48. Spend analysis and forecasting
Ten minutes every Monday on one page catches the surprises the agents miss.
You now have the pieces — cloud spend, Daalu’s subscription, and inference. This chapter is about understanding the total, catching the surprises early, and forecasting well enough to plan.
The one page that matters
Usage & Pricing (Chapter 27). Spend time on this page once a week. Almost everything below is just a structured way to read it.
A weekly habit
Ten minutes every Monday. Three questions:
- Are we on track? Compare this month’s run-rate against last month’s at the same date.
- Anything weird? Scan the per-service breakdown for any line that’s more than 2× last month.
- Are we forecasting OK? Glance at the projected month-end.
Tip. The cost-anomaly agent (Chapter 20) flags the obvious bumps in real time. The weekly skim catches the slow drifts its thresholds are tuned to ignore. You want both.
Per-account drill
When something looks off, drill into the cloud account responsible:
- Cloud spend → [account] → Top services — which AWS / GCP / Azure service spiked.
- Top regions — sometimes a developer spun something up in the wrong region.
- Top tags — which team or project owns it.
The Assistant can do the legwork: “Show me the spend deltas for AWS account X between this week and last week, grouped by service.”
Forecasting
The page’s projected month-end is a straight-line extrapolation from month-to-date. That’s right for stable spend and wrong for:
- Newly-deployed services. Their first partial month under-counts; next month is higher.
- Seasonal patterns. Holiday weeks, end-of-quarter, batch cycles.
- One-off events. A backup job that ran heavily once.
For tighter forecasting, build a saved query on Reports → Query (Chapter 23) that aggregates by week-of-month or day-of-week, then layer in your own adjustments.
Tag-based attribution
If your cloud resources are tagged consistently with team / project, Daalu surfaces tag-based breakdowns directly. If they aren’t, untagged spend shows up as “untagged” — and the real fix is the tagging discipline upstream.
Two patterns work:
- Mandatory tagging at create-time — enforced by IaC (Terraform policy, AWS Config rules). Best.
- Periodic untag sweep — an automation in Daalu that lists untagged resources and opens change proposals to fix them.
Either way, aim to get untagged spend below 5%.
When the cost-anomaly agent fires
Thresholds live in Settings → Briefings → Cost anomaly. Defaults:
- Daily total — alert if today’s spend exceeds 1.5× the trailing-7-day average.
- Per-service — alert if a single service exceeds 2× its trailing average.
When it fires, the alert payload includes the responsible service / region / account, the size of the anomaly in dollars, and the Assistant’s inferred cause when it’s obvious (e.g. a new resource appeared). Click through to the billing page filtered to that slice to investigate.
Inference spend specifically
Inference is metered separately and shown under Usage & Pricing → Inference (Chapter 47). Patterns worth watching:
- Commercial usage growing faster than seat count. Usually your own GPU is offline more than you think — calls are spilling to the Daalu-hosted and external tiers. Check the card’s health on the AI Factory page.
- A sudden jump in long-context quality calls. Usually an automation you authored is summarizing too much. Inspect its run history, and check the by-source breakdown to confirm which one.
- An embedding/index spike. Usually a re-index job. Fine if expected.
Why it matters. The single biggest lever on your inference bill is the share of calls served by your own GPU. If that share is dropping, the bill is rising — and it’s almost always a health problem on the card, not a usage problem in your team.
Cost-saving recommendations
The recommendations agent surfaces savings opportunities on Operations → Recommendations:
- Idle compute — EC2 / GCE instances under 5% CPU for 14 days.
- Unattached disks — EBS / persistent disks not connected to any instance.
- Old snapshots — RDS / disk snapshots past your retention policy.
- Oversized databases — instance classes far larger than utilization warrants.
- Commitment savings — when steady-state spend crosses the threshold where reserved / committed-use pricing pays off.
Each shows an estimated dollar saving. Click to open a change proposal that applies it, or dismiss with a reason.
Forecasting beyond Daalu
Daalu’s forecasting is built for the Monday glance, not enterprise FinOps. For real financial planning:
- Export the CSV from Cloud spend → [account] → Export.
- Pipe it into your finance team’s spreadsheet, FinOps tool, or warehouse.
- Model it properly there.
We don’t try to replace Apptio, CloudHealth, or Vantage. The goal is operational awareness, not enterprise cost analytics.