How does GeForce Now stream games?

GeForce Now runs the actual game binary on an NVIDIA RTX GPU inside an edge data center, encodes the rendered frames with NVENC into AV1 or HEVC, and streams the bitstream over UDP using a proprietary low-latency protocol layered on top of the NVIDIA Streaming SDK. Player input flows in the reverse direction over the same session. Each session is pinned to one of roughly 30 global regions, and the client only handles decode, display, and input capture, which is why even a Chromebook can run Cyberpunk 2077 at 60 fps.

What backend does Xbox Cloud Gaming use?

Xbox Cloud Gaming (xCloud) runs custom Xbox Series X server blades inside Azure datacenters: real console silicon, not virtualized PCs. The backend stack is Azure App Service for the session broker, Cosmos DB for entitlements and save metadata, Azure Front Door for global routing, and a proprietary streaming protocol over UDP that ships H.265 video at up to 1080p60. Save data round-trips through the same Xbox Live backend (PlayFab) the consoles use, which is why progression carries across console and cloud.

What's the latency budget for cloud-gaming backends?

The industry target is sub-50 ms input-to-photon end-to-end. That budget breaks down roughly as: 5-15 ms one-way network from client to edge GPU, 8-16 ms render frame on the server, 6-10 ms encode (NVENC AV1 or HEVC), 5-15 ms one-way return, plus 8-16 ms decode and display on the client. GeForce Now and xCloud target dense regional footprints (sub-30 ms RTT for 80% of users) because anything past 80 ms feels like wireless lag to the player.

How does cloud gaming handle save data?

Save data lives in the platform's backend, not on the streaming VM. GeForce Now stores PC saves in a per-user Steam-style cloud blob synced with the publisher's own backend (Steam Cloud, Epic Online Services, or a custom save service). Xbox Cloud Gaming reuses Xbox Live cloud saves backed by PlayFab. The streaming VM is stateless: it mounts the player's save at session start, writes deltas during play, and is wiped on session end so the next user gets a clean machine.

Can I run my own game on GeForce Now?

Yes, but only via the GeForce Now Cloud Game Onboarding program. The publisher submits the title for compatibility certification (NVIDIA verifies Steam, Epic, Ubisoft Connect, or other launcher integration, anti-cheat behavior, save sync, and DRM), then opts in per SKU. There is no public self-serve upload like Steam Direct. Once approved the game runs in NVIDIA's RTX Server pool and is billed against the publisher's distribution agreement. Indie titles typically route through Steamworks for the cleanest onboarding path.

How GeForce NOW & Xbox Cloud Gaming Actually Work

A technical look at cloud gaming latency, encoding, edge GPU infrastructure and integration boundaries behind GeForce NOW and Xbox Cloud Gaming.

Written and reviewed by Supercraft Engineering

Cloud gaming is projected to grow from a $0.99B market in 2026 to $4.86B by 2035 (19.2% CAGR). Behind services like GeForce Now, Xbox Cloud Gaming (xCloud), and PlayStation Now lies a radically different backend architecture-one designed for real‑time video encoding, input‑stream synchronization, and global CDN delivery. This guide explains how cloud‑gaming backends work, how they differ from traditional multiplayer hosting, and what developers need to know to integrate their games.

The streaming shift: Cloud gaming isn’t just “remote desktop for games.” It’s a specialized backend stack that must deliver 60‑120fps video with under 50ms input‑to‑photon latency, while scaling to millions of concurrent players across dozens of regions.

Why Cloud‑Gaming Backends Are a Separate Category (2025‑2026)

Market validation: GeForce Now surpassed 25 million users in 2025; xCloud is bundled with Xbox Game Pass Ultimate (34 million subscribers).
Technology maturation: AV1 encoding, WebRTC streaming, and edge‑GPU virtualization have made streaming viable for all but the most twitch‑sensitive genres.
Business model evolution: Revenue‑share agreements (typically 70/30) are now standard, making cloud‑gaming partnerships financially attractive for studios.
Hardware democratization: NVIDIA’s Grace Hopper superchips and AMD’s Instinct MI300X have driven GPU‑hosting costs down by 40% since 2023.
Player demand: 42% of gamers have tried cloud gaming; 18% use it weekly (Newzoo 2025).

How Streaming Backends Differ from Dedicated Game Servers

Traditional multiplayer backends focus on game‑state synchronization. Cloud‑gaming backends focus on video‑frame generation and delivery.

Aspect	Traditional Game Server Backend	Cloud‑Gaming Backend
Primary output	Game‑state updates (JSON, binary)	Compressed video stream (H.265, AV1)
Network protocol	UDP (custom), WebSocket	WebRTC, RTMP, SRT
Latency target	20‑100ms (input‑to‑state)	30‑50ms (input‑to‑photon)
Compute resource	CPU‑bound (game logic)	GPU‑bound (rendering + encoding)
Scaling unit	Virtual machines (4‑8 vCPU)	GPU instances (A10, A100, MI300X)
Cost driver	CPU‑hours, bandwidth	GPU‑hours, encoding licenses, CDN egress

Rule of thumb: If your game already runs on dedicated servers, you’re 30% of the way to supporting cloud gaming. The remaining 70% is video‑encoding infrastructure and input‑stream synchronization.

Key Components of a Cloud‑Gaming Backend

1. GPU Virtualization & Orchestration

Each player gets a dedicated GPU slice (virtual GPU or full physical GPU) for the duration of their session.

NVIDIA vGPU (M10, A10): Time‑sliced GPUs, 1:8 to 1:16 ratio (8‑16 players per physical GPU).
AMD MxGPU (MI300X): Hardware‑partitioned, 1:4 to 1:8 ratio.
Orchestrator: Kubernetes with GPU‑scheduling extensions (NVIDIA K8s‑Device‑Plugin, AWS Neuron).

2. Video Encoding Pipeline

The rendered frames must be compressed in real‑time with minimal quality loss.

Codec	Bitrate at 1080p60	Encoding Latency	Hardware Support
H.264 (AVC)	5‑8 Mbps	2‑5ms	Universal (NVENC, QuickSync)
H.265 (HEVC)	3‑5 Mbps	3‑7ms	Modern GPUs (2018+)
AV1	2‑4 Mbps	8‑15ms	NVIDIA Ada Lovelace, Intel Arc

Encoding architecture: Each GPU runs a parallel encoder (NVENC, AMF) that compresses frames as they’re rendered. The encoded stream is packetized and sent to the streaming gateway.

3. Input‑Stream Synchronization

Player inputs (keyboard, mouse, controller) must reach the game with sub‑10ms latency and be correctly matched to the video frame they affect.

// Simplified input‑to‑photon flow
1. Client sends input packet (timestamp t0)
2. Edge gateway receives at t0 + network_latency (≈5ms)
3. Input‑router forwards to correct GPU instance (≈1ms)
4. Game engine processes input in next simulation tick (≈16ms at 60Hz)
5. Rendered frame includes input effect (≈8ms render)
6. Frame encoded and sent (≈5ms)
7. Video reaches client at t0 + ~35ms total

Critical detail: Inputs are batched per‑frame and time‑warped if they arrive late. Prediction algorithms (dead‑reckoning) fill gaps when network jitter occurs.

4. Streaming Delivery Is a Platform Concern

If you are publishing through GeForce NOW, Xbox Cloud Gaming or Luna, do not assume the platform’s CDN topology, commercial split or per-title integration contract from a generic architecture article. Those terms are platform- and agreement-specific. Treat the streaming platform as an external runtime and keep your own auth, saves, entitlements and multiplayer backend reachable through documented APIs.

Use Published Network Requirements as a Test Boundary

NVIDIA’s current GeForce NOW system requirements publish concrete client-side network guidance: for example, 15 Mbps for HD/60 FPS, 25 Mbps for FHD/60 FPS, and less than 80 ms network latency from an NVIDIA data center. Those are service requirements, not a guarantee of your game’s input-to-photon latency.

For your own integration, test the parts you control:

startup/login without assuming local durable disk;
cloud-save conflicts when a player switches between local and streamed sessions;
controller and remapping behavior through virtualized input;
multiplayer region selection when the stream host and the player are in different network locations;
support telemetry that separates game-server latency from streaming-path problems.

Getting Started: Should You Build or Partner?

Most game studios should integrate with an existing streaming platform rather than build video-streaming infrastructure themselves. Consider owning the streaming stack only when it is itself part of the product or you already operate the GPU, encoder, transport and regional-capacity layers required to run it reliably.

Partner‑First Roadmap

Evaluate platforms: GeForce Now (broadest reach), xCloud (Game Pass integration), Luna (Twitch integration).
Review partnership terms: Revenue‑split, marketing support, technical requirements.
Allocate engineering time: 4‑12 weeks for integration, depending on game complexity.
Plan launch marketing: Cross‑promote your cloud‑gaming availability to capture new players.
Monitor performance: Track streaming‑specific metrics (input latency, stream quality, churn).

Related in This Hub

Edge‑Computing Game Backends – Low‑latency infrastructure for streaming.
AI/ML‑Powered Game Backends – AI‑enhanced streaming (upscaling, bandwidth prediction).
Multiplayer Backend Architecture Patterns – How streaming fits alongside traditional multiplayer.
Serverless Game Backends – Managing live‑ops for cloud‑gaming titles.
Game Server Backend hub – All backend guides.

Cloud gaming is no longer a futuristic concept-it’s a viable distribution channel that can expand your player base by 10‑20% with minimal upfront investment. By understanding the backend infrastructure and partnership models, you can make informed decisions about when and how to embrace game streaming. If your streaming stack is also absorbing AI workloads such as upscaling or prediction, Supercraft AI is a natural companion layer.