What type of data centre actually needs to get built in Australia?
A credit lender looking at the data centre investable market in Australia is asking the question – is the artificial intelligence (AI) boom real and should we be lending into it?
To explore this, Oxford Economics Australia’s latest research notes the Australian Energy Market Operator received 44 GW of data centre connection requests from Network Service Providers. Oxford Economics Australia posit 6 in every 7 MWs of connection requests are estimated to be ‘phantom demand’. That is, they are not expected to materialise under AEMO’s Step Change scenario. So, of the 44 GW of connection requests received, only 6 GW of prospective project capacity is expected to meet demand[1].
This churn of speculation leads to incredible confusion for all stakeholders – regulators, government, developers, utilities and investors – who are grappling with trying to understand the ‘shape’ of the data centre forecast. And to take one stakeholder perspective, it is critical to know precisely, so the transmission network planning team can accurately forecast the electricity infrastructure required to accommodate the demand.
To look at this pipeline another way, of the data centre projects that are to be energised, what type of data centre actually needs to be built in Australia?
Or to put it more bluntly:
We should start with first principles thinking – why would compute live in Australia at all?
Strip away the frothy announcements and there are exactly two structural reasons for a data centre to exist in Australia rather than somewhere with cheaper and more abundant power.
Captive data
The Australian Government’s Security of Critical Infrastructure Act imposes security obligations on critical infrastructure operators such as asset registration, mandatory cyber incident reporting and risk management programs. This is supported by the Whole-of-Government Hosting Strategy and the hosting certification framework for government data and dataset specific rules. This means government data such as from Department of Defence, or private and public sector data such as health records is captive demand that must be domiciled in Australia and not arbitraged away to a cheaper site in say the United States or India.
Latency
Enterprise applications, real-time payments and other consumer, cloud-based functions must sit adjacent to the users they serve – hence the established data centre market that is already operating in metropolitan Sydney and Melbourne – it is where the demand is.
If cloud is the base case, why then should AI training and inference be built in Australia?
JLL’s report notes that despite AI gaining daily users, it only represents about a quarter of all data centre workloads in 2025, with training driving most of that demand. JLL posit a significant shift is anticipated in 2027 when inference workloads could overtake training as the dominant AI requirement.
While an AI model represents a one-time or periodic investment, once the model is created, inference generates ongoing revenue through actual application usage. The theory being every AI model deployment creates sustained inference demand that grows with user adoption. This growth trajectory however, depends on the emergence and rapid adoption of inference applications that does not yet exist.[2]
Australia has no structural advantage for AI training
AI training is not bound by latency requirements and is the most location indifferent workload in computing. It chases the cheapest firmed power available and that contest is being won in the United States and the Middle East, not on the NEM where wholesale prices sit well above those markets and metropolitan connection timelines run to years.
To put some numbers to this claim, the average NEM wholesale spot price for 2025 was USD$58/MWh (AEMO CY25 mean converted to AUD/USD) and Texas ERCOT was USD$35/MWh. Meaning NEM energy for data centres is 66% more expensive than Texas, a key data centre growth corridor. (AEMO, US Energy Information Administration).
AI inference inherits both structural tailwinds as cloud
Enterprise and transactional inference are latency-sensitive and increasingly carry the same data-residency obligations as cloud – so it must be served from within the country. This structure favours the existing metropolitan data centres where the users and existing cloud infrastructure already sit, not a regional, remote campus in say a renewable energy zone adjacent to renewable generation.
Existing metropolitan-based cloud data centres have the right location and tenant relationships to capture inference workloads but the facilities will need significant power and cooling upgrades to host it.
Inference is the slice of the AI boom Australia could capture and it arrives through the metropolitan data centre locations, serving the same hyperscale and enterprise tenants as the cloud build out that preceded.
So sovereign cloud and co-location can win, and the investable Australian share of AI is largely inference leveraging the same infrastructure.
But because inference must sit in metropolitan areas, it stacks onto the most congested area of the high-voltage transmission network. Therefore, the unlock is not necessarily wholesale power price, it is metropolitan transmission and distribution connection capacity. And as per a previous article, metropolitan Sydney has 16 GW of data centre connection enquiries against a network forecast expectation of 800 MW of actual near-term data centre load.
And, the case for inference rests on demand that is a forecast, not locked-in contracts.
What is perhaps speculative for Australia and carries the most risk is the data centre being built to only capture training demand.
What a credit lender should take away from this article
If you lend or are looking to lend against these assets, the first-principles thinking can collapse into four underwriting rules:
Tenancy quality is the asset. A facility anchored by hyperscaler pre-commitments, with contracted leases and approvals in hand is fundamentally a different proposition from a tenantless build in a secondary location. The former is infrastructure. The second is a land bank with a generator or battery attached.
The energisation date is a claim not a fact. The information memorandum will assert an energisation date however almost none of them can be independently verified by the lender against connection queue positioning, the augmentation costs and upstream augmentation impacts. A date that slips by 18 months reshapes the entire conversation.
Distinguish the workload thesis. ‘AI demand’ can mean anything from contracted sovereign inference capacity for a hyperscaler or a hope that someone, eventually wants to train a model in Australia.
Watch the wholesale price forecasts with great interest. When the finite grid capacity is seen to be absorbed by data centres which in turn, spikes the wholesale price, the regulatory and political response will be swift. This is a live risk and the curtailment and system security impacts are not well understood.
Not investment advice, general information only. As ever, if you would like to discuss this further, or wish to stress-test a specific project, I can be contacted at: info@followthebottleneck.com
[1] 2025-11_Oxford-Phantom-Demand-Research-Briefing.pdf
[2] 2026 Market Outlook for Global Data Centers | JLL Research