# Example memo based on synthetic data

> Northwind Analytics is a fictional company, and every figure below comes
> from the synthetic FOCUS dataset in this folder (`results.json`). The memo
> shows what an analyst might write. It is not advice about any real bill.

**To:** Finance lead, Northwind Analytics
**From:** Cloud cost analyst
**Re:** Cloud cost review, 2026 Q3 (July to September)

## Headline

In Q3, BilledCost totalled $332,678.83 while EffectiveCost, the cost we
actually consumed, was $213,887.47 (Q1). In this file Google's BilledCost is
shown before credits, so the actual invoices are lower by Google's x_Credits.
The gap between the two totals comes mainly from commitments:

- We prepaid the AWS Savings Plan: $149,650.00 billed in July. Only
  $37,720.00 of it has been recognized so far, which is $410.00 a day for 92
  days, leaving $111,930.00 for future quarters (Q2b).
- Budget and unit-cost reporting should use EffectiveCost. Cash planning
  should start from BilledCost, net of any provider credits held in separate
  columns.

## Finding 1: a fifth of the AWS commitment is now unused

The Savings Plan was fully used until 2026-08-14 (Q2c: $18,450.00 recognized,
$0.00 unused). On 2026-08-15 part of the EC2 fleet it covers was retired.
Since then:

- $3,917.00 of the $19,270.00 recognized was unused: 20.3%, or $83.34 a day
  (Q2c).
- The unused amount was $1,439.92 in August and $2,477.08 in September (Q2a).

The commitment cost does not go down when the fleet does. We pay for it
whether or not we use it.

## Finding 2: a runaway backfill job in the data-lake account

The anomaly check flagged AWS sub-account 100000000102 on all five days from
2026-09-14 to 2026-09-18 (Q4a, z-scores 18.8, 4.0, 3.6, 3.6 and 3.6).
Breaking those days down by the `team` tag puts almost all of the increase on
`ml-research` (Q4b):

- The cost came from S3 replication tagged `job: embedding-backfill`.
- It grew from $407.58 on Sep 14 to $3,317.14 on Sep 18, against a prior
  7-day average of $8.59 a day before it started.
- The `analytics` and `data-eng` teams stayed within $11.00 a day of their
  own 7-day averages.

## Finding 3: unit cost

Effective cost per 1,000 orders was $54.84 in July, $41.16 in August and
$44.61 in September (Q5). Two adjustments are needed to read these:

- **Tax and one-off items.** July includes tax on the upfront payment, which
  FOCUS does not amortize. August includes a $2,500.00 credit. On Usage rows
  only, the figures were $45.84, $40.82 and $42.61 (Q5). August improved as
  orders grew. September rose again, in the same month as the backfill spike.
- **Billed cost.** On a BilledCost basis, July shows $154.51 per 1,000 orders
  against $34.50 in August (Q5). Billed cost is the wrong basis for unit
  economics.

## Recommendations

1. **Put the unused commitment back to work before buying more.** Before we
   consider any new commitment, steer new or migrating compute that is
   eligible for the Savings Plan onto it.
   - Caveat: this data cannot tell us whether that much eligible usage
     exists. A Savings Plan also cannot be cancelled, so some of the unused
     $83.34 a day may simply be the cost of the downsizing decision.
2. **Confirm the backfill with ml-research and add guardrails.** Ask the
   ml-research team whether the embedding-backfill job has finished. Then set
   a budget alert on sub-account 100000000102 and require a `job` tag on
   batch workloads.
   - Caveat: our data ends on 2026-09-30, so we cannot see whether the spend
     stopped after Sep 18. The z-score method is also less reliable once a
     spike sits inside its own 7-day baseline. It flagged all five days here
     only because the cost kept climbing.