Action figure: tabulating and managing streaming TCO

The challenge was, “Write a story about total cost of ownership (TCO) in 3,500 words.” Why is this a challenge? Because TCO is a book-length topic, if not a whole encyclopedia series. Cover everything, and you’ve got the proverbial “mile wide and inch deep” story, entry-level coverage, and obvious suggestions ill-suited for a knowledgeable audience. The solution? Carve it up, and focus on a few discrete elements. There are literally millions of ways to do this. Here’s the approach I took:

• AI presents lots of cost-saving options. What are the ones all producers should be leveraging?
• You want to cut costs fast. What are the best options? Hint: per-title encoding, packaging, and CDN efficiency.
• You want to improve viewer QoE. What are the most important layers here? Monitoring and multi-CDN.

So, that’s the truth-in-advertising description. Let’s get started.

Lens 1: AI cost savings all producers should leverage

AI has changed the basic math behind captions, subtitles, and dubbing. Rather than paying humans for 100% of the work, you can use AI for most of it and pay humans where needed for quality or accuracy.

When audio is ideal, AI captioning is accurate enough for some use cases. For perspective, human captioning services usually charge a few dollars per minute of video, while AI tools cost pennies per minute. In practice, this means savings of 60%–80% compared with manual captioning, in some cases even higher.

However, with premium content, audio is rarely ideal, and AI captions are rarely sufficiently accurate for standalone deployment. Music, sound effects, overlapping dialogue, and off-screen voices make automatic captioning less reliable. Accuracy drops, and timing problems increase. For flagship shows and other high-stakes content, accuracy below 99% or so is unacceptable.

A hybrid workflow offers a middle path on cost and quality. AI generates a first pass, while human editors focus on the difficult scenes and correct errors. Translated, a provider of AI-driven language solutions, calls this “human-AI symbiosis.” This approach is still much cheaper than fully manual captioning, while increasing accuracy to the required level.

Localisation and dubbing

Localisation and dubbing hit TCO even harder. Industry guides estimate that traditional studio dubbing for a 1-hour video in a single language costs between $5,000 and $15,000 (or the local equivalent). AI dubbing providers report costs as low as a few hundred dollars, or 90% cheaper.

Not surprisingly, most service providers offer human-AI hybrid approaches. Notably, some services, such as Deepdub (see Figure 1), “combine AI technology with the expertise of native-speaking human adapters and an in-house production team of industry professionals to ensure authentic, high-quality dubbed content.”

Figure 1. Deepdub’s AI-assisted dubbing services ensure accuracy and quality while delivering significant savings.

AI metadata and enrichment

Particularly for AVOD and FAST services, scene-level metadata is often the missing piece for higher CPMs. The manual effort required to create and maintain it means most publishers barely scratch the surface. AI metadata and enrichment change the economics of discovery and monetisation.

Video intelligence platform vendors like TwelveLabs and MixPeek analyse streams to identify who appears, what happens, and which themes or moods are present in each segment. Then they generate rich, standardised metadata at the programme, segment, and scene levels. Content data platforms like IRIS.TV structure that metadata around persistent content identifiers and distribute it into ad tech and measurement systems. Together, these tools make granular metadata economical to apply across an entire catalogue rather than only a handful of flagship shows, thereby supporting discovery, programming, and contextual ad targeting. This allows advertisers to connect with the right prospects in the right content at the right moment, improving campaign performance and return on ad spend. When campaigns work better, buyers will pay more for inventory that carries strong contextual signals, which is how scene-level metadata translates into higher CPMs for the publisher.

Lens 2: cut external costs (encoding, packaging, and CDN)

Encoding, storage, and CDN delivery are often the largest controllable OpEx bucket in a video P&L, so this lens focuses on levers that change how many bytes you ship and what you pay per gigabyte.

Cut the bits delivered with per-title encoding

Traditional ABR workflows start with a fixed-bitrate ladder for all content. Per-title encoding, also called content-adaptive encoding (CAE), an­alyses each asset and designs a custom ladder for that specific title. For easy-to-encode content, per-title encoders will reduce bitrates and sometimes drop rungs entirely. For hard-to-encode content, they may preserve or even raise bitrates at the top end to achieve the necessary quality.

Note that per-title is codec-agnostic, although some techniques, like capped constant rate factor (CRF) encoding, aren’t available for all codecs. Per-title works for live and VOD, although VOD techniques are typically more advanced.

How much per-title saves

The savings per-title delivers depends on multiple factors, including your existing ladder structure, the encoding complexity of your content, and which ladder rungs your customers typically consume. Let’s take these in order.

If your current ladder is conservative—say, 6Mbps–8Mbps for the 1080p top rung—savings could be significant. However, if your current ladder is aggressive—say, 4Mbps–5Mbps for the top 1080p rung—savings will be significantly less. The same theory goes for 4K, just with higher numbers.

If you’re delivering easy-to-encode content, such as animations, talk shows, drama series, game shows, and the like, per-title should be fertile ground. This approach is less applicable if most of your programming is hard-to-encode sports content like basketball or football.

Finally, per-title encoding saves the most bandwidth in the top rungs, perhaps knocking your top rung from 6Mbps to 4Mbps, a 33% savings. But if you’re mostly distributing middle rungs—say, 2Mbps–3Mbps to mobile viewers—bandwidth savings will be minimal. You’ll still be delivering a 2Mbps–3Mbps rung, although it likely will be higher quality.

Putting these together for 1080p video, if RTL Super were starting with an 8Mbps top-rung bitrate delivered mostly to living room audiences, bandwidth savings could be in the 35%–50% range. If DAZN’s current football ladder peaked at 4Mbps for H.264, per-title likely won’t deliver any benefit at all, particularly if mobile viewers comprise a significant swatch of the audience.

Choosing a per-title technology

You have three basic options for deploying per-title encoding. The first is DIY, using capped CRF or similar techniques. The second is to use features built into your encoding or cloud provider. For example, Bitmovin (see Figure 2), Tencent, and other services offer per-title configuration as a checkbox option in VOD and sometimes live workflows. Finally, many third-party encoders, like those from Capella Systems and Harmonic (which was recently acquired by Media­Kind), also offer per-title functionality.

Figure 2. Bitmovin’s per-title encoding promises improved QoE and reduced costs.

When choosing a per-title technology, keep a few features in mind. First, the most primitive technologies use a fixed number of rungs at specific resolutions and adjust only the bit­rate. This is how capped CRF works and how most live per-title approaches work. More advanced technologies adjust the number of rungs, their resolutions, and the bitrates and deliver the most savings. The most advanced technologies also use variable GOP sizes and insert I-frames at scene changes, working in concert with the packaging application to maintain the integrity of the adaptive bit­rate playback.

From a TCO perspective, you may spend a little more on analysis and smart encoding, then recoup that and more through lower CDN egress and lower origin storage, especially at scale. If your typical asset is watched more than a few thousand times, it’s probably worth the expense. If not, you may not recoup the encoding costs through bandwidth savings.

Package in CMAF for CDN efficiency

Most streaming producers support both HLS and DASH for maximum compatibility, but this is inefficient from a storage and caching perspective. The Common Media Application Format, or CMAF, allows both HLS and DASH to use a single set of fragmented MP4 segments. This significantly reduces storage overhead because you don’t need duplicate assets across different protocols.

It also improves caching. HLS and DASH clients request the same segment URLs, allowing caches to share footprints across devices rather than splitting them. This increases cache hit ratios and reduces bytes pulled from the origin.

While the CMAF-related math sounds simple, there are several important caveats. Some legacy devices still need TS HLS fallbacks. Migrating to CMAF can require encoder and packager upgrades. Achieving very low latency still needs careful tuning of chunk sizes, player buffers, and CDN cache behaviour. All of this aside, it’s long past the time to start seriously considering CMAF for dual format workflows.

Cutting CDN transport costs

You’ve already cut encoded bits per title via content-aware ladders and storage via CMAF. The next step is to reduce transferred bits by having the CDN move fewer bytes more intelligently. To start, understand that cost isn’t just your headline price per gigabyte; it’s that rate multiplied by the volume of data you actually ship. A cheap CDN with poor cache efficiency can easily end up costing more overall if it forces traffic to take longer paths or constantly fetch files from your origin.

Your first sanity check should be geographic. Does the CDN have points of presence (PoPs; see Figure 3) and direct peering in the locations where your viewers actually live? If not, you’ll see a higher time to first byte (TTFB), more origin traffic per session, and a worse user experience. To find these routing inefficiencies, monitor your per-region latency alongside the split between edge and origin bytes.

Figure 3. CDN77 has POPs around the world, but is particularly strong in Europe and Africa.

True efficiency means keeping traffic on the edge and cutting origin and backbone volume. For a mature VOD service, a healthy cache hit ratio (CHR) for media segments can reach the high 90s. For live streaming, segment ratios in the low-to-mid 90s are common, while live manifests naturally tend to have lower hit rates because they update every few seconds. If your segment CHR drops below these benchmarks, you’re paying to move bits from your origin that should already be cached. Track your CHR by object type (separating manifests from segments) and by region to pinpoint leaks.

Next, audit your time-to-live (TTL) settings and cache keys. Live manifests need lifetimes measured in seconds so players stay synchron­ised. Conversely, video segments and artwork need long lifetimes, measured in hours or days, so the CDN can reuse these heavy objects.

You also need to ensure the CDN handles the transactional pattern of modern video. Adaptive streaming requires constant small retrievals of manifests, short chunks, and partial segments. Video-oriented CDNs optimise their HTTP versions, connection reuse, and small-object caching for this exact behaviour, whereas generic static-file CDNs often struggle, showing latency spikes as viewer concurrency rises. Watch your latency underload and segment error rates to see if your provider is keeping up.

If the edge remains inefficient, consider an origin shield or a tiered cache. This adds a centralised caching layer between your edge PoPs and your origin to collapse duplicate requests. While some providers charge small fees for shield-level requests, that internal CDN traffic is significantly cheaper than cloud data transfer out (DTO) fees from your storage or compute buckets. Before and after turning it on, compare your origin request counts and egress gigabytes to verify the drop.

An origin shield also protects against thundering-herd events. When a popular file expires, request coalescing ensures that the CDN turns hundreds of parallel edge misses into a single upstream fetch. Without this, cache expirations cause sudden origin stampedes and traffic spikes. Monitor your origin hits during asset expiration windows; if dozens of identical requests hit your servers simultaneously, coalescing isn’t working.

Only after you’ve squeezed maximum efficiency out of a single CDN should you layer on a multi-CDN strategy. Splitting traffic across multiple networks initially lowers cache ef­fi­­ciency because individual caches warm up more slowly. Optimising one CDN first ensures you aren’t masking fundamental configuration flaws with complex multi-CDN steering logic.

Lens 3: improve QoE

The next lens I chose to explore is improving viewer QoE. While a great QoE doesn’t guarantee a successful service, poor QoE can absolutely doom a service to underperformance or failure.

The first step toward improving QoE is to measure it. The Streaming Video Technology Alliance QoE working group has already published a common set of key network delivery metrics that give the industry a shared vocabulary for how delivery affects viewer experience. Most commercial analytics products and platform dashboards report at least this core group of metrics, such as startup time, buffering, bitrate, and playback failures, even if they use slightly different names or formulas.

• Video start time: Time from the initiation of the play event until the first frame of network-delivered video is rendered
• Video start failure: A session in which the first chunk of video is not fully delivered within a defined cutoff, and the viewer never sees a first frame
• Total rebuffering time: The total duration after the first frame is rendered when the video is not being smoothly played
• Rebuffering ratio: Total rebuffering time divided by total playing time plus total rebuffering time, often expressed as a percentage
• Average media bitrate: The time-weighted average bitrate actually rendered during playback, which reflects the quality the viewer sees
• Bitrate stability: How frequently and how far the bitrate changes during playback over the session
• Playback failure rate: The percentage of sessions that end in a fatal playback error, including failures from delivery, DRM, ads, or player bugs

If you are not collecting these signals from real players, you are flying blind. Once you start, it helps to have targets to interpret the data. NPAW publishes its annual Video Streaming Industry Report, but the data is directional, indicating changes in values rather than absolute figures. One useful source for both benchmarks and troubleshooting information is a Qualabs blog post titled “Video Quality of Experience (QoE) on OTT: Metrics, Benchmarks, and How to Improve Them." The data is shown in Figure 4.

Figure 4. QoE datapoints from Qualabs

Once you have data and targets, prioritise problems based on how likely they are to cause viewers to leave. First, address startup time and video start failures by using analytics to identify the worst-performing regions, ISPs, and devices.

Begin with your ABR ladder, making sure the bitrates of your lowest rungs are retrievable by viewers on slow connection speeds. Check your manifests to ensure that a mid- to low-quality rung is retrieved first to ensure responsiveness. Check the player to see how many segments it buffers before starting playback. If it’s too long, reduce the buffer.

Next, look at rebuffering and bitrate stability. If your player enables it, consider extending the interval between rung shifting to maintain a steady quality (also called hysteresis). Check your cache hit ratio to ensure that local or regional caching is supplying the vast majority of your viewers.

Finally, address device-specific issues. QoE analytics almost always surface outlier platforms. A particular smart TV family may have high error rates. A console app might have very slow starts. These problems often require tailored solutions, such as a different startup bit­rate for that device, a slightly different ladder, or a custom player configuration that matches its CPU and memory constraints. Without per-device QoE, these issues hide in the aggregate. With per-device QoE, you can treat them as focused engineering tasks.

The reality of Multi-CDN: strategy by scale

Multi-CDN is frequently marketed as a universal insurance policy for streaming uptime and performance. The reality is far more nuanced. Implementing a multi-network architecture introduces immediate operational penalties in cache efficiency, engineering overhead, and origin egress costs. As such, the multi-CDN decision is often a function of publisher sizs.

Small publishers: the case for a single premium CDN

In most cases, small publishers should avoid multi-CDN. Splitting a modest traffic footprint across multiple networks introduces structural inefficiencies that actively harm streaming performance. The major exception is when multiple CDNs are needed to effectively serve different customers in disparate regions. Here’s why. CDNs rely on volume to keep content hot at the edge. When traffic is low and split across multiple vendors, the lower per-CDN volume means fewer objects remain cached near users. This triggers several issues, including reduced cache efficiency, commitment floor penalties, and increased origin egress costs.

Regarding multi-CDN for smaller publishers, Katerina Dobnerova, CDN77’s head of technical integration, comments, “For a smaller publisher, it may not make sense because you lose the volume-based discount and end up with a higher per-gig price simply from spreading traffic across more CDNs.” The best option for smaller publishers is to choose a CDN with multiple PoPs in the regions where their viewers actually live, since that keeps content cached close to the audience and improves both startup time and overall QoE. Here are some potential CDNs by region:

• Europe: CDN77, bunny.net, KeyCDN
• Japan: Cloudflare, KeyCDN
• China: Alibaba Cloud CDN (Aliyun), Tencent Cloud CDN, Huawei Cloud CDN, and BaishanCloud
• Middle East and Africa: EdgeNext, CDNetworks

Medium-to-large publishers: managed orchestration

Once you satisfy the minimum volume commitments of multiple networks, a multi-CDN strategy becomes viable. The core technical decision centres on traffic steering. While DNS-based steering works well for coarse regional failover, it lacks the granularity required for real-time QoE-sensitive steering inside an active playback session. For dynamic midstream adjustments, publishers at this layer should leverage commercial orchestration platforms such as NPAW’s CDN Balancer or Conviva’s Precision.

Multi-CDN at this tier can be used for QoE improvement and/or cost optimisation (see Figure 5). You can configure the orchestrator to route traffic to the lowest-cost CDN by default, shifting to a premium provider only when performance drops below a specified QoE threshold.

Figure 5. NPAW’s MidStream CDN Balance directs viewers to the CDN that delivers the best experience.

One underappreciated benefit of a managed orchestrator is observability. Specifical­ly, these services centralise per-CDN logs and QoE signals so your ops team can troubleshoot incidents in one place instead of stitching together dashboards and log feeds from multiple providers.

Companies at this scale should consider deploying a multi-CDN architecture that pairs a premium global network with a cost-optimised or regional provider. Delegate the routing logic to a commercial third-party management platform to avoid heavy engineering costs.

Very Large Enterprise Publishers:
Proprietary DIY Systems

Companies operating at this scale are not reading articles like this for technical direction. For completeness purposes only, the base­line advice here is that tier-one global services almost always build proprietary, server-side manifest manipulation engines to avoid the cost-prohibitive volume fees of commercial orchestrators.

At this extreme volume, the engineering focus shifts entirely from basic routing to complex infrastructure normalisation. This requires
deploying a dedicated origin-shielding layer to protect the storage origin from duplicate cache
misses across networks and writing custom integration code to ensure that token authentication, DRM, geo-fencing, and dynamic ad insertion behave identically across all edge providers.

Balancing the Video P&L

While total cost of ownership spans an encyclopedic number of operational touchpoints, managing a video P&L ultimately comes down to making targeted, high-impact architectural choices. This article focused on a few of the most critical approaches: leveraging human-AI hybrid workflows to cut localisation costs, utilising content-adaptive encoding and CMAF to compress the distribution footprint, and establishing a multi-CDN strategy that scales strictly alongside monthly traffic volume. By executing on these specific areas, publishers can move past generic cost cutting and build a highly optimised, predictable delivery pipeline that protects the bottom line.

Free

for qualified subscribers

Subscribe Now Current Issue Past Issues

The State of CMAF: The Ultimate Format or Just Another Acronym?

It's all the buzz on the conference circuit, but deployment challenges mean CMAF hasn't taken over the streaming world yet. Here's why it's getting so much attention.

02 Jan 2020