How Much Storage Does 4K Video Use? File Size per Hour and AI Dataset Planning

A 4K video dataset can range from a few terabytes to several petabytes because resolution alone does not determine file size. Bitrate and runtime drive the encoded source size, while codec, frame rate, bit depth, retained renditions, and preprocessing determine the infrastructure footprint.

For teams building or licensing AI training datasets, this distinction matters before footage is transferred or staged for training. “10,000 hours of 4K” describes visual resolution and runtime, but it does not tell an infrastructure team whether the source package will be closer to 100 TB, 500 TB, or several petabytes.

Start with measured bitrate or GB per hour from representative files. Then add the working renditions, extracted frames, caches, temporary processing space, backups, and archives that will coexist. That separates the delivered source from the working footprint the training pipeline will actually consume.

The fastest estimate comes from bitrate. At 100 Mbps, encoded 4K video uses about 0.75 GB per minute or 45 GB per hour. That is approximately 4.5 TB for 100 hours, 45 TB for 1,000 hours, and 450 TB for 10,000 hours before working copies or backups are added.

At 400 Mbps, one hour is about 180 GB, so 1,000 hours is about 180 TB and 10,000 hours is about 1.8 PB. These are bitrate-based reference points, not universal 4K file sizes.

If representative files are available, measure them directly. For variable-bitrate or mixed-source collections, real sample files are usually a better planning input than a nominal resolution label.

​

Values use decimal GB, TB, and PB and are calculated from average encoded bitrate. Actual package size can differ because of variable bitrate, audio, container overhead, metadata, and clip-level variation.

For constant or average bitrate video, the basic calculation is:

Storage (GB) ≈ bitrate (Mbps) × duration (seconds) ÷ 8 ÷ 1,000

For one hour, the shortcut is even simpler: multiply Mbps by 0.45. A 50 Mbps stream is about 22.5 GB per hour, 100 Mbps is about 45 GB, and 200 Mbps is about 90 GB.

For variable-bitrate video, use the measured average from a representative sample. If the collection combines cameras, codecs, frame rates, or production sources, calculate each technical group separately and add the results. A single clip may not represent an entire archive.

​

Wavebreak Media’s current 4K video dataset lists 287,265 clips and 1,571.6 total hours, with 4K/UHD bitrate of approximately 87–93 Mbps and HD at about 22 Mbps. Using 90 Mbps as a planning midpoint, the 4K rendition is roughly 63.7 TB for 1,571.6 hours.

If both the 4K and HD renditions are retained for the same 1,571.6 hours, the HD copy adds about 15.6 TB at 22 Mbps, bringing the two encoded renditions to roughly 79.2 TB before container overhead, manifests, working copies, temporary processing, or backups.

This is the planning distinction that generic 4K estimates miss: resolution describes the frame, bitrate determines the encoded data rate, and the delivery package determines how many versions of that data must exist.

​

4K UHD is 3840 × 2160 pixels, four times the pixel count of 1920 × 1080 video. 8K UHD is 7680 × 4320, four times the pixel count of 4K. Pixel count is useful for understanding potential spatial detail, but it is not a direct storage multiplier unless bitrate and other encoding choices are also comparable.

For many computer vision tasks, 1080p may be sufficient. 4K becomes more useful when the pipeline needs fine spatial detail, aggressive crops, small-object visibility, or source material that may be reused for higher-fidelity tasks. 8K is more specialized because decoding, transfer, and storage costs rise quickly.

Choose resolution for the information the model needs. Size storage from the files that will actually be delivered.

​

“4K” fixes frame dimensions, but several technical choices determine how much data those frames occupy.

Bitrate and Codec

Bitrate is the clearest direct predictor of encoded file size because it expresses how much data is used per second. Codec determines how efficiently that bitrate represents the image. H.264 remains broadly compatible; HEVC and AV1 can provide higher compression efficiency in suitable workflows, while decoding support and compute cost differ by environment.

Two videos can therefore have the same 4K resolution and runtime but very different sizes. If they are encoded at the same average bitrate, however, their storage footprints will be similar even if the codecs differ.

Frame Rate

Higher frame rates give the encoder more frames per second. The file size does not necessarily double when frame rate doubles because compression can exploit temporal similarity, but 50 or 60 fps material should not automatically be sized with the same assumptions as 24, 25, or 30 fps footage.

Bit Depth and Chroma Subsampling

Professional video may use 10-bit or 12-bit precision and 4:2:2 or 4:4:4 chroma rather than the 8-bit 4:2:0 combinations common in delivery video. These representations preserve more color and tonal information and often require higher data rates or less aggressive compression.

Audio

Audio is usually smaller than the video stream, but it still belongs in the storage model for speech-heavy and multimodal datasets. Multiple language tracks, multichannel audio, lossless audio, transcripts, diarization outputs, and alignment data create their own storage and processing layers.

Runtime and File Count

Once format is stable, storage scales approximately with runtime. File count matters operationally as well: hundreds of thousands of short clips create more object, checksum, manifest, indexing, and request overhead than a few long files with the same total duration.

Source and Camera Type

Smartphones, mirrorless cameras, cinema cameras, and aggregated device pools can all produce nominally 4K footage with very different codecs, bitrates, bit depths, and frame rates. Plan around the agreed delivery format unless camera-native originals are part of the licensed package.

​

How many versions are retained

Production storage often exceeds the downloaded package because teams keep more than one representation: an untouched source, normalized working copies, extracted frames or proxies, training shards, temporary outputs, annotations, and backups. Storage planning should count the versions the workflow needs at the same time.

​

It is useful to group video by operational role rather than assume that all 4K footage has one typical file size.

Compressed delivery formats

H.264, HEVC, AV1, and similar codecs are designed to make distribution and storage practical. For many computer vision, video understanding, retrieval, and multimodal workflows, compressed delivery offers a sensible balance between fidelity, decoding support, and infrastructure cost. The key question is whether the bitrate preserves the visual information required by the model.

Production and mezzanine formats

Formats such as Apple ProRes and Avid DNxHR are common in production and post-production workflows where repeated editing and re-encoding are expected. They typically use substantially higher data rates than compact delivery encodes, so a dataset delivered in a mezzanine format should be budgeted separately rather than estimated from consumer 4K examples.

Camera-native and RAW formats

RAW becomes relevant when the task needs camera-pipeline information, high bit depth, wide dynamic range, or flexibility before color processing. It can be appropriate for restoration, computational imaging, generative-video research, or other fidelity-sensitive work, but it should not be requested by default.

If a training pipeline immediately converts everything to resized or compressed inputs, a much larger source format may add storage and transfer cost without adding information the model ultimately receives. Source fidelity should be justified by the task.

​