Stocks

Forget the Conspiracy Theories. Nvidia’s Real Problem Is the Credit Loop It Built Under Its Customers.

Forget the open-model conspiracy. The story NVDA holders should care about is the credit loop Nvidia built under its customers, and Colette Kress is now the one deciding whether to keep it going.

MrFlourish · Stocks desk
Nov 15, 2026 · 10:10 AM CST · 5 min read

By Chief Editor Clark for The Market MBA · Opinion / analysis

Two stories are fighting for the “real reason” behind Nvidia’s leadership shake-up. One is a thriller and the other is accounting. If you own NVDA🔔 MFI alert: Buy, the accounting story is the one that matters.

Nvidia (NVDA) ended its first week without Jensen Huang near $163, down about 13% for the week, its sixth straight weekly loss and roughly 30% off the early-October high. The S&P 500 had its best week since August. Investors didn’t dump AI. They dumped one company’s version of it. (Week wrap: the discount rate explains the week)

Two explanations are going around online. One says Huang was pushed out because his open-model strategy threatened the closed AI labs heading toward IPOs. The other says a vendor-financing machine grew too big to ignore, and the executive who questioned it now runs the company. The first is mostly noise. The second is the real investment debate.

The open-model theory: real facts, wrong conclusion

The ingredients are real. Nvidia has committed about $26 billion over five years to building open-weight models. It released Nemotron 3 Ultra, a roughly 550-billion-parameter open model. It formed a coalition of eight outside labs to co-develop open frontier models. In September it agreed to buy Hugging Face for about $12.9 billion, a deal still waiting on regulators and expected to close in the first half of 2027.

The theory breaks at the next step. On independent benchmarks, Nemotron 3 Ultra leads U.S. open models but trails the best Chinese one and sits well below the top closed systems. It’s a fast, cheap workhorse, not a Claude Opus killer.

The money is modest too: about $5 billion a year, next to $279 billion of supply commitments. And every Nemotron model is tuned for Nvidia hardware. Open models are how Nvidia sells GPUs to everyone who can’t afford a closed-lab contract, so a board protecting its customers has little reason to kill a program built to create more of them.

Our base case is that Kress keeps it. It’s cheap, it blunts Chinese open models built for rival chips, and walking away from Hugging Face would burn developer goodwill for little savings. If she cuts it anyway, read that as a company-wide cost purge, not a plot.

The darker claims riding along with this theory, about who leaked what, are unverified, and we won’t repeat them. Motive-guessing makes a good post, not a valuation.

The guarantee loop: where the risk actually sits

Here’s what should keep NVDA holders up at night. Nvidia’s quarterly filing for the period ended July 26 shows:

  • $366 billion of future commitments. Supply and capacity commitments alone went from $119 billion to $279 billion in a single quarter.
  • $36 billion of agreements to buy back capacity from AI cloud partners, who can resell it elsewhere at better prices. Nvidia is the buyer of last resort.
  • Up to $108.5 billion of guarantees, $105 billion of it for OpenAI’s Ohio campus, phasing in from fiscal 2029 and covering defined lease and power payments if the tenant defaults.
  • About $91 billion of total liabilities on the actual balance sheet.

Add up the commitments and guarantee caps and you get close to half a trillion dollars, roughly five times the liabilities on the balance sheet. These aren’t debts, and most only cost Nvidia money if a customer stumbles. But this kind of exposure stays quiet while capital is cheap and gets loud all at once when it isn’t.

Reporting this week describes how the loop worked. Nvidia offered credit support to small cloud providers and data-center developers for a cut of their revenue, some of it pushed by a recently promoted vice president from Huang’s family. Kress kept asking how the guarantees were priced and what Nvidia got back. She reportedly cut the Ohio guarantee from a proposed $250 billion to $105 billion, and some deals were paused in August.

Put simply, Nvidia can be the investor, the lender, the supplier and the backup customer in the same deal. We flagged that circular risk in our AI debt brief, and the market is pricing it now. About half of this year’s roughly $80 billion in U.S. data-center bonds are junk, and interest bills at Nvidia-backed AI clouds are climbing, as CoreWeave’s quarter showed.

The Cantor headline: check the date before the drama

A headline this week said Cantor Fitzgerald cut its Nvidia holdings by nearly half. The filings back the number, not the timing most readers will assume. Cantor’s quarterly 13F reports show about 3.49 million Nvidia shares (roughly $652 million) at the end of September 2025 and about 1.78 million (roughly $310 million) at the end of March 2026, a 49% drop. By June 30 the count was flat at about 1.81 million. The Sept. 30, 2026 report isn’t filed yet. It’s due Monday.

Two cautions. A 13F is a snapshot of long positions on a quarter’s last day, filed up to 45 days later. It isn’t a live trade and says nothing about this month’s sell-off. And the firm says the shares mostly hedge client swap trades. The same filings list large Nvidia puts, covering 3.6 million shares in September 2025 and 1.5 million in March. That looks like a dealer’s hedge book, not a conviction call.

So don’t trade the headline. At most it’s a read on mood: a hedge book that halves can mean clients wanted less leveraged Nvidia exposure well before the CEO drama, which fits our view that sophisticated money began de-risking the AI financing trade early. A clue, not evidence. Monday’s filing will be fresher, and still seven weeks old.

Our take: Kress is the right person for the wrong moment

The open-model crowd sees Kress as a placeholder. That’s a misread. The finance chief who kept questioning the loop now decides whether it continues: good for the balance sheet, painful for the growth story in the short run. If Nvidia pulls back credit support, smaller customers must borrow in a junk market already demanding close to 10%. Some will order fewer chips. That isn’t a scandal. It’s what happens when a company stops funding its own demand.

The leadership drama is a short-term overhang. The guarantee book could be a multi-quarter one. The board’s reported approach to Arm’s Rene Haas (our Nov 13 mid-day brief) points to an engineer-operator as the permanent pick. What matters more is whether that person inherits a disciplined credit policy or a pile of open promises.

What to watch Tuesday, Nov 17

  1. The commitments table, not headline revenue. The quarter ended Oct. 25, before the scandal, so a beat on the roughly $108 billion guide is mostly priced in. Did the $366 billion grow or shrink? Did guarantees spread beyond Ohio?
  2. The AI cloud buyback line. If the $36 billion keeps growing, Nvidia is still buying back its own demand.
  3. A formal credit-support policy. Clear limits, board sign-off and a return hurdle would be the most bullish thing Kress could say.
  4. Gross margin and the Q4 guide. Margin is guided to bottom at 71–72% in Q4. A deeper dip or softer guide would suggest the loop was propping up demand.
  5. Hugging Face and the open-model roadmap. One sentence reaffirming both ends the conspiracy theory.

A cheaper stock with a more honest balance sheet is a better long-term owner’s deal than an expensive stock that pays for its own demand. Tuesday is when we find out which one Nvidia is.

Sources hidden.

This is opinion and analysis, not financial advice, and not a recommendation to buy or sell any security.

Chief Editor Clark
The Market MBA

NFA. This article is for educational purposes only and is not financial advice.

For education and entertainment only. Not financial advice.

Trade on Liquid with MrFlourish — save 10% on fees with code DL6N7COO

Referral link