It usually shows up as an argument about jobs, or local suppliers, or what gets left behind.

The government wrote its version down in March, in a set of expectations for data centre and AI infrastructure developers. Meeting them, it said, "will be the foundation of their social licence to operate in Australia".

The fifth expectation is the one about us. Providers of large-scale compute are expected to enable "access to compute for Australian start-ups, innovative small businesses, researchers and not-for-profits on favourable terms".

That last phrase is doing the work, and there's no number in it. It never says which tenant, for how many hours, over what term, or at what price.

But in the Financial Review, writing about the AI standards still to come, Jim Chalmers said they "will build on the work we've already done to secure favourable terms for Australian AI start-ups". Thirteen days earlier, at the ANU, the assistant minister Andrew Charlton described the same thing as unfinished.

Charlton has been making the case since May, and he is the clearest voice in the government on it. Here he is at the ANU on 18 August:

"If we provide the land, the power and the buildings while foreign firms own the chips, the models and the customer relationship, we will capture some of the wealth while the largest rents accrue somewhere else."

Also from the same lecture: "We can build the ecosystem around the infrastructure, deliberately, by making it a condition of the welcome. Minister Ayres and I are working on how we can make that expectation a requirement of hyperscalers, to support the Prime Minister's National AI Standards."

What favourable terms cost in the credit market

Last week IREN borrowed US$2.4 billion to buy Nvidia's newest chips for a campus at Mackenzie in British Columbia, at 9 per cent fixed over about two and a half years. Blue Owl led it, with PIMCO alongside. The lenders are covering roughly 90 per cent of the chip cost, with no investment-grade customer behind them.

IREN was founded in Australia. That's about all the Australian content in the deal.

The same company finances the same generation of Nvidia hardware at about 6 per cent when Microsoft is the customer standing behind it. CoreWeave writes that rule into its own loan documents, where advance rates are set "based upon the credit rating of the applicable customer".

A credit committee reading the fifth expectation sees an operator asked to hold capacity for start-ups, researchers and not-for-profits, who are the weakest credit in the building. The operators do not appear to be the ones refusing.

The condition is worth having. Writing the rules at the start of a build rather than a decade into it is a genuinely good instinct, and March was early to be doing it.

But until somebody writes a number into it, favourable terms remains a statement of intent. It may have stayed words for five months because nobody has worked out yet who wears the difference.

A confusing week from the same government

Treasury gave Chalmers a briefing note on 27 August, previewing the AI chapter of the Intergenerational Report. It puts the data centre pipeline at around $150 billion to 2030, and that's roughly 5 per cent of today's nominal GDP.

As the Financial Review reports it, Treasury's advice is that most of the economic gains from AI will come from importing the technology from overseas and embedding it in business practices. On the same advice, Australia has to match American adoption rates to see much of it.

Charlton has spent four months warning that we risk becoming "a permanent renter of intelligence from abroad", which is more or less what Treasury has now described as where the gains are, and both of them speak for the same government.

And on Monday he flew to the United States with the deputy prime minister, to meet the six companies we'd be importing from: Anthropic; OpenAI; Microsoft; Google; Amazon and Nvidia.

Chalmers' own piece asks for both at once, growing our AI sovereignty and encouraging adoption and diffusion. As an aspiration that's fine. Few would argue with it. But it's hard to see what a chief information officer choosing a vendor next quarter does with it, when nothing in front of her says whether the policy wants the best tool or the local one.

OpenAI on measuring what AI is worth

In Capital Brief today, Satya Tammareddy published the case for the other side. Stop counting adoption, she says, and look at whether the work got better, because otherwise "AI adoption charts are just another dashboard".

Her four questions are good ones. Her two customer examples are real, and her byline says she's OpenAI's head of go-to-market for Australia and New Zealand.

But OpenAI's own economists published a working paper three weeks ago. It links ChatGPT Enterprise records, more than seventeen million messages, to worker roles, task classifications and public company financials.

When it tested whether usage intensity moved with revenue per employee, the estimates came back "too imprecise to support a strong conclusion". Page 23 recommends that future work connect enterprise AI telemetry to measures of output.

The paper and the piece can both be true at once. The people holding the telemetry are also the people selling the seats.

All of them are probably right about something. Five months on from the expectations, there is still no figure in them that somebody could be wrong about.

In the March quarter, labour productivity went backwards by 0.6 per cent. The Bureau of Statistics publishes the June quarter national accounts at half past eleven tomorrow morning.

That won't settle the argument. It will at least give us another figure somebody can be wrong about.

One more thing

In the first episode of The Hollowmen, the prime minister announces a National Childhood Obesity Offensive before the government has one. Officials ask for eighteen months to work it out. Tony’s answer is: “The Biggest Loser gets it all done in ten weeks.”

The March expectations did something useful: they put access to compute on the table before the campuses were built. Five months later, the government still has not said what that access amounts to - how much, for whom, for how long, at what price, or who pays the difference.

Five months is more than twice Tony’s television timetable. If favourable terms is to be a condition of the welcome, it needs to become something a provider can cost, a recipient can claim and a minister can be held to.

A number would be a start.

‍

Author: Matt Vitale