All Roads Lead to America: The $490 Million AI Tax Australia Doesn't Know It's Paying

The new landlords

CLAIRE BENNETT: In the old world, all roads led to Rome. Grain from Egypt, silver from Spain, slaves from Gaul. The empire collected tribute from everywhere, built infrastructure that served the centre, and extracted wealth from the provinces for centuries.

The mechanics of corporate restructuring don't change much. When Atlassian announced 1,600 job cuts this month and said it would 'self-fund further investment in AI,' the pattern was familiar: announce a strategic pivot, cut costs in one column, redirect spending to another.

What makes this restructure different is where the redirected spending goes. Every Australian company that adopts AI is signing a perpetual lease with a landlord in California. The rent is usage-based, paid per token, per API call, per compute hour. Unlike hiring an Australian worker who pays tax here, spends locally, and retires into our superannuation system, this rent clears in San Francisco every month for as long as the company operates.

Following the money

MARCUS CHEN: The unit economics of an AI-driven layoff work like this.

When a company like Atlassian fires 1,600 workers at average tech salaries of $110,000 per head, it removes $176 million per year in wage costs from the business. That money doesn't disappear. It moves through three channels:

First, subscription fees. Enterprise AI contracts with Microsoft Azure, Amazon Web Services, or direct OpenAI API access run between $50,000 and $500,000 annually, scaling with usage. These are perpetual operating costs, not capital expenditure. They show up every month until the company stops using AI, which is to say, indefinitely.

Second, compute costs. Every AI query burns tokens. GPT-4o costs $15 per million input tokens. Claude Sonnet costs the same. At enterprise scale, large companies process millions of tokens daily across customer service, code generation, document analysis, and internal tooling. Each token is a micro-payment to Anthropic in San Francisco, OpenAI in San Francisco, or Google in Mountain View.

Third, infrastructure dependency. Australia will spend $172 billion on IT in 2026 according to Gartner, with server spending up 30% year-on-year. Those servers run software licensed from American vendors, trained on American compute clusters, optimised for American cloud providers. The hardware might sit in Sydney, but the value chain runs back to the same three California addresses.

The workers got fired, the salary savings became subscription fees, and the subscription fees left the country.

What the numbers show

CLAIRE: In the first ten weeks of 2026, Australian tech companies cut 4,450 roles. Every termination was attributed to AI transformation, according to data compiled by RationalFX from TrueUp, TechCrunch, and Layoffs.fyi.

At conservative salary estimates for software developers and engineers, those cuts removed roughly $490 million in annual wages from the Australian economy. The cascading effects include approximately $147 million in lost income tax to the ATO, $56 million in superannuation contributions that won't compound over careers, and $240 million in reduced local consumer spending.

The companies receiving that redirected spending pay almost nothing in Australian tax. According to ATO Corporate Tax Transparency reports analysed by the Australian Computer Society, Microsoft collected $6.3 billion in Australian revenue and paid an effective rate of 1.9%. Apple paid 1.5%. Google paid 4.8%.

The Australian worker who got replaced was paying about 30% in income tax. The American company that replaced them pays less than two.

The subsidy loop

MARCUS: The extraction looks worse when you trace where these companies started.

Atlassian claimed $220 million in R&D Tax Incentive refunds in fiscal 2022-23, according to Startup Daily, the largest claim of any Australian company. The R&D Tax Incentive is funded by Australian taxpayers and designed to encourage local innovation and local employment.

Atlassian used those subsidies to build a company now capitalised at tens of billions of dollars. Co-founder Mike Cannon-Brookes has a net worth estimated at $15 billion. This month, the company fired 1,600 Australians so it could 'self-fund' AI investments that flow primarily to Microsoft Azure.

Microsoft pays 1.9% effective tax here. Meanwhile, the Australian government signed a five-year deal with Microsoft this year to expand AI adoption across the public sector.

Australian taxpayers subsidised Atlassian's growth. Atlassian fires Australian workers. The savings go to American vendors. Those vendors pay almost nothing in Australian tax. The government signs more deals with those vendors. Anyone who models business processes recognises a closed loop when they see one.

The CBA precedent

CLAIRE: Some will point to failures as evidence that AI replacement is overhyped. Commonwealth Bank's AI chatbot, deployed in August 2025, actually increased call centre volumes rather than reducing them. The bank was forced to reverse 45 planned job cuts after the Finance Sector Union took the dispute to Fair Work.

The FSU's statement was blunt. CBA had claimed the chatbot 'led to a reduction in call volumes' by 2,000 a week. Fired workers told the union 'this was an outright lie.' Call volumes had been increasing at the time they were dismissed, with the bank 'scrambling,' offering overtime, and redirecting management to answer phones.

At the tribunal, CBA admitted 'they didn't properly consider that an increase in calls would continue over a number of months.' The bank apologised and offered to rehire the affected workers.

But that was seven months ago.

MARCUS: Since CBA's chatbot failed, the capability landscape has shifted substantially. OpenAI shipped Codex, a coding assistant that writes production-quality software and navigates complex codebases autonomously. Anthropic released Claude 3.5 and Claude Code. Cursor, an AI coding editor, reached $100 million in annual recurring revenue by helping developers ship code at roughly ten times their previous velocity. Nvidia announced inference chips at GTC this month that cut token costs by another 40%.

The technology that failed CBA's implementation last August will likely work in a competent deployment within the next twelve months. And the generation after that will be substantially more capable. The question for Australia is whether any of the productivity gains stay in the country or whether every efficiency dividend gets exported as subscription revenue.

What other countries are doing

MARCUS: Other countries have noticed this problem.

The UK published a comprehensive AI infrastructure strategy in July 2025 through the Tony Blair Institute. The report, titled 'Sovereignty, Security, Scale,' argues that AI data centres are fundamentally different from traditional data centres, requiring cheap baseload power and creating geopolitical dependencies. The UK strategy calls for AI Growth Zones with streamlined planning, nuclear power purchase agreements, and sovereign compute capacity.

The European Commission has launched a shared AI infrastructure initiative, recognising that no single European country can compete with US hyperscalers alone. France and Germany are building sovereign AI compute facilities. Saudi Arabia and the UAE are investing tens of billions in AI infrastructure as part of post-oil economic diversification.

Australia has no equivalent strategy. The May budget will apparently focus on 'supply-side reforms' and 'digital skills' according to Treasurer Jim Chalmers. Training programs for jobs that may not exist while the infrastructure gap widens.

The structural gap

MARCUS: AI data centres require cheap, reliable baseload power. The model training clusters and inference farms that generate tokens run 24 hours a day at massive scale. They need power that doesn't fluctuate with cloud cover or wind patterns.

Australia is closing coal plants, locked in political conflict over nuclear energy, and watching renewable intermittency push wholesale electricity prices into unpredictable ranges. US hyperscalers, meanwhile, are signing 20-year power purchase agreements with nuclear facilities. Microsoft is restarting Three Mile Island. Amazon is buying data centre campuses adjacent to nuclear plants.

The big AI infrastructure builds are happening in Oregon, Texas, and Virginia, where power is cheap and regulations accommodate 24/7 industrial loads. Australia has a handful of enterprise data centres in Sydney and Melbourne serving local workloads. Our contribution to global AI compute capacity is negligible.

The compound problem

MARCUS: The urgency comes from the rate of change. OpenAI's inference costs dropped 90% over eighteen months. Nvidia's new chips cut costs by another 40%. Run the compound curve forward and the economic case for replacing any given $150,000 worker with AI equivalent improves by roughly 15% each quarter.

Today, for complex knowledge work, a senior engineer might still be more cost-effective than equivalent AI compute. Run the numbers forward two years and the equation flips. At five years, the comparison won't be close.

Every quarter Australia delays building sovereign compute capacity, the gap widens. Every year without a value capture mechanism, more wealth flows offshore. The curve is compounding against us.

Where this leads

CLAIRE: Jensen Huang said something notable at Nvidia's GTC conference this month. He proposed giving engineers annual token budgets worth half their base salary, on top of regular pay. His framing: 'How many tokens comes along with my job?' is now a recruiting question in Silicon Valley.

Compute is becoming a compensation category alongside salary, bonus, and equity. Access to AI capability is becoming a form of workplace wealth. The entities that control token allocation, the ones who own the data centres and the models and the fabrication plants, are concentrated in a single country.

MARCUS: Sam Altman has suggested publicly that tokens could eventually become a form of universal basic income, a world where people are allocated compute credits rather than currency, distributed by whoever controls the infrastructure.

Australia doesn't control the infrastructure. We don't own the models. We don't fabricate the chips. The trajectory from here looks like a province in someone else's empire, paying tribute through subscription fees.

CLAIRE: Professionals Australia, the union representing Atlassian workers, said hundreds of Australian employees had joined seeking a say in AI decisions before the cuts were announced. They got a four-minute video from the CEO and an email confirming their accounts had been deactivated.

Corporate restructures have always promised investment in the future. The future arrived. The investment went somewhere else, and it's going to keep going there until someone builds a road that runs in the other direction.