumetric telepresence, in communication, is already showing its ability to communicate non-verbal cues, spatial proximity and collaborative manipulation of 3D objects in ways flat video cannot. More and more teams will meet up in persistent virtual workplaces – digital twins of real offices or purpose-built creative spaces – instead of the grid-based video calls (Deloitte Insights, 2025).
The spatial computing market in work and industry is projected to reach $3.98 billion in 2025 and is expected to reach $23.45 billion by 2030 at a CAGR of 42.53%, fuelled by enterprise digital twin deployments, cloud rendering services, and managed XR fleets (Mordor Intelligence, 2025). Infrastructure providers (telcos, cloud platforms, CDN operators) are transforming from pipe-and-cache utilities to spatial compute platforms.
And in entertainment, the shift from passive viewing to spatial participation will require a reimagination of every layer of the production and delivery stack. Content will be navigable environments, not chronological timelines. Streaming is about synchronising shared state, not pushing identical bytes to disconnected viewers.
The shift in infrastructure from streaming video to spatial presence is ultimately a shift from a broadcast model to a networked reality model. The engineering challenges – latency, bandwidth, synchronisation, and security – are formidable but surmountable. The real work is in creating the governance, standards and trust architectures that will enable billions of people to share and rely on a common spatial layer of the world.
