At a glance
The AFR reports Anthropic is seeking a copyright arrangement with the Australian Government, alongside its 1.4GW data centre tender, so it can legally train its Claude models on Australian data.
The reported trade is a text-and-data-mining exemption in exchange for a data centre investment worth tens of billions and a fund for Australian creatives paid on a subscription basis.
The Australian Government ruled out a text-and-data-mining exception in October 2025, choosing to pursue licensing and disclosure through the Copyright and Artificial Intelligence Reference Group instead.
Attorney-General Michelle Rowland said on 7 July that there are no plans to weaken copyright protections, and urged the tech industry to reach solutions directly with the creative sector.
The copyright question now gates the training half of Anthropic’s plan, since contracting the data centres does not by itself give it a legal basis to train on local data.
The capacity tender is only half the deal
Anthropic’s 1.4GW Australian data centre tender has a second condition attached that the megawatt figures do not show. According to the Australian Financial Review on 7 July, the company is seeking to sign its first Australian data centre contracts this month at the same time as a copyright deal with the Albanese government that would let it train Claude on Australian data legally.
The two are linked because Australian copyright law, as the AFR notes, does not currently give large AI developers a clear basis to train on local copyright material without a licence. Anthropic can contract the capacity and build the campuses, but neither step gives it the legal basis to train Claude on Australian data. That is what the company is reported to want resolved before its US initial public offering, and it is the part of the plan that sits with government, beyond the operators’ remit.
What Anthropic is asking for
The AFR reports that the copyright talks involve a text-and-data-mining exemption, the mechanism that would allow AI developers to use copyright works to train models, offered in exchange for a data centre investment worth tens of billions of dollars and a fund for creatives paid through annual subscription-type payments.
Anthropic’s tender documents did not specify whether the Australian capacity would be used for training models or for inference, the process of generating outputs from an already-trained model. Sources told the AFR the company is seeking to train Claude in Australia, making the country its second home for training outside the United States. An Anthropic spokeswoman said the local push responds to requests from Australian enterprises and government agencies, particularly those with data residency requirements.
Anthropic comes to the negotiation with recent copyright history. In 2025 it agreed a US$1.5 billion settlement, reported as the largest in a US copyright case, to resolve a class action by authors over roughly 500,000 pirated books it had downloaded to build training datasets. A US court had found that training on the books was fair use, and the liability was for how the copies were acquired. Australia has no equivalent fair-use defence, so the training question here turns on an exemption or a licence rather than on how the data was sourced.
The government has already declined an exemption
The arrangement Anthropic is reported to be seeking is the one the Australian Government ruled out nine months ago. The exception traces to the Productivity Commission, which proposed a text-and-data-mining exception in a 2025 interim report. On 26 October 2025, Attorney-General Michelle Rowland ruled it out, saying the government would not introduce such an exception into the Copyright Act, and would instead pursue licensing models through the Copyright and Artificial Intelligence Reference Group, a standing body convened by the Attorney-General’s Department to work through AI copyright questions with industry and the creative sector.
That reference group is examining collective and voluntary licensing, disclosure of the material used to train AI systems, and lower-cost enforcement options. The government’s stated position is that creators should be compensated and able to negotiate terms.
Rowland restated the position on 7 July. She said the government had repeatedly stated there are no plans to weaken copyright protections when it comes to AI, and encouraged the tech industry and the creative sector to find workable solutions that support innovation while ensuring creators are compensated. On the reported figures, Anthropic is offering a licensing-style creatives fund alongside the exemption, which sits closer to the licensing path the government has favoured than to a free-use exception.
Why copyright gates the data centre build
For Australian data centre operators, the copyright question changes what the 1.4GW tender is worth to them. The build proceeds on financing and power, as the tender’s own checklist makes clear. The workload that fills it depends on the legal question.
If training on Australian data is not permitted, Anthropic can still run inference and other workloads from local capacity, which is the basis on which hyperscalers already operate here. The stated goal, training Claude in Australia, is the part that a copyright resolution unlocks. The company’s National AI Plan memorandum with the government and its early Australian Government customers, including the Fair Work Commission, sit on one side of that question; the copyright negotiation sits on the other.
Two positions, one unresolved question
The dispute is genuine on both sides, and the creative sector’s response has been forceful. Nicholas Pickard of the music rights body APRA AMCOS called Australia a flashpoint in the global fight over whether tech companies can train on copyright works without paying, and the Australian Writers’ Guild welcomed the October 2025 decision as letting creators negotiate as partners rather than be dictated to. Critics have cast the reported investment-for-exemption offer as trading creators’ rights for infrastructure dollars. The underlying argument is that using copyright works to train commercial AI models without payment removes the value of the work and the livelihood behind it, and that licensing is the fair path.
AI developers argue that training requires large volumes of data, that a workable exemption or streamlined licensing lowers the cost and legal risk of building models in a country, and that the alternative is to train elsewhere and import the result. Anthropic’s reported offer, an investment and a creatives fund tied to access, is an attempt to bridge the two by paying for what an exemption would otherwise take for free. Whether that meets the government’s compensation test is the open question, and it is one for Canberra and the creative sector rather than for the operators bidding on the capacity.
What to watch
Watch the Copyright and Artificial Intelligence Reference Group for any move from principle to a licensing mechanism; whether Anthropic and creative-sector bodies strike a direct arrangement of the kind Rowland has urged; whether Anthropic signs its capacity contracts before the copyright question is settled or holds the training workload back until it is; and the detail of the reported creatives fund, which is the part that would show whether the offer meets the government’s stated bar of fair compensation.