By Brad Ferris · 5 September 2026
On 31 August, iTnews carried the Reuters report that OpenAI will stop supplying its models to Cursor from 12 November. The trigger was not performance, price, or a breach of contract. SpaceX had acquired Cursor's parent company Anysphere for US$60 billion, and OpenAI said it could not trust the new owner to honour its terms of service. Anthropic responded inside a few hours by offering extra compute to support Claude models inside the same product.
Most Australian operators have never used Cursor and never will. It is a coding tool. The reason it is worth your attention is what it demonstrates about the arrangement underneath it. Cursor is one of the most commercially successful AI products built to date, with the resources and the relationships to match, and it lost a major model supplier overnight because of a decision made in someone else's boardroom. If the terms can move for a customer that size, they can move for a business with forty staff in Rocklea.
The Cursor decision did not happen on its own. Two other moves in the same window point the same direction.
Nvidia agreed to buy Hugging Face for about US$13 billion, its largest acquisition, and pointedly not a semiconductor deal. Hugging Face is where open-weight models are distributed. It is the thing people are gesturing at when they say "we can always fall back to open source if a vendor gets difficult". Nvidia has said it will keep the platform open, and there is no reason yet to doubt that. The point is narrower: the fallback now has a strategic owner with its own hardware agenda, which makes it a supplier rather than a commons.
Then Atlassian told customers it will start charging for AI usage above a fixed Rovo credit allowance from 3 December, following a jump in adoption. That is the third move, and for most businesses it is the one that lands first. The period where AI capability arrived bundled inside a seat licence is closing. What was a fixed cost becomes a variable one, and variable costs need somebody watching them.
Withdrawal, ownership, and repricing. None of the three involved a customer doing anything wrong.
Most mid-market businesses in Australia are past the experiment stage now. There is a quoting workflow, or claims triage, or first-line support, or a drafting step in the proposal process, and it has been running for months. It works. Nobody thinks about it, which is the sign that it has become infrastructure.
That workflow was almost certainly written against one provider. Not as a decision, but as a default: someone picked what was in front of them, it worked, and the next four workflows were built the same way. The result is a supplier concentration position that has never been examined as one, because it arrived as a series of technology choices rather than a procurement decision.
You would not run your business with a single freight provider, no alternative quoted, no idea what the switching cost was. Same discipline, different category.
Which workflows stop if the supply stops? Write the list down. When Telstra built an internal control tool to bring discipline to its AI use, it identified 380 internal use cases. A business a hundredth that size will still find more than the two or three the leadership team can name from memory. You cannot assess an exposure you have not counted.
Where does the work actually live? The value in a mature AI workflow is rarely the model. It is the prompts, the context you have assembled, the examples of good and bad output, the rules the business argued about for a fortnight. If that material sits inside a vendor's product with no export, you are renting your own operating knowledge. If it sits in your systems in a portable form, changing suppliers becomes an afternoon rather than a project.
What is your switching time, and has anyone tested it? "We could move to another provider" is an assumption until somebody runs the workflow on a second model and compares the output. That is a half-day exercise for most single workflows. Do it on the one that matters most, and you will have a real number instead of a comfortable belief.
What do you pay per unit of work, not per seat? Metered pricing rewards businesses that know their cost per quote, per ticket, per document. It punishes the ones that find out at renewal. Work out the unit now, while the allowance is still generous.
Anthropic moved on Cursor within hours of the cut-off news. That is what a competitive supply market looks like, and it is good news for buyers. A business that can move between providers gets to use that competition, on price, on capability, and on contract terms including data handling. A business that cannot move gets whatever terms it is offered.
The lesson from the last fortnight is not that AI suppliers are unreliable. It is that they are suppliers. Treat the relationship the way you already treat every other input your business depends on, and the next time the terms change, it will be an inconvenience rather than an incident.
Where does your business stand? The free AI Scorecard takes three minutes and shows you. If you want a straight steer from a person, book an AI Opportunity Call.