I run GPUPerHour.com — it's a pricing comparison site for GPU cloud rentals. I built it because the same H100 costs anywhere from $1/hr to $3+/hr depending on where you rent it, and I got tired of checking 30 provider pages manually.
While scraping all those pricing pages, I kept seeing another line item that gets way less attention than it should: data egress. So I pulled the numbers from every provider I could find.
44 providers. The cheapest charges $0/GB. The most expensive charges $0.55/GB. That's a 127× difference for moving your own data off their servers.
Full breakdown here: gpuperhour.com/reference/data-egress
Half of GPU clouds don't charge egress at all
This caught me off guard. I figured GPU providers would gouge on egress since they already charge a premium for compute. Nope.
6 out of 12 GPU cloud providers charge zero egress. RunPod, Lambda, Vast.ai, Salad Cloud, Voltage Park, VERDA — all free, unlimited.
Then you've got CoreWeave at $80/TB and Paperspace at $120/TB for the same thing. If you're saving model checkpoints regularly — and those can be 10-50GB per save — that gap hits different.
Hyperscaler egress is exactly what you'd expect
AWS is $90/TB after 100GB free. Google Cloud is $120/TB. Azure is $87/TB.
Hetzner? $1.18/TB with up to 20TB free included. OVHcloud? Free. Cloudflare R2? Free.
AWS charges 76× more than Hetzner for egress overages. I run my own infra on Hetzner dedicated boxes and I basically never think about egress costs. On AWS that same traffic would be a real line item.
Developer platforms are the worst and it's not close
This one was weird to me. The platforms that market themselves to indie devs and small teams have the most brutal overage rates:
Netlify: $550/TB. That's the single most expensive number on my entire list.
Render: $300/TB.
Vercel: $150/TB.
Firebase: $200/TB.
Yeah, they bundle bandwidth into plan tiers so you're not paying that from day one. But once you go over? Those "my Netlify bill was $104K" posts on Hacker News start making a lot of sense.
This matters more for ML work than most people realize
GPU pricing comparisons (mine included) usually focus on the hourly rate. Fair enough, that's the biggest number. But real ML workloads move a lot of data around:
Training datasets are regularly 100GB+. You save checkpoints every N steps. You push models to registries or other clouds. You serve inference results.
On Lambda or RunPod, all of that data movement is free. On AWS, moving 1TB of checkpoints costs $90 on top of whatever you're paying for the GPU. That's not nothing.
I added an all-in cost calculator that factors egress, storage, and ingress into the total so you can compare on actual cost, not just the GPU sticker price.
How I actually got this data
There's no API for this. Every provider structures their pricing page differently. Some bury egress in footnotes. Some call it "bandwidth" or "data transfer" or "network egress." A few just don't mention it until you get the bill.
I went through each provider's docs, pulled the standard rate for internet-bound traffic from NA/EU regions, and normalized everything to $/TB. It took longer than I expected.
Full table with all 44 providers, free tiers, and per-TB rates: gpuperhour.com/reference/data-egress
Data is current as of February 2026, licensed CC BY 4.0 — use it however you want.
Bottom line
Check egress pricing before you pick a provider. The difference between free and $90/TB doesn't matter when you're moving a few gigs. It matters a lot when you're training models and shuffling terabytes between services.
I'm Faiz — I built GPUPerHour.com to track GPU pricing across 30+ cloud providers in real time. I also run a Twitter bot that tweets GPU deals. If you're renting GPUs, check it out.
Top comments (1)
Render is $150/TB; please update your calculator.