Video is the part of our stack we are least willing to hand to someone else. It sits on the critical path of a first date that happens over a screen, and the difference between a call that feels present and one that feels broken is measured in milliseconds. So we build and operate this layer ourselves, end to end, for our own brands.
The three architectures we run
Peer-to-peer carries our two-party calls beautifully and falls apart past three. NAT traversal pushes a meaningful share of sessions onto our own TURN relays, so we run that relay fleet ourselves and watch its cost like a hawk.
Selective Forwarding Units route streams without re-encoding them. They scale linearly with participants and are the backbone of our group experiences. We operate our own SFU cluster rather than rent one, because the calls our members make are exactly the workload we tune for.
Multipoint Control Units composite streams server-side. They cost dramatically more in CPU but deliver a single low-bandwidth stream to each participant. We reserve them for the broadcast and bridging moments where they earn their keep, and nothing else.
Why we own this layer
Owning the video stack means our unit economics, our reliability, and our roadmap are ours. We are not waiting on a vendor's median customer to justify a fix in a geography we care about, or absorbing a price change we did not choose.
It is harder this way, and that is the point. The codepath that survives a Friday-night load spike is the one we wrote, instrumented, and stand on call for. For the brands we intend to run for decades, that ownership is not overhead. It is the product.