You’re paying for bandwidth you don’t need.
Every time someone streams a 360p video on your platform, H.264 demands 1 Mbps. Multiply that by millions of viewers, and your cloud bill looks like a phone number. You’ve probably accepted this as the cost of doing business — video is heavy, bandwidth is expensive, and codecs are what they are.
Microsoft just broke that assumption wide open.
The dirty secret of video streaming isn’t that we lack better codecs — it’s that we’ve been trapped between quality and cost for two decades.
The open-source release of ML Video Codec (MLVC) does something that shouldn’t be possible: it delivers the same visual quality as H.264 at roughly 122 kbps for 360p video at 30 fps. That’s an 8x bitrate reduction. Under real-time conditions. With inference compute kept approximately equal to traditional codecs.
Read that again. The ML model isn’t trading compute for bandwidth. It’s winning on bandwidth without losing on compute.
This is where most people get it wrong. They hear “machine learning video codec” and immediately think: sure, but the inference cost will eat your savings alive. That’s been the curse of every ML compression paper since 2018. Beautiful PSNR numbers, useless in production because the model needs a data center to run in real time.
MLVC sidesteps that trap entirely.
The breakthrough isn’t that AI can compress video better — we’ve known that for years. The breakthrough is that it can finally do it on devices people actually own.
Think about what this means at scale. Not for the Netflixes of the world — they’ll be fine. Think about the 2.6 billion people who live in regions where 1 Mbps is a fantasy, not a baseline. Where mobile data is metered in pennies per megabyte and a single video buffer can burn through a day’s wages.
122 kbps. That’s a trickle. That’s a connection so thin most modern web pages wouldn’t even load on it. And yet MLVC streams watchable, 30 fps video through it.
For engineers building streaming infrastructure, this is a cost earthquake. An 8x reduction in bitrate doesn’t mean 8x savings — it means you can serve 8x more users on the same infrastructure. It means the edge device in someone’s pocket becomes a viable streaming target without compromising quality. It means the economics of video delivery fundamentally shift.
Every codec war in history was won by whoever made quality cheaper. MLVC just made it nearly free.
But here’s the part that should keep traditional codec teams awake: MLVC is open-source. Microsoft didn’t lock this behind a licensing wall or bury it in a research paper. They released it. The same approach that gave us H.264’s dominance — broad availability, industry adoption — is now being used to potentially replace it.
The contrarian take? This isn’t just a better codec. It’s the beginning of the end for the traditional codec development model. For 20 years, we’ve relied on standards bodies moving at geological speed, shipping incremental improvements every half-decade. MLVC suggests a different path: ML-based codecs that improve continuously as models get better, deployed through software updates, not hardware refreshes.
The traditionalists will push back. They’ll talk about artifacting, edge cases, the gap between benchmark numbers and real-world performance. Fair concerns. But they’re the same concerns raised against H.264 when it threatened MPEG-2, and against VP9 when it threatened H.264.
The codecs that win aren’t the ones with the best specs on paper. They’re the ones that make quality accessible to people who couldn’t afford it before.
MLVC does exactly that. Not by inching the efficiency needle. By redefining what efficient means.
If you’re in video engineering, content delivery, or streaming infrastructure, this is your signal. The bandwidth tax you’ve been paying isn’t inevitable. It’s just outdated.
FAQ
Q: Isn't ML inference too expensive to make this practical at scale?
A: That was the assumption, and it's exactly what MLVC breaks. The team kept inference compute approximately equal to traditional codecs for 360p and 540p. This isn't a research toy that needs a GPU farm — it's engineered for real-time, low-power deployment. The old tradeoff between compute and bandwidth doesn't apply here.
Q: What does an 8x bitrate reduction actually mean for my business?
A: It means you can serve 8x more concurrent viewers on the same bandwidth budget, or deliver higher quality to users on weak connections without upgrading infrastructure. For any platform paying for egress, CDN, or mobile data, this directly hits the bottom line. It also unlocks markets where bandwidth was previously a hard blocker.
Q: Is this really the end of traditional codecs like H.264 and H.265?
A: Not tomorrow, but the writing is on the wall. Traditional codecs improve on a 5-7 year standards cycle. ML-based codecs improve every time models get better, deployable via software updates. MLVC being open-source accelerates adoption the same way open standards killed proprietary video formats before. The trajectory favors ML.