The word open gets attached to almost every model release in 2026. For an Australian business planning a budget, the implied word free is doing a lot of quiet damage. Open weights and zero cost are not the same thing, and the difference shows up the moment a model moves from a laptop demo into production.
This post walks through what the free label actually hides: the infrastructure bill, the people bill, and the compliance bill. None of them are reasons to avoid open source. All of them belong in your plan before you commit.
What open weights actually give you
When DeepSeek, Qwen, or Llama publish open weights, you get the trained model file and a licence that lets you run it. That is genuinely valuable. What you do not get is anywhere to run it, anyone to keep it running, or any guarantee it behaves safely with your customer data.
Downloading the weights for DeepSeek V4 or Qwen 3.5 costs nothing. Running them so they reliably serve your business is a different story. The bill simply changes shape, from a licence fee into infrastructure, people, and risk.
Cloud GPU rental, often $3 to $12 per hour for each instance you keep running
An engineer who genuinely understands inference, scaling, and security
Monitoring, logging, and incident response for when a node falls over mid-task
Compliance work to satisfy the Privacy Act for any personal data the model touches
None of these appear on the download page. All of them appear on your invoices.
Where the money actually goes
A modest self-hosted setup for a Sydney SMB can reach $60,000 a year once you add compute, on-call cover, and tooling. That figure arrives before the model produces a single dollar of value for the business. The economics get worse, not better, when usage is uneven.
Idle GPUs still cost money overnight, on weekends, and over the holidays
Bigger context windows demand more memory and push the spend higher
Each model upgrade can mean fresh testing and integration work
Security patching is a recurring task, not a one-off
Compare that with a usage-based API. A team running 50,000 requests a month through a managed model might spend $800 to $2,500 a month depending on context size and model tier, with no idle cost, no patching, and no on-call roster. At low and medium volume the managed path usually wins on raw dollars, not just convenience.
The people cost is the one that hurts
Hardware prices show up in calculators. Salaries do not, and they are the bigger line. A machine learning engineer in Sydney or Melbourne commands $150,000 to $220,000 a year, and self-hosting reliably needs at least a meaningful slice of one. If your open source plan assumes the existing dev team will absorb inference infrastructure on top of their day jobs, the plan has already failed; it just has not told you yet.
There is also a continuity risk. If the one engineer who understands your inference stack resigns, your AI capability walks out the door with them. A managed provider spreads that risk across an entire company.
Compliance is part of the bill
Australian businesses handling personal information carry Privacy Act obligations regardless of where the model runs. Self-hosting moves the entire compliance surface onto your team: access controls, audit logging, breach response, and data retention all become your build. With a managed provider you still own your obligations, but you inherit mature controls instead of writing them from scratch. For APRA-regulated firms the bar is higher again, and the documentation burden of a self-built stack is real money in consultant hours.
The licence small print
Open also hides licence variety. Apache 2.0 and MIT weights are usable in commercial products with few strings. Llama-style community licences add conditions on user counts and acceptable use. Some releases restrict commercial use entirely or require separate agreements. Before any build starts, someone needs to read the actual licence against your actual use case. Twenty minutes of reading can prevent a rebuild later, and a short legal review at $400 an hour is far cheaper than discovering a restriction after launch.
A simple test for your own numbers
The fastest way to cut through the free label is to run your real figures, not the vendor's.
Take your expected monthly request volume and average context size
Price it against a usage-based API and against a fixed GPU node running all month
Add the realistic staff time each option needs, costed at real salaries
Add a line for compliance work if any personal data is involved
Most Australian SMBs find the fixed-cost path only wins at volumes they do not have. Until then, free open weights cost more than a managed model that bills for what you use.
When self-hosting genuinely wins
To be fair to the open source path, there are cases where it earns its keep: data that contractually cannot leave your infrastructure, sustained high volume that keeps GPUs busy around the clock, or a product where the model is the product and unit economics demand it. If you are in one of those positions, self-host with confidence and budget honestly. If you are not, the maths rarely gets there.
We size these decisions for Australian SMBs so the number on the plan matches the number on the invoice. We cost both paths in plain figures and have no stake in which one you pick. Book a brainstorm session and bring your volumes.
