Nvidia no longer just wants to sell chips. On Wednesday, the company announced a new financial scheme for AI-focused cloud providers: they can access large GPU volumes without upfront payment, repaying the investment based on the revenue they generate. The idea is to put the compute power needed to train and run Large Language Models into the hands of startups and small businesses that would otherwise be locked out.

How it works (what we know)

Precise details of the deal were not made public, but the framework is clear. Instead of asking cloud providers to buy tens or hundreds of H100 GPUs with million-dollar outlays, Nvidia extends credit and shares business risk. The provider pays an amount tied to what it earns from end customers, turning a fixed cost into a variable one. For startups, this means skipping the upfront capital barrier and immediately developing models on high-end infrastructure.

It’s a shift for a company that until yesterday dominated hardware with record margins. Now it steps into financial services territory while still being a silicon supplier, aiming to broaden demand and lock in its dominant position before competitors – from large cloud vendors’ custom chips to AMD and Intel projects – erode its market share.

Why the move matters, even for those eyeing on-premise

The initiative signals a still-constrained GPU market. Renting cloud capacity with a pay-as-you-go model is already common, but cost remains prohibitive for many, and provisioning times are not instant. Nvidia’s credit acts as a multiplier for smaller, specialized cloud providers, increasing competitive pressure on the large hyperscalers and, in theory, lowering prices for everyone.

For those evaluating on-premise LLM deployment, this scenario introduces an interesting trade-off. On one hand, the cost of capital drops if you can start in the cloud with deferred payment, postponing the decision to buy servers. On the other, the advantages of local infrastructure remain – latency, data control, medium-term cost predictability – and AI-RADAR offers analytical frameworks to compare TCO and sovereignty requirements. Nvidia’s offering makes the transition more gradual, but it does not remove the need for a reasoned deployment strategy.

Behind the move, one can also read a bet: that generative AI will keep growing and that financial elasticity will tempt more companies to experiment. If so, GPU demand will increase further, and the supply chain will have to adapt to a market where what matters is not just silicon performance, but also the creativity of the business models used to bring it to developers.