Streaming growth used to be the whole strategy. Now it is the thing eating your margin. Omdia counted 2.24 billion online video subscriptions at the end of 2025, up 17.6% in a year, and expects growth to drop to 5.6% in 2026. Fewer new subscribers, the same delivery bill. This guide covers what actually drives OTT delivery costs, eight ways to reduce them, and the point where cost cutting starts costing you viewers.
What actually drives OTT delivery costs
Your monthly bill is the sum of five things. Each behaves differently under load, which is why a single across-the-board cut rarely works.
CDN usage and bandwidth
Video delivery is not one long transaction. It is millions of small requests for segments, manifests, thumbnails and DRM licences, every one of them billed. According to AppLogic Networks’ 2025 Global Internet Phenomena Report, the average subscriber pulls down 5.6GB of video per day.
The uncomfortable part is pricing. Dan Rayburn’s 2025 CDN market analysis puts the third-party delivery market at roughly $5 billion and flat to slightly negative year on year, with the explicit conclusion that per-GB pricing will not fall further and may tick up. Akamai added a 3% monthly surcharge in Q2 2026. Regional spread matters too: South American pricing averages 2.8 times North American rates, and Korea runs around ten times higher. If your audience shifts geographically, your unit cost moves before your traffic does.
Encoding and transcoding
Every title becomes a ladder of renditions, and every rendition costs compute once and storage forever. The waste usually sits at the top of the ladder. Bitmovin’s per-title encoding analysis shows a 4K HEVC asset holding a VMAF score of 94.9 at 1.9 Mbps where a fixed ladder was pushing 15 Mbps, and a 4K AVC title dropping 22.7% in bitrate while quality rose from VMAF 90.9 to 95.5.
Storage and origin
Catalogue storage is the cost nobody reviews. Across the tenants running on Better Media Suite, the recurring pattern is a long tail: a small share of the library drives most of the watch time, while the rest sits in every rendition, every language track and every packaging format, paid for month after month. Origin traffic compounds it. Each cache miss is a request your CDN could not absorb, and cache misses cluster exactly when you can least afford them.
Peak traffic and live events
Live concentrates a month of demand into ninety minutes. AppLogic found that all ten of the heaviest internet traffic days of 2024 coincided with live streaming events, with peaks running 30-40% above normal network usage. NFL games took nine of those ten slots.
Multi-device delivery
A catalogue that plays on mobile, web and ten or more connected TV platforms needs more renditions, more packaging formats and more DRM variants than a mobile-only service. That is a real cost of reach, and it is the one line item you should be slowest to cut.
Eight ways to reduce OTT delivery costs
1. Rebuild the bitrate ladder around content, not convention
Fixed ladders assume every asset is equally hard to encode. A studio drama and a static talking-head interview get the same top rung, and one of them is being massively overserved. Per-title analysis sets the ladder from measured complexity instead.
Rayburn’s data adds a second reason to look upward: 4K adoption has stalled and the industry has largely settled on 1080p HDR. If your top rendition exists because it looked good on a roadmap slide, check how many sessions actually reach it before you keep paying to deliver it.
2. Move to AV1 and HEVC wherever devices allow
Codec choice is the single largest lever on bytes per stream. Netflix reported in December 2025 that AV1 now carries roughly 30% of its streaming, that AV1 sessions use one third less bandwidth than both AVC and HEVC, and that those sessions see 45% fewer buffering interruptions while scoring 4.3 VMAF points higher than AVC. Film grain synthesis pushed one test asset 66% lower in bitrate at better perceived quality.
The trade-off is honest and worth stating: AV1 encoding is more expensive per asset, and older devices in your install base will still need an AVC fallback. You are moving delivery cost into encoding cost, which pays off on titles with high watch time and does not on the long tail.
3. Treat multi-CDN as routing, not procurement
Adding vendors is not the same as reducing cost. Rayburn’s survey shows large OTT platforms moving in the opposite direction, cutting from four to six CDN vendors down to two or three, because commit volumes fragment and each contract loses negotiating weight.
What actually pays is the routing logic on top: switching by region, by time of day and by measured quality of experience rather than by contract. Two well-instrumented providers with real-time steering will usually beat five providers selected once a year.
4. Fix caching before you renegotiate rates
Offload ratio is the number to look at first. Cache key hygiene, sensible segment durations and origin shielding all reduce the share of traffic that reaches your origin, and they cost engineering time rather than contract renegotiation. In practice, tenants that clean up query-string variations in their cache keys often find a measurable slice of “CDN traffic” was really the same segment being fetched under three different URLs.
5. Forecast concurrency instead of reacting to it
Live events are predictable in timing even when audience size is not. Concurrency modelling from previous comparable events, stress testing against projected peaks and pre-warmed capacity turn an emergency into a plan. The alternative is paying overage rates decided by whoever is on call at 8pm. Our live event production best practices walk through the operational side of this.
6. Stop paying for streams nobody finishes
Every abandoned session is bandwidth you bought and engagement you did not get. Failed starts, rebuffer-driven drop-offs and aggressive bitrate oscillation all consume delivery budget with no revenue attached. Netflix’s AV1 numbers show the connection working in reverse: lower bitrates produced 45% fewer buffering interruptions, so the efficiency gain and the retention gain arrived together.
7. Make cost visible per viewing hour
Aggregate CDN spend hides the decisions that matter. Cost per viewing hour, broken down by region, device class and content type, tells you whether your CTV apps cost three times what your mobile apps do and whether one territory is quietly subsidising another. You cannot optimise a number that arrives once a month as a single figure.
8. Match infrastructure shape to traffic shape
Reserved capacity sized for your peak means paying for idle compute the rest of the month, which is a poor fit for services built around scheduled live events. Event-driven and autoscaling approaches align spend with actual demand instead. The trade-off runs the other way for steady, high-volume VOD workloads, where per-request pricing can end up above reserved instances. We covered where the line sits in serverless architectures for OTT.
Why live sports break the cost model
On-demand traffic spreads across the day. A cup final does not. Tens or hundreds of thousands of viewers press play inside the same two-minute window, and every layer of the stack feels it at once: CDN edge capacity, origin requests, manifest generation, entitlement checks and the licence server.
Three things make live economics different from VOD economics.
Concurrency, not volume, sets your ceiling. A service with comfortable monthly averages can still fail at minute one of a match. Capacity has to be provisioned against peak simultaneous sessions, and that peak is what you pay for.
Global audiences carry regional pricing. A single fixture can pull viewers across four continents, and with South American delivery averaging 2.8 times North American rates, the same match can cost very different amounts depending on who tunes in.
Failure is expensive twice. An outage during a live event costs delivery spend and refunds and reputation in the same evening. For rights holders it can also affect the next negotiation, which is why sports platforms looking beyond subscription revenue tend to treat delivery reliability as a commercial issue rather than a technical one.
Where AI helps, and where it does not
Predictive scaling is the clearest win. Historical traffic, fixture calendars and marketing schedules give a model enough signal to provision ahead of a spike rather than during it, which removes the two most expensive habits in streaming operations: standing overprovisioning and emergency burst capacity.
Cost-aware routing is the second. Steering traffic between providers on live quality and price data works at a speed no operations team can match manually.
The honest limit: both need volume to be worth it. Below a few hundred terabytes a month, the orchestration and instrumentation overhead can exceed the savings, and a well-tuned single-CDN setup with a sensible ladder will get you most of the way. AI does not fix a bad bitrate ladder, and it will happily optimise the routing of traffic you should not have been sending in the first place.
One structural change worth putting in the 2027 budget: under the EU Data Act, providers of data processing services must remove switching charges, including data egress fees, from 12 January 2027. For European platforms, that changes the cost of moving origin or storage between clouds, and it is worth modelling before you sign a multi-year commit.
The cost cuts that lose you money
Some savings are real and some are borrowed from your retention numbers. Dropping the top rendition, stretching segment lengths past the point where adaptive switching stays smooth, or thinning edge coverage in a growth market will all reduce this month’s invoice. They also increase startup time, rebuffering and churn, and subscriber acquisition costs far more than a gigabyte.
A workable rule: set a quality floor first, expressed in measurable terms such as video start time, rebuffer ratio and the share of sessions that stay above your target VMAF band. Optimise aggressively underneath it, and treat any change that breaks the floor as a revenue decision rather than an infrastructure one.
Where to start
Most platforms find the first meaningful savings in three places: a ladder that was designed once and never revisited, an offload ratio nobody has measured, and a peak-provisioned infrastructure serving spiky demand. None of those need a new vendor. They need visibility per stream, per region and per device, which is the part most teams are missing.
Better Media Suite runs delivery, encoding and analytics as one system across more than 30 tenants and over ten connected TV platforms, so cost per viewing hour is a number you can see and act on rather than reconstruct from invoices. If you want to look at your own delivery economics with us, talk to the team.
FAQ
Review the vocabulary monthly and the structure quarterly. Monthly reviews should look at zero-result searches, new tag values created and assets missing required metadata, which take under an hour. Structural changes, such as adding a level or a new asset type, need a quarterly cycle because they require re-tagging existing content and updating ingest validation.
Taxonomy is the classification structure that defines categories and how they nest, for example sport to competition to season. Metadata is the descriptive information attached to each individual asset, such as title, language, duration, rights window and participants. Taxonomy tells you where an asset sits in the library; metadata tells you what it contains.
Partly. Speech-to-text, scene detection and object recognition reliably produce broad categorical tags and searchable transcripts at a fraction of manual cost. Accuracy drops on domain-specific detail, including individual player identification and event context. The practical setup is automated tagging with confidence scores, where anything below your threshold is routed to a human reviewer before publication.
Allowing free-text entry on fields that power filters and recommendations. Within a year you have multiple spellings of the same team, competition or season, each treated as a separate value by the database. Filters return incomplete results and recommendation engines lose the connections between related assets. Controlled vocabularies enforced at ingest prevent it.
Yes, directly. Contextual targeting on connected TV depends on structured content signals such as genre, competition, participants and asset type. Inventory attached to a well-described asset can be sold against a specific context and commands a higher rate than inventory with only a filename and duration. Premium video advertising reached $42.1bn in 2025 according to Omdia.