Nvidia CFO Colette Kress Defends AI Customer Financing Strategy

The finance chief behind Nvidia's AI pitch

Jensen Huang usually gets the attention when Nvidia talks about artificial intelligence. Over the past two months, though, the most important explanations of where Nvidia's money goes have come from Chief Financial Officer Colette Kress. On one August earnings call she gave a revenue forecast well above what analysts expected, warned that profit margins would get thinner, explained a large increase in supply commitments, and defended Nvidia's practice of investing in the AI labs that buy its chips. The sentence most people repeated was a short one: "We recognize the scale of this support, and we know some will call this circular financing. We see it differently."821

That sentence has kept coming back. In October, a report on OpenAI's revenue pulled Nvidia shares down, and once again investors are asking whether the money behind the AI boom comes from end customers or partly from Nvidia itself. Kress and investor-relations chief Toshiya Hari held an investor meeting on October 13, which shows the company is still making its case in person.25 This article argues that Kress's view of the financing deals is now the main thing holding up Nvidia's valuation, and that it matters more right now than any product launch.

What Kress actually promised

Coverage of the August numbers broadly agrees. Second-quarter fiscal 2027 revenue was $96.22 billion, up about 106% from a year earlier. That beat consensus, and adjusted earnings of $2.22 per share also came in ahead of estimates.1015 Data center revenue was $89 billion, up 117%, and made up about 92% of total sales.10 Guidance for the third quarter was about $108 billion, which would be Nvidia's first quarter above $100 billion.10

The bigger story was Kress's look ahead. She said Nvidia expects revenue to grow about 70% in fiscal 2028, compared with the roughly 44% analysts had been modeling. She added that customer forecasts suggest growth could double, and that the 70% figure is limited by supply, not demand.1018 She also said capital spending by the top five hyperscalers should rise from about $800 billion in 2026 to $1.3 trillion next year.10 Benzinga reported that Nvidia stock turned higher after hours once she gave the outlook, so the CFO's guidance moved the market more than the quarterly results did.10

Kress also pointed to a growing list of customers. She said the Vera Rubin platform, which follows Blackwell, has orders from every major hyperscaler, AI cloud provider and system maker.8 On China she was cautious: she said the forward outlook includes no China data center compute revenue because of geopolitical uncertainty.18

The margin and supply trade-off

Kress was direct about costs. Nvidia expects gross margin to fall before bottoming out in the fourth quarter of fiscal 2027 at 71% to 72%, mainly because memory prices are rising.10 "We want to be direct about this, rather than let it linger as an open question," she said, and argued that the memory shortage is largely caused by the AI buildout itself.10

The balance sheet shows what that costs. Nvidia's supply and capacity commitments through 2032 rose to $279 billion from $119 billion, mostly for memory.8 Kress also said supply would probably remain a bottleneck at least through the end of fiscal 2028.8 She has made this argument before. In November 2025 she said Nvidia's balance sheet is a competitive advantage because suppliers trust the company's purchase orders and forecasts.5 The August disclosures show how far Nvidia has taken that approach: it is using its cash to secure scarce parts before competitors can.

The practical message is that Nvidia is trading some margin for supply. At 71% to 72%, gross margin would be noticeably lower than the 74.9% GAAP figure reported for the quarter ended April 2026.1 If demand holds up as Kress expects, that is a reasonable trade. If it weakens, Nvidia will be left with large fixed purchase obligations.

"Paid twice" and the concentration question

Kress has also explained how Nvidia plans to deal with its reliance on a few large buyers. Three direct customers made up 44% of Nvidia's $177.8 billion revenue in the first half of fiscal 2027, up from two customers making up 35% a year earlier.16 Her answer is a revenue-sharing model. "In this model, we get paid twice—once on the hardware sale and again through the share of rental revenue," she told analysts.16

The idea sounds attractive, but it ties Nvidia's income even more closely to how its customers' businesses perform. Critics have noticed. Steve Eisman, citing Nvidia's 10-Q, said the top five direct customers accounted for 70% of accounts receivable in the quarter.16 Kress has also said the compute Nvidia ships goes to investment-grade customers or to customers backed by one.3 That is meant to reassure investors, but it also concedes that credit quality is now a fair question.

Where the reporting diverges on financing

Coverage of Kress's "we see it differently" argument splits most clearly here. Her case is that frontier AI labs are growing faster than their balance sheets can fund. She describes Nvidia's investments as a way to fill that gap and says the returns should be excellent while the risk is limited.21 She has put Nvidia's commitments to frontier labs at nearly $50 billion and described that as a small share of expected free cash flow.21 Nvidia has also lined up six large asset managers, Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, to unlock about $500 billion of financing for chip buyers.21

Critics count the exposure differently, and the figures vary from outlet to outlet. One investor commentary puts Nvidia's guarantee obligations as high as $108.5 billion.22 An analysis in The Conversation describes a guarantee of up to $105 billion for an OpenAI data center in Ohio.27 Bloomberg reporting, as summarized in a reference compilation, discussed talks that could exceed $750 billion in total, including a possible $250 billion guarantee for OpenAI's Ohio lease. That compilation notes the terms were unconfirmed.26 Huang has said Nvidia might cover up to 25% of a project's cost through a residual-value support arrangement. That would likely keep the obligation off the balance sheet at first, but Nvidia would still be on the hook if a customer defaulted.26

The gap between "nearly $50 billion" and these much larger figures matters most. Kress is counting equity invested in labs. Critics are adding guarantees, leases and residual-value promises. Both can be correct, because they are measuring different things. Still, the narrower figure lets Nvidia describe its risk as modest, and that is a choice about how to present the numbers.

The October stress test

In October, outside events put these arguments to the test. On October 8, Nvidia fell 2.94% to $230.48 after the Financial Times and Reuters reported that OpenAI's annualized revenue run rate was about $50 billion. Figures circulating earlier had been around $68 billion to $70 billion. Oracle and CoreWeave fell more sharply. Shortly before that, Nvidia and SoftBank reportedly put in a final $20 billion to close OpenAI's pre-IPO round. Venture commentators read this as the investor group getting smaller rather than larger.28

The Conversation laid out the risk in plain terms: if OpenAI-type customers cannot earn enough to meet their commitments, Nvidia loses twice, first through lower sales and then through its financing exposure.27 That is the reverse of Kress's "paid twice" model. The structure that lets Nvidia collect two income streams in good times also leaves it exposed twice when conditions worsen. Regulators have said something similar. The IMF and the Bank for International Settlements have both described AI circular financing as a systemic risk.26

Reading the CFO's own signals

Kress's personal stock sales have been routine. On September 17 she sold 34,918 shares, about $7.6 million, under a Rule 10b5-1 trading plan she adopted on June 16. On September 16, 40,747 shares were withheld to cover taxes on vested stock.1112 After the sale she held 765,103 shares directly, which means the sale cut her direct stake by about 4.4%.15 Because the plan was set in advance, the sale tells investors little about her current view.

Valuation figures vary depending on the source and the measure. MarketBeat listed a P/E of about 28, while one investor column cited a trailing P/E of 46 and about 17 times forward earnings.153 Either way, the stock is not priced like a company under serious doubt. It traded near $228.52 on October 14.2

The verdict

Kress has done well at explaining Nvidia's strategy. She told investors about the margin pressure early, gave detailed reasons for the larger supply commitments, and defended the financing deals in public instead of avoiding the topic.1021 Her 70% growth forecast, however, depends on things Nvidia does not control: hyperscalers continuing to spend, memory prices easing, and labs like OpenAI earning enough to pay for the computing capacity they have committed to buy.10 In practice, Nvidia has become partly a financing company for AI as well as a chipmaker. That means the numbers that matter most in the coming quarters are likely to be customer creditworthiness and guarantee exposure, not just chip sales.