evidoria

← Back to browse

Good practice Imported

USPS's Edge Compute Infrastructure Program — AI-Powered Package Scanning Across 195 Mail-Processing Centers

United States of America · Washington, D.C. · See the United States of America profile · See the Washington, D.C. profile

Evidence: Descriptive / self-reported Top 66% 53/100 · Ask Evidence Copilot about this practice

USPS equipped 195 mail-processing centers with Nvidia GPU edge servers that scan barcodes and addresses on packages, cutting missing-parcel investigations from days to hours. USPS says the tool flags anomalies for troubleshooting only, without a published formal error rate.

195 centers
Mail-processing centers equipped with edge AI servers
1,000+ machines
Sorting machines feeding camera data
20 TB/site/day
Package images processed
129 billion pieces/year
Annual mail volume
7.3 billion packages/year
Annual package volume
231 packages/second
Package processing rate
~800 CPUs / 2 weeks reduced to 4 GPUs / ~20 minutes
Compute task time (self-reported)
8-10 staff over several days reduced to 1-2 staff over a few hours
Lost-package tracing time (self-reported)
USPS's Edge Compute Infrastructure Program — AI-Powered Package Scanning Across 195 Mail-Processing Centers

Details

Maturity
Established
Promoter
United States Postal Service (USPS)
Period
2019-2026
Keywords
postal services, logistics, public administration, e-commerce

Context

USPS operates at massive scale (about 129 billion mail pieces and 7.3 billion packages a year, roughly 231 packages per second) and in 2019 launched the Edge Compute Infrastructure Program (ECIP), installing HPE edge servers with four Nvidia V100 GPUs inside 195 mail-processing centers to keep pace with volume.

Objectives

The program aims to speed up computer-vision processing of package images — reading barcodes and addresses, verifying postage, and flagging hazardous-materials symbols or damaged labels — directly at the processing-center edge rather than in a central data center.

Activities

Cameras on more than 1,000 sorting machines feed roughly 20 terabytes of package images per site per day into the edge servers, which run computer-vision models; a compute task that once needed about 800 CPUs and two weeks now runs on four GPUs in about 20 minutes.

Results

USPS staff report that tracing a lost or misrouted package, once taking eight to ten staff several days, now takes one or two people a few hours; however, USPS itself has said the system does not produce a formal "error rate," describing its main use as flagging anomalies for troubleshooting rather than a measured accuracy benchmark.

Conclusions

The USPS Office of Inspector General's 2026 white paper, which catalogued more than 35 AI use cases across the agency, flagged this lack of a published, independently verified performance metric as a transparency gap, even though the program has been replicated identically across all 195 sites.

Implementation
Indicative cost
High (€500k–€5M)
Time to results
Long (> 3 years)
Staffing & skills
Package-tracing investigations shifted from 8-10 staff over several days to 1-2 staff over a few hours (self-reported by USPS staff)

Conditions for success

  • Camera-equipped sorting machines (1,000+) generating package images at each site
  • Edge GPU servers (HPE, 4x Nvidia V100) co-located at each processing center to handle ~20 TB/site/day
  • High package throughput (~231 packages/second) that justifies edge rather than centralized compute

Common failure modes

  • USPS itself states the system does not produce a formal 'error rate', limiting independent verification of accuracy
  • USPS Office of Inspector General's 2026 white paper flagged this as a transparency gap across the agency's 35+ AI use cases

Where it fits

Governance type
federal government agency (quasi-independent, self-funded)
Scale
national, 195 mail-processing centers
Income level
high-income

Commonly funded by

National / regional programmes Own resources / municipal budget

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful