The Egress Fee Era Is Winding Down. The Lock-In Isn't.

Cloud egress fees remain a major migration cost even after hyperscaler exit waivers, while EU switching rules and local compute shift the calculus. Here's what to measure before moving data.

Cloud data center servers with network cables and billing spreadsheet

The Egress Fee Era Is Winding Down. The Lock-In Isn't.

Apple announced an M6 14-inch MacBook Pro this month, and Ghostty 1.4 hit the front page of Hacker News with 960 points. Neither of those things has anything to do with cloud billing. Both are part of why I've spent the last three weekends with a spreadsheet open, re-measuring what my AWS bill is actually buying me.

Here's the asymmetry that's been eating at me. The cost of compute I own falls every year — the M6 is faster than the M1 Max that's still serving half my homelab, and it draws less power. The cost of getting my data out of someone else's datacenter has been $0.09/GB for a decade. Storage gets cheaper, compute gets cheaper, and the exit door stays exactly where it was.

Forty terabytes in S3 is a $3,600 bill the moment you decide you'd rather it lived somewhere else. That's not a network cost. Transit has gotten cheaper over the same period. It's a number chosen to make you not ask the question.

Except now the regulators are asking it for you, and the hyperscalers are suddenly very generous about a thing they spent ten years defending.

What actually changed

Three things, in roughly this order.

The unilateral policy reversals. Google Cloud announced in January 2024 that it would stop charging egress fees when a customer leaves entirely. AWS followed in March 2024. Azure made similar noises. Read the fine print and you'll find these are "exit" waivers, not "operational" waivers — they apply when you're moving all your data off, you typically have to request it, and your normal day-to-day egress is untouched. It was a preemptive strike against regulation, and it worked well enough that most people now believe egress fees are gone. They aren't.

The EU Data Act. Regulation (EU) 2023/2854. The switching provisions became applicable on 12 September 2025, and the hard deadline for eliminating switching charges entirely lands in January 2027. It covers IaaS, PaaS, and SaaS, requires providers to give you exportable data in a structured format, caps the notice periods, and — this is the part people miss — phases out the egress fees you pay to leave. I'm not a lawyer and this isn't legal advice, but if you have any EU footprint at all, this is now a negotiating lever you can name out loud in a vendor call. That alone changes the temperature of the conversation.

Local compute crossed a threshold. This is the quieter one. Two years ago, the "just run it on your own hardware" argument only worked for storage and static sites. Today it works for inference, for CI runners, for a decent chunk of your batch processing. If you can read a 30B parameter model off an M6 at a reasonable token rate, the calculus on "do I need to pay for this GPU time plus the egress to get the results out" genuinely shifts. I wrote about the memory-price trap in GPU buying — the same logic applies here, just with an API bill instead of a card.

The fees didn't disappear. They moved.

This is the part that frustrates me most, because it's the part nobody puts in the announcement blog post.

Waiving egress for customers who leave doesn't help you if you're not leaving. And it does nothing about the fees you pay inside the cloud, which are where the money actually went. If you're on AWS, go look at these lines in your bill:

  • NAT Gateway. $0.045/hour per gateway, plus $0.045/GB processed. Every byte your private subnets send to the internet pays this twice — once in, once out.
  • Cross-AZ traffic. $0.01/GB in each direction. A chatty microservice mesh spread across three AZs is a standing bill you never approved.
  • Transit Gateway. $0.05/hour per attachment plus $0.02/GB. The "fix" for spaghetti VPC peering is itself a metered product.
  • PrivateLink / VPC endpoints. Per-hour per-endpoint, plus per-GB processed. The recommended way to avoid NAT gateway charges has its own charges.
  • Cross-region replication. The price of your DR plan, billed continuously, forever.

None of these are covered by the exit waivers. None of them are covered by the Data Act's switching provisions either — those are about leaving, and you're not leaving, you're just operating. The lock-in didn't get removed; it got layered. The door is open, but the hallway now has six more turnstiles.

Measure before you move

Every migration I've watched go badly started the same way: someone estimated the egress and was wrong by an order of magnitude. Don't estimate. Query.

# What am I actually paying for? Group by usage type, filter to the
# things that are secretly data transfer, sort by cost.
aws ce get-cost-and-usage \
  --time-period Start=2026-07-01,End=2026-10-01 \
  --granularity MONTHLY \
  --metrics UnblendedCost \
  --group-by Type=DIMENSION,Key=USAGE_TYPE \
  --output json \
| jq -r '.ResultsByTime[].Groups[]
    | select(.Keys[0] | test("DataTransfer|NatGateway|Bytes|Regional"))
    | [.Keys[0], (.Metrics.UnblendedCost.Amount | tonumber)] | @tsv' \
| sort -k2 -nr | column -t

Run that against a year, not a month. Seasonal traffic will lie to you otherwise. What you're looking for is the ratio between "egress to internet" and "egress that never left the provider." In most accounts I've poked at, the second number is larger, and it's the one with no exit waiver attached.

Once you know the size, rclone is still the right tool for the actual move, and it's dramatically better than it was five years ago:

# Sanity check the volume first.
rclone size s3-prod:my-bucket

# Copy, don't sync. sync deletes on the destination — you want that
# nowhere near a migration you're still validating.
rclone copy s3-prod:my-bucket r2:my-bucket \
  --transfers 32 --checkers 64 \
  --s3-chunk-size 32M --s3-upload-concurrency 8 \
  --fast-list --progress \
  --log-file /var/log/rclone-migrate.log

Then run it a second time right before cutover to pick up the delta, because the first pass on a live bucket will always be stale by the time it finishes. Verify with rclone check --one-way and only then flip the application config.

If you'd rather not pay anyone for the destination at all, MinIO on your own iron is a genuinely viable target now — I run one alongside the rest of my homelab automation stack, and the S3 compatibility is good enough that rclone doesn't care where it's pointed.

Where to put the data

List prices, roughly us-east-1 tier, and they move constantly. Check before you commit.

Provider Egress to internet Storage /TB-month S3 API The catch
AWS S3 ~$90/TB (100GB free) ~$23 Native Inter-AZ and NAT charges are separate and larger than you think
Google Cloud Storage ~$120/TB (200GB free) ~$20 Interop Exit waiver only if you leave entirely
Azure Blob ~$87/TB (100GB free) ~$18 Partial Hot/cool/archive tier transitions bill as operations
Cloudflare R2 $0 ~$15 Native Class A/B operation limits; no lifecycle tiering worth using
Backblaze B2 $0 via Bandwidth Alliance partners ~$6 Native Free egress usually means "through Cloudflare"
Wasabi $0 ~$7 Native 90-day minimum storage duration — delete early and you pay anyway
Hetzner Object Storage Included allowance, then cheap ~€5 Native EU-only regions; not for latency-sensitive US workloads
MinIO (self-hosted) Your bandwidth bill Your disk bill Native You are now the on-call rotation

The honest read on that table: the storage price differences are noise. The egress column is the whole decision. A 20TB working set that moves 30TB/month of egress costs you roughly $2,700/month on S3 and roughly nothing on R2 or B2. That's the entire argument, and no amount of "but the ecosystem" survives contact with it.

And if data residency is part of your problem — which it increasingly is for anyone with EU customers — the provider question stops being about price at all. I wrote about the residency side of this in the context of Ansible and DNS, and the short version is that "where is the bucket" is only half the question. Where the control plane runs matters just as much, and repatriating storage while leaving your orchestration in us-east-1 doesn't actually solve the problem you told your compliance team you solved.

What I'd do this quarter

Not migrate. Not yet.

Instrument first. Run the Cost Explorer query above over twelve months. I'd bet money that more than half your "egress" never crossed a provider boundary and is therefore immune to every waiver and regulation currently on the books. You can't re-architect what you haven't measured, and I've watched teams spend six months on a storage migration that addressed 15% of the bill.

Then re-architect the internal traffic. Collapse chatty cross-AZ service calls. Put a VPC endpoint where the NAT gateway is. Move batch workloads to the same AZ as their data. This is unglamorous work with no vendor announcement and no regulatory deadline, and it's usually the highest-return thing available.

Then, and only then, talk about leaving. And when you do, name the Data Act in the call. If you're an EU customer, you now have a legal basis to demand export tooling and a removal date for switching charges. Vendors know this. The tone changes when you say it.

The caveat

I still run production workloads on AWS. I'm not moving them, and I don't think most people should. RDS, IAM, and the fact that I don't have to think about a rack at 2am are worth real money, and the "everything to Hetzner" crowd tends to underestimate what they're signing up for. See the MinIO row in that table — that's not a joke, that's a pager.

The point isn't to leave. It's to be able to leave, cheaply enough that the option has value. A vendor that knows you can walk gives you better pricing, better terms, and returns your emails. A vendor that knows your egress bill is a moat does not.

The M6 in that MacBook Pro is roughly as fast as the instance class I was paying for in 2021. That's not a coincidence worth ignoring — it's a reminder that the floor on what "good enough" costs keeps dropping. Your architecture should be able to follow it down.

Do this week

Pull twelve months of usage-type cost data. Find the line item where data moves and you weren't expecting a charge. That's the one to fix first, and there's a decent chance it's never leaving your provider's network at all.

Disclosure: some links below are affiliate links. If you buy through them I may earn a commission at no extra cost to you. I only list tools I actually run.

ToolWhy it is here
Amazon - SSDs & NVMeBoot drives and storage for your homelabVisit
Amazon - Mini PCsLow-power always-on nodes (ProDesk, EliteDesk, N100 boxes)Visit
Amazon - PC ComponentsCPUs, RAM and cases for a buildVisit
Amazon - GPUsConsumer GPUs for local inference (16GB VRAM is the sweet spot)Visit
Amazon - RAM & MemoryMore RAM is the cheapest upgrade for VMs and containersVisit
WasabiFlat-rate hot storage, no egress feesVisit