Open source AI is not always the expensive trap that cost-focused critics describe. In the right conditions it saves Australian teams real money, and in the wrong conditions it quietly doubles their costs. The skill is in telling those conditions apart before you commit, so you self-host when the maths works and choose a managed model like Claude when it does not.
This post sets out the conditions that favour open source, the conditions that punish it, and the rough numbers an Australian business should check before building a detailed business case either way.
The conditions that favour open source
A few factors push the economics firmly toward running your own model. None of them are about which model tops a benchmark this month. All of them are about your workload and your team.
Very high and steady request volume that would run up large API bills month after month
A narrow, stable task you can tune once and then mostly leave alone
An existing engineering team with real spare capacity to own the infrastructure
A workload that keeps your GPUs busy through the day rather than idling between bursts
When most of these hold at once, a self-hosted open model can undercut a managed API by a wide margin. A classification pipeline that processes millions of documents a day on a fixed schedule is the classic case. The volume is predictable, the task barely changes, and the hardware never sits idle.
The conditions that work against it
When the following appear instead, open source usually costs more than the managed alternative once everything is counted.
Spiky or low traffic that leaves expensive GPUs idle most of the day
Sensitive or regulated data that demands serious compliance and audit work under the Privacy Act
A small team already stretched across other priorities, with no one to carry the pager
Frequent changes to the task or the model, forcing repeated rounds of testing and retuning
The pattern is consistent. Open source rewards stability and scale. It punishes variability, thin teams, and workloads that never keep the hardware busy.
Putting real numbers on it
Start with the hardware. A single production GPU node hosted in Australia runs around $40,000 a year before anyone writes a line of application code. That figure covers the node itself, not the engineering time to provision it, secure it, monitor it, and patch it.
Then add the people. An engineer who can run inference infrastructure well costs $160,000 or more a year in Sydney or Melbourne, and even a part-time slice of that person is a real cost. If self-hosting consumes a quarter of their year, that is another $40,000 on the bill.
Against that, price the managed path at your actual volume. A team spending $50,000 a year on API calls for a well-matched, high-volume workload might cut that figure by half with a tuned open model. A team spending $12,000 a year on spiky, low-volume calls cannot save anything by self-hosting, because the floor cost of the hardware alone is several times their entire bill.
Profile your real usage over at least a month before deciding
Cost the hardware and the people together, never the hardware alone
Recheck the maths as your volume grows, because the answer can flip
A worked example
Consider an Australian logistics business processing 80,000 delivery documents a day through an extraction model. The task is narrow, the volume is steady, and the team already runs its own infrastructure. Their API bill at that volume would clear $90,000 a year. A self-hosted open model on two GPU nodes, plus a slice of an engineer they already employ, lands near $55,000. Open source saves them roughly $35,000 a year, and the saving grows with volume.
Now consider a 20-person professional services firm using AI for drafting, summarising, and client research. Usage is spiky, the tasks change weekly, and there is no infrastructure team. Their managed bill is about $14,000 a year. The cheapest credible self-hosted setup would cost them more than triple that before a single document is processed. The same technology, applied to two different businesses, produces opposite answers.
A rough screen before you build the business case
A quick screen tells you whether open source is even worth pricing in detail. It takes an afternoon, not a quarter.
Steady high volume plus spare engineers usually favours self-hosting
Spiky or low volume, sensitive data, or a thin team usually favours managed
Anything in between deserves a proper side-by-side costing on your real tasks
The screen saves time because it stops you building a detailed case for a path the basic shape of your usage already rules out. Most Australian SMBs fall into the second line, which is why a managed model is the sensible default for most of them.
Where Claude fits
Our default recommendation for Australian SMBs is Claude, because for most workloads it delivers the lowest total cost once people, reliability, and compliance are counted. But the costing comes first, and where the numbers point to open source we say so plainly. The goal is the cheapest path to the outcome you need, not loyalty to either camp.
If you want the maths done on your actual usage, we run side-by-side costings for Australian teams and report back in plain English. Book a costing review and bring your last three months of usage data.
