That line captures the immediate dilemma Australia now faces: rapid uptake of AI-driven services and hardware, but persistent risk that the economic rewards of those services will flow offshore. The government’s National Cabinet will meet this week to set AI policy, but current public discussion is focused on physical data-centre capacity and not on who controls the outputs or what those outputs will cost. The difference could determine whether Australia is a seller of AI products and services that keep returns onshore, or merely a host for machines that generate value elsewhere.
Why token pricing decides who captures value
AI outcomes are priced in “tokens,” small units that average about four characters each. More powerful models tend to have more expensive tokens and also consume more of them when producing answers. Whoever sets token prices exercises considerable influence over the economics of any industry running on AI.
Goldman Sachs projects global token consumption will multiply 24 times by 2030, to 120 quadrillion tokens a month. Analysis by Singapore-based aggregation company Ai.cc, covering 2.4 billion enterprise API calls, put the blended cost at US$6.07 (A$8.47) per million tokens — down 67 percent from the prior year. But procurement choices matter: organisations routing tasks to cheaper-capable models paid a median US$2.31 per million tokens, while those sending everything to frontier models paid US$18.40. The same intelligence, at eight times the price.
Security agencies are already concerned this dynamic pushes enterprises toward open-weight models that offer near-frontier capability at far lower cost. Many of these models originate in China, raising a national and geopolitical risk — to Australia and to the United States, described in the source as “our dominant technical alliance partner” — that Australia could remain a physical digital-infrastructure winner without controlling or creating value from the outputs that run on it.
Investment spike, weak GDP impact
Recent capital flows underline the danger. Machinery and equipment investment, led by digital infrastructure hardware and software, jumped 16.3 percent in the first three months of 2026. In the same period, GDP grew just 0.3 percent, a gap explained in the source by net trade: most of that equipment was imported. In short, the quarter’s biggest capital story “barely touched the economy it was meant to drive.”
Reporting in the Australian Financial Review cited organisations that have reached or exceeded their annual AI budget allocations far earlier than expected. Two cost drivers noted: models increasingly “show their working” before answering — a behaviour that consumes more tokens — and pilots have expanded into broader internal demand. Volume is rising faster than prices fall, shifting boardroom conversations from pilot cases to ensuring acceptable returns on investment.
This site is the portfolio.
OSINTSights runs on Cloudflare Workers, D1, R2, and Vectorize, with an AI pipeline on Hetzner ARM. Nubivance designed, built, and operates it. We do the same for clients.
See what we buildData-centre geography, and computation moving to the edge
The supply-side picture is uneven. Demand to run AI models is projected to exceed demand to train frontier models by 2028, and that operational demand is increasingly satisfied “at the edge” — on local devices and computers rather than distant data centres. Yet more than 80 percent of Australia’s planned or active data-centre capacity is concentrated in Canberra, Sydney and Melbourne — locations the source says are “far from the mines, ports, communities or border posts the computing power will serve.” Concentration risks mismatches between where raw computing sits and where industry domain expertise and operational needs are located.
Three government options to keep returns onshore
- Build and maintain a domestic reserve of AI computing capacity, with guaranteed access for government agencies and industry if supply the country does not control is interrupted.
- Diversify where that capacity sits, putting computing closer to the places it will serve rather than concentrating it in a few capitals.
- Require any vendor bidding for government or critical-infrastructure work to disclose cost per outcome, not cost per GPU, so departments can tell whether they are paying something like US$2.31 or US$18.40 per million tokens and whether public money is buying real value.
The source argues confidential deals with tech giants “have never benefited Australia and won’t start doing so now,” and that the National Cabinet’s current agenda — prioritising land, building, power, energy and data — leaves the decisive questions of output control and price largely unaddressed.
What this means for mining, agriculture, defence and health
Mining, agriculture, defence and health are singled out as areas where Australia already holds domain data and operating expertise “no frontier model can buy.” If policy and procurement capture the value created by those assets, Australia could export what runs on the hardware rather than merely hosting it. The source notes a past success to underline the point: in 2020, Australia’s exports of mining equipment, technology and services were worth A$17 billion.
Conversely, the source warns, getting this wrong would render ambitions for AI “moot” and cement Australia’s reputation as a “well-regulated price taker.”
The prime minister’s call to support digital infrastructure and its beneficiaries is reflected in the proposals above. The practical test for National Cabinet this week is whether Australia will treat computing merely as infrastructure, or whether it will insist on structures — reserves, geographic diversity and outcome-based pricing — that keep the repeatable benefits of AI onshore.
