At a glance

  • In early July, investors sold off the neoclouds, firms that rent out GPU power such as CoreWeave and Australia’s IREN, on reports Meta may resell its own spare compute. Many read it as the AI boom peaking.

  • The prices operators charge say otherwise. Rent for the newest chips is still rising, with one-year H100 contracts up about 40% since October and B200 up about 24% in early 2026, while only older chips are getting cheaper.

  • The market is splitting in two: a scarce frontier of new, contracted capacity, and a cheaper pool of older, interchangeable capacity. What decides survival is contracts, not size, with operators on long-term deals safe and those renting capacity out short-term on heavy debt exposed.

  • Australia’s operators, Firmus, Sharon AI and IREN, mostly sit on the safe side, building capacity backed by signed NVIDIA and hyperscaler contracts, secured power and blue-chip capital like Blackstone.

Confidence in the AI build-out cracked in June

Through June the AI trade cracked. The Nasdaq had its worst day since April 2025 on 4 June, and a second global leg on 23 June sent Korea’s KOSPI down 10% and wiped more than US$1 trillion from world markets. Coverage turned anxious: NPR asked whether AI is “one big bubble”, Fortune aired the view that “2026 is looking like 1999”, and Capital Economics noted that selloffs this sharp had previously come only in bear markets like the dot-com crash and the Great Financial Crisis. On 28 June the Bank for International Settlements put an AI capex bust and opaque circular financing among the top risks to the financial system, likening the build-out to the canal and railway manias that ended in reversals.

Then on 1 July the alarm found a target. Bloomberg broke the news, carried the same day by CNBC, that Meta plans to resell surplus compute through a business called Meta Compute, and the neoclouds it buys from slid at once, CoreWeave and Nebius off about 15% and Australia’s Nasdaq-listed IREN dropping with them. 24/7 Wall St asked whether it was the end of the neocloud boom.

The prevailing view is that the boom is topping and the neocloud model is breaking, though some, including NVIDIA’s Jensen Huang, called the month’s selloff a buying opportunity. The prices operators charge tell a more divided story: the compute market is splitting in two.

The frontier is still scarce

At the frontier there is no glut. On SemiAnalysis’s index, one-year H100 contract prices have risen about 40% since October 2025. Silicon Data puts the B200 rate up about 24% across the first quarter of 2026, concentrated in a March spike, with capacity for immediate hire booked out into the third quarter. Hyperscaler capex is still climbing: combined 2026 guidance across Microsoft, Alphabet, Amazon and Meta runs near US$700 billion, up sharply on last year.

The binding constraint has moved rather than eased. Microsoft’s Satya Nadella frames the bottleneck as power, not silicon, describing chips sitting in inventory with no energised shell to plug them into. With interconnection queues and multi-year lead times on grid and transformers, that constraint cannot loosen on a 2026 timeframe, so scarcity is migrating from chips to memory and megawatts rather than ending.

Only older capacity is getting cheaper

The falling prices are real, but they are confined to older, last-generation chips rented out by the hour, where a glut of previous-generation Hopper GPUs has pushed rates down. The newest Blackwell chips, and capacity locked into long contracts, have held firm. Meta reselling spare capacity would add supply to that cheaper, interchangeable pool, not to the scarce frontier that is still booked out.

The plan to resell compute is reported rather than confirmed, and the surplus framing that travelled with it does not hold up. AWS co-founder Benjamin Black has said the idea that Amazon built AWS from spare retail capacity is a myth, and today’s would-be sellers are themselves short of compute, with Microsoft telling investors it expects to stay capacity-constrained through 2026. A hyperscaler rationing its own capacity looks more like an offensive move into cloud, and a signal to investors on capex, than a clear-out of idle hardware. We set out the market reaction in our coverage of Meta’s move to become its own neocloud.

What decides survival is contracts, not size

The fault line runs between capacity locked into long-term customer contracts and merchant capacity, meaning GPUs rented out on the open market with no committed buyer. Independent neoclouds carry more than US$20 billion of GPU-collateralised debt, on PitchBook’s tally, at rates that have reached low double digits, while hyperscalers fund from cash flow. That financing gap is the genuine structural asymmetry.

The disciplined neoclouds have answered it by locking in long-term revenue. CoreWeave reported a revenue backlog near US$99 billion in the first quarter, with nearly all of its quarterly revenue coming from take-or-pay contracts, so a softer rental rate does not reprice the signed book. For that tier the real exposure is refinancing depreciating hardware at a moment of doubt, plus concentration on a handful of large customers. The rental rate matters less. The operators that carry real risk are the ones renting undifferentiated capacity on the open market with debt to service and no committed customer, a distinction we drew from the financing panels at DCD Connect in Bali.

The capital is already pricing the split

Blackstone’s US$10 billion facility for Firmus, closed in February 2026 with Coatue, is long-dated infrastructure debt drawn against executed long-term contracts with hyperscalers and blue-chip AI customers rather than merchant spot revenue. Private credit is funding the contracted side of this market and steering clear of speculative capacity.

The equity signal points the same way. Leopold Aschenbrenner’s Situational Awareness fund, built on the thesis that compute and power are the binding constraints on AI scaling, has concentrated its bets on energy and infrastructure rather than the model builders, and is reported among the backers of Sharon AI’s July ASX raise. Debt and equity are converging on the same view: fund contracted infrastructure and secured power, not merchant compute.

Australia’s independents sit on the contracted side

Australia’s independents mostly sit on the right side of the line, and they sit there on physical build. Firmus is rolling out its AI Factory platform, Project Southgate, across multiple Australian sites toward 1.6 gigawatts of energy-efficient, NVIDIA DSX-based capacity by 2028, financed by the Blackstone and Coatue facility that draws against executed long-term customer contracts. Sharon AI is scaling sovereign AI factory capacity in Australia behind an NVIDIA agreement worth up to US$4.88 billion. IREN, the largest Australian-founded operator at a market value near US$15 billion, has run its halls in the US and Canada, but in June it signed a transmission connection for an 800MW campus at Bundey in South Australia, reported at around A$10 billion and targeting first power from 2028, the clean-power pitch we covered in our South Australia data centre strategy report. It sits on contracts with Microsoft and NVIDIA valued at US$9.7 billion and US$3.4 billion, set out in our analysis of that agreement. These are contract values and customer commitments rather than guaranteed floors, but that offtake, plus the power and land behind it, is what keeps an operator off the merchant tail.

Sovereignty adds a second layer, and it is a real premium segment rather than a marketing line. NVIDIA reports sovereign AI revenue passing US$30 billion in its last financial year, and the Australian Government’s Hosting Certification Framework gates government and regulated workloads to in-country, Australian-controlled capacity, with its separate March 2026 national expectations for data centre and AI infrastructure developers pushing in the same direction. It is a moat with a ceiling, though, because the hyperscalers are building their own sovereign regions, with AWS’s European Sovereign Cloud live from January and its announced A$20 billion Australian investment. The durable edge for Australia is the combination the market is now paying for: contracted demand plus secured power, with sovereignty as the premium layer on top. Our neocloud market report tracks where each operator sits.

The state of play, and the strains to watch

The build-out is not slowing, it is concentrating. Demand for frontier compute still outruns supply, and the binding constraint has moved from chips to power: grid-connection queues and multi-year lead times on transformers now set how fast new capacity can come online, and will shape the next two years more than chip availability does.

By our tracking of the Australian market, three supply and demand patterns give real cause for concern. Demand is concentrated, the long-term contracts that make a neocloud bankable rest on a handful of AI labs, so the order book holds only as long as those buyers keep paying, the circular-financing risk the Bank for International Settlements flagged. Hardware depreciates fast and debt is dear, so operators building capacity to rent on the open market face refinancing against falling collateral if sentiment turns. And if the resale model spreads beyond Meta, the older, undifferentiated layer is where prices would compress first. For Australia the strain is physical rather than financial: demand has clustered on the eastern seaboard, firming and transmission are the gating items, and the operators that come through will be the ones that bring their own contracted power, as IREN’s Bundey campus shows.