Anthropic Signs First Australian Data Centre Lease at Proposed 2.16-Gigawatt Queensland Campus

Four days after Anthropic's chief executive publicly urged the artificial intelligence industry to slow down, his company was revealed to be doing something that, at first glance, looks like the opposite: locking in a vast new reservoir of computing power on the far side of the world. Reuters reported on Wednesday that Anthropic has signed its first Australian data-centre lease, anchoring a proposed 2.16-gigawatt campus in rural Queensland that would rank among the largest digital infrastructure projects ever attempted in the country.

The deal, reported by Reuters on September 16 citing two sources familiar with the matter, puts Anthropic at the centre of a project that was already one of Australia's most ambitious — and most contested — technology bets. The site, a 726-hectare stretch of farmland known as the Western Downs Digital Park, sits about 37 kilometres northwest of the town of Dalby and roughly 250 kilometres from Brisbane. If built as described in the developer's technical filings, the campus would eventually draw as much power as several million Australian homes, with the first phase expected to come online in 2027.

Neither Anthropic nor the developer, Zerra DC, commented on the report. But the contours of the deal are visible in planning documents already on the public record — and they sketch a picture of a company racing to secure the physical foundation for the next phase of its business, even as its CEO argues the industry should think harder about where that race is headed.

Anthropic CEO Dario Amodei. Days after Amodei publicly urged the AI industry to slow frontier-model development, his company signed its first Australian data-centre lease. (Photo: Kimberly White/Getty Images for TechCrunch via Wikimedia Commons, CC BY 2.0)

The deal at a glance

What Reuters established is straightforward: Anthropic has taken a lease at the Zerra DC campus, its first data-centre commitment in Australia. The sources did not disclose the size of Anthropic's portion of the site, the length of the lease, or its value — and the company declined to comment. But the scale of the project around it is a matter of public record, and it is enormous.

According to a detailed report by the Asia-Pacific real estate publication Mingtiandi, Zerra DC — the data-centre arm of Singapore-based AGP Sustainable Real Assets, unveiled in January — has filed plans for the Western Downs campus at an estimated construction cost of A$14.5 billion (about $10.4 billion). The operative master plan calls for four 360-megawatt data-centre buildings to be built in stages, totalling 1.44 gigawatts, while the technical appendices in the filing describe an electrical design for six 360-megawatt buildings totalling 2.16 gigawatts — the figure cited in both the Reuters report and the announcement of the deal by Queensland Premier David Crisafulli, who described the agreement as historic.

The main planning report estimates the proposal would deliver A$31.9 billion in private capital investment overall. Construction would likely take four to six years, with the completed campus expected to support 1,400 direct operational jobs and contribute A$760 million a year to Queensland's economy, according to an economic assessment by the consultancy Urbis cited in the filing.

Why far-west Queensland

To understand why one of the world's leading AI labs would lease capacity on farmland northwest of Dalby — a town of about 12,000 people, best known for agriculture — it helps to look at the map through the eyes of a data-centre developer. The Western Downs region sits inside an energy zone hosting an estimated 5.4 gigawatts of generation and storage capacity, according to the project website. The proposed site lies near three gas-fired power stations, a solar farm, and the Braemar transmission substation.

Under the plan, the campus would connect directly to Queensland's high-voltage transmission network through customer-owned switching stations operating at 275kV and 330kV, rather than drawing from the local distribution system that serves homes and businesses. That design is deliberate: it is meant to answer, in advance, the objection that has shadowed hyperscale data centres from Virginia to Dublin — that they strain the grids ordinary people depend on.

Water, the other flashpoint of data-centre politics, has been handled with equal care in the pitch. The project website describes the facilities as 100 percent air-cooled, eliminating the need for evaporative cooling entirely, with a closed-loop system designed to minimise demand for clean water. The filing identifies rainwater, recycled water, processed water from nearby gas operations, and an existing dam among the proposed sources. Still, the numbers are striking: construction would require about 522 kilolitres of water a day, against regular operational consumption of just 16.5 kilolitres, according to ABC News reporting cited by Mingtiandi. Western Downs mayor Andrew Smith has said water, power, and geography are the key issues residents are raising as the project enters consultation. "Those questions around water are very real and the applicant will need to deal with those," he told ABC News.

The project also arrives in the middle of a genuine policy fight. Australia's federal government wants new large AI data centres to underwrite additional renewable generation and firming capacity, pay the grid costs they create, and minimise water consumption; Queensland favours a technology-neutral framework that would allow continued reliance on coal and gas. Reuters reports that Zerra DC would source renewable power purchase agreements and bear the grid connection costs itself — a structure that appears designed to satisfy both levels of government.

File photo: the roof of a large data centre in Arizona, covered in rooftop cooling equipment and backup generators. The proposed Queensland campus would instead use 100 percent air cooling, with no evaporative water demand. (Photo: Rsparks3 via Wikimedia Commons, CC0 1.0)

Inference, not training: why that distinction matters

The single most revealing detail in the Reuters report is a technical one: the Queensland capacity is intended for inference, not training. The distinction is worth unpacking, because it is the key to understanding the whole deal.

Training is the enormously expensive, one-off process of building a frontier AI model — the weeks-long runs on tens of thousands of specialised chips that produce systems like Claude. Inference is what happens after: the comparatively cheap, endlessly repeated work of actually answering users' questions. Every time someone asks Claude to draft an email, debug code, or analyse a document, that is inference — and it happens on data-centre capacity that has to exist somewhere, close enough to users to keep responses fast.

Anthropic leasing inference capacity in Australia is, in that light, less a bet on building bigger models than a bet on serving more users. It is the physical footprint of Claude's growth: as adoption spreads across the Asia-Pacific, the company needs computing power in the region rather than routing every query back to the United States. The closed-loop, air-cooled design and the renewable power agreements are the packaging that makes a gigawatt-scale commitment politically survivable in a country where data-centre water and power use have become live political issues.

It also reflects a broader industry turn. The AI infrastructure boom that has powered a historic stock-market rally is increasingly about deployment, not just research — about the unglamorous work of building the capacity to run models for hundreds of millions of people. CBRE forecasts that Australia, with roughly 1.4 gigawatts of live data-centre capacity in 2025, could face a supply shortfall of between 700 megawatts and 1.7 gigawatts by 2028 as cloud and AI demand outpaces construction. Reuters estimates Australia's potential data-centre investment at A$150 billion (about US$107 billion) by 2030. Anthropic is not early to this thesis; it is arriving with the biggest names in the industry already circling.

The slowdown paradox

And yet the timing is impossible to ignore — which is why the deal deserves to be reported with the tension stated plainly rather than smoothed over.

On September 12, Dario Amodei published an essay, "We Must Pace the Frontier," arguing that the AI industry should deliberately slow the rate at which it improves the capabilities of its most advanced models. "We must slow the pace at which we improve the capabilities of AI models," he wrote. "Progress will still seem fast, and we must make wise use of the time we gain." The essay won rapid backing from Elon Musk ("Dario is right"), OpenAI's Sam Altman, and Google DeepMind's Demis Hassabis — and a volcanic response from President Donald Trump, who called the safety concerns a "HOAX" in a string of Truth Social posts on September 14. We covered that collision in detail here.

So is the Queensland lease a contradiction? The honest answer is: not exactly — but the optics are genuinely awkward, and the awkwardness is instructive.

Amodei's call was specifically about pacing frontier capabilities — the research frontier of building ever-more-powerful models — not about freezing the deployment of existing ones. Leasing inference capacity to serve Claude's current users does not, on its own terms, violate that argument. A company can believe the next generation of models should be developed more carefully while still building the infrastructure to run the current generation for paying customers. Indeed, the distinction between training and inference is doing real analytical work here: the lease is about running today's models at scale, not about accelerating the training of tomorrow's.

But the deal also reveals something Amodei's essay did not dwell on: the commercial engine underneath the safety debate. Anthropic is a business with soaring usage and soaring infrastructure needs, and infrastructure at this scale cannot be summoned overnight — leases are signed years before the first server racks arrive. The company is planning for a world in which AI use keeps growing explosively, even as its CEO argues the industry should be more deliberate about what it builds next. Both things can be true. Whether they can both be true indefinitely is the question the industry's critics will keep asking.

There is a geopolitical undertow as well. The lease is Anthropic's first in Australia, and it extends the company's physical footprint beyond the United States at precisely the moment when Washington is treating AI infrastructure as a strategic asset — Trump has framed the AI race as a contest America must win against China, and his administration has positioned data-centre construction as a matter of national competitiveness. Building inference capacity in a close ally's territory hedges against concentration risk: regulatory, political, and physical. It is the kind of quiet diversification that rarely makes headlines but shapes where the industry's centre of gravity sits a decade from now.

What still has to go right

For all its scale, the Queensland project remains, for now, a proposal with an anchor tenant rather than a finished fact. The Reuters report says the deal is subject to approval by Australia's Foreign Investment Review Board — the gatekeeper for significant foreign investment — and the campus itself still faces local and state planning approvals plus a public consultation process in which water and power are already the dominant concerns.

Zerra DC has yet to disclose a financing package or a construction start date for the campus, and the planning documents describe delivery in stages "as demand materialises" — developer language for the reality that a multi-gigawatt build-out lives or dies on signed customers. Anthropic is the first named; more will need to follow for the full 2.16-gigawatt vision to be realised. The gap between the 1.44-gigawatt master plan and the 2.16-gigawatt electrical design in the technical appendices is itself a reminder that the project's final shape is still being negotiated, on paper and in public hearings.

And the political weather is unsettled. The federal-state clash over data-centre standards, the resident concerns in the Western Downs, and the global backlash against AI infrastructure — over electricity prices, water, and jobs, an issue now central to America's November midterms — form the backdrop against which every new campus will be judged. Anthropic's lease does not settle any of those fights. It simply raises the stakes: the company that asked the industry to slow down is now one of the largest bets on the physical infrastructure of the AI age. How it reconciles those two positions, in public and in practice, will be one of the stories to watch through 2027 and beyond.

Sources: Reuters (Sept. 16, 2026); Mingtiandi (2026, citing Zerra DC planning filings, Urbis economic assessment, and ABC News); TechNode Global (Sept. 16, 2026); TechPulse USA reporting on Amodei's "We Must Pace the Frontier" essay (Sept. 12, 2026).