NVIDIA Corporation

On 27 September we published NVIDIA at $225.07 and concluded "Still too hard". The business had beaten every forecast, but NVIDIA had become a large investor in, lender to and guarantor of the companies that buy its chips, and we could not underwrite the extra assumption. For readers inside the circle we offered a number we said we believed in: about $180, nineteen times that year's expected earnings, "where you would be paid for a pause in AI spending rather than exposed to it". The shares closed at $166.19 yesterday. Here is the account.

The uncomfortable grade is the last one. When a price you named is reached and you then name a lower one, readers may fairly ask whether you will ever buy anything. Our answer is that the number was attached to facts, and two have changed. In September, $180 assumed that Huang would keep running NVIDIA and that any pause in AI spending would be for business reasons. Since then the company has lost its founder overnight, and the spending has met politics: after the midterms, the governor-elect of Texas, with about a quarter of planned data-centre capacity, promised a moratorium on approvals. Neither proves the business is worth less. Both make it harder to know.

The fair criticism of September is the management grade. A founder on record as not believing "in succession planning", a proxy silent on it and a board without a chair made a fragility we described and did not score. We score it now.

Start with what NVIDIA itself has put on the record, because it is short. On 11 November the board "accepted the resignation of Jen-Hsun Huang, as NVIDIA's President, Chief Executive Officer and member of the Board, effective immediately". It appointed Colette Kress, 59, executive vice president and chief financial officer since 2013, as interim CEO "while the Board conducts a search"; Kress also stays CFO. It made Harvey C. Jones, 73, a director since 1993 and a former chief executive of Synopsys, the company's "first-ever Chair of the Board". The search will "consider internal and external candidates". The third-quarter results stay on 17 November, "as previously scheduled".

The filing does not give a reason. It describes no separation agreement, no severance, no treatment of unvested shares, and no future role. The 2026 proxy says NVIDIA has no employment or severance agreements with its executives, so there may be nothing to disclose. A company spokesperson told NBC News that Huang "will not work with Nvidia in any capacity going forward".

What is not established. No court, regulator or NVIDIA statement has established that any espionage took place, that any secret was compromised, that Huang or NVIDIA broke any law, or that the reports about employees are accurate. The investigation has not reported. We name no one but the company's officers and directors and will not repeat the details: an owner does not need them to decide.

What an owner does need are three facts not in dispute. The founder who set NVIDIA's strategy for thirty-three years is gone and will not advise it. The board expressed confidence on a Sunday and accepted the resignation on a Wednesday. And the company faces an investigation, a lawsuit and Congress in the month of its most important results of the year.

Huang's shares. On the last Form 4, Huang held about 811.5m shares through trusts and family companies: about 3.4% of NVIDIA, roughly $135bn, with no special votes. No selling plan has been disclosed since one expired in December 2025. Large sales would still need Form 144 notices, and the market will watch every one.

The headlines put the whole fall on the founder. The prices do not. From the 6 October high of $239.24 to the 6 November close of $186.64, before the board said anything, NVIDIA had already lost $52.60 a share, about 72% of the total fall to yesterday. The first press report about Huang's contacts came on 29 October, so the split is not perfect, but the largest single day, −9.3% on 5 November, had nothing to do with the founder.

5 November was the first trading day after the midterms, and investors read the results as a turn against data centres. In Texas, which Wells Fargo estimates has 27% of planned capacity, the governor-elect had promised "a moratorium, a ban, on all data center approvals". The semiconductor index fell more than 12% that day, AMD 15.6%, NVIDIA 9.3%.

Before that, October was about money. Lenders to the newer AI clouds that rent out NVIDIA's chips grew cautious as rates rose. Microsoft trimmed its 2026 capex plan from $175bn to $170bn, blaming delays at outside GPU providers, and Broadcom's CEO said the biggest constraint on AI might be "capital". None of this says demand has fallen. It says the bill has become a political and financial question.

Why the distinction matters. A scandal at the top can heal if the company runs well without its founder. A change in how voters and lenders see the AI build-out changes the conditions for every one of NVIDIA's customers, whoever runs NVIDIA. This is the part of the fall that tells you most about the next five years.

Who Kress is. For thirteen years Kress has explained NVIDIA's numbers to the world: the quarterly CFO commentary, the hardest questions on the calls, the signature on the filings. Revenue rose from about $4bn to $216bn in that time. Crediting the growth to the finance chief would be overreach. What the record shows is credibility: analysts at Bernstein, KeyBanc and Futurum, who disagreed about the stock this week, agreed about the appointment.

What Kress is not is Jensen Huang, who was also the chief architect of the strategy, the chief salesman to heads of state and cloud chiefs, the keynote speaker who set the industry's roadmap, and the chief negotiator with Washington. No one person replaces that. The likely result is a company run by a team: Bill Dally and Michael Kagan on technology, Ian Buck on the large clouds, Debora Shoquist on supply, Parker on sales, and Kress at the centre.

Three precedents. In June 2018 Intel's CEO resigned over a relationship with an employee that broke company policy; the CFO, Bob Swan, became interim CEO and then permanent. The finances stayed in order, the technology lead slipped further, and in 2021 the board brought in an engineer, Pat Gelsinger. In 2019 McDonald's removed its CEO on similar grounds and promoted Chris Kempczinski, an insider operator; the company barely broke stride. And in 2011 Apple passed from its founder to Tim Cook, the operations chief, who turned it into the most valuable company in the world.

The lesson is narrow. An operator can succeed a founder when the moat is institutional and the strategy is set. The danger comes when the next turn in the technology needs a bet only a founder would make. For NVIDIA the turn is now: from training AI models to running them, against the customers' own chips, within limits of power and capital.

Our expectation, in one sentence: Kress will keep the machine running and the numbers clean, will not try to be Huang, and will probably hand over to a permanent CEO, possibly from outside, within about a year. Good for the next four quarters; it leaves open the question that matters for the next ten.

A $4 trillion company without a chair or a plan. Until this week NVIDIA had a lead independent director, Stephen Neal, 77, and a founder who was the only director who was not independent. The 2026 proxy does not contain the word "succession"; the annual report mentions it once, in a generic risk factor. The founder's stated view was that succession planning was unnecessary. The board let that stand.

Long tenures. Jones and Tench Coxe have been directors since 1993, Brooke Seawell, the audit chair, since 1997. Long service brings knowledge of the company, and familiarity with one person's way of running it. In June, shareholders cast 2.2bn votes against Neal out of 16.8bn, a large protest by NVIDIA's standards.

The three days. On 8 November the board opened an investigation and expressed confidence in Huang. Three days later it accepted the resignation. Either it learned something important in those three days, or it expressed confidence before it knew enough. The new class action is built on that gap, and the board has not said which it was. In fairness, once it acted it acted quickly: it kept the results date, separated chair and chief executive, and hired investigators with real independence.

Buffett's first test of management is integrity, and it is a veto. It is tempting to apply it to this story and stop, but that would be a mistake of logic: the person whose conduct is in question no longer runs the company, and the security allegations are unproven. The test now applies to those who remain, the board above all: will they tell owners the truth about what the investigation finds? If they publish a clear account of what was found and what the board knew and when, they pass. If they say little, or Congress or the courts later reveal what the company knew and did not disclose, the veto applies. We put Management and Capital at 5 out of 10, from 8, and will move it again either way.

In September we described the moat in four layers: CUDA, the programming system almost every AI developer writes for; the network that makes thousands of chips work as one; the rack, a whole computer sold as a unit; and the cadence, a new generation every year. Which of them depend on the founder?

CUDA does not: millions of developers and twenty years of libraries are the most institutional part of NVIDIA. The network and the rack rest on engineering teams built since the Mellanox purchase and on suppliers bound by $279bn of commitments. The cadence is different. Whole racks, Mellanox before networking mattered, CUDA a decade before it paid: these were bets by a founder who could commit billions on conviction. A manager answering to a board mid-investigation is less likely to make the next one, and the next one is due. Running AI models cheaply is a different problem from training them, and it is where Google's TPUs, Amazon's Trainium and AMD compete best.

The last quarter we have. In the three months to 26 July, revenue was $96.2bn, up 106%, at a 75.0% gross margin; GAAP earnings were $2.46 a share, including about $7.8bn of gains on stakes. Vera Rubin was "in full production". Two numbers moved the wrong way: inventory rose to $31.6bn, and customers took longer to pay, with days of sales outstanding up from 45 to 60 on "extended payment terms". Operating cash flow halved, to $24.1bn.

The guidance for the quarter to late October is $108bn, plus or minus 2%, with no China data-centre sales and a 74% gross margin. LSEG's consensus is $108.9bn. The question on 17 November is less whether NVIDIA meets it, which it usually does, than what it says about the quarter after.

At the end of July NVIDIA had $279bn of supply and capacity commitments, mostly memory, up from $119bn six months earlier; $56bn of cloud agreements and leases; and guarantees with a maximum exposure of $108.5bn, $105bn of them for an Ohio campus leased to OpenAI, which has agreed to reimburse NVIDIA. Against that it held $56.6bn of cash and debt securities, owed $33.5bn of bonds, and generated about $70bn of free cash flow in six months.

How to read it. None of this is hidden, and none is debt in the usual sense: the supply will be sold at today's demand, and guarantees are called only if a customer cannot pay. The risk is that all three depend on one thing, the AI clouds and laboratories continuing to raise money. The market priced that in October, and again when the CEO who negotiated many of these deals left.

Concentration. Three unnamed direct customers bought 44% of first-half revenue, and one AI laboratory bought "a meaningful amount" more through other clouds. Five customers owed 70% of the receivables. Each is building or buying an alternative, and the person who held those relationships is gone.

For most of 2026 NVIDIA's closeness to Washington looked like an advantage: its founder had direct access to the President, some export licences to China were granted, and the company was treated as a national champion. This month the closeness became a problem. The House asked to interview Huang on 1 December and General Counsel Tim Teter on 2 December; the Senate Banking Committee invited Huang to a public hearing on 3 December and will hear the Commerce Secretary on 10 December. Senators Wyden and Neal opened an inquiry into export approvals for China-linked companies.

The Justice Department. Semafor reported, citing two officials, three inquiries that could touch NVIDIA, including an antitrust probe, with no indication of criminal charges. Neither NVIDIA nor the department has confirmed this. We treat it as possible, not established, and will look for disclosure in the Q3 10-Q.

China is already outside the guidance: licensed H200 shipments were under 1% of data-centre revenue because Beijing blocked them. The new risk runs the other way, with chip-security bills proposed for the defence budget and calls for tighter controls, which China said this week would threaten supply chains. For a company that sells 8% of its products to customers in China and Hong Kong and makes nearly all of them in Taiwan, the politics of chips is not a side issue.

The finances are what make this a governance crisis and not a financial one. In the twelve months to July NVIDIA had $303bn of revenue and $193bn of GAAP net income, a 64% margin; in fiscal 2026 it turned $102.7bn of operating cash into $96.7bn of free cash flow. Interest cover is above 400 times.

The dividend is safe in any year like the last three and would survive a pause, paid from cash. At 0.60% it is a courtesy, not a reason to own the stock. The buyback is the larger channel: $39.8bn in the first half, at about $202–214 in the second quarter, with $99.3bn of authority left. A company that bought at $210 should be buying more at $166; the Q3 10-Q will say whether it did.

The model. We value NVIDIA as we did in September: revenue in the year to January 2030, an owner-earnings margin after stock pay, an exit multiple, discounted about 3.2 years at 10%, plus the dividends on the way and the net cash. We ignore the equity stakes in both directions. The forecasts on the screen run much higher than our middle case; we use them only for the third scenario.

The middle case moves to about $146 at a 14 times exit multiple and $207 at 20 times; to $182 at a 9% discount rate and $172 at 11%; to $137 if fiscal-2030 revenue is $500bn and $217 if it is $800bn. The range runs from less than half of today's price to more than double. That width is what "too hard" means in numbers.

What is priced in? At $166.19, on our middle margin and multiple, NVIDIA needs fiscal-2030 revenue of about $610bn: half as much again as this year's ~$410bn, then roughly flat. Not heroic for a company growing 90% a year, and not cautious either: it assumes this autumn's doubts delay the build-out without stopping it, and that the change of leader costs no share.

Why the number moves to $125. In September our DCF and the price agreed at about $225, and $180 was 20% below. Our central value is now about $177, because we assume a later, lower plateau and a modestly lower multiple for a company searching for its second chief executive. For a business we cannot value confidently the margin must be wide: about $125 is 30% below the middle case, 13.5 times this year's expected earnings and 8 times next year's. That is the honest reading of our own rule, even though it means we did not act when our first number was reached.