Nvidia's chief executive Jensen Huang has characterised the deal as creating an entirely new asset class. Just_Super/iStock.
17 Aug 2026, 2:28 pm
Last week, Nvidia announced a landmark $500 billion financing arrangement under which technology companies will be able to lease rather than purchase its chips outright, with financing provided by some of the world's largest private capital firms – including Apollo Global, KKR, Brookfield, BlackRock and Goldman Sachs. Nvidia's chief executive Jensen Huang has characterised the deal as creating an entirely new asset class, backed by GPU hardware and accessible to the $22 trillion private capital industry at a time when tech investors are wary of the uncertainty over how AI will impact on business models.
The question for businesses currently wrestling with the procurement challenges we recently highlighted – allocation queues, capacity reservation models, pricing volatility and deposit requirements – is whether a mature leasing market might address some of those structural difficulties. The answer is nuanced.
Before examining the implications for GPU procurement, it is worth briefly setting out the mechanics of the deal.
Rather than technology companies purchasing chips outright – which requires substantial upfront capital – the Nvidia arrangement allows them to lease GPUs from a speciality finance vehicle funded by Wall Street institutions. In broad terms, the structure operates as follows:
Loan terms are currently structured over three to five years, reflecting assumptions about how quickly chip values decline – though, as discussed below, the Nvidia deal is explicitly premised on the proposition that those assumptions may be too conservative.
The core appeal of the GPU leasing model, from a supply-side perspective, is that it disaggregates the problem of access from the problem of capital. Under a leasing structure, a speciality finance vehicle purchases chips from Nvidia and leases them to the end user, which pays a regular rental fee rather than committing to a large upfront capital expenditure. For businesses that have struggled to commit the funds required to secure future allocation rights – one of the key dynamics we identified in our earlier analysis – this could reduce the initial financial barrier to obtaining GPU access. Leasing also offers greater flexibility in the event that a business's AI infrastructure requirements change over the course of a project.
However, the leasing model does not directly increase the volume of new chips entering the market. Nvidia, as the chip designer whose products are fabricated by foundries such as TSMC, faces the same production capacity limitations regardless of whether downstream customers lease or purchase. Businesses should therefore not expect a leasing model to eliminate allocation queues or remove the need for the advance planning and contractual commitments we previously described.
That said, the deal does carry an indirect supply-side implication that is worth noting. If Nvidia's contention holds – that older chip generations retain commercial utility for longer than previously assumed, remaining productive for less compute-intensive workloads such as inference and query processing – then the effective pool of available computing capacity could expand over time, as hardware that would previously have been retired is instead redeployed within leasing pools.
There is a secondary effect worth considering here: if earlier-generation hardware absorbs demand from applications that do not require the highest-specification chips, this could in turn free up capacity on newer high-performance GPUs for the training and high-complexity workloads that genuinely require them.
This would not resolve acute near-term shortages of the latest hardware, which remain the primary bottleneck for businesses training cutting-edge AI models. It may, however, offer meaningful relief for organisations whose workloads can be served by earlier-generation chips, and – through the reallocation of demand across chip generations – contribute modestly to easing pressure on the newest and most constrained parts of the market.
Businesses and their legal advisers should not treat a leasing model as a straightforward solution to the challenges we previously identified. In several respects, the risks are not eliminated but rather repackaged and, in some cases, intensified.
The central tension in the Nvidia financing deal – and the point on which its entire structure pivots – is the question of how quickly GPUs lose their value. Conventional wisdom, and current lender behaviour, has assumed that chips depreciate rapidly as technology advances, with lenders typically requiring full repayment within three to five years. The Nvidia deal, however, is explicitly premised on a different and more optimistic assumption: that GPUs are holding their value for longer than previously anticipated.
Nvidia points to several factors in support of this view.
As pointed out above, demand for computing power across all chip generations has remained intense, with older hardware remaining commercially viable for less demanding AI tasks such as inference and query processing, even as newer chips handle cutting-edge model training. Nvidia's proprietary Cuda software platform – the ecosystem that enables GPUs originally designed for graphics to accelerate AI workloads – is continuously updated, extending the functional life of existing hardware and reinforcing customer dependency on Nvidia's architecture. Nvidia also cites observed market data: its six-year-old A100 chips remain in active commercial use, and rental pricing for newer chips has increased rather than declined.
Nvidia has backed this conviction with its balance sheet, guaranteeing that leased chips will retain at least 25% of their value throughout the lease term – absorbing first losses on any depreciation beyond that floor.
The risks, however, are real. A significant improvement in model efficiency – where AI systems achieve equivalent results with substantially less compute – could rapidly erode demand for older hardware and undermine residual value assumptions. The Cuda ecosystem, while currently dominant, faces competitive pressure from alternative software platforms, and the rental pricing data on which Nvidia relies reflects an exceptionally strong demand environment that may not persist.
Businesses entering into GPU leases on the basis of current residual value projections should be alert to the possibility that those projections embed assumptions that have not yet been tested across a full market cycle.
Various technology analysts have warned of the risk of overbuild – a scenario in which data centre construction races ahead of actual AI demand, leaving both lessors and lessees exposed to underutilised capacity. For businesses structuring AI projects around leased GPU capacity, this risk cuts both ways: either demand for their own AI applications falls short of projections, or broader market oversupply depresses the value of the capacity they have committed to.
The leasing model introduces a new layer of contractual complexity on top of the issues we previously identified. Rather than focusing mainly on allocation rights, delivery commitments and cancellation provisions, businesses will now need to assess:
The points we made in our earlier article regarding export controls, sanctions compliance and end-user restrictions remain equally applicable in a leasing context. However, a leasing structure adds a further dimension: where title to the chips remains with the lessor throughout the lease term, questions may arise as to which party bears responsibility for ongoing regulatory compliance, how end-user certificates are structured, and how re-leasing or redeployment of chips at the end of one lease term is treated for export control purposes. These issues should be addressed expressly in any leasing documentation.
The emergence of GPU leasing as a mainstream financing model is a significant market development, but it does not fundamentally alter the legal analysis that procurement teams and their advisers need to undertake. If anything, it adds further layers of risk to be allocated carefully.
Businesses evaluating GPU leasing arrangements should ensure that their legal teams focus on:
The Nvidia deal signals that the GPU market is maturing rapidly and that sophisticated financial structures are beginning to emerge around AI hardware as an asset class. For businesses navigating GPU procurement, a leasing model may over time offer a more accessible and flexible route to securing the computing capacity their AI projects require – particularly for those that have been priced out of the upfront capital commitments that current allocation models demand.
However, the fundamental challenges we recently identified – allocation risk, pricing volatility, technological obsolescence and contractual complexity – have not disappeared. They have simply migrated into a new set of documents and a new set of counterparty relationships. The leasing model redistributes risk; it does not eliminate it. Legal and commercial analysis remains as important as ever
Written by
