How much does AWS Wavelength cost per month? Breaking down the four charges, plus 5 cost-reduction measures

Conclusion first
- A Wavelength bill is never "how much does one server cost." EC2 instances, EBS volumes, Carrier Gateway hourly fees, and traffic in and out of the Wavelength Zone—four charges stacked together. If you budget by looking only at the instance unit price, you will underestimate by at least 30–50%.
- Typical bill for three m5.xlarge instances running continuously in
us-east-1-wl1-bos-wlz-1: the instances themselves are roughly USD 450–550 per month; adding traffic and Carrier Gateway, the total falls at USD 600–800, equivalent to RMB 4,000–6,000. It varies by region, instance type, and carrier; refer to the AWS official pricing page. - Cost-reduction ROI from high to low: reduce cross-zone traffic > stop instances during off-hours > change instance type or negotiate discount commitments > storage cleanup > downgrade configuration. The first two can save more than 30%, while the last three combined usually save less than 15%.
Cost-reduction measures ranked by ROI; the first can save 30%
| Rank | Measure | Typical reduction | Implementation cost | Time to take effect |
|---|---|---|---|---|
| 1 | Reduce Wavelength→Region cross-zone traffic | 25–40% of bill | Medium, requires architecture changes | Next billing cycle |
| 2 | Stop instances during non-business hours | 20–35% (depends on uptime) | Low, EventBridge + Lambda | Same day |
| 3 | Switch to Graviton instance types or buy discount commitments | 15–20% | Low, but confirm support scope first | Immediately / 1 month |
| 4 | EBS slimming, switch gp2 to gp3, clean up snapshots | 3–8% | Very low | Same day |
| 5 | Downgrade instance type (m5.xlarge → m5.large) | Depends on load | Low | Next cycle |
The number one item is cross-zone traffic. The reason is simple: it is priced per GB, and traffic out of the Wavelength Zone and traffic back to the region are two separate prices; many people only calculate the egress side. The approach is to push static assets, logs, and image repositories back to the region side for centralized processing. Keep only the compute that truly needs to run at the edge on the Wavelength side, and change origin fetches to batched asynchronous.
The second-ranked stop/start strategy has one prerequisite that must be stated clearly first: Wavelength Zones do not have Spot. The idea of saving money with Spot does not work here; the only option is to stop/start according to business hours. For scenarios with clear peaks such as live streaming and cloud gaming, running 10 hours a day versus 24 hours a day makes a bill difference of more than double.
The pitfall with the third-ranked instance-type change is that Graviton's support scope in Wavelength is narrower than in regions. Do not buy according to the instance type list in the region; check first:
aws ec2 describe-instance-type-offerings
--region us-east-1
--location-type availability-zone
--filters Name=location,Values=us-east-1-wl1-bos-wlz-1
--query 'InstanceTypeOfferings[].InstanceType'
--output text
The output shows which instance types this zone actually offers. I have seen people hard-code `m6g.xlarge` based on the region list, and the apply fails outright. Whether discount commitments (Savings Plans / RI) can be applied to Wavelength also needs to be checked separately by selecting the zone on the pricing page; do not assume the region's discounts carry over.
Storage and configuration downgrade are tasks you can knock out along the way. gp3 is cheaper than gp2 and has better baseline performance; for a 100GB volume, the monthly difference is only a few dollars, but with many volumes it is no longer a small amount. Instance-type downgrade is placed last because workloads running on Wavelength are often latency-sensitive; once CPU is reduced, the money saved earlier may be eaten back by experience issues.
## How are the charges calculated: pricing dimensions of each of the four charges
| Cost item | Pricing dimension | Order of magnitude (using `us-east-1-wl1` as an example) | Easily missed point |
|---|---|---|---|
| EC2 instance | Billed per second, minimum 60 seconds, invoiced hourly | m5.xlarge about USD 0.22/hour, around USD 160 per month | Noticeably more expensive than a normal AZ in the same region |
| EBS | Per GB-month + provisioned IOPS | gp3 100GB about USD 8/month | Still charged after stop; snapshots stored incrementally |
| Carrier Gateway | Hourly fee + per-GB processing fee | A few dollars for a whole month when idle; goes up per GB when traffic is high | It is not the same thing as a NAT Gateway |
| Cross-zone traffic | Per GB, different prices for inbound and outbound directions | Depends on origin-fetch volume | Wavelength→Region priced separately |
Amounts vary by region and instance type; refer to the AWS official pricing page.
Billing granularity is basically the same as in regions, but there are two details that can trip you up. First, EBS continues to be billed after an instance is stopped; if you do not clean up the volumes, you keep paying. Second, if a Carrier IP is bound to an ENI and not released, even if the instance is stopped, that hourly fee still continues. The amount is very small—less than 1 cent per hour—but it stays on the bill, making you suspicious when reviewing costs.
## Hands-on test and pitfalls: NAT Gateway could not be created, and after switching to Carrier Gateway the bill actually went up
In 2025-11, I helped a team doing overseas real-time voice chat deploy on a Verizon Boston node. The first version of Terraform directly copied the private-network template from the region: private subnet plus NAT Gateway for outbound internet access.
apply was stuck for more than 40 minutes, and the original error was:
Error: creating EC2 NAT Gateway: UnsupportedOperation:
NAT Gateways are not supported in Wavelength Zones.
The basis for the judgment at the time was that the only public internet egress path for a Wavelength Zone is Carrier Gateway; NAT Gateway and some ELB types simply do not exist in this environment. The verification method was crude but effective: first confirm the zone type, then look through the Wavelength quotas documentation, not guess.
```bash
aws ec2 describe-availability-zones
--region us-east-1
--zone-names us-east-1-wl1-bos-wlz-1
--query 'AvailabilityZones[0].{Zone:ZoneName,Type:ZoneType,Group:GroupName,Status:OptInStatus}'
--output table
+---------------------+-----------------------------+
| Group | us-east-1-wl1 |
| Status | opt-in |
| Type | wavelength-zone |
| Zone | us-east-1-wl1-bos-wlz-1 |
+---------------------+-----------------------------+
Once you see the Type `wavelength-zone`, you should go check the limitations list instead of continuing to struggle with NAT. The fix was to delete NAT, create a Carrier Gateway, and add `0.0.0.0/0` in the route table pointing to it. The service came up, and voice round-trip latency dropped by around 20 milliseconds compared with going through the region; the customer was satisfied on the spot.
Then the bill came out on the 3rd of the following month: the EC2 line was just over USD 500, and Carrier Gateway data processing fees plus cross-zone traffic were more than USD 300. Only after grouping by Usage Type in Cost Explorer did it become clear that the bulk was not in compute at all—object storage reads and writes were all placed on the region side, so every request crossed a zone.
Finally, two weeks of VPC Flow Logs aggregated by destination IP located the issue: more than 60% of traffic was images and logs going to the S3 endpoint in the region. After changing thumbnails to local caching on the Wavelength side and logs to batched asynchronous uploads, cross-zone traffic dropped by 70%, and the total bill came back to around USD 550.
There is only one lesson: Wavelength architecture review should be done before writing Terraform, not after the bill arrives.
## When not to use Wavelength
- When the business's own latency requirement is at 50
Leave a Reply