Comment by 0cf8612b2e1e
9 hours ago
Somewhat relatedly, how do the economics for Huggingface work? They must be hosting petabytes of models and datasets by now. I have downloaded quite a few “just in case”, only to replace them with the later iteration months later.
Does the file hosting actually cost peanuts when you do it yourself and the cloud has shattered my understanding of what it actually costs to deliver so much data?
HF uses chunk-level deduplication that greatly reduces the amount of physical storage needed for a given amount of logical storage.
https://huggingface.co/docs/xet/en/deduplication
File hosting is pretty cheap. Egress traffic is about $90/TB in the cloud, but around $1/TB in the real world. Storage is in the realm of $5/TB/month after adding redundancy
At the scale of Huggingface, that still amounts to a lot load of money. Significantly less than if you did the same in AWS, but still a lot
That said, they do have a deal with AWS to make the data available in AWS ip space. Maybe they got some cheap hosting out of that too
It's less than $1/TB. If you have settlement-free peering it's like fractions of a penny at scale. Cloud provider DTO pricing is literally the biggest scam on Earth.
My company spends $8,000,000/month on storage and $4,000,000/month on data with aws... In the financial reports we do we round this number away. Make of that what you will
What does DTO mean in this case? And settlement-free peering? I didn't know the price for S3 could be so cheap with AWS
1 reply →
Bandwidth is really cheap when you run your own infra
It's not only cheap, it's actually free since if you are not using your 95th percentile you are losing money.
Some enterprising ISPs get into the hosting business because most of their traffic is ingress, so adding some egress revenue is effectively zero cost.
when trying to download models using browser, the DL domain for me us.aws.cdn.hf.co which points to bunch of singapore AWS IPv4 for me, despite it saying US.
Maybe download numbers are relatively low
Download numbers are reported on the site, and they aren't low.
CDN costs are pennies compared to inference and training though, HuggingFace will just get another 100 million and be set.