Key takeaways
- Graphics Processing Unit (GPU) clusters / high-performance computing (HPC) infrastructure, are emerging as a new, institutional-grade infrastructure asset class, underpinned by long-term contracted revenues and hard-asset collateral.
- Private credit is leading artificial intelligence (AI) infrastructure financing globally, with commercial banks – particularly in Australia – yet to materially participate.
- Australia is increasingly attractive for AI infrastructure investment due to export controls, sovereign positioning and regulatory alignment.
- Power, cooling and data centre capacity remain critical constraints – creating a near-term advantage for early movers with secured infrastructure.
- Regulatory expectations – particularly around energy transition and grid impact – are shaping how AI infrastructure is developed and financed.
Financing AI Infrastructure – biggest deals in Q1 2026
Private credit has emerged as the primary source of debt capital for AI Infrastructure/GPU cluster financings globally, underpinned by the asset class's attractive fundamentals: long-term, take-or-pay customer contracts with investment-grade hyperscalers and blue-chip AI counterparties, hard-asset collateral in the form of GPU infrastructure, and contracted revenue streams that lend themselves to an infrastructure-style credit analysis.
In Australia, Firmus Technologies signed a USD10 billion GPU financing facility in February 2026 – one of the largest private credit financings in Australia's data centre and AI sector – backed by cornerstone lender Blackstone. Structured to support the staged acquisition and deployment of high‑performance GPU clusters (NVIDIA chips) and associated networking, storage and data‑centre infrastructure, the debt facility will be drawn against executed long-term customer contracts with leading hyperscalers and other blue‑chip AI customers.
Read more about this deal here.
In the US, CoreWeave has been the pace-setter, securing approximately USD28 billion in combined equity and debt commitments over the past twelve months — including a landmark USD8.5 billion delayed-draw term loan facility closed in March 2026, achieving the first investment-grade rating (Moody's A3/DBRS A (low)) ever assigned to a GPU-infrastructure-backed financing.
These transactions illustrate private credit's appetite for a rapidly institutionalising asset class, filling a gap that traditional commercial bank capital has not yet occupied.
While the underlying credit fundamentals – long-dated contracted cashflows from creditworthy counterparties, identifiable hard-asset collateral, and risk parameters consistent with those commercial banks have historically been comfortable underwriting in other infrastructure asset classes – commercial banks (and in particular, Australian banks) have not yet featured prominently in the debt capital stack for GPU cluster financings, with debt capital in this sector to date being predominantly deployed by private credit providers.
As institutional familiarity with the asset class deepens, there is a clear opportunity for commercial banks – in particular, Australian banks – to participate in these financings alongside private credit, leveraging their existing infrastructure lending expertise to underwrite an asset class whose credit profile increasingly aligns with the essential infrastructure sectors in which they have deep institutional experience.
GPUs and GPU clusters – what are they?
Graphics Processing Units are specialised chips with many parallel processing cores that can perform many tasks at once, making them the workhorse for tasks requiring high computational power, including training or running highly complex models. Evolving from their initial use in computer gaming, GPUs are now essential components of AI as they accelerate both training and inference, shortening time to deployment for increasingly complex AI solutions.
To achieve the necessary speed and scale, GenAI and large language models such as ChatGPT rely on GPU clusters to train models, run inference at scale and process massive datasets. In practical terms, clusters are composed of GPU nodes that bundle one or more GPUs with CPUs, memory and storage, tied together over high-speed networks. The result is large scale parallelism that enables AI training to run materially faster than on traditional computing architectures.
While GPUs remain the key hardware powering AI, demand is outpacing supply due to manufacturing delays, supply chain issues, geopolitical uncertainty and export controls. Supply bottlenecks have created a significant opportunity for GPUs-as-a-Service offerings from hyperscalers to specialised AI infrastructure companies.
AI infrastructure
AI infrastructure companies (sometimes referred to as neoclouds) are a new breed of specialist digital infrastructure provider that deliver scalable, high‑performance GPU‑accelerated computing for AI training and inference. They operate dense GPU clusters with high‑speed interconnect fabrics and, increasingly, liquid cooling, and monetise this by renting that capacity to customers such as Meta and OpenAI for critical AI workloads. Unlike hyperscalers such as Microsoft Azure, Google Cloud and Amazon Web Services, which offer a broad set of cloud services, AI infrastructure providers primarily focus on purpose‑built GPU compute and leaner delivery models that can enable lower cost to the customer for comparable GPU capacity and performance.
Hyperscalers secured most advanced-chip allocations, but struggled with the rising demand for GPUs, leading to soaring prices. Many businesses were unable to access GPU capacity at the speed they required. AI infrastructure providers emerged to fill this gap, offering tailored contracts, faster provisioning, and dense racks of GPUs in an AI optimised environment.[1]
AI infrastructure providers are often regionally focused in certain jurisdictions enabling them to offer important alternatives to hyperscalers in highly regulated markets where data sovereignty and data localisation are critical (such as the public sector and healthcare). Their regional focus and ability to offer jurisdiction-specific hosting and contractual commitments may better align with certain regulatory and policy requirements.
GPUs need to be hosted in data centres
AI infrastructure companies will need to have their GPUs hosted in data centres (DCs). Modern GPU hosting – especially for AI and high-performance computing – typically demands much more power and cooling infrastructure than traditional CPU-based servers. In practice, accommodating powerful GPU servers often means deploying them in purpose-built data centres with robust power delivery, advanced cooling (frequently using water or liquid cooling), and sufficient physical space for all the necessary equipment. For example, to put things into perspective:
- Power usage requirements – a large AI cluster of 30,000 GPUs can draw ~35 MW of power (roughly the output of a small power plant).
- Cooling and water requirements – virtually all of a GPU’s electrical power (≈99%) is ultimately converted into heat. Cooling high-density GPU hardware is therefore a major challenge. Conventional air-cooled data-centre designs are typically effective up to around 20–25 kW per rack, while at rack densities above ~30–40 kW, liquid-based cooling solutions (such as direct-to-chip or immersion cooling) are generally required to dissipate heat reliably and efficiently.
- Physical space and infrastructure – while GPUs pack more computing power into each server, supporting infrastructure (power transformers, backup generators, cooling distribution units, coolant piping, heat exchangers, etc.) increases the overall space requirements. GPU-dense servers operate at much higher power and thermal densities than traditional CPU-based systems, resulting in substantially greater heat loads per rack. Additional floor space is therefore commonly required for mechanical and electrical plant, including chillers, pumps, cooling distribution units (CDUs), or coolant reservoirs, as well as to support airflow management or accommodate liquid-cooling configurations.
Australian DCs are an ideal destination for these high-powered GPUs.
Why we think Australia will be an attractive destination for GPU cluster investment
A shift in the US regulatory landscape is pushing more AI infrastructure investment toward Australia. Recent export controls under the Export Administration Regulations (EAR) sharply limit the supply of advanced U.S.-origin AI chips to China and other restricted destinations. These measures–reinforced by a broad ‘General Prohibition Ten’ (GP10) rule forbidding any knowing evasion of export controls– have severely restricted China's access to cutting-edge GPUs, making allied markets more attractive.
As a trusted AUKUS partner and APAC destination outside the China- and Macau-focused licence requirements, Australia holds a strategic advantage and is attracting increased investment from hyperscalers and vendors (such as Nvidia and its authorised OEM and server partners and distributors).
In tandem with these controls, the US Commerce Department’s Bureau of Industry and Security (BIS)has issued detailed guidance clarifying the industry’s ‘compliance obligations across the AI ecosystem’. This recent guidance makes clear that every link in the AI supply chain – from chipmakers, to cloud/data center providers, to distributors – should step up due diligence to avoid export-control violations.
Key BIS guidance:
- PRC-designed chips – chip manufacturers are expected to verify the legality of any Chinese-designed advanced chips they handle. BIS now presumes that high-performance chips meeting ECCN 3A090 (like Huawei’s Ascend 910B/910C/910D series) were developed or produced using U.S. technology without authorisation, meaning that any dealings with such chips could violate U.S. export law unless a BIS license was obtained.
- Diversion and due diligence:
Cloud and data-centre providers should implement robust 'know-your-customer' screening procedures-denying or terminating service if they know or have reason to know that their advanced computing resources will be used by restricted Chinese or other Country Group D:5 parties (countries subject to a US arms embargo under the Export Administration Regulations).
BIS's policy statement on AI model training warns that providing state-of-the-art GPUs or AI services to train models for or on behalf of parties headquartered in China or another D:5 country creates a red flag that may trigger a licence requirement under the catch-all controls– particularly where the provider is aware (or should be aware) that the AI models will support military or weapons of mass destruction uses
Intermediaries and distributors are also put on alert by an extensive list of red flags for chip diversion. BIS’s guidance highlights warning signs such as a customer with no prior history of buying high-end chips, unclear or residential installation addresses, a minimal online footprint, proximity to sanctioned entities, or claims by buyers that raise questions (e.g. data centres without adequate power/cooling infrastructure, or the customer providing Infrastructure as a Service does not or cannot affirm that its own end-users are not headquartered in the PRC, whether or not such customer is located inside or outside of China and Macau).
In short, companies at every level of the AI supply chain are now expected to verify end users, end uses, and ownership of advanced computing tech – and to adhere strictly to licensing requirements – in order to avoid severe penalties or even being added to the ‘Entity List’ (restricted party list that effectively bars most U.S.-origin exports to listed entities) for aiding prohibited activities involving China, Macau, or other restricted destinations.
To mitigate adverse exposure to these regulations, reputable data centre operators and AI cloud infrastructure providers contract with quality customers and seek robust contractual protections in their master services agreements from these customers. These typically include covenants from customers requiring compliance with applicable trade controls (including export control laws and sanctions) in their and their end-users’ use of the services, together with express carve-outs from force majeure relief where a party becomes subject to targeted restrictions under trade controls.
Lenders financing AI infrastructure similarly require extensive protections in their credit documentation – including (i) representations that the borrower has implemented policies and procedures designed to achieve export control compliance, (ii) representations that the borrower's activities have not been used for prohibited end-uses, (iii) customer eligibility criteria that expressly exclude counterparties domiciled in or primarily operating from China, Russia, or any other Country Group D:5 jurisdiction (or ultimately controlled by entities in those jurisdictions), (iv) affirmative covenants requiring borrowers to implement contractual controls in their customer agreements making it a terminable breach if services are used by customers or their end-users in violation of export control laws and (v) ongoing notification obligations in respect of any governmental investigation relating to potential violations.
Key considerations – developing and financing AI Infrastructure
Several key considerations arise for AI infrastructure providers, investors and financiers seeking to develop, or to support the development and deployment of, GPU clusters. These considerations centre on risk profile and margin protection – and, ultimately, on securing a consistent and robust revenue stream from key customers:
- Utilisation of GPUs– depending on the proposed contractual structure revenue may fluctuate based on the customer’s use of GPU hours – coupled with a cost base which may be largely fixed (that is: debt service, colocation fees and power), margins may be squeezed.
- Key risk mitigants to be considered:
- ‘Pay-whether-used’ contractual model – fixed payments under customer contracts, where customer reserves a fixed amount of capacity and must pay for this regardless of whether such capacity is used.
- Key lender covenant protections in credit documentation – CPs to draw, maintenance covenants and/or distribution conditions linked to realised contract revenue or GPU “average hourly uptime” (supported by periodic reporting).
- Key risk mitigants to be considered:
- GPU depreciation and stranded-asset risk – GPUs are depreciating assets subject to rapid technological succession. Each new chip generation can significantly reduce the rental value of its predecessor within roughly five years.
- Key risk mitigants to be considered:
- Facility utilisation and/or maintenance covenants ––tested against the value of net capital expenditure (such capex value reducing in accordance with an agreed "GPU Depreciation Schedule"). This keeps leverage aligned to declining collateral value (i.e. as the "depreciated amount" increases over time, the net asset base falls and permitted borrowings step down accordingly).
- Facility amortisation ––scheduled repayments ensure that outstanding debt reduces over the life of the GPU fleet, complementing the lender protections described above and reducing residual exposure as hardware ages.
- Key risk mitigants to be considered:
- Operational and technical performance –GPU infrastructure must be kept at optimal performance and to reduce hardware and software failures and cooling outages which reduce GPU hourly uptime.
- OEM Agreement – robust OEM support (in other words: warranty package and onsite remediation / post sales service) - extending usable life and preserving residual value.
- Facility Agreement – evidence of full delivery to the deployment site prior to funding and proof of warranty services for the full life of customer MSAs.
Key considerations for AI infrastructure more broadly
On a more macro / structural level operators, developers and investors will need to think about the following:
- Power and data-centre capacity constraints– the physical supply side remains a binding constraint. Demand for AI compute outstrips available data-centre capacity and grid-connected power; there is not enough built-out, cooling-ready rack space to accommodate the GPU deployments the market demands. Providers that secure scarce power and purpose-built facilities hold a near-term advantage, though it may narrow as new capacity comes online.
Commonwealth Government’s Expectations of data centres and AI infrastructure developers
Beyond securing physical capacity, operators and developers must contend with clear and directive regulatory expectations from the Commonwealth. On 23 March 2026, the Australian Government published its “Expectations of data centres and AI infrastructure developers”, a framework that effectively codifies the social licence to operate for data centres and AI infrastructure providers.
“Expectation 2 – Supporting Australia's energy transition” – is particularly significant. It provides that new data centres and AI infrastructure should not place upward pressure on energy prices and should make a positive contribution to Australia's energy transition. Specifically, operators are expected to work with energy regulators and suppliers to: secure new and additional clean energy generation and/or storage to offset demand; cover their share of transmission and distribution infrastructure costs; minimise energy demand and emissions by adopting industry-leading efficiency measures and technologies; and improve the overall security and stability of the energy grid, including by enhancing demand flexibility and opportunities for peak-load management and appropriate sharing of consumption data. The Commonwealth has made clear that energy-intensive data centre proposals not closely aligned with these Expectations will not be prioritised by Commonwealth regulatory assessments – and conversely, proposals most closely aligned with the Expectations will be prioritised.
Market response
Some market participants have already been responding to this regulatory environment.
Active grid participant rather than a passive load
Firmus Technologies, for example, has developed proprietary software — including its Synert platform and AI FactoryOS – that enables its AI factories to operate as active grid participants rather than passive loads, providing frequency control ancillary services (FCAS), load modulation, and firming capacity to the grid.
This approach directly addresses concerns that Australian Energy Market Operator –(AEMO) has raised about new AI load types. In its 2025 Transition Plan for System Security, AEMO identifies large data centres (including those supporting AI workloads) as large inverter-based loads (LIBL) and warns that, without coordinated planning and careful connection, they present system security risks for the National Electricity Market (NEM), including through power system oscillations, sudden load loss, and erratic demand profiles.
Reflecting the scale and distinct behaviour of these facilities, AEMO has also moved – for the first time – to forecast data-centre demand as a standalone component in its updated Electricity Demand Forecasting Methodology, rather than grouping it with other commercial loads. AEMO has noted rapid growth driven by artificial intelligence, cloud computing and global data services, with data centres consuming around 4 TWh across the NEM in FY2025 (approximately 2.2% of total grid demand) and consumption forecast to reach around 12 TWh by 2029–2030.
The Australian Energy Market Commission (AEMC) has followed suit, publishing a draft determination and draft rule on 12 March 2026 (AEMC - National Electricity Amendment (Improving the NEM access standards – Package 2) Rule 2026) that introduces new access standards and a clear regulatory framework for LIBLs connecting to the NEM, stating expressly that "data centres and other large IBLs cannot be treated as passive loads, and their potential impacts on power system stability and security must be considered when seeking connection to the NEM.”
AEMO has also commenced a targeted review of technical requirements for LIBL connections, with the 2026 General Power System Risk Review – the final report for which is due by 31 July 2026 – to investigate the risks associated with increasing large load connections as a priority area.
Energy, water efficiency and lower PUE
As discussed above, among the Commonwealth's Expectations, the requirement that operators minimise energy demand and emissions by adopting industry-leading efficiency measures and technologies will be a strategic advantage for those already operating to this standard. Companies such as Firmus Technologies have been actively implementing measures across their infrastructure platform, energy sourcing and site selection consistent with these Expectations.
Firmus has invested in a vertically integrated physical infrastructure platform – its proprietary ‘HyperCube’ technology. Each HyperCube is a modular AI factory building block designed to house GPU racks in a primarily liquid-cooled, high-density configuration, with a reported power usage effectiveness (PUE) of 1.10 at greenfield sites – materially below industry standard, translating to energy cost reductions designed to deliver materially lower energy costs per AI token compared with more traditional data centre designs.
Firmus' chief technology officer has described the broader platform as a "Model-to-Grid" architecture that integrates model behaviour, GPU performance, thermal management and grid conditions into a single optimisation framework – making its AI factories both model-aware and grid-aware, with real-time responsiveness to energy pricing and broader grid signals.
Firmus has positioned its Tasmanian Green AI Factory Zone as a geographic and energy differentiator. The Southgate Tasmania site draws on a predominantly renewable grid (hydro, wind and solar) and is located in a region with access to significant existing transmission infrastructure and industrial-grade power connections that are being freed up as end-of-life metal smelters release capacity back to the grid. Firmus has stated that Project Southgate could have the capacity to support the development of up to 5.1 GW of new solar, storage, hydro and wind generation projects across Tasmania, Victoria, New South Wales and the ACT – a proposition that aligns squarely with the Commonwealth's Expectation 2, which requires new data centres to secure new and additional clean energy generation to offset demand, and to make a positive contribution to Australia's energy transition.
For operators, investors and financiers evaluating significant GPU cluster deployments, the signal is clear: the projects most likely to navigate the regulatory pathway efficiently – and to maintain their social licence over time – will be those that are designed from inception to operate as grid-responsive, energy-transition-aligned infrastructure. - Moving up the AI stack and hyperscaler competition– a strategic tension sits at the heart of the model: remaining at the bare-metal layer (that is GPU as a service only) exposes providers to offering a more commoditised service, yet moving up the AI stack into training orchestration, distributed inference platforms and managed machine learning services risks direct competition with hyperscalers on whom many neoclouds depend for anchor demand. Resolving that tension without alienating key customers or overextending capital is the defining strategic challenge for the sector.
McKinsey's analysis identifies three durable paths for neoclouds seeking to resolve this tension: carving out defensible niches where hyperscalers are less effective or less welcome – such as sovereign compute, staying focused on start-ups and growing with them as they scale into large AI companies and building loyalty that hyperscalers find difficult to replicate.
In practical terms, the AI infrastructure providers most likely to endure are those that differentiate across multiple dimensions simultaneously – not merely by offering cheaper GPU hours, but by embedding themselves into the energy, regulatory and sovereign infrastructure fabric of the jurisdictions in which they operate.
Firmus Technologies illustrates what this multi-dimensional differentiation can look like in practice. Rather than attempting to replicate hyperscaler application platforms, Firmus is building structural advantages across sovereign compute (see below), physical infrastructure and energy (see above) that are difficult to commoditise and closely aligned with the Commonwealth Government’s regulatory expectations.
Sovereign AI capacity
Firmus is developing sovereign AI capacity – a niche where hyperscalers may face structural disadvantages. Its Sustainable Metal Cloud already provides GPU computing infrastructure to AI Singapore (a national AI programme) to support the development of SEA-LION, an open-source large language model suite for Southeast Asian languages, and the company has signed a memorandum of understanding with an agency under Singapore's Ministry of Home Affairs, to explore AI infrastructure design for public safety applications.
In Australia, Project Southgate is designed to serve Australian research, government and sovereign users alongside hyperscalers and global software companies, with its colocation partner CDC (whose major shareholders include Infratil, the Future Fund and CSC) bringing a proven track record in government, national critical infrastructure and regulated industries.
Supporting domestic supply chains
This sovereign positioning is reinforced by a commitment to domestic supply chains: Firmus has reportedly committed more than $300 million to Australian manufacturing with partners, including Benmax (HyperCube production) and Maas Group (electrical and power systems manufacturer) together establishing advanced manufacturing capacity for AI factory components – cooling, power and integrated system modules.
For governments and regulated industries where data residency and operational control are paramount, this combination of in-country infrastructure, sovereign-compute capability and local industrial capability is a proposition that US-headquartered hyperscalers cannot easily match.
What clients need to know
- AI infrastructure – particularly GPU clusters – is rapidly maturing into a bankable infrastructure asset class, but financing remains dominated by private credit.
- Australia presents a compelling investment environment, driven by geopolitical tailwinds, data sovereignty demand and supportive regulatory positioning.
- Securing power, data centre capacity and compliant customer contracts will be critical to project viability and financing outcomes.
- Regulatory scrutiny is increasing – particularly around export controls, energy usage and grid impact – requiring robust contractual and compliance frameworks.
- Investors and developers who align early with energy transition expectations and infrastructure constraints will be best placed to capture long-term value.