A SpaceX compute deal is funding new open-source AI models. When open-source AI saves money for Australian SMEs, and when a commercial API still wins.
In October 2025 SpaceX agreed a compute deal worth roughly A$6.3 billion with Reflection AI, securing access to Nvidia’s GB300 systems to train advanced open-source models through to 2029. That is a number no small business will ever spend, and no SME is going to buy a compute cluster. The reason it matters is what happens downstream. Deals of this size exist to produce capable open-weight models, and those models get published for anyone to download and run.
For a cost-conscious Australian SME, that is the interesting part. The frontier-grade capability that used to sit behind an expensive subscription is increasingly available as a model you can host yourself or run cheaply. This piece looks at what open-source AI actually gives a smaller business, where it falls short, and how to decide between open-source and a commercial product without getting caught up in the spending headlines.
What open-source AI gives a smaller business
When people say open-source AI they usually mean open-weight. The model’s weights are published so you can download it, run it on your own hardware and adapt it, even if the full training recipe is not released. For an SME that opens up three things that a rented commercial API does not.
The first is cost control at volume. Commercial APIs charge per call, which is fine when usage is light but adds up quickly once a tool is embedded in daily work for a whole team. A downloaded model has no per-call fee. Once it is running, the marginal cost of another query is close to nothing, so a high and steady workload can be much cheaper to serve yourself.
The second is control. You are not exposed to a vendor changing prices, deprecating the model you built on, or shifting its usage rules. The model you downloaded keeps working the way it did the day you tested it, which matters when you have wired it into a real process.
The third is privacy. Running a model in your own environment means sensitive data never leaves the business. For a firm handling client records, health information or commercial-in-confidence material, that is often the deciding factor, and it sits more comfortably with your obligations under the Privacy Act 1988 than sending the same data to a third party offshore.
The catch
None of that is free, and the honest version of this story includes the bill that the word “open” tends to hide.
Hosting is the first cost. A capable model needs real hardware, usually a GPU server or a cloud GPU instance, and that runs whether you are using it heavily or not. For low or spiky usage you can end up paying for idle capacity, which is exactly the situation where a pay-as-you-go commercial API is cheaper.
Maintenance is the second. A self-hosted model is now your software to run. Updates, security patches, monitoring and the occasional 2am fix are yours, and they need a person who knows how. Commercial providers absorb all of that into the subscription, which is part of what you are paying for.
Governance is the third, and it does not disappear because you own the model. You still need access controls, logging, and someone accountable for what the model produces, especially for anything customer-facing or regulated. An unmaintained self-hosted model is a liability, not a saving.
And sometimes the commercial API simply wins. For the hardest reasoning tasks the leading proprietary models still tend to be ahead. For a small team with no infrastructure and modest, irregular usage, the speed of signing up and paying per call beats the effort of standing up and running your own. Open-source is a strong option, not a default answer.
How an Australian SME should decide
The decision is easier when you treat it as a workload question rather than a values question.
Start by sizing the usage. If a tool will run constantly across a team, model the monthly API cost at that volume and compare it to the cost of hosting a model plus the time to run it. High, steady volume tends to favour self-hosting. Low or unpredictable volume tends to favour a commercial API.
Then check the data. If the work involves sensitive or regulated information that should not leave your environment, that pushes hard towards a self-hosted open-source model regardless of the raw cost comparison.
Then be honest about capability and skills. Test both options on your actual task, not a demo, and see whether an open model is good enough. Ask who in the business will keep a self-hosted model patched and running, and if the answer is nobody, either budget for that help or stay with a managed option.
A practical middle path suits a lot of SMEs. Use a commercial API to prove the workflow quickly, and once the usage and value are clear, move the steady, high-volume or privacy-sensitive parts to a self-hosted open model. You are not obliged to pick one for everything.
If you want help mapping this to your own setup, our artificial intelligence work covers building and integrating these models, while tailored solutions covers fitting them to a specific workflow rather than buying off the shelf. You can also see the technologies we work with to understand what self-hosting an open model involves in practice.
What this means for you
The spending behind the SpaceX and Reflection AI deal is not something an SME competes with. It is something an SME benefits from, because it keeps producing capable open-weight models that anyone can run. The result is genuine choice. Enterprise-grade AI you can host cheaply, with your data kept in-house, alongside commercial tools that are faster to start and still lead on the hardest work.
The businesses that come out ahead will not be the ones that pick a side on principle. They will be the ones that match the tool to the workload, count the full cost of hosting and maintenance honestly, and keep a person accountable for what the model produces. Decide on volume, data and skills, and the open-source question answers itself.
