"Move all agents onto a single host to save infrastructure cost" — the numbers

Short answer: consolidating agents saves ~A$0/month, because the agents are not what costs money. Measured this morning, every figure below probed, not estimated.

1. The three hosts, and which ones bill

Two of three hosts cost nothing to run. They are owned hardware — moving agents off them saves nothing, because there is no bill to reduce.

2. The agents themselves cost almost nothing

Both hosts are idle. The agents use ~16 MB each. Consolidating them onto one host relocates a rounding error.

3. Why prod cannot be retired by moving agents

prod's real job is the public websites, not the agents:

solos-docs.service  → docs.youngle.au   (8091)  active
pulse-site.service  → pulse.youngle.au  (8092)  active
v3-heart-prod       → 16.8 MB, the only agent there

Move every agent off prod and prod still has to run, because it serves the two public sites. The infrastructure cost does not change by one cent.

4. What consolidation would actually cost us

5. Where the money actually is (my lane — the real lever)

The infrastructure is not the cost centre. Measured:

The saving worth chasing is the $11.37 on qwen3.8-flash, not a 16 MB agent process. Free models did 8.8× the tokens at $0.00 in the same period — that is a real number with a real lever, and it is already my lane's work.

6. What I would recommend instead

1. Leave the topology as-is. It is already the cheap shape: owned hardware for agents, one billed node for public services.
2. Close the A2A gap (Sol/Tam, in flight) — that is the actual blocker Young named, and it is a wiring defect, not a cost problem.
3. Confirm the Oracle burn (Young's console check) — that is the only infra number that can move, and it is still open.
4. Attack model spend — the free-first ladder and cache ratio (91.8% today) are where the money is.

Honest limits

• I did not measure the Pulse laptop's specs (it is not reachable for a shell from here — ping only). Its cost is $0 as owned hardware either way.
• Oracle billing remains unconfirmed — my two-row infra table still collapses to one row when Young's number lands.
• "Save infrastructure cost" may mean something I have not measured (e.g. retiring prod entirely and hosting the sites elsewhere). If so, that is a different question and I would need to price the alternative.