What Is the Most Expensive Part of a Live Streaming App’s Infrastructure?

Bandwidth. Specifically, the cost of moving video segments from your origin servers to every concurrent viewer. It accounts for 50–70% of total infrastructure spend, and unlike transcoding or storage, it scales linearly with audience size.

This is why Twitch has never turned a profit despite generating roughly $2 billion in annual revenue, and why it exited South Korea entirely over network fees.

So, where does this money go, how do costs scale, and what architectural decisions separate sustainable platforms from money-losing ones?

Why Does Bandwidth Cost So Much?

Live video creates a uniquely expensive delivery problem. Unlike VOD, where popular content gets cached at CDN edge nodes and amortized across millions of views, live streams generate fresh segments every 2–6 seconds that must be pushed to every viewer simultaneously. There's very little caching benefit because the content didn't exist moments ago.

The raw numbers explain the dominance. A single 1080p viewer at 5 Mbps consumes 2.25 GB per hour. AWS CloudFront offers flat-rate plans starting at $0/month, but even the $1,000/month Premium tier caps data transfer at 50 TB, which a single popular 1080p stream burns through in under two days.

For live streaming at any real scale, you're on pay-as-you-go ($0.085/GB for US/Europe) or negotiating a private pricing agreement. At PAYG rates, 100,000 concurrent viewers watching a four-hour stream costs $76,500 for delivery alone.

The gap between cloud PAYG pricing and owned infrastructure is roughly 67x. Cloud providers charge $0.05–$0.12/GB for egress, while wholesale IP transit at competitive US and European hubs costs $0.05–$0.10 per Mbps per month (Q2 2025 data), which converts to roughly $0.0003/GB. That 300x markup is why every platform at scale eventually builds its own CDN.

An analysis of Twitch's costs estimated that running at AWS retail pricing would consume $2.4 billion per year, nearly matching their entire revenue. By operating their own CDN with direct ISP peering, Twitch brings actual bandwidth costs to an estimated $650 million annually, roughly 70% cheaper than cloud retail but still enormous.

Where Does Transcoding Fit in the Cost Picture?

Transcoding is the second-largest line item, but it's an order of magnitude smaller than delivery at scale, since it occurs once per stream regardless of audience size. A stream with 10 viewers and a stream with 10 million viewers require the same transcoding compute.

The job is to convert a streamer's single-ingest feed (typically 1080p60 H.264 via RTMP) into an adaptive bitrate (ABR) ladder of multiple renditions (1080p, 720p, 480p, 360p, 160p) so that viewers on different connections get appropriate quality. On managed cloud services, this is expensive:

• AWS MediaLive: A standard channel with 2 HD HEVC inputs, 5 AVC outputs, and advanced audio runs $3.94/hour on-demand. Reserved pricing with a 12-month commitment drops the rate to $0.56–$0.78/hour, depending on single-pipeline vs. standard configuration.
• Amazon IVS: Video input costs $2.00/hour for a standard channel, or $0.50/hour with Multitrack Video enabled (where the streamer's GPU handles encoding). Cheaper tiers exist: Advanced HD ($0.85/hour), Advanced SD ($0.50/hour), and Basic ($0.20/hour). These are input costs only. Output delivery is billed separately per viewer-hour, from $0.036 (SD, North America) to $0.25 (Full HD, South Korea).

Self-managed hardware changes the economics dramatically. NVIDIA T4 GPUs on AWS g4dn.xlarge instances handle four simultaneous ABR ladder streams at a total of $0.526/hour, bringing the per-stream cost to roughly $0.13/hour. Purpose-built ASICs push it further: NETINT's Quadra T2 cards achieve amortized costs of $0.02–$0.04 per stream per hour.

Mayflower illustrates the progression. For 10,000 simultaneous streams, their infrastructure evolved from 500 CPU servers ($10M CAPEX, 325 kW power) to 105 GPU servers ($3.36M, 112 kW) to just 38 ASIC-equipped servers ($1.44M, 34 kW). Per-stream capital cost dropped from $1,000 to $141.

Twitch's own evolution in transcoding is worth noting. The platform could initially transcode only 2–3% of channels due to cost. After building custom hardware-based transcoders using AMD Xilinx FPGAs (30x the performance of a CPU for VP9), they extended transcoding to every stream.

How Do Platforms Reduce These Costs at Scale?

The optimization roadmap follows the cost hierarchy: attack bandwidth first, then transcoding, then everything else.

1. Building a custom CDN delivers the largest single cost reduction. Netflix's Open Connect network deploys hardware directly within ISP networks at 1,000+ locations, eliminating reliance on third-party CDNs. Twitch runs bare-metal video-edge servers in colocation facilities (not on AWS) and connects them via a private backbone. Twitch's estimated bandwidth costs of $650M/year are roughly 70% lower than the equivalent AWS retail pricing.
2. Codec efficiency compounds these savings. As of December 2025, AV1 powers 30% of Netflix streaming, delivering roughly one-third less bandwidth than H.264 with 45% fewer buffering interruptions. The tradeoff is encoding compute: AV1 requires roughly 5x more processing power than H.264.
3. Client-side encoding is an emerging approach that shifts transcoding compute to the streamer's machine. Amazon IVS Multitrack Video has the streamer's GPU produce multiple quality renditions locally, cutting the cloud transcoding cost from $2.00 to $0.50 per hour, a 75% reduction.
4. Peer-to-peer offloading and multicast ABR are gaining traction for live events. Operators such as TIM (Italy) and BT Group have reported an 84% reduction in traffic compared with unicast delivery. P2P mesh networks built on WebRTC claim to offload 50–90% of CDN traffic for popular content, though reliability and consistent quality remain challenges.

What About Storage, Chat, and the Other Infrastructure Costs?

These costs are real, but they don't break the bank the way bandwidth and transcoding do.

• VOD storage: A 1080p60 stream produces 2.7 GB/hour. At 10,000 active streamers averaging four hours daily, that's \~3.24 PB/month, roughly $69k/month on S3 Standard. Twitch controls this by retaining VODs for only 14–60 days, and by using tiered storage (S3 Standard at $0.023/GB, down to Glacier Deep Archive at $0.00099/GB), which cuts archival costs by 96%.
• Origin and ingest: AWS MediaPackage charges $0.03/GB for ingest and $0.05/GB for origination, roughly $0.25–$1.35 per hour per channel. Twitch operates \~100 PoPs globally, running their proprietary Intelligest media proxy.
• Real-time chat: A stream with 100K concurrent chatters at 1,000 messages/second requires 100M message deliveries per second for that single channel. Twitch rewrote its chat system from Python to Go to handle this. Self-hosted infrastructure at that scale runs $3,000–$5,000/month; managed services like Stream Chat handle it at a comparable cost with less operational burden.
• Low-latency delivery: Standard HLS (10–30s latency) is the cost baseline. LL-HLS (2–5s) costs more because partial segments generate more HTTP requests and reduce CDN cache hit ratios. Sub-second WebRTC is significantly more expensive again, as it requires dedicated SFU (Selective Forwarding Unit) servers rather than CDN-based delivery. Amazon IVS Real-Time prices sub-300ms video at $0.036–$0.25/hour, workable for interactive features but not viable as the primary delivery method at scale.

What Does This Mean for Developers Building a Live Streaming Product?

The cost structure creates a natural progression. At small scale (hundreds of concurrent viewers), cloud-managed services are the right choice. AWS IVS, Mux, or Cloudflare Stream handle transcoding, packaging, and delivery for $2–5 per stream hour. The operational simplicity is worth the premium.

At moderate scale (thousands of concurrent viewers across many channels), the managed service bill starts to dominate your budget. This is where negotiated CDN contracts, reserved transcoding capacity, and aggressive VOD retention policies become necessary. Switching from on-demand to reserved MediaLive pricing alone saves 80%+.

At large scale (hundreds of thousands of concurrent viewers), the platform must own its delivery infrastructure to survive. The build vs. buy decision for CDN is existential. Every platform that reaches this scale (Netflix, YouTube, Twitch, Facebook) builds its own delivery network.

The architectural decisions you make early determine whether your cost curve remains sustainable at 10x or 100x your current audience:

• Codec choice matters now. AV1 halves bandwidth costs but requires 5x more encoding compute. If you start with H.264, plan your migration path.
• Design for CDN portability. Use standard HLS/DASH packaging so you can switch CDN providers or add your own edge servers without rewriting the delivery stack.
• Separate your real-time layers. Chat, presence, and live reactions have different scaling characteristics than video delivery. Using a managed service like Stream for chat and messaging lets you focus engineering effort on the video pipeline, where the cost pressure is highest.
• Instrument everything. Per-stream, per-viewer cost visibility is the only way to catch runaway spend before it compounds. Track bitrate, concurrent viewers, cache hit ratios, and transcoding utilization as first-class metrics.

The most counterintuitive takeaway: cloud egress pricing marks up wholesale transit rates by roughly 300x. That gap defines the entire economics of live streaming at scale, and every architectural decision should be made with it in mind.