Improve your 24-month delivery cost projection with realistic assumptions about your true encoding efficiency

• Video delivery costs can compound quietly inside a pipeline that never changes. Most platforms model revenue growth, but not what CDN and storage spend will look like 24 months from now.
• Uniform encoding applied to variable content is a financial problem at scale. Every wasted bit can multiply across millions of uploads and billions of views.
• ML models can produce files that are an average of 60% smaller at matched perceptual quality. The savings compound because every efficiently encoded asset stays efficient for its entire delivery lifetime.

---

Per-Stream Costs Seem Much Smaller Than They Are

A few cents per gigabyte feels manageable in isolation, until it starts to multiply across growing ingest volumes, higher resolution ladders, expanding global concurrency, and an asset library that never shrinks.

Storage and CDN costs rarely spike overnight, rather they accumulate quietly until a line item demands explanation. By the time someone notices, the inefficiency is structural and deeply embedded in every stored asset and every delivery path.

Most platforms model revenue growth carefully, but often ignore what delivery spend looks like with annual audience growth if encoding efficiencies remains unchanged.

This post will attempt to run the math nobody wants to run.

The 24-Month Projection

The following model uses a single set of assumptions. Every number in the table is derivable from these inputs, so you can verify the math and substitute your own values.

Baseline assumptions:

• Starting ingest: 1,500 hours per month of new content (approximately 5 TB/day at mixed resolutions)
• Average output bitrate: 8 Mbps across all renditions
• Annual ingest growth: 30% (compounding monthly at approximately 2.2%)
• Average views per asset: 300 per month
• Blended CDN rate: $0.02 per GB
• Content-aware encoding reduction: 60% smaller files at matched perceptual quality

How the math works: Each month’s new content, once encoded at 8 Mbps, produces 5.4 TB of output (1,500 hours × 3,600 seconds × 1 MB/s). That output is stored and then delivered repeatedly. Each month, each content cohort generates delivery volume equal to its encoded size multiplied by 300 views. With ML-powered encoding, file sizes are 60% smaller and every delivery event and storage month costs proportionally less. Ingest volume grows at 30% annually, meaning monthly output rises from 5.4 TB to approximately 7.0 TB by month 12 and 9.1 TB by month 24.

Cumulative delivery cost over 24 months (new content only)

	Fixed-Bitrate Encoding	ML-Powered Encoding	Difference
Month 6	~$700K	~$280K	$420K saved
Month 12	~$2.7M	~$1.1M	$1.6M saved
Month 18	~$6.3M	~$2.5M	$3.8M saved
Month 24	~$11.6M	~$4.6M	$7.0M saved

These figures reflect delivery costs for new content only, excluding the existing library. Storage savings are additional and follow the same 60% reduction on every new asset.

Savings more than quadruple between month 6 and month 18. This is the compounding effect of every new content cohort adding its delivery volume on top of all prior cohorts, each one growing at 30% annually. The sooner your content is encoded with this approach, the faster the savings accumulate.

Scaling Up a Larger Platform

To illustrate the same dynamics at higher volume, consider a platform ingesting 10,000 hours of content per month at an average output bitrate of 10 Mbps.

At those rates, each month’s content produces 45 TB of encoded output (10,000 hours × 3,600 seconds × 1.25 MB/s). Assuming an average of 1,000 views per asset, that single month of content generates approximately 45 PB of total delivery volume over its delivery lifetime.

With ML-Powered encoding, the same content produces approximately 18 TB of encoded output, reducing lifetime delivery volume to roughly 18 PB.

At a blended CDN rate of $0.02 per GB, the lifetime CDN cost difference for each month’s content cohort is approximately $540,000. That saving applies to every month of content ingested, which continues to grow as the platform grows.

The magnitude changes at each scale, but the underlying math is the same.

Three Scenarios Worth Comparing

The projection above models new content only. In practice, most platforms also have an existing content library that continues to generate delivery volume. How you handle that library determines your total cost trajectory.

The following scenarios use the same baseline assumptions as the 24-month projection, plus an existing library equivalent to 12 months of prior content generating 200 views per asset per month.

Scenario A: New content only. Apply content-aware encoding to all new ingest going forward. The existing library remains at current file sizes. Savings grow over time as the proportion of efficiently encoded content increases, but the library continues generating delivery costs at its original file sizes.

Scenario B: New content plus library re-encoding. Apply content-aware encoding to new ingest and re-encode the existing library. Savings are immediate and comprehensive across all delivery volumes.

Scenario C: Status quo. Encoding efficiency remains unchanged. Every new asset and every incremental viewer carries the current inefficiency forward. Costs compound at the rate of audience and catalog growth combined.

Cumulative total delivery cost by scenario (new content + existing library)

	Scenario A: New Content Only	Scenario B: New + Library	Scenario C: Status Quo
Month 6	~$1.8M	~$900K	~$2.3M
Month 12	~$4.2M	~$2.3M	~$5.9M
Month 18	~$7.2M	~$4.4M	~$11.0M
Month 24	~$10.9M	~$7.1M	~$17.8M

Note that these figures are higher than the new-content-only table because they include delivery costs for the existing library, which can be quite a significant ongoing cost driver for most platforms.

Scenario A savings accelerate over time as efficiently encoded content becomes a larger share of total delivery. Scenario B delivers the full savings rate immediately because the entire catalog benefits from day one. Both diverge sharply from Scenario C, and the gap widens every month.

For Scenario B, the one-time re-encoding cost depends on library size and the compute resources allocated to the job. For most platforms, the CDN savings recoup that investment within one to three months, depending on library size and average view frequency. The payback is fastest for libraries with high ongoing viewership, since archival content with low view rates takes longer to recoup, but still benefits from reduced storage costs. After the payback period, the savings are pure margin improvement.

Why It’s Important to Consider This Today

ML-powered encoding has been in production at the largest streaming platforms for nearly a decade, with principles that are proven and results that are well-documented.

Until recently, building a content-aware encoding pipeline required three things most platforms did not have: a dedicated ML research team to develop and maintain scene analysis models, custom GPU infrastructure to run inference at production throughput, and the engineering capacity to integrate both into an active encoding workflow. All of this was previously possible only inside organizations large enough to build and operate it end to end.

GPU inference infrastructure, especially NVIDIA’s production-grade inference stack, has allowed us to bring the per-frame cost and speed of ML analysis down to a level where it pays for itself in CDN savings, even at moderate scale. Additionally, the model architectures for perceptual analysis have matured to the point where scene-level intelligence can be deployed reliably without per-customer tuning.

For a deeper look at how Qencode’s ML pipeline and NVIDIA GPU infrastructure make this work technically, see Content-Aware Encoding at Production Scale.

What Does All of This Mean For Me?

If your encoding pipeline applies uniform settings to variable content, structural inefficiency is deeply embedded in your cost base. Production deployments across major platforms have already established that a customized encoding approach for each piece of content produces the best results.

Run it on your own content. Upload a sample at Qencode and compare file size and quality output side by side with your current pipeline.

Want us to model your specific scenario? Feel free to contact our team directly to model a cost projection based on your specific ingest volume, content profile, and growth assumptions at support@qencode.com.

Try it here -> cloud.qencode.com

---

Qencode is a member of the NVIDIA Inception Program.