At a glance

  • Global advisory firm Alvarez & Marsal’s June 2026 report Rise of the Neoscalers maps where AI infrastructure gets built, contracted and financed.

  • Its core claim: contract quality now sets a project’s cost of capital, with the most bankable financing spreads down about 900 basis points in three years.

  • Sovereign and dedicated cloud sit at the top of that bankability ladder, where Australia’s HCF certification gives local operators an edge.

  • Inference leads the demand story at 75%+ of future compute, which is Australia’s real lane.

  • So Australia reads as an inference and sovereign market, with Firmus, Sharon AI and IREN on the funded side.

Rise of the neoscalers

In June 2026 the global advisory and restructuring firm Alvarez & Marsal, better known for turnaround and diligence work than market commentary, published Rise of the Neoscalers, a global read on the AI-infrastructure cycle aimed at investors, lenders and data centre operators. A&M’s term, the neoscaler, covers the challengers to hyperscaler dominance, split into model neoscalers, the frontier AI labs that mostly lease infrastructure, and infra neoscalers, the GPU cloud platforms and AI-grade landlords that own the physical capacity. It is the same tier we track as the neocloud, set out in our neocloud market report.

The report is framed around geography, but its real subject is bankability. Its through-line is that contract quality, not asset quality, is now the dominant driver of a project’s economics, and that the single largest variable in a project’s cost of capital is the credit rating of the anchor tenant. A&M’s backdrop is that the binding constraint is now supply and delivery, with the big five US hyperscalers guiding to more than US$750 billion of capital expenditure for 2026, and every one of them pointing to power and delivery as the limit rather than customer demand.

Contract quality now sets the cost of capital

A&M sorts neocloud contracting into four models, from multi-year anchor take-or-pay through enterprise reservations and mid-market commitments to on-demand spot, with customer concentration running from 60% to 95% of revenue in the top one-to-three customers at the anchored end. Investment-grade, anchored, contracted structures can raise senior project finance at terms competitive with core infrastructure debt. On-demand and short-tenor exposure stays in dearer private credit and equity.

The gap between the two has become structural. A&M puts the spread compression on the most institutional, contract-backed structures at roughly 900 basis points over three years, and cites major US banks projecting US$30 billion to US$40 billion of annual data centre securitisation by 2026 and 2027. It is the same fault line we set out in AI compute is splitting in two, where contracts, not size, decide survival. The rental rate on merchant capacity matters far less than whether a backlog is genuinely take-or-pay and termination-protected.

Australia’s float queue is that thesis in the flesh

Sharon AI and Firmus show the thesis in how they approach the public market. Both locked in large private raises and contracted revenue before bringing a float to the ASX. Sharon AI is targeting an ASX listing in the second half of July 2026, a line of CHESS Depositary Interests alongside its Nasdaq stock with Macquarie Capital running the roadshow, after its oversubscribed US$1.6 billion June raise anchored by Leopold Aschenbrenner’s Situational Awareness fund and Oaktree. Firmus, the larger float, has slipped past September while it builds committed revenue, including a new Indonesian campus. In A&M’s terms they are sequencing the float behind the contracts, because contract quality sets the cost of capital, the same logic we drew from the financing panels at DCD Connect in Bali and from Sharon AI’s Aschenbrenner-backed raise.

Sovereignty sits at the top of the bankability chart

A&M’s contracting hierarchy puts dedicated and sovereign cloud in the concentrated, long-tenor anchor zone, the most bankable position on its chart, and it says sovereignty commands a pricing premium that lets DC-owning operators anchor to Private AI demand in sovereign, health and finance. That is a third party pricing the same sovereign premium we have tracked, where the Australian Government’s Hosting Certification Framework gates government and regulated workloads to in-country, Australian-controlled capacity. NVIDIA reports sovereign AI revenue passing US$30 billion in its last financial year, so the segment is real rather than a marketing line. It is a moat with a ceiling, because the hyperscalers are building their own sovereign regions, with AWS’s European Sovereign Cloud live from January and an announced A$20 billion Australian investment. The durable Australian edge is the combination the market now funds: contracted demand plus secured power, with sovereignty as the premium layer.

Where Australia actually sits on A&M’s map

A&M splits the world into three market types: AI powerhouse markets, low-cost offshoring hubs, and domestic AI and inference markets, the last defined as demand-rich but power-constrained geographies focused on sovereign and local inference workloads. On A&M’s own definitions Australia reads as a domestic AI and inference market, its power expensive at scale and its grid slow to connect, the very time-to-power A&M names as the swing factor. The report’s does list Australia among training-favourable regions, but it concedes the United States still hosts the largest share of frontier training, and independent reads point the same way, with JLL’s 2026 outlook expecting inference to overtake training globally around 2027 and naming speed to power the primary driver of site selection. The training that does land here is a firmed-renewable niche, such as Firmus’s Tasmanian and South Australian sites, not the bulk of the US training overflow.

The operators built for the funded side

The three Australian names are building on the funded side of A&M’s line, and their capital shows it. Firmus is rolling out its Project Southgate AI Factory platform toward 1.6GW of energy-efficient, NVIDIA-based capacity by 2028. It is funded by a US$10 billion Blackstone and Coatue debt facility and a US$505 million Coatue-led equity round at a US$5.5 billion valuation, and in July it added a 600MW South Australian energy supply deal with Gunvor to underpin planned campuses at Tailem Bend and Stirling North. Sharon AI, listed on the Nasdaq as SHAZ and hosting its clusters across NEXTDC facilities, is building 132MW with 102MW already contracted, behind a six-year NVIDIA cloud agreement worth up to US$4.88 billion, the funding structure we set out in how Sharon AI funds 56,000 NVIDIA GPUs on Australian soil. IREN, the Sydney-founded operator at a market value near US$15 billion, has secured a transmission connection agreement for an 800MW campus at Bundey in South Australia, targeting first power from 2028 and reported in trade press at around A$10 billion, on Microsoft and NVIDIA cloud contracts worth US$9.7 billion and US$3.4 billion that we covered in our analysis of its NVIDIA agreement and its South Australian clean-power siting. IREN runs most of its live compute in the United States and Canada, which keeps the point sharp: the Australian edge is contracted power and sovereignty, and its heavy training sits offshore.

The obsolescence clock is the risk under all of it

A&M’s sharpest structural warning is about the hardware. It puts GPU residuals down roughly 75% within three years of launch, and argues that the powered, contracted building outlasts the chips inside it. The best-insulated operators, on A&M’s read, keep GPU financing separate from facility financing and step fleets through frontier training, production inference and batch work as each generation ages. A&M benchmarks AI-grade capacity at US$10 million to US$15 million per MW against US$5 million to US$12 million for conventional cloud, at rack densities of 50kW to more than 200kW, which is why CapEx discipline reads as the yield differentiator, and why efficient, power-led builds like Firmus’s compress the premium. For an operator, firmed power and long contracts matter more than headline GPU counts.

What to watch

Five markers will show whether Australia converts the contracted position A&M’s framework rewards. The first is Sharon AI’s second-half-July ASX listing and the timing of the larger Firmus float. The second is grid-connection and firming timelines on the eastern seaboard and at IREN’s Bundey site, the binding constraint on how fast contracted capacity energises. The third is the first neoscaler contract-renewal cycle, which A&M flags as untested and which will show whether contracted demand holds when leases roll. The fourth is the NSW data centre inquiry due to report on 30 September 2026. The fifth is whether the US$30 billion to US$40 billion in data centre securitisation A&M expects reaches Australian issuers and lowers the cost of capital for contracted local capacity.