The person who has to sign off a 2027 cluster budget is looking at a demand case built from a stack of announced buildouts by named buyers, and the useful test on each line is not how big the order is but where the buyer got the money.
The composition of Nvidia's holdings is the part that changes a spreadsheet. Business Insider's breakdown splits the book into roughly $48bn of publicly traded stocks and other marketable securities, roughly $48bn of private and non-marketable holdings, and about $3bn of equity-method investments [5]. That puts about 48 percent of the total in positions no market prices [1]. Funding rounds price them, and as the reporting puts it, nobody exits those on a bad morning [8].
The legible half is not diversified either. Nvidia's disclosed US stock positions as of 30 June included $30bn in Intel, from an investment that cost $5bn [6]. The $25bn of gain on that one holding is about 52 percent of the entire marketable bucket [2]. Worth noting for anyone reconciling the two disclosures: Intel at $30bn and SpaceX at $21bn sum to $51bn, which is $3bn more than the whole marketable half reported four weeks later [3]. The dates differ, so at least one of those positions sits outside that bucket, and the material does not say which.
The mechanism does something different from what the earnings call calls it. CFO Colette Kress told analysts that frontier labs outgrow their own balance sheets and credit profiles and cannot secure AI factory infrastructure alone, so Nvidia is needed to help power this flywheel [15]. Naveen Chhabra of Forrester told CNBC the same money gives infrastructure financiers, specialised clouds and model labs the balance-sheet strength to buy tens of thousands of Nvidia GPUs [16]. At the component layer, Chhabra argues that funding Coherent and its peers keeps their tooling and design work optimised for Nvidia's architecture, raising switching costs and defending the CUDA moat against AMD and against the clouds' own custom silicon [18]. Ian Fogg of CCS Insight put the softer version of the control question: the investments help companies innovate, and they also give Nvidia influence over whether the innovation takes an Nvidia-shaped path [17].
None of this material shows revenue coming back around. Michael Burry says Nvidia is overreaching by financing and investing in the customers for its own chips, and Mark Cuban called the dependence truly scary, but neither has produced a number [19]. CoreWeave, taking its $2bn, said the proceeds would go to land, power, infrastructure, research and hiring rather than to buying Nvidia chips [20]. Equity that arrives does not have to leave as a purchase order.
The forcing function for a demand model has two axes rather than one. First, is this buyer's purchasing capacity independent of Nvidia's capital. Second, is the position behind it marked by a market or by a round. Independent buyer, market-priced mark, and the order is clean signal. Nvidia-funded buyer, round-priced mark, and you are modelling Nvidia's willingness to keep writing checks. The visible size of that willingness is more than $40bn committed to financing deals during 2026 plus $25bn of investment commitments still outstanding, or about $65bn [4][7]. And the $12.93bn Hugging Face purchase counts as an acquisition rather than a stake, so it sits on top of the $99bn instead of inside it [12][14].
