11 min read6 viewsinvesting

Odd Lots: He Changed None of the Numbers, and Downgraded the Stock Anyway

A long chip-company conference room late at night, in the foreground the chair at the head of the table pushed back from a cold cup of coffee and an open notebook, dozens of chairs still tucked in along both sides, the room running away into darkness with a half-erased timeline on the far whiteboard

Notes on the Odd Lots episode released on 14 November 2026, where Joe Weisenthal and Tracy Alloway talk with Bernstein Research semiconductor analyst Stacy Rasgon about NVIDIA after Jensen Huang's departure. Recorded 13 November at 12:10 New York time, with the stock down about 4%. Rasgon had downgraded NVIDIA from outperform that week and halved his target price without cutting a single financial estimate, and said he doesn't expect customers to buy less data center hardware next year; what he repriced was who makes the long-term architectural calls. Every number is as stated at the time of recording. Educational listening notes only — no investment advice, no price targets, and the judgements here may be wrong.

  • NVIDIA
  • podcast-notes
  • key man risk
  • export controls
  • semiconductors
Contents
  1. A downgrade that took hours to decide and 20 minutes to write
  2. The thing he downgraded has no line on the income statement
  3. Where this reasoning breaks
  4. Why China policy and the competition don’t explain it
  5. A few other things worth remembering
  6. Further reading
  7. The one thing worth taking with you

A long chip-company conference room late at night, in the foreground the chair at the head of the table pushed back from a cold cup of coffee and an open notebook, dozens of chairs still tucked in along both sides, the room running away into darkness with a half-erased timeline on the far whiteboard

Thrice he was sought out, and with him the strategy of the whole realm was weighed; through two reigns the old minister’s heart held the work together.

—— Du Fu, “The Minister of Shu” (Tang, 760; my own translation)

In the Odd Lots episode released on 14 November 2026, Joe Weisenthal and Tracy Alloway talk with Stacy Rasgon, semiconductor analyst at Bernstein Research, about NVIDIA after Jensen Huang’s departure. They recorded on 13 November at 12:10 New York time, with the stock down about 4%. Rasgon had downgraded NVIDIA from outperform that week and halved his target price, and he did it without cutting a single financial estimate, saying he doesn’t think customers will buy less data center hardware next year because of this; what he repriced is who makes the handful of architectural calls five and ten years out. He puts the risk past next year — the near-term revenue trajectory, in his view, is not where the problem is.

A downgrade that took hours to decide and 20 minutes to write

Start with how that week ran. The reporting described in the episode first covered his relationship with a customer, and the intelligence community’s concern that she might be working for China; Rasgon says he wasn’t convinced then and still wasn’t convinced at the time of recording that she was a spy, and the note he published on Sunday evening still had him surviving — he also says judging anyone’s personal conduct isn’t his job. What turned it was the later disclosure about relationships with subordinates, and misleading the board. Rasgon calls that the kiss of death: he has had CEOs in his coverage fired for exactly that, and for less egregious versions of it. The board confirmed the dismissal the next day, and his question became what to do with the rating.

One detail stuck with me. Rasgon said the team spent hours discussing whether to downgrade, and writing the note took 20 minutes — “there’s not a whole lot of analysis to do.”

So start with what he left alone. He didn’t cut his numbers, he sees little risk to the near-term revenue trajectory, and he doesn’t expect customers to order less next year. He said the stock was already cheap and has gotten cheaper (“cheap” is his word), so you can argue the price already has plenty of bad news in it. He was explicit that he has sympathy for that case for keeping the buy.

The revenue estimate bars on the left keep their original height, while the rating and price-target bars on the right stand at half height inside dotted outlines, showing that what got cut was not the financial numbers.

Then why downgrade? The line he gave is the core of the episode: if he can’t sit in front of a client and say you should absolutely be buying this dip, he finds it hard to maintain the buy.

The thing he downgraded has no line on the income statement

Separate what he worries about from what he doesn’t. He said he isn’t worried about operations — the company has very smart people and great technologists. What worries him is who calls the big turns five and ten years out. The near roadmap looks set to him: Rubin is shipping, Rubin Ultra sounds baked, Feynman is next and he doesn’t know how fully baked that is — and after that? What the architecture looks like, which memory architecture they use, how much HBM goes in. Those are major inflections somebody has to call.

A timeline runs from a solid-line Rubin onward to a dashed Feynman, and at the far end it splits into two unlabeled dashed paths with a question mark hanging over the fork.

Why does he think that is worth a rating? Because there’s precedent, and the detail of the precedent carries more weight than the conclusion. He points at CUDA: the company poured money into the software ecosystem when, he says, nobody really wanted it or cared about it, people were telling him to stop, and he didn’t — and it turned into a moat. Rasgon adds a yardstick. The stock sat between $10 and $20 for years, and there have been 40-for-1 splits since, which on today’s share count works out to 25 to 50 cents. Those were the years he was spending into the software while being told to quit.

A share price line sits flat at low levels on the left for years, marked as the years nobody cared and someone told him to stop, then turns sharply upward on the right into a moat.

The shape of the organisation amplifies all of this. Tracy, citing reports, says he had something like 60 direct reports and no real succession plan; Rasgon describes a very flat organisation where most of the key folks reported to him directly and he made all the key decisions. Then he adds the most honest line in that stretch: that has a positive side, because it let them move at a pace most companies could not. People are coming out this week saying their company was in dire straits, they emailed him, and the problem was fixed in seven minutes.

Dozens of lines radiate from a single central point down to a row of small dots, with the hub labeled both for speed and for the day he leaves.

Several other things hang on the same person. A lot of the supply chain was arranged on handshakes; customers and suppliers added capacity because they trusted him and he talked them into the leap, and Rasgon thinks that will be hard for a new CEO to carry. On the “bank of AI” vendor-financing reputation, he says the push came from the force of Jensen’s will — and he points out where attention lands wrong: most of the physical cash that went out the door went into securing supply, locking up memory, wafers and CoWoS, with backstops and financing a much smaller share, and the supply deals rest on his personal partnerships too. Then there’s talent: Rasgon says a large share of employees have net worths in the tens or even hundreds of millions, don’t need to work, and may have been there because he was. And there have been stories that Colette Kress, the CFO now serving as interim CEO, may not have been fully on board with the magnitude of those commitments; he notes that is a CFO’s job, and that the appetite of whoever becomes permanent CEO is an open question.

Two columns of blocks with their sizes exactly swapped: in the attention column the backstop financing block is largest, while in the actual cash column locking up supply is largest.

All of it runs off one source, and that is how I read the episode. A cheap valuation prices what this set of cash flows is worth; the person who decides what the next set looks like is the one who left.

This is where I get stuck most often myself. A company stumbles, the numbers hold, the price gets cheaper — is that an overreaction or a structural change? After this episode my first question is whether the damage hit this set of cash flows or the machinery that produces them. Time repairs the first. The second gets repaired when somebody new gets a call right, and you don’t know how many years that takes. Rasgon’s self-test strikes me as more useful than any multiple: can you sit in front of a specific person and lay out the whole case for buying? I tried it once, and halfway through I heard myself saying “it’s down a lot.”

Where this reasoning breaks

It has several failure modes. Three of them he states in the episode; the one about time scale is mine.

The first is a successor who makes calls just as good. The internal name floated in the episode is Ian Buck, who Rasgon says was the driving force behind CUDA and now runs the hyperscale business, and he thinks the technical bench there can keep things going. If the next architectural call gets caught cleanly, the downgrade was wrong and “the stock is cheap” was the right read.

The second is a news flow that doesn’t arrive. The other half of his reason is the expectation of more tape bombs from inside and outside the company: he says there seem to be elements in the US government that want them gone, that calls for investigation will multiply, and that when somebody subpoenas texts and emails, digging hard enough into anyone’s history turns something up — which will then be spun in the worst possible light. If that line closes out in a few months with nothing new, the reason expires on its own.

The third is the time scale, and this one is my extension of his. He puts the risk explicitly past next year and possibly past the year after. A judgement like that takes two or three years to settle, and over those years the price will be busy reacting to other things.

On a timeline, neither next year nor the year after shows the problem; the answer sits at the far right two to three years out, while the price line above the axis keeps jumping in reaction to other things.

The fourth is the bind he admits to: he thinks the board and the company will find that letting him go was a mistake, and that keeping him might have been a bigger one, because anything that now comes out of the investigations can at least be argued to belong to the individual rather than the company.

Why China policy and the competition don’t explain it

I tried three other causes for the downgrade. None of them carries the weight.

China policy is the easiest one to reach for, and the quantities don’t support it. Rasgon says China including Hong Kong was roughly 8% of sales last quarter, with a big chunk of that being gaming and automotive, and that effectively nothing in data center is shipping directly into China. So tighter policy changes the political temperature, while next year’s revenue line there is small. The blockage also sits on the Chinese side: NVIDIA has US licences to ship older parts, and the Chinese government won’t let it, because it wants to build up local players. He walks into the paradox from there — reporting suggested someone wanted him lobbying for more China sales, except he was already doing that, and China won’t allow it.

Only a small segment at the far right of the full revenue bar is China including Hong Kong; that segment is enlarged below, where most of it is gaming and automotive and directly shipped data center is close to a thin line.

The remote-access hole is real, and it can’t serve as the explanation here, because he can’t size it. Today a company in China can legally reach cloud capacity outside China, and the government wants that stopped. Joe asked him to steel-man the status quo — they turn it into “silicon man” and “chip man” in the episode — where you can’t sell the most advanced GPUs into China but you can sell them to a Singaporean company operating in Malaysia that lets Chinese AI firms log in remotely. Rasgon’s answer is that this isn’t a consistent position unless the remote access is limited to older parts, and he separates it from the question of whether lower-end chips should be sold into China at all; both questions are still open.

The competition is the interesting one, because it loops back to the same person. His 2027 numbers: NVIDIA’s own guidance implies revenue growth of at least 70%, putting it in spitting distance of $700 billion; Broadcom has suggested data center revenue doubles to around $115 billion and doubles again in 2028 to $230 billion; AMD’s GPU sales look set to nearly triple into the low $40 billions; Qualcomm has said $5 billion. His test is whether the pie is bigger — if it is, there’s room for everybody, and if it isn’t, everybody is in trouble. Not one of those figures moved because of this. As for competitors rallying this week, his reading sends it back to the same cause: the market is betting that a weaker long-term NVIDIA roadmap strengthens everyone else.

An arrow heading straight into China is blocked with a cross, while a dotted line below routes through a Singapore company and a Malaysian data center around the block to the same endpoint.

A few other things worth remembering

  • What he wants from next week’s earnings. The numbers had better be technically perfect, the interim CEO has to convey that the ship is stable and still sailing, and then come the questions on talent retention. Joe’s addition has a human edge: tech people don’t have a reputation for saying “I’m good, I’m ready to enjoy life.”
  • How to read the geographic split. NVIDIA and a number of semiconductor companies now report by the location of the customer’s headquarters; shipment destination is no longer the basis. He thinks who owns and runs the data center matters more — which sets up the funniest line of the episode: he assumes Larry Ellison isn’t sneaking into Shenzhen with a Vera Rubin shoved up his behind.
  • How to size the smuggling. He does it by elimination: US customers over 60% and Taiwanese customers in the low 20s (both figures given from memory and roughly right), Singapore and Malaysia together about 1%, Chinese customers unable to buy much of anything, so the Taiwanese block is the likeliest source. Some people have floated 20% of revenue; he doubts it’s that big, and says it isn’t zero either.
  • People are working around the CUDA moat. AMD and others have started saying that AI is getting good at writing software, and are experimenting with using AI to port things over. He still calls CUDA a moat, and notes that the big decisions about where the software platform goes next are also unmade.
  • The White House fight spilled onto television. On the morning of the recording, Treasury Secretary Scott Bessent said on air that this was Commerce Secretary Howard Lutnick’s fault. Tracy’s description had me laughing for a while: after throwing him under the bus, the driver backed up and ran him over a few more times, then blocked the ambulance.
  • One meme, two readings. Joe says he used to meet people who told him their company got an NVIDIA investment and they got to meet Jensen Huang, and he thought that was cool; going back over those stories now, he hears the double reading — “we have a super involved CEO,” said with pride the first time and with worry the second.

Further reading

  • The Odd Lots episode released 14 November 2026 and recorded 13 November, the source for this piece
  • Bernstein Research’s downgrade note on NVIDIA from that week
  • NVIDIA’s quarterly results and earnings call next week
  • The Anthropic registration statement mentioned in the episode, and what it says about using multiple chip sources
  • The public guidance from Broadcom, AMD and Qualcomm on data center and AI chip revenue

The one thing worth taking with you

When an organisation concentrates its key judgements in one head, its speed and its fragility are the same bill.

Sixty people reporting to him directly and every key decision his meant that someone could email him and have a problem fixed in seven minutes, and that the company could turn at a pace rivals couldn’t match. It also meant that the day he left, nobody knew whether the next big turn would be called right. Those two facts are two faces of one design, and you only get the invoice when the person goes.

Here’s something I’ve tried that holds outside investing too. Pick something a small group you belong to is doing — a household, a club, a team at work — and write down three steps only one person knows how to do: the tax form, the key to the equipment cupboard plus the supplier’s phone number, the script that has to run every month. Then take the smallest one and spend an afternoon walking somebody else through it, until they can do it once on their own. The moment you write the list down, you’ll notice a couple of the items are you.

This article is an educational discussion of investment method. It is not advice to buy or sell any individual security, offers no target prices, and does not analyze any current holding. Investing carries risk; make your own decisions or consult a qualified professional.

Comments

Loading comments…

Sign in with Google before posting. Only your name and profile picture are shown.