Google just recently lowered it's egress rates Through the release of standard networking.

Also don't expect snap to be paying list rate.

https://cloudplatform.googleblog.com/2017/08/introducing-Net...

(Disclaimer I work for Google cloud. I also have no direct knowledge of snaps gcp rate but have worked on other offline contracts)

Yes they have a giant contract with GCP https://www.recode.net/2017/3/1/14661126/snap-snapchat-ipo-s...

That cost is definitely not bandwidth. Content delivery is around $0.005-0.01 per GB with an external provider. Less if yiu have the scale to do it in house. My guess is that the bulk of that spend is compute + storage/io.

The cost is definitely bandwidth, Snapchat doesn't have much of a storage problem as snaps have an actual TTL.

Assuming half of it is bandwidth costs, that means that a daily active user is using ~380 MB/day (assuming $0.01/GB). That seems pretty hefty as an average. Especially if you consider the encoding snapchat uses puts each 10s video at about 1MB in size. 380MB is a little over an hour of snapchat quality video, that seems excessive as an average. If I were a major stockholder I'd be pushing to see why these hosting costs were so high.

I think that’s $0.60 per quarter not per day. 380MB for the quarter per daily active user doesn’t seem that unreasonable.

$0.60 for 380MB seems unreasonable, though. I think the GP was calculating based on the quarterly period.

Oh, you are right. I was thinking $0.006 not $0.60 when I did that math in my head.

CDN? Cross-region replication? DB replication/failsafe/backups? No idea, just spitballing.

I bet bandwidth is a large cost, but I bet compute is a big component too. Snapchat uses Google App Engine, an awesome platform but expensive (comparable to Heroku).

Regardless, my understanding is that Snap now gets some sort of flat rate deal from Google, so not sure how important the individual components are.

Could also be incompetence in regards to managing servers and load testing with laziness for not shutting down instances and ignorance of using reserved instances?

Ahem, I mean it could be.

Hehe! You will be surprised how often this happens in best of the companies. A large fintech startup I work with had 5 GPU servers lying around idle before the CTO realized that they only need them while training the Machine Learning models and not during inference stage.

Been there. Remember getting a call from a data center I'd never heard of asking to speak to the old sysadmin who'd quit 18 month ago. They asked something about the rack of servers we had in their data center, I replied "what rack? what data center?" Turns out that 2-3 years previous the old sysadmin had set up a dozen servers in their data center for some project (that I'd never heard of) that very shortly afterwards first got pushed forwards and then canceled. Everybody involved in that project promptly moved on to other things and no remembered to cancel the deal with the data center. So for two years we'd been paying quite a lot of money to keep a completely idle rack of pretty expensive servers and apparently no one noticed.

Google Cloud does reserve instances automatically...

For every user with the app open, they have to constantly calculate new information too send out, not just snaps but ads and news content. It's not mining bitcoin, but not trivial. Probably a few dozen database reads/writes per user every minute.