Huang says the risk is low. He also says his borrowers are not investment grade.

What Huang described on Wednesday is credit arbitrage, and he described it accurately. Nvidia can raise money cheaply. Its customers cannot. He said the frontier labs are "not investment grade" and lack the capital track record to secure funding at low cost, "and this is where Nvidia could be helpful" ^[7]. That is not a short seller's characterization. It is the seller of the equipment explaining that the buyers cannot finance the purchase on their own terms.

The defense against being left holding the assets is redeployability: the infrastructure can be moved across customers and workloads, which limits exposure to any single investment ^[8]. That works when a failure is specific to one company. It does not work when the failure is about the price of frontier compute. If a lab cannot service the obligations behind a leased campus, the likely reason is that training and inference revenue came in below the cost of the racks, and that is also the reason the next tenant bids less for the same racks. Counterparty quality and collateral value run on one variable, and they move together.

The scale, in Nvidia's own units. The Ohio backstop of $105 billion, with OpenAI as the tenant, is about 1.09 times the $96.2 billion the company booked in the quarter ^[5][2][13]. The arrangement with Wall Street firms to line up as much as $500 billion for data centers is roughly 1.3 times the annualized run rate of that same quarter ^[6][14], though Nvidia is arranging that money alongside banks rather than writing all of it ^[6]. Both sit on top of equity positions in model makers including OpenAI and Anthropic, and in neoclouds that rent out Nvidia-powered capacity ^[10].

Then there is the number nobody publishes. Data centers produced $89 billion of the $96.2 billion, about 93 percent of the company ^[3][2][15]. The CNBC account contains no split between customers who paid with capital they raised independently and customers who paid with capital Nvidia helped arrange ^[17]. Without that split, the guide of roughly 70 percent revenue growth for fiscal 2028 ^[4] cannot be stress-tested from outside. An analyst has no way to say which portion of it survives a funding market that closes.

Shareholders have already been charged something for the ambiguity. Revenue more than doubled from a year earlier while the shares are up 12 percent this year ^[2][11]. The two windows do not align exactly, so treat this as directional, but a doubling of sales against a 12 percent move in the price puts the trailing price-to-sales multiple near half of where it was ^[16]. The 4 percent after-hours reaction to the print does not disturb that arithmetic ^[11].

"The risk is low" ^[1] is a statement about probability. It is not a statement about size, and Nvidia's equity owns the size. Holding the stock at this level means underwriting sub-investment-grade compute buyers at Nvidia's cost of capital ^[7], and being paid for it in a multiple the market has spent the year cutting ^[16].