Answering “anything new?” cost more than the news.

The bill averages $231,133 a month and, on the current design, reaches $331,638 in two years. We identified $114,731 a month (50% savings totalling $1,405,695 over the next 12 months). Three quick wins take $69,637 off with no design change. Three investments take $45,094 off today, growing to $76,324 in year two.

Already going well, no change recommended:

• Compute Savings Plan. The steady compute floor is already covered by a 1-year plan, sized at what the investments leave running.
• Non-production. Staging runs at a fraction of production, dev scales down on its own, and the partner sandbox stays up because partners test around the clock.

How we will know it worked: cloud spend per shipment falls from 0.61¢ to 0.31¢.

Page 1 of 10

Proposed roadmap

Three decisions sit outside engineering.

Page 2 of 10

The levers, priced per unit

Three levers lower what the carrier pays per unit, worth $69,637 a month with no design change. Each is configuration or a schedule.

QW-1. “Out for delivery” text

A text is billed in segments. Plain GSM-7 text fits 160 characters in one segment. A single character outside that set, such as a curly apostrophe, switches the whole message to UCS-2, which fits 70 characters, or 67 a segment once a message splits. The “out for delivery” message is 103 characters, so it bills as two.

The fix is a straight apostrophe, plus a template check that blocks any character outside GSM-7 so it cannot come back ($28,845).

Per “out for delivery” text: about 2.4¢ (2 segments) to about 1.2¢ (1 segment) · $28,845 a month

QW-2. Partner-poll logs

Every successful partner poll writes seven log lines, each repeating about 1.2 KB of request context. Four of them are old debugging lines, left on from the partner migration and the encryption rollout, that no team or system uses today:

{"ts":"2026-08-14T13:02:11.484Z","level":"INFO","svc":"tracking-api","req":"7f3c9a1e","partner":"p-0412","msg":"Entering legacy partner compatibility handler", ...}

{"ts":"2026-08-14T13:02:11.485Z","level":"INFO","svc":"tracking-api","req":"7f3c9a1e","partner":"p-0412","msg":"Skipped response cache lookup: cache disabled", ...}

{"ts":"2026-08-14T13:02:11.486Z","level":"INFO","svc":"tracking-api","req":"7f3c9a1e","partner":"p-0412","msg":"Starting encryption of tracking response payload", ...}

{"ts":"2026-08-14T13:02:11.487Z","level":"INFO","svc":"tracking-api","req":"7f3c9a1e","partner":"p-0412","msg":"Finished encryption of tracking response payload", ...}

The four lines are about 4.9 KB of every poll (4,894 bytes), about 41,403 GB a month. Removing them takes that off the log ingest ($19,280) and the endpoint the logs pass through ($385): $19,665 a month.

Log ingest per million partner polls: about $3.93 to about $1.67 · $19,665 a month

QW-3. Delivery photo storage

The tracking page shows the photo for 14 days, the delivery day and 13 more; after that, it loads only when a customer asks for it. Legal confirmed the carrier only has to produce a delivery photo when a customer asks for one. The site’s access logs show how rarely that happens late: after the 13th day from delivery, photos account for 0.0003% of all photo loads.

So photos stay in S3 Standard for 13 days, then move to Glacier Instant Retrieval, and the API keeps serving them from there in milliseconds.

Keeping the API on Instant Retrieval is an intentional choice. Storage drops to about a sixth: $0.004 a GB-month in Instant Retrieval against $0.023 in S3 Standard. The catch is a retrieval surcharge every time an older photo loads, about $0.0001 a photo ($0.03 a GB plus $0.01 per 1,000 requests), so a user who reloads one again and again adds cost. That risk is small next to the savings, and next to support handling every retrieval, so it is a worthwhile trade. We recommend an alarm on Glacier Instant Retrieval retrievals, so a run of reloads is caught early.

Per delivery photo, stored for a year: about 0.074¢ to about 0.019¢ · $21,127 a month

Page 4 of 10

Overview

Left alone, the bill reaches $331,638 a month in two years; with every recommendation in place it is $152,911. That is $151,318 a month less in year one and $178,727 less in year two, against $114,731 today: the saving grows faster than the business, 1.56 times against 1.44 times more shipments, because more partners move to the push feed each year.

Proposed architecture updates

Page 6 of 10

I-1. Resize photos at capture · $5,199 a month

What changes: a driver app update resizes each delivery photo on the handheld before upload, from about 3.4 MB to about 0.33 MB, still clear enough to settle a dispute.

Why do it: smaller photos cost less to store and download: $5,199 in the first month, rising to $9,921 a month a year in as the full-size photos age out. The resize runs on the handheld, so it adds no cloud work, even at the holiday peak.

Risks: a photo too small to settle a dispute, or a faulty app build. The team tests sizes on real deliveries and starts with one depot before every handheld.

Business impact: a per-unit cut. Every delivery’s photo costs less to keep, so the saving grows with volume, and customers see the same doorstep.

Page 7 of 10

I-2. Tracking table, one year of orders · $21,458 a month

What changes: a DynamoDB table answers tracking polls. It holds every shipment in transit and those delivered in the last 13 days, kept current by the scan events. The primary keeps one year of orders and archives older months to S3, so it steps down from a db.r6i.24xlarge to a db.r6i.16xlarge, moves to gp3 storage, and is then reserved.

Why do it: polls stop landing on the database that takes orders. Its IOPS fall from 32,769 to 7,958 and its storage from 7,893 GB to 2,237 GB. The table scales with the requests on its own, even at the holiday peak.

Risks: in the minutes after a failover, finding an order by tracking number is slower until memory warms, a risk the plan accepts. The table runs beside the database for a week before partners move over one at a time, and the resize lands in one blue/green switchover, with a db.r6i.24xlarge as the fallback.

Business impact: scale. Partner polling no longer grows the database, a partner burst can’t slow orders and scans, and every partner benefits with no change on their side.

Page 8 of 10

I-3. Push feed, three largest partners · $18,437 a month

What changes: the carrier pushes every tracking update to its three largest partners through EventBridge instead of answering their polls. One reconciliation poll per package stays.

Why do it: these partners make 42% of all tracking queries. The feed removes about 3.58 billion polls a month, saving $8,519 of API Gateway, $6,271 of logs, $2,148 of WAF and $1,499 of everything else. Partners get each update as it happens and can retire their pollers.

Risks: it needs the partners’ buy-in and engineering time, so dates follow their roadmaps. If the feed fails, deliveries retry for 24 hours and partners can still poll.

Business impact: a per-unit cut. Each package those partners track takes 8 pushes instead of 253 polls, and the saving grows as more partners follow, reaching about 51% of tracking queries by year two.

Page 9 of 10