AV1 encoding economics — when AV1 actually saves money vs HEVC

AV1's compression advantage over HEVC is real (~30% bandwidth reduction at equivalent quality) but it's not free. The encoding cost is meaningfully higher, and the hardware decoder install base is smaller. Whether AV1 pays off depends on the math: bandwidth saved × audience reachable vs additional encode compute. This page is the engineering reference for the AV1 economics question — when does AV1 actually save money, and when does it not?

#The bandwidth equation

AV1 vs HEVC at equivalent VMAF (1080p mixed content):

• AV1 (SVT-AV1 preset 6) — typically 25-35% lower bitrate.
• AV1 (libaom-av1 cpu-used 4) — typically 30-40% lower bitrate.
• AV1 (NVENC AV1) — typically 15-25% lower bitrate (hardware quality lags software).

For a streaming service shipping 1 PB of HEVC content per year, switching to AV1 would save ~250-300 TB of bandwidth at equivalent quality. At typical CDN egress costs ($0.005-0.05 per GB depending on tier), that's $1,250-15,000 per year per PB.

Scale this up:

• 10 PB/year HEVC streaming → switching to AV1 saves $12,500-150,000/year.
• 100 PB/year → $125,000-1,500,000/year.
• 1 EB/year → $1.25M-15M/year.

The bandwidth savings are real. But they're proportional to streaming volume, not encoding volume. A service with high streaming volume per encoded asset (popular content) gets more value than a service with low streaming per asset (long tail).

#The encode compute equation

AV1 encoding is significantly more compute-expensive than HEVC:

• SVT-AV1 preset 6 vs x265 medium — roughly 1.5-2x slower per frame.
• SVT-AV1 preset 4 vs x265 slow — roughly 1.5-2x slower.
• libaom-av1 cpu-used 4 vs x265 medium — roughly 5-10x slower.
• NVENC AV1 vs NVENC HEVC — similar speed; quality similar.

For software encoding of 1 PB of source content per year:

• HEVC at x265 medium: maybe 100,000 CPU-hours.
• AV1 at SVT-AV1 preset 6: maybe 200,000 CPU-hours.
• AV1 at libaom cpu-used 4: maybe 800,000 CPU-hours.

At cloud compute prices (~$0.05-0.20/CPU-hour depending on instance and discount), the additional AV1 compute cost is:

• SVT-AV1: $5,000-20,000 per PB encoded vs HEVC.
• libaom-av1: $40,000-160,000 per PB encoded vs HEVC.

For pipelines that encode each asset once and stream many times, the compute overhead is amortized. For pipelines with high content velocity, compute overhead matters more.

#The break-even calculation

The AV1 ROI math:

Annual savings = (HEVC bitrate × streaming volume × CDN cost) -
                 (AV1 bitrate × streaming volume × CDN cost) -
                 (additional encode compute cost × content volume)

Simplifying:

Net savings = (bandwidth savings × streaming volume × CDN cost) -
              (additional compute × content volume)

For AV1 to break even:

streaming volume × CDN cost × bandwidth savings ≥
content volume × additional compute cost

Or:

(streaming volume / content volume) ≥
(additional compute cost) / (CDN cost × bandwidth savings)

The left side is "average streams per encoded asset" — how many times each piece of content is streamed.

The right side is the break-even threshold. For typical numbers:

• Additional compute: $10/asset (SVT-AV1 over HEVC for a 60-minute asset)
• CDN cost: $0.01/GB
• Bandwidth savings: 1 GB per stream (rough)
• Break-even streams/asset: $10 / ($0.01 × 1) = 1,000 streams per asset

So for AV1 to pay off, each asset needs to be streamed ~1,000+ times. For premium VOD with popular content, this is easy. For low-traffic content (educational, niche), it's harder.

#When AV1 wins

The cases where AV1 pays off clearly:

Premium VOD with high streaming/asset ratio — Netflix, Apple TV+, Disney+ premium tiers. Popular content streamed millions of times per asset. AV1 saves bandwidth at scale.

UGC platforms at scale — YouTube, TikTok-class. Massive aggregate streaming volume. Even modest per-asset streaming counts add up across billions of assets.

Mobile-heavy distribution — bandwidth on mobile networks costs more (carrier-paid networks have higher egress). AV1's bandwidth savings translate to more value per byte saved.

Bandwidth-constrained markets — services serving regions with expensive bandwidth. The savings ratio is higher when the cost-per-byte is higher.

#When AV1 doesn't pay off

The cases where AV1 doesn't make economic sense:

Low-traffic content — corporate training, archive content streamed rarely. Encode cost dominates; bandwidth savings are minimal because aggregate volume is low.

High content velocity, low per-asset streaming — local news platforms, daily-shows networks. Each asset is streamed modestly; encoding cost compounds across high asset volume.

HEVC-only ecosystems — services targeting only Apple Safari (no AV1) and older devices. AV1 isn't reachable for the audience; encoding it is wasted.

Constrained encoding budgets — startup/small services where compute cost dominates economics. Cheaper HEVC encoding wins until streaming scale justifies the upgrade.

#Hardware AV1 — changes the equation

Hardware AV1 encoders (NVENC AV1 on Ada GPUs, NETINT Quadra, AMD MA35D, Intel Arc) significantly change the encode-cost economics:

• Hardware AV1 throughput — ~50-100 1080p streams concurrent per chip (NETINT Quadra T1A, AMD MA35D).
• Per-stream cost — for live, hardware AV1 cost is typically $0.05-0.20 per stream-hour vs CPU AV1 at $5-15 per stream-hour.
• Quality — hardware AV1 quality is roughly SVT-AV1 preset 8-10 equivalent. Lower than software preset 4-6 but acceptable for live and throughput VOD.

For live AV1 specifically, hardware AV1 is the only economically viable option at scale. CPU live AV1 at quality presets requires more compute than is practical.

For throughput-bound VOD (UGC, mid-tier streaming), hardware AV1 makes the encode-cost equation favorable even for moderate streaming volumes per asset.

#Mixed AV1/HEVC ladder economics

Most premium streaming services in 2026 ship mixed ladders:

• AV1 top tiers (4K, 1080p) — for AV1-capable audiences.
• HEVC mid/floor tiers (1080p, 720p) — for non-AV1 audiences (Apple Safari, older Android).
• H.264 floor (540p, 360p) — universal.

Economics of mixed ladders:

• Encode cost: encode each tier with appropriate codec. Top AV1 tier costs ~2x equivalent HEVC; mid HEVC tier costs same as before; floor H.264 tier same as before. Net encode cost ~25-40% higher than HEVC-only.
• Streaming cost: AV1-capable audience streams AV1 (saves bandwidth); other audience streams HEVC (no change). Average bandwidth savings depends on AV1-capable audience share.

For typical 2026 audiences (~50-60% AV1-capable for streaming services with broad reach):

• Net bandwidth savings: ~12-18% (about half the AV1-only savings, weighted by audience share).
• Net encode cost increase: ~25-40%.
• Break-even streams per asset: ~2,000-4,000 (higher than AV1-only because savings are diluted).

For content with adequate streaming volume, mixed ladders pay off. For content with low volume, the math gets tighter.

#Reference numbers from major streaming services

Approximate publicly-shared numbers from major services:

• Netflix — has reported 30%+ bandwidth savings on AV1 streams; AV1 represents a growing share of streamed content. Compute cost amortized across high streaming volume.
• YouTube — uses AV1 extensively for popular content; per-asset streaming volume is massive.
• Meta (Facebook video) — switched substantial AV1 production for similar reasons.

For services with similar scale, AV1 ROI is positive. For services 1-2 orders of magnitude smaller, the math is tighter.

#The 2026 production answer

For most premium streaming services in 2026:

1. Encode AV1 for top tiers of the ladder. The bandwidth savings on the high-bitrate tiers compound; the encode cost is amortized.
2. Use SVT-AV1 preset 6 for production VOD. Hardware AV1 (NVENC, NETINT) for live and throughput VOD.
3. Calibrate against actual content — your specific content's AV1-vs-HEVC compression ratio may differ from generic benchmarks.
4. Verify AV1-capable audience before committing — if your audience analytics show <20% AV1 capability, the mixed-ladder math weakens significantly.
5. Don't AV1-encode low-traffic content — for archival or rarely-streamed content, HEVC remains economically optimal.

#Operational considerations

Things that affect AV1 economics in production:

• Encoder version evolution — SVT-AV1 keeps improving. Re-encode your benchmark every 6 months.
• Hardware AV1 availability — Ada GPUs, NETINT, MA35D are recent; capacity availability matters for large-scale operations.
• Content reuse patterns — content with high re-watch value (movies, popular series) gets more AV1 ROI than transient content (news clips).
• CDN tier negotiations — AV1 bandwidth savings × CDN cost is tier-sensitive; favorable CDN deals make AV1 less attractive (smaller cost-per-byte to save against).
• Storage cost — AV1 reduces storage costs proportionally to bandwidth (~25-30% smaller files).

#What MpegFlow does with AV1 economics

MpegFlow's FfmpegExecutor worker image runs SVT-AV1 (software) on the CPU pool; hardware AV1 (NVENC AV1 on Ada-class GPUs, where the GPU pool has them) is configurable per rendition. The DAG runtime expresses each AV1 rendition as a parallel stage on job_stages with explicit dependency tracking; per-stage retry handles transient failures; the KEDA-driven autoscaler sizes the GPU and CPU pools independently to the queued workload.

For customers evaluating AV1 adoption, the typical analysis we run during onboarding:

1. Measure their actual AV1-vs-HEVC compression ratio on representative content. Generic benchmarks rarely match specific content.
2. Project audience AV1 capability based on customer's audience analytics.
3. Calculate break-even streams per asset based on actual CDN costs and content velocity.
4. Recommend mixed-ladder strategy based on the analysis.

For pipelines committed to AV1, the partitioner places AV1 stages on AV1-capable workers and other codec stages on standard pools — heterogeneous compute is the runtime's job, not the workflow author's; sibling cancellation propagates fatal failures across rendition stages so dependents don't waste compute.

The strict-broker security model handles AV1 work like any pipeline payload — workers carry no ambient credentials; content access flows through short-lived presigned URLs scoped per stage; access is disposed on completion. AV1's compute requirements affect instance sizing, not security posture.

The general guidance: AV1 is the codec for the future, not necessarily for today's specific workload. The economics depend on your content, audience, and scale. For premium streaming with broad audiences and meaningful streaming volume per asset, AV1 pays off in 2026. For smaller-scale or niche content, HEVC remains economically optimal until streaming volumes grow.