Apple's Secret Weapon for Streaming: From WaveOne to PICO

Apple bought a neural video compression startup in 2023 and said nothing about it. In May 2026 it published the result.

This article originally reported the acquisition and guessed at what Apple might do with it.

Three years on there is an answer, and it is not quite the one anyone expected.

The 2026 Update: Apple Published PICO

In May 2026, Apple Machine Learning Research published PICO, a learned image codec.

Learned means the compression itself is done by a neural network, rather than by a conventional algorithm with machine learning bolted onto the side of it. That distinction is the whole reason this is interesting, and it separates PICO from most things marketed as AI compression.

Apple reports that it delivers between 2.3 and 3 times the bitrate saving of AV1, AV2, VVC and JPEG AI at matched quality, and 20 to 40% against the best competing learned codecs.

The speed matters more than the ratio. On an iPhone 17 Pro Max, Apple reports encoding a 12-megapixel image in roughly 230 milliseconds and decoding in 150 - faster than most machine-learning codecs manage on a datacentre GPU.

Apple also released the code publicly, which is not something the company does often.

Three Things to Keep in Perspective

It is an image codec, not video. WaveOne built video compression, and the speculation at the time was all about Apple TV+ bandwidth. The first public output handles stills.

The numbers are Apple's own. They come from large-scale subjective user studies run by Apple, measuring perceived quality rather than a standard objective metric, and nobody outside Apple has independently reproduced them.

Nothing has shipped. This is a research paper and a code repository, not a feature in iOS or a format your phone writes today.

What Was Reported in 2023

Apple acquired WaveOne, a Mountain View startup working on AI algorithms for compressing video, and never announced it. The company's website went dark around January 2023 and several staff, including a co-founder, turned up in Apple's machine learning groups.

Former WaveOne head of sales Bob Stankosh confirmed the sale in a LinkedIn post in February 2023, and TechCrunch reported the acquisition the following month.

WaveOne was founded in 2016 by Lubomir Bourdev and Oren Rippel, with the stated aim of making video codecs AI-powered. Bourdev was a founding member of Meta's AI research division, and both men worked on Meta's computer vision team.

The company had raised around $9 million from Khosla Ventures, Vela Partners, Incubate Fund, Omega Venture Partners and Blue Ivy.

What WaveOne Was Actually Building

WaveOne's approach was content-aware compression that ran on the AI accelerators already present in phones and increasingly in PCs.

The system used scene and object detection to work out what mattered in a frame, spending bits on faces and taking them from everything else. That is a very different idea from conventional codecs, which treat every part of an image by the same rules.

Its second claim was graceful degradation. Where a conventional stream freezes when bandwidth collapses, WaveOne's would keep playing at reduced detail.

Both ideas are visible in what Apple published, though applied to a single image rather than a moving one.

Apple's Other Codec Moves Since

This was not the only thing Apple did about video compression in that period.

Six months later it shipped the A17 Pro chip with an AV1 decoder, finally putting hardware AV1 support into iPhones, and the M3 generation of Macs followed.

Those two moves point in opposite directions and both make sense. Adopt the industry standard for what you have to play today; research something better for what you might ship in a decade.

What Everyone Else Was Doing

Apple was not alone. In 2022, Alphabet's DeepMind applied MuZero, a reinforcement learning system originally built to play board games, to the problem of choosing encoding parameters for YouTube video, reporting around a 4% reduction in the data YouTube needs to stream.

That result is worth understanding because it is the pattern that actually shipped. It uses machine learning to make better decisions inside a conventional codec rather than replacing the codec.

Netflix's VMAF is the measurement half of the same idea, and unlike most of this field it is free to download and use.

What It Means for You

Honestly, nothing yet. There is no PICO decoder in any operating system, no file you can make, and nothing to install.

The video codecs that matter for anything you watch or encode this year are still HEVC and AV1.

Our comparison of which codec to stream in covers that practical choice, and generative AI is changing multimedia in ways that have arrived faster than compression has.

What has changed is the credibility of the idea. In 2023 neural compression was a research curiosity with a startup attached; it now has a published standard in JPEG AI and a major platform holder publishing working code.

For where this sits in the wider picture, see where codecs are heading, and for the on-device cost of any decoder there is what AV1 does to your battery.

Quick questions

A three-year-old acquisition story with an ending attached: the technology was real, the team delivered, and it landed somewhere nobody was looking.

Browse more codec news, or tell us what you make of it.