Amazon Web Services (AWS) multiplies each high-definition output minute by 22.97 in its own billing example before applying the per-minute rate [1]. Your render ladder sets the final bill, not the length of your raw footage. A standard six-rung delivery ladder puts one hour of source video at $4.95 on AWS Elemental MediaConvert, or $9.90 on Google Cloud [1, 2]. Meanwhile, a current-generation graphics card encodes that same 1080p output at 506 frames a second for about half a cent of electricity per hour [3, 4, 5].

The big picture:

The headline price is the least important number on either rate card.

Elemental MediaConvert quotes $0.0075 per “normalized” minute on its basic tier. That price covers the first 100,000 minutes a month, then drops to $0.0053 and $0.0038 [1]. The professional tier starts at $0.0120 [1]. Google’s transcoder quotes $0.015, $0.03, and $0.06 per output minute for standard, high, and ultra-high definition output [2].

A normalized minute is not a flat minute of video. It is a minute of video multiplied by the resolution, the frame rate, and the codec quality you request [1].

Take plain, single-pass Advanced Video Coding (AVC). The multiplier runs 1x for standard definition, 2x for high definition, and 4x for 4K at 30 frames a second [1]. Push the frame rate past 30 and those same three outputs cost 1.25x, 2.5x, and 5x [1]. Ask for the VP9 video codec in multi-pass mode, and high definition jumps from 2x to 4x, then to 7x at 60 frames [1].

Now build a delivery ladder. Three standard tiers, two high-definition tiers, and one 4K tier is an everyday delivery set.

That ladder bills 11 normalized minutes for every single source minute: 3 + 4 + 4 [1]. At $0.0075, one source hour costs $4.95 [1]. The same six output versions cost $9.90 on Google’s card, which prices each output by its raw resolution and gives no ladder discount [2].

What strikes me here: AWS’s own math is blunter than any example I could invent. They model ten thousand 45-minute videos through two ladders, 11.48x for standard definition and 22.97x for high definition, and bill 15.5 million normalized minutes in a single month [1].

I keep coming back to what that does to a small shop. Forty hours of source video a month is a daily two-hour stream, or eighty half-hour episodes. On the 11x ladder, that is $198 a month on AWS and $396 on Google [1, 2].

The local hardware math looks entirely different. Nvidia’s hardware encoder (NVENC) runs 977 frames a second on 1080p H.264 video at its fastest preset on Blackwell silicon [3]. It runs 506 at a mid-tier preset and 227 at its slowest. Those are 8-bit 4:2:0 measurements at 1,920 by 1,080 pixels, with the video clocks stated [3].

At 506 frames a second, one wall-clock hour yields nearly 17 hours of 30-frame output. A 575-watt card running on standard commercial power burns just 7.7 cents in that hour [4, 5]. That is 0.46 cents per output hour.

So the local box does not compete with the transcode fee at the margin. It competes with the raw hours you feed it.

A $2,500 local workstation divided by that $4.95 cloud fee pays for itself after roughly 506 source hours [1]. That is twelve and a half months at 40 hours a month, and over four years at 10 hours a month.

By the numbers

  • 22.97x — Ladder multiplier: AWS multiplies one high-definition 60-frame VP9 output minute by 22.97 before charging the $0.0120 professional rate [1].
  • $4.95 vs $9.90 — One source hour: Six output versions — three standard, two high, one 4K — cost $4.95 an hour on MediaConvert’s basic tier and $9.90 on Google’s transcoder [1, 2].
  • 1,134 fps — High-efficiency encoding: The fastest preset runs 1,134 frames a second on 1080p High-Efficiency Video Coding (HEVC), and 181 at the slowest [3].
  • 0.46 cents — Local power cost: A 575-watt card at 13.41 cents a kilowatt hour, running 506 frames a second, spends under half a cent of power per output hour [3, 4, 5].

What I’d watch:

The pipelines I watch go local always count output versions before they count minutes. The teams that stay on cloud transcode usually ship just one or two outputs per source file.

Concurrency is where the local plan usually breaks. Nvidia limits standard consumer graphics cards to eight simultaneous sessions per system [3]. That hard cap counts every consumer card in the box. Adding a second card buys processing speed, not more concurrent sessions.

Nvidia’s support matrix lists twelve sessions for the newest flagship card, so their two documents disagree [7]. I would verify the graphics driver limits before anyone sizes a server farm on either number.

The quality preset is the dial nobody prices. Between the fastest and slowest settings, a single card loses 77% of its H.264 throughput, and 84% on HEVC [3].

Short clips carry a hidden billing tax. Six-second clips in six renditions bill as six ten-second outputs. That means 36 seconds of content is charged as 60, carrying a 67% overhead [1].

Then there is file size. HandBrake’s benchmark notes say hardware encoders run faster at some cost in picture quality and usually in larger files [6]. Its test found standard software encoding (x264) strikes a better balance of speed, quality, and size [6].

Bigger files move your costs off the transcode invoice and onto cloud storage and network egress bills. Neither transcode rate card includes those data transfer fees. That is the missing variable I have not seen anyone publish in full.

The catch

That massive 22.97x figure applies to a professional-tier VP9 ladder at 60 frames a second, not a standard default setup [1]. Quoting it as typical would be the same trick as quoting the cheapest per-minute rate as the average. The plain single-pass ladder in this piece is 11x.

My read is that the transcode fee stays small for most teams. It only hurts shops pushing ladder-heavy, high-frame-rate video. Below roughly 40 source hours a month, the cloud bill is cheaper in pure cash than buying any physical box [1]. A local rig must then justify itself on fast turnaround times or data privacy, not pure arithmetic.

The local calculation above is also incomplete in the vendors’ favor. It prices the graphics card’s power and nothing else. It skips the host machine, cooling, failed drives, and the engineer’s afternoon lost to a driver upgrade.

What the evidence actually supports is narrower and more useful than the usual claim. The cloud prices the output version. The box prices the hour. The two only meet at a volume the reader has to measure for their own pipeline. I could be wrong about how many shops sit above 500 source hours a month, but the ones that do are easy to spot: their transcode invoice is a four-figure line.

At a glance

  • The Big Shift: Cloud transcoding bills per output minute, scaled by a multiplier table that rarely enters a buying conversation. The delivery ladder decides the invoice, not the raw footage.
  • Why It Matters: At $4.95 a source hour on AWS and $9.90 on Google, a many-rendition pipeline pays a real monthly bill. The same output on one graphics card spends about half a cent of power per hour of output.
  • What I’d Watch: Where the volume line falls, and what the vendors do there.
  • Ladder composition: How many renditions each source minute is delivered in, since every tier and frame-rate band carries its own multiplier.
  • Session limits: The cap on simultaneous encode sessions per system, which decides whether a local farm scales by adding cards.
  • File-size trade: Whether the faster hardware encode costs more in storage and network egress than it saves on the transcode line.
  • The Catch: The 22.97x example is a professional-tier VP9 ladder, the cheapest per-minute rate is not the average, and the local figure here prices only the card’s power.

Related reading

Sources

[1] AWS Elemental MediaConvert — On-Demand Pricing: https://aws.amazon.com/mediaconvert/pricing/ [2] Google Cloud — Transcoder API pricing: https://cloud.google.com/transcoder/docs/pricing [3] NVIDIA — NVENC Application Note, Video Codec SDK 13.0: https://docs.nvidia.com/video-technologies/video-codec-sdk/13.0/nvenc-application-note/index.html [4] NVIDIA — GeForce 5090 specifications: https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/ [5] U.S. Energy Information Administration — Electricity explained: Prices and factors affecting prices: https://www.eia.gov/energyexplained/electricity/prices-and-factors-affecting-prices.php [6] HandBrake Documentation — Performance: https://handbrake.fr/docs/en/latest/technical/performance.html [7] NVIDIA — Video Encode and Decode Support Matrix: https://developer.nvidia.com/video-encode-and-decode-gpu-support-matrix-new