Cloudflare Stream Opens Up Low-Latency Playback

Cloudflare has made its Stream product capable of sub-second latency video playback in open beta. The service now supports delivery over SRT and RTMPS protocols, which are designed for real-time video transport, rather than relying exclusively on HTTP-based formats like HLS and DASH. Pricing remains consistent with other delivery methods: $1 per 1,000 minutes delivered, regardless of encoding settings.

The company describes the service as akin to an HDMI cable for the internet—users can connect a live stream from one source and display it on screens anywhere in the world without significant delay.

Defining Sub-Second

Latency is the time between a camera capturing an event and a viewer seeing it on their display. Even traditional cable TV runs about five seconds behind the live action. On the internet, the gap varies widely depending on the technology used.

Streaming standards like HLS and DASH, which power most major video platforms, typically introduce 10 to 30 seconds of delay. That range becomes problematic for interactive use cases: a live esports broadcast with text chat, a Q&A in a remote webinar, or any scenario where the viewer needs to feel present in the same room as the broadcaster. For those applications, latency needs to drop to one or two seconds or less.

Cloudflare's approach routes video primarily within its own network. A broadcaster connects to the nearest Cloudflare point of presence (PoP), and viewers connect to their nearest PoP, with the Cloudflare backbone carrying the stream between them. That avoids long, unreliable internet paths and lets the company actively measure both latency and reliability at every point from the source to the last mile.

Dynamic Routing for Minimal Delay

Rather than routing through a central hub, Stream treats each PoP as a peer. Content flows directly between Cloudflare locations, meaning the latency is largely determined by physical distance. The topology for each stream is constructed dynamically and expands as the audience grows, so it handles both small broadcasts and large events.

This design also means any server in the Anycast network can receive or publish low-latency video, requiring no configuration on the customer's end to find the nearest ingest point. Traffic takes advantage of the same network telemetry that powers Argo smart routing to avoid congestion and reduce delay.

Early testing has measured roughly 500ms of glass-to-glass latency between San Francisco and London over residential connections. A beta tester achieved 300ms by broadcasting with OBS over RTMPS and using ffplay over SRT to pull the stream down. As the service matures, Cloudflare says it will continue working on both latency and playback reliability.

BLOG-1081 Embedded Image - 0w4IT9

Why Not Just HTTP?

HTTP dominates internet video delivery because it plays well with existing caching infrastructure and browser compatibility. But it wasn't built for video, and attempts to push its latency lower over the years have only gotten streaming down to a few seconds. WebRTC and QUIC-based delivery hold promise but face inconsistent support across devices and platforms today.

RTMPS and SRT are purpose-built for video and sidestep many of HTTP's latency issues, but historically traded that performance for a lack of browser support and a need for custom client libraries. Cloudflare said it is exploring other delivery methods going forward, potentially including WebRTC ingest or playback over QUIC, HTTP/3, or WebTransport. The goal, the company said, is not to pick a single winner but to support a mix of protocols so customers can tailor delivery to their needs—receiving over RTMPS and delivering over SRT today, for instance, or ingesting via SRT while publishing HLS.

Interactive Use Cases

Service should unlock new kinds of real-time interactions. A remote cycling instructor could react to a rider's heart rate instantly, or an esports league could feed live video to commentators without lag before they voice their reactions.

Other applications include second-screen augmentation at concerts or sporting events—readers can see synchronized replays as actions happen live—as well as security camera monitoring from anywhere, with no open ports or relay servers. The company expects the protocol mix to evolve, and the current support span ingest and output will allow developers to choose the path that best fits their applications.

To get started, Cloudflare Stream customers can simply use existing inputs already connected to the Service; they will immediately be available for RTMPS or SRT playback URLs without any reconnection. New accounts start at $5 per month. Push and delivery still require an encoder like FFmpeg or OBS, and players like VLC and FFplay are needed for playback testing. For native iOS apps, FFmpeg wrappers such as ffmpeg-kit are available. The functionality also works with custom ingest domains, letting brands keep their own domain names in playback URLs.