At a glance
CoreWeave, one of the world’s largest neoclouds (specialist providers that rent out NVIDIA GPUs as cloud computing for AI), is the listed name investors treat as a proxy for the sector. Its shares have fallen about 58% from their June 2025 peak as investors mark down debt-funded AI infrastructure, and it has borrowed about US$30 billion to fund 49 data centres.
In a 21 July interview with the Wall Street Journal, chief executive Michael Intrator rejected the “neocloud” label, arguing the moat is the software, security and operational expertise layered on top of the hardware, and said he is “on the right side” of a generational change.
CoreWeave’s reframing puts software and platform at the centre of neocloud competition, a dimension apart from the contracted offtake, secured power and sovereignty that Australia’s operators compete on today.
IREN, the largest Australian-founded operator, answered the same week with US$2.8 billion of new contracts across nine named AI customers and prepayments covering about 45% of GPU capital expenditure, and its shares jumped about 20%.
Australia’s operators compete on contracted power and sovereign, in-country capacity that global platforms cannot build offshore, while the software-platform layer is contested by CoreWeave, the hyperscalers and Meta.
CoreWeave falls 58%, rejects the neocloud label
CoreWeave is one of the world’s largest neoclouds, the specialist operators that rent out NVIDIA GPUs as cloud computing for training and running AI models. US-based, NVIDIA-backed and Nasdaq-listed since its March 2025 float, it is the listed name investors treat as a proxy for the sector, and the reference point for Australia’s own GPU-cloud operators. Its shares are down about 58% from their all-time high in June 2025, dragged by industrywide concern about the capital spending and debt behind the global AI build-out. The trigger for the latest leg was Meta confirming Meta Compute, a commercial cloud on its own GPU fleet, which turned CoreWeave’s largest customer into a competitor. We set out that reaction in our coverage of Meta’s move to become its own neocloud and the split it exposed in AI compute is splitting in two.
Intrator answered in a 21 July interview with the Wall Street Journal’s CIO Journal. “AI is a generational, maybe multigeneration, change,” he said. “We know what we’re building. We know why we’re building it. I’m pretty comfortable that I’m on the right side of this.” His plan is to build the infrastructure, then layer on “the greatest possible software solution for making it usable, easy, accessible and efficient.”
CoreWeave now rejects the “neocloud” tag, arguing it describes only what came before, and prefers NVIDIA chief executive Jensen Huang’s description of it as a “new generation AI-native cloud” built from the ground up for this era of computing. Neocloud is the category Australia’s operators still sit in.
CoreWeave spends US$1.7bn moving up the software stack
CoreWeave’s pitch is that the value is no longer in the silicon. “Our customers don’t come to us to rent GPUs. They partner with us to run production AI on our purpose-built stack,” executive vice-president Chen Goldberg wrote on 9 July. The company spent 2025 building toward that claim, acquiring the machine-learning tooling firm Weights & Biases for about US$1.7 billion, and its architecture now spans analytics, observability, a training system for researchers, application software and liquid cooling for NVIDIA’s latest hardware. Intrator frames the moat the same way: “Access to compute gets organizations started, but what helps them succeed is the software, security and operational expertise to build and operate AI at scale.”
CoreWeave is climbing the stack as margin at the bottom thins. As previous-generation GPUs come off contract and flood the hourly rental market, undifferentiated capacity gets cheaper, a dynamic we traced in the two-market split. Renting raw silicon becomes a commodity exposed to whoever has the newest chips and the cheapest power, while a platform that customers build on is stickier and harder for a hyperscaler to undercut. CoreWeave is defending that position with a diversifying customer base, from Microsoft and Meta to trading firm Jane Street, and with take-or-pay contracts under which customers pay for booked capacity whether they use it or not. Those contracts back its debt, including an US$8.5 billion loan closed in March, the first investment-grade-rated facility secured on GPUs and a customer contract. Depreciation and interest reached 81% of revenue in the first quarter, up from 77% a quarter earlier. We map where each operator sits in our neocloud market report.
Australia’s operators sit at the infrastructure layer
Australia’s three main operators carry serious contracted demand. IREN, the largest Australian-founded operator, holds customer commitments from Microsoft and NVIDIA analysed in our coverage of the IREN and NVIDIA agreement, and in June signed a transmission connection for an 800MW campus at Bundey in South Australia. Firmus is building its Project Southgate platform toward 1.6 gigawatts of NVIDIA-based capacity by 2028, funded by a US$10 billion Blackstone and Coatue debt facility drawn against executed contracts. Sharon AI is scaling sovereign capacity behind an NVIDIA agreement worth up to US$4.88 billion.
None of the three has a software platform. Firmus’s Hypercube is an energy-efficiency and reference-architecture play. Sharon AI’s 600-petabyte sovereign data backbone with VAST Data is data infrastructure. Both sit a layer below the model-development tooling CoreWeave now sells. These are infrastructure and power companies at their core, built to secure grid connections and pour concrete. Competing at CoreWeave’s platform layer calls for software talent, an installed enterprise customer base and the balance sheet for acquisitions the size of its US$1.7 billion Weights & Biases deal, none of which the three have assembled.
Secured power and land anchor Australia’s operators
The software-platform layer is globally contestable. Secured domestic power, land and grid connections are not: they take years to assemble and cannot be acquired from offshore. That physical base is what IREN’s 800MW Bundey campus in South Australia, Firmus’s Project Southgate and Sharon AI’s Australian build represent, and Firmus and Sharon AI market theirs as sovereign, in-country capacity.
Counterparty profiles differ too. An Australian-domiciled operator on long-dated contracts, secured power and blue-chip infrastructure debt carries a different risk profile than one whose backlog rests on a handful of AI labs continuing to pay.
IREN signs US$2.8bn as customers prepay the chips
The same week as the CoreWeave interview, IREN signed US$2.8 billion of new AI cloud contracts across nine named customers, including Microsoft, NVIDIA, Perplexity, Figure AI and Together AI, lifting its 2026 annualised revenue target above US$4 billion with about 85% already under contract. Its shares jumped about 20%, after falling more than 40% across the prior month. Customers are prepaying about 45% of the cost of the GPUs up front, which cuts the debt IREN carries before revenue starts. IREN is targeting 480MW of AI cloud capacity in 2026, scaling toward roughly 1.2 gigawatts in 2027.
IREN won that week on the infrastructure playbook: diversified customers, contracted revenue, capital risk pushed onto buyers through prepayments, and secured power and land underneath. That contracted, sovereign-leaning path drew a bid at the same moment the market marked down CoreWeave’s debt.
What to watch
Three markers will show which way the local market turns. The first is whether any Australian operator signals platform ambition, through a software acquisition or a managed-service launch. The second is the pricing of the Firmus and Sharon AI ASX floats, marked against repriced US comparables, which will test whether public investors pay a premium for contracted, sovereign capacity. The third is whether IREN’s prepayment financing spreads to the ASX-bound operators, which would ease the debt-wall worry hanging over the tier.