Australia has settled on a comfortable story about artificial intelligence: the country is building fast, spending big and catching up. Ministers point to fresh data centre approvals, banks trumpet the number of staff issued a Copilot licence, and universities line up to report double-digit jumps in AI budgets. On almost every input measure you could name, the graph points up and to the right. The question that keeps getting lost in the celebration is whether any of it is the right thing to count.
That is the uncomfortable prompt behind a recent piece in the Australian Financial Review, which asked whether the nation is measuring the right things as it races to build out its AI capacity. It is a deceptively simple challenge. Governments and boards love metrics that move, because moving metrics make good announcements. Yet the metrics that move most easily, dollars committed, servers switched on, pilots launched, are the ones that tell you the least about whether Australia is genuinely becoming more productive, more capable or more sovereign in the technology.
Counting inputs, not outcomes
The distinction matters more than it sounds. Building AI capacity is an input. A gleaming hyperscale facility outside Wagga Wagga, a sovereign large language model sitting on a government tender, a training program run through a university short course: all of these are things you buy or build. What they are supposed to buy you is an output, faster drug discovery, cheaper logistics, better public services, more high-value jobs. The trouble is that inputs are visible within a news cycle while outputs take years to show up in the national accounts, if they show up at all. So the incentive is to keep announcing the inputs and quietly hope the outputs follow.
Australia has plenty of the inputs. Capital expenditure on AI infrastructure has surged, regional towns are pitching themselves as data centre hubs, and enterprise spending indices keep climbing quarter on quarter. What is far harder to find is a credible, shared way of measuring whether that spending is translating into productivity, the one number that has been stubbornly flat in this country for the better part of a decade. Productivity is precisely the problem AI is meant to solve, and precisely the thing nobody can yet prove it is solving.
Two ways of keeping score
There are broadly two camps in this argument, and both have a point. The first says the inputs are the honest measure for now, because you cannot get outputs without them. On this view, a country that hesitates on compute, energy and skills will simply be locked out of the frontier, and worrying about productivity statistics today is like fretting about petrol receipts before you have bought the car. Build the capacity, this camp argues, and the returns will come, as they eventually did with electricity, broadband and the cloud. The lag is real, but so is the payoff.
The second camp is more sceptical, and it is where much of the research community sits. Bodies such as the CSIRO and a string of Australian universities have spent the past two years trying to quantify not how much AI the country is buying, but how well it is being used, whether models are trustworthy, whether workers can actually apply them, whether the promised gains survive contact with real workflows. Their concern is that a scoreboard built entirely on spending rewards activity over impact, and that a lot of the activity is theatre. A pilot that never reaches production still counts as adoption. A licence that sits unused still counts as a deployment. Measure the wrong things and you can look like a winner while standing still.
This is not an abstract worry. If the official story is that Australia is surging ahead, based on inputs, then the pressure to check whether the spending is working evaporates. Good metrics are uncomfortable precisely because they can tell you the emperor has no productivity gains.
Why this lands hard in Australia
The debate has a distinctly Australian edge. This is a mid-sized economy with a small domestic market, a heavy reliance on imported technology and a well-earned reputation for being an early and enthusiastic adopter rather than a builder. That combination makes measurement more important here, not less. A larger economy can afford to spray money at AI and let scale sort out the winners. Australia cannot. Every dollar of public and enterprise capex that goes into infrastructure without a matching return is a dollar that could have gone into skills, research or the handful of areas where the country might actually build something the world wants.
There is also the sovereignty dimension that has dominated so much of the local conversation, from government procurement of home-grown language models to worries about who owns the compute the nation runs on. Sovereignty, too, is easy to announce and hard to measure. A sovereign model that no agency trusts enough to deploy is a line item, not a capability. If Australia is going to spend heavily on doing AI its own way, it needs a scoreboard that can tell the difference between genuine independence and expensive symbolism.
The energy question sharpens all of this. Data centres are hungry, and the national grid is already carrying the weight of the transition to renewables. Justifying that extra load on the basis of jobs and productivity only works if someone is honestly counting the jobs and the productivity. Right now, the load is easier to measure than the return.
What comes next
The practical fix is not glamorous, but it is clear enough. Australia needs measures that follow AI spending through to outcomes: productivity per worker in firms that have deployed it, the share of pilots that reach production, the durability of the gains once the consultants leave, the trustworthiness of the systems in the wild. Some of that work is already under way in the research sector, which is exactly why the CSIRO and the universities keep turning up in this story. The harder task is getting governments and boards to adopt those measures when the flattering ones are so much easier to put in a media release.
None of this argues against building. Australia probably does need more compute, more skills and more infrastructure to stay in the game. The argument is narrower and more pointed: if the country is going to race, it should at least agree on where the finish line is. Counting the fuel you have burned is not the same as knowing whether you are winning, and for a nation with a productivity problem it cannot afford to ignore, that difference is the whole point.
Sources: Australian Financial Review.
