For years the neat way to describe Nvidia was that it sold shovels in a gold rush: it did not have to dig for anything, it just took money from everyone who did. On July 1 the company quietly changed the deal. In a blog post co-authored by chief financial officer Colette Kress, Nvidia announced a revenue-sharing and credit-support programme for AI cloud operators. From now on it will help smaller providers buy its chips, and in return it will take a recurring slice of whatever those chips earn. Nvidia is no longer content to sell the shovels. It wants a cut of the gold.
The mechanics are more interesting than the branding. The first two partners, Australian firms Sharon AI and Firmus Technologies, have committed to roughly 210,000 of Nvidia's Grace Blackwell GB300 GPUs between them, with Firmus building a 360-megawatt data centre on Batam Island in Indonesia. Deals of that size run into a wall that most people outside finance never see: nobody wants to lend against the hardware. A cluster of top-end GPUs that costs hundreds of millions today may be worth far less in eighteen months, when the next architecture ships. Only Nvidia knows how steep that fall will be, because only Nvidia controls when the successor arrives.
That is the knot the new programme cuts. By guaranteeing to buy back idle capacity at a set price, Nvidia puts a floor under the value of its own chips, and a floor is exactly what a nervous lender needs. Suddenly a loan that no bank would touch becomes bankable, because the bank is really lending against Nvidia's balance sheet rather than the machines. On top of that, cash-strapped startups can draw token credits against future compute and start training immediately, with Nvidia collecting on the back end as usage piles up.
For Nvidia the appeal is obvious. It earns once when it sells the chip, then again every time a token is generated on it. That turns lumpy, cyclical hardware sales into something closer to recurring, usage-linked income, the kind of revenue Wall Street rewards with a richer valuation. It also extends a pattern the company began with CoreWeave in 2025, when it agreed to buy that firm's unsold capacity through 2032, and reporting suggests a similar backstop is being negotiated for OpenAI's data-centre plans.
Here is where the coverage splits. Trade and startup outlets framed the move as a democratising one: smaller operators who could never front ten billion dollars in chips can now scale quickly, and access spreads. Tech Times was blunter, reaching for the phrase the blog post carefully avoids. This is vendor financing, the same structure telecom-equipment makers used to prop up their customers during the late-1990s fibre boom. When those customers could not fill the capacity they had been lent the money to build, the losses flowed straight back up the chain, and almost everyone was hurt.
The bet, in the end, is a bet on demand. If the AI cloud keeps filling up, Nvidia has invented a machine that compounds its own growth and locks partners into its stack. If demand stalls or capacity sits idle longer than expected, Nvidia has quietly agreed to absorb the shortfall. The company that spent the boom insulated from everyone else's risk has just tied its fortunes to theirs. Over the next two years that decision will look either brilliant or reckless, and there may not be much room in between.