150
Tbps of video at the global evening peak
Assume an hour a day per membership, a 3× evening peak and 4 Mbps a stream. Yet only about 14k plays start a second.
Where do the bytes come from?
Bad: Rent a pull-through CDN
A new episode misses every edge at once, across busy interconnects.
Good: Own boxes inside ISPs
Filled overnight from a forecast. Bytes never leave the ISP.
The files move while you sleep
A storage appliance fills from 2 AM to 2 PM local time, up to 6.5 TB a night. A whole new season is about 224 GB.
Mid-episode, the appliance streaming to your TV stops answering.
The trick
Your TV already holds three or more ranked URLs. It asks the next for the same byte range; the buffer hides the gap.
18 PB
saved by 20% fewer bits on one series
One series streamed for 50M hours at 4 Mbps. So give every title its own ladder: the compute is a rounding error.
Losing a region costs capacity, not correctness
- Every service on the call path is stateless
- Each region uses only local resources
- No cross-region calls; replicate data asynchronously
- Rehearse it: Chaos Kong evacuates a whole region
Launch night is a flag, not a deploy
The episodes reached the ISPs days ago. At launch, a gate opens once copies, rights and time line up. No bytes move.
Remember
Decide in an active-active cloud; stream from your own boxes inside ISPs, filled overnight. Each play gets a ranked list of appliances, and every title gets per-shot encoding.