Hands-on AWS knowledge base for developers Managed Ops
Talk to Us

Wavelength vs. Outposts: How to Choose for Edge Computing — Start with the Order of Magnitude of Latency

Export PDF Print-friendly layout with a copyright watermark; choose Save as PDF in the print dialog
Wavelength vs. Outposts: How to Choose for Edge Computing — Start with the Order of Magnitude of Latency

Conclusion first

  • If the bottleneck is the tens of milliseconds from "device to server": the devices are on a 5G network, users are concentrated in one or two cities, and the workload is stateless — choose Wavelength. End-to-end RTT can usually be reduced to the order of single-digit to low-teens milliseconds.
  • If the bottleneck is "data cannot leave this building": factory MES, medical imaging, in-store local processing, compliance residency — choose Outposts. RTT within the same LAN is generally under 1 ms, at the cost of a 4–8 week deployment lead time and on-site operational responsibility.
  • If you just want users a bit closer and can accept around 10 ms: do not use either of these; Local Zones plus CloudFront is an order of magnitude cheaper. Before choosing a solution, measure RTT first — do not choose based on gut feeling.

What exactly is the difference between the three kinds of "edge": location determines the order of magnitude of latency

Many people call both Wavelength and Outposts "edge," but they fundamentally do not solve the same problem. Wavelength is AWS putting compute inside the carrier's 5G facility; it is an extension zone of the parent Region, named like us-east-1-wl1-nyc-wlz-1. Outposts is AWS moving a rack or 1U/2U servers into your own data center. One sends outward, the other pulls inward.

The difference in latency magnitude is intuitive. A device in mainland China connecting directly to us-east-1 has an RTT of around 200 ms; this is determined by the physical distance across the Pacific, and no amount of CDN layers can change the first packet. Local Zones are placed in metro data centers, with metro RTT on the order of about 5–15 ms. Because Wavelength is inside the carrier network, the 5G air interface itself accounts for 10–20 ms, and the round trip seen by the application is usually on the order of the teens to thirty milliseconds. Outposts is right under the same switch, 0.2–0.5 ms, no different from a local data center.

A note on mainland China reality: Wavelength currently has zones only in regions where AWS partners with carriers, such as the US, Japan, South Korea, and Europe; there is no zone in mainland China. For Outposts, ordering availability in the China Regions is also subject to the official service list. If you are building business overseas, this is worth calculating carefully; if you are building business in mainland China, you can basically skip this section.

Magnitude comparison and testing methods: do not copy other people's numbers, measure them yourself

I generally do not look at latency numbers on vendor slides, because link quality, device models, and carrier interconnection can all shift the result by an order of magnitude. The test method has just three steps and is very low cost.

Step one: from a real device (not office broadband), run mtr against the edge-side instance, at least one hundred packets, and look at average RTT and packet loss:

# Run on the 5G device side, 100 packets, display statistics
mtr -rwzc 100 198.51.100.24


Step two: check the throughput ceiling. The Wavelength side is limited by the carrier link; do not plan based on the intuition of bandwidth within a Region:

```bash
iperf3 -c 198.51.100.24 -t 30 -P 4


Step three: run an application-layer concurrency test, focusing on P99 rather than the average:

```bash
wrk -t4 -c200 -d60s --latency https://api.example.com/healthz


After these three tests, the magnitudes in the table below should roughly match the actual situation of your own business.

| Dimension | Wavelength | Outposts | Local Zones |
|---|---|---|---|
| Physical location | Carrier 5G facility | Your own data center | AWS metro data center |
| Typical RTT | Teens to thirty ms (including 5G air interface) | <1 ms (same LAN) | 5–15 ms (same metro) |
| Who manages the hardware | AWS + carrier | AWS manages servers and switches; you manage power, cooling, and access control | AWS |
| Deployment lead time | Days to weeks, requires carrier-side provisioning | From 4–8 weeks, including logistics, racking, and activation | Hours to days |
| Available instance types | Few, and varies by zone (common t3, r5, and several g4dn specs) | A limited number of families, capacity calculated locally | Medium, slightly fewer than in a Region |
| Billing model | On-Demand/Reserved; monthly bill per instance on the order of thousands to tens of thousands of RMB | 1-year or 3-year commitment; monthly bill for a full rack starting in the tens of thousands of RMB | On-Demand, unit price slightly higher than in a Region |
| Typical scenarios | Cloud gaming, interactive live streaming, AR/VR, connected vehicles | Factory production lines, medical imaging, data residency | Game servers, video transcoding, hybrid cloud access |

Amounts are given only as orders of magnitude; specifics vary by region and instance type, and the AWS official pricing page prevails.

## Hands-on testing and pitfalls: the time an instance would not start in a Wavelength Zone

In March 2025, I was helping a team building interactive live streaming for overseas markets deploy a pre-streaming authentication service in `us-east-1-wl1-nyc-wlz-1`. The first version took the easy route: copied the launch template from the parent Region directly, kept the instance type as `m5.large`, switched the subnet to the Wavelength subnet, and changed nothing else. RunInstances failed immediately:


An error occurred (Unsupported) when calling the RunInstances operation: Your requested
instance type (m5.large) is not supported in your requested Availability Zone
(us-east-1-wl1-nyc-wlz-1). Please retry your request by not specifying an instance type
or choosing the following size: t3.medium, t3.xlarge, r5.2xlarge, g4dn.2xlarge.


The first reaction was quota. I checked with `aws service-quotas get-service-quota --service-code ec2 --quota-code L-1216C47A`, and usage was far below the limit, so that was ruled out. The second suspect was the AMI architecture, also ruled out. Finally I went back to the phrase in the original error text: "not supported in your requested Availability Zone" — Wavelength Zone instance types are allowlisted, and the specs available in each zone differ. Verification command:

```bash
aws ec2 describe-instance-type-offerings 
  --location-type availability-zone 
  --filters Name=location,Values=us-east-1-wl1-nyc-wlz-1 
  --region us-east-1 
  --query "InstanceTypeOfferings[].InstanceType" --output text


The output was just seven or eight instance types, and `m5.large` was indeed not among them. Switching to `t3.medium` brought it up immediately.

The same incident included a second stumble: after the instance was Running, the partner's callback could not get in. Security groups and NACLs were both allowing traffic, and packet capture showed only the SYN in the request direction with no return path. The reason is that instances in a Wavelength subnet do not have public addresses by default; you must separately allocate a carrier IP and associate it:

```bash
aws ec2 allocate-address --domain vpc 
  --network-border-group us-east-1-wl1-nyc 
  --region us-east-1


That time it took about forty minutes from alert to localization, and half the time was wasted ruling out quota and AMI. The lesson: the answer was already written in the original error text; do not rush to guess based on experience.

On the Outposts side, the most common failure point is local gateway routing. In the VPC route table, you must explicitly add a 0.0.0.0/0 route pointing to `lgw-xxxxxxxx`. Many people's symptom is "the same subnet can communicate, but traffic cannot get out," and then they go looking at security groups, when in fact the route was never configured at all.

## Beyond performance: operational responsibility, deployment lead time, and billing are the real deciding factors

Once the latency gap reaches a certain point, it is good enough; what really causes hesitation is the other three things.

**Who bears the responsibility.** For Wavelength, the hardware, facility, and power are all AWS's and the carrier's responsibility; you only touch the console. For Outposts, although AWS manages the servers and switches, rack space, dual utility feeds, cooling, access control, on-site escort, and disk degaussing are your job. Without a 7×24 on-call team, Outposts will become a long-term burden.

**How long until it can be used.** Local Zones can be opened in a few hours. Wavelength depends on carrier-side coordination, from days to weeks. For Outposts, from order to being able to run workloads, our production case took about six weeks, including logistics, on-site racking, and activation, with property management and carrier line installation also needing coordination in between. If the business must go live next month, neither path will make it in time.

**At what tier the bill starts.** Edge instance unit prices are generally higher than Region instances of the same spec; be mentally prepared for that. Wavelength is billed per instance, with a monthly bill per instance on the order of thousands to tens of thousands of RMB depending on the instance type; Outposts is a one-year or three-year commitment, with the monthly bill for a full rack starting in the tens of thousands of RMB, and if the business volume is not large enough, it cannot be amortized.

There is also a point that teams in mainland China easily overlook: after switching the console language to Chinese, the translations of Wavelength-related pages and quota items do not match the English documentation. When checking errors, it is recommended to search directly with English keywords.

### When not to use these two solutions

- **Users are mainly in mainland China.** There is no available Wavelength zone, and cross-ocean deployment will only be slower; just use Region + CloudFront.
- **You just find the Region a little slow.** Reducing RTT from 200 ms to 15 ms is enough with Local Zones, and the cost differs by an order of magnitude.
- **The data volume is large and ultimately must be sent back to the Region.** When the edge only performs computation and does not store data, the backhaul bandwidth becomes the new bottleneck; calculate the cost clearly in TB/day first.
- **You have no on-site operations capability.** With unstable power, no dedicated cabinet, or a facility that people cannot enter, installing Outposts is just planting an incident source.
- **You have no 5G private network or carrier relationship.** Wavelength access depends on the carrier; if this does not hold, the solution does not hold.

## FAQ

**Q: Can Wavelength and Outposts be used at the same time?**
Technically they are both VPC extensions under the same account, but one is in a carrier facility and the other is in your own data center, so their physical locations are too far apart. In real projects they almost never appear at the same time; evaluate them separately.

**Q: Can RDS and ElastiCache run directly in a Wavelength Zone?**
The services you can select in that zone are far fewer than in the parent Region, and stateful managed services are basically not available. What lands on Wavelength is generally the stateless compute layer, while the database remains in the parent Region, so the RTT of cross-zone calls must be counted separately in the total latency.

**Q: How should EBS capacity on Outposts be planned?**
Local disk capacity is limited and cannot be shared across Outposts. Plan based on the usable capacity of a single unit and leave 30% headroom. Sending images and snapshots back to the parent Region also consumes bandwidth.

## References

- [AWS Wavelength Developer Guide — What is AWS Wavelength](https://docs.aws.amazon.com/wavelength/latest/developerguide/what-is-wavelength.html)
- [AWS Outposts User Guide — What is AWS Outposts](https://docs.aws.amazon.com/outposts/latest/userguide/what-is-outposts.html)
- [AWS Local Zones User Guide — What is AWS Local Zones](https://docs.aws.amazon.com/local-zones/latest/ug/what-is-aws-local-zones.html)
- [AWS Wavelength product page](https://aws.amazon.com/wavelength/)

Before choosing a solution, do one thing first: take a real device and run `mtr` once, measuring to the Region, to the same metro, and to the 5

Leave a Reply

Your email address will not be published. Required fields are marked *