Nvidia now agrees to rent back unused GPU capacity from neocloud operators if customer ...

Sharon AI’s six-year, 40,000-GPU deployment in Australia was the first deal built this way, followed swiftly by Firmus Technologies committing up to 170,000 GPUs across a 360MW facility in Batam, Indonesia. The mechanism is elegant on paper. A neocloud that can’t get a $500 million GPU loan on its own credit gets one once Nvidia is standing behind the residual value. But look closely at what that actually means, and Nvidia isn’t just derisking a niche corner of the GPU cloud market. It’s becoming lender of last resort, chip supplier, equity holder and demand backstop across an increasingly large share of the same infrastructure, all at once. How the backstop actually works The structure itself is straightforward. Nvidia guarantees a minimum utilisation floor on deployed GPUs, in Sharon AI’s case a six-year commitment, and shares in the resulting cloud revenue in return. If a neocloud’s customer demand underperforms, Nvidia commits to renting the idle capacity at a predetermined rate, absorbing the downside that would otherwise sit with the lender. That’s precisely the mechanism Capacity flagged in its own coverage of CoreWeave’s Q2 debt figures , where total borrowing reached $35 billion by the end of June, describing the backstop as aimed “squarely at the smaller merchant GPU cloud operators carrying the refinancing wall”. It isn’t confined to smaller players, either. Nvidia has separately been in talks over a reported $250 billion backstop to help OpenAI lease a 10-gigawatt data centre campus in Ohio being developed by SoftBank’s SB Energy, a deal that would sit alongside a further $350 billion in prospective chip financing for the same project. The same logic that makes a 40,000-GPU deployment in Australia bankable is now being scaled to underwrite some of the largest single infrastructure projects in the AI sector. Nvidia frames this as simply extending what compute already is. Chief executive Jensen Huang has argued the shift reflects a genuine change in how the industry should value its hardware: “These are revenue-generating assets now. They’re productive, they’re long-lived, they’re fungible, they’re flexible.” That framing underpins the much larger $500 billion Wall Street financing platform Nvidia confirmed with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR this month, of which Nvidia may itself backstop up to $125 billion. CNBC’s David Faber, moderating the joint announcement interview, noted the appeal for lenders directly: “that conceivably will lower the cost of the capital.” The circularity problem hiding in plain sight Here’s where it gets harder to untangle. This isn’t the first time Nvidia has sat on both sides of a neocloud’s balance sheet. CoreWeave’s own 2025 IPO filing disclosed that Nvidia held an equity stake of roughly 1.21% in the company while simultaneously serving as both a major customer and a capacity backstop. The new AI Compute Partnership formalises and extends that same pattern across a wider set of smaller operators: Nvidia sells the GPUs, guarantees the revenue that makes the loan bankable, and, in some structures, earns a share of the resulting cloud revenue too. The practical effect, as Capacity’s reporting on the CoreWeave figures pointed out, is that vendor-financed capacity and independently financed capacity are no longer directly comparable products, since a backstopped neocloud is carrying Nvidia’s revenue-share obligations into its own pricing. For enterprise buyers shopping between GPU cloud providers on headline rate alone, that’s a distinction that isn’t visible on the invoice. For the lenders extending credit against that hardware, it means they’re not really underwriting the neocloud’s business at all. They’re underwriting Nvidia’s willingness and ability to make good on the backstop if utilisation falls short, across potentially hundreds of these arrangements at once. That’s a subtly different risk to the one being priced. A neocloud defaulting on a conventional loan is a contained, idiosyncratic event. A widespread GPU residual value shock, triggered by a faster-than-expected hardware refresh cycle or a genuine slowdown in AI demand, would hit Nvidia’s backstop commitments across Sharon AI, Firmus and every subsequent deal simultaneously, at the same moment Nvidia’s own equity, its chip sales and its Wall Street financing platform would all be under pressure too. The company isn’t just correlated with the sector’s fortunes anymore. In this structure, it effectively is the sector’s fortunes. What this means for the refinancing wall None of this is abstract for operators watching the so-called neocloud refinancing wall building through 2026 to 2028. The backstop model exists precisely because that wall is real: debt taken on since 2023 to finance GPU fleets, much of it at loan-to-value ratios in the region of 60 to 70%, is now maturing into a market where lenders have grown considerably more cautious about pricing residual value on their own. Nvidia’s willingness to step in as guarantor has genuinely unlocked capital that likely wasn’t available on comparable terms even six months ago. That’s not a small thing for operators trying to finance gigawatt-scale build-outs on compressed timelines. But unlocking financing and removing risk from the system are not the same thing. Whether the backstop model resolves the refinancing wall or simply relocates the same risk onto a single, highly concentrated balance sheet, one that also happens to control chip supply, pricing and the pace of the next hardware generation, is the question serious lenders and data centre financiers should be asking before they treat a backstopped deal as meaningfully de-risked. If GPU values hold, the arrangement performs exactly as designed. If they don’t, the company sitting behind the guarantee is the same one that determined how fast the collateral would depreciate in the first place. RELATED STORIES 8 sessions to watch at Metro Connect Fall x Datacloud USA 2026 Data centres could use 2.7% of Texas’s water by 2030: Where they land matters Datacloud USA 01 September 2026 Register now See event details Nadine Hawkins Director of Content and Insights Share Share on LinkedIn Share on Twitter Copy Link URL Top Articles Top articles --> News Lenders start pricing community opposition into data centre financing as $130bn in projects stall 14 August 2026 3 minutes data centre Nebius to deploy AI capacity at Vantage’s Newport campus in first South Wales AI Growth Zone commitment 14 August 2026 3 minutes News Digital infra’s 300,000-worker gap: Veterans could fill it, report finds 14 August 2026 4 minutes News Nvidia now agrees to rent back unused GPU capacity from neocloud operators if customer demand falls short 14 August 2026 5 minutes data centre North America data centres absorbed a record 25GW in H1 2026 13 August 2026 3 minutes