# The Compound and Friends: An NVIDIA Without Its Founder, and the Queue of Companies Volunteering to Be Second Source Source: Realpha Blog (blog.getrealpha.com) Original article and charts: https://blog.getrealpha.com/en/blog/compound-2026-11-18-nvidia-blow-up-beneficiaries-retail-sells-the-dip-/ > Notes on The Compound and Friends recorded on the evening of NVIDIA's earnings, 18 November 2026, where Josh Brown and Michael Batnick put the quarter, the governance vacuum left by the founder's resignation, and the shifting balance of compute purchasing side by side: revenue beat by roughly $6bn, data-centre revenue 30x its fiscal 2023 level, another $150bn of buyback authorisation — against a 36% drawdown from the high and about 11x forward earnings, while NVIDIA's share of announced frontier-lab compute agreements has gone from 80–90% two years ago to 28%. The episode also cites Adam Parker's work: when results miss, cheap stocks are punished harder than at any point in 25 years. 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. Published: 2026-11-18 Locale: en Tags: NVIDIA, podcast-notes, semiconductors, governance risk, positioning ![The main aisle of a data centre late at night, an empty office chair pushed back in the foreground with a leather jacket draped over it, the screen in front of it dark, and the status lights of the server racks still burning all the way to the far end of the aisle](/covers/compound-2026-11-18-nvidia-blow-up-beneficiaries-retail-sells-the-dip--cover.png) > Law cannot stand on its own; precedent cannot apply itself. With the right person they survive; without them they perish. > > —— Xunzi, "The Way of the Ruler" (Warring States period, my own translation) This episode of The Compound and Friends was recorded on the evening of 18 November 2026, as NVIDIA's results landed, with Josh Brown and Michael Batnick reading the after-hours numbers live. Revenue beat by roughly $6bn, data-centre revenue came in 30 times its fiscal 2023 level, and the company authorised another $150bn of buybacks — while the stock sat 36% below its high of three weeks earlier, market cap back from over $5tn to $3.5tn, around 11x forward earnings. Both hosts land somewhere outside the income statement: the founder resigned last week, the board's independent investigation is still running, he testifies before Congress on 1 December, and NVIDIA's share of announced frontier-lab compute agreements has fallen from 80–90% two years ago to 28%. From this quarter on, how good the numbers are and how much pricing power the company has left are two separate questions. ## From $27bn to $293bn: look at that slope first When Michael pulled up the annual revenue chart, Josh asked to go back to the fiscal 2023 column — the year before ChatGPT. NVIDIA did $27bn that year. Three quarters into this fiscal year the total is $293bn, with the fourth quarter still to come, already above the whole of fiscal 2026 — and fiscal 2026 was itself a monster, so the easy-base explanation doesn't hold here. ![Two bars side by side: the left one, for fiscal 2023 revenue, is tiny, while the right one, covering the first three quarters of this fiscal year, is about ten times taller, showing this line did not grow off a low base.](/figures/nvda-revenue-270-to-2930-en.svg) Josh drew a line from that I want to keep: with this drawdown included, the fundamentals have outrun the share price, and the stock hasn't been expensive for a long while. There was a reason it traded above $200, and what the shorts are now betting on is "this is as good as it gets, and everything from here deteriorates." Both of those sentences hold at once. So why, back at a $6tn market cap, did the market still pay a lower multiple for it than for the index? Michael says they circled that question for a year and a half. Part of the answer is arithmetic: priced off its then-growth rate, it would have swallowed half the S&P 500. The other part is the shape of the risk — the bigger a stock gets, the more asymmetric it becomes; good news moves it less and the sudden arrival of bad news moves it a long way. Josh's line was that nobody has this map, because there has never been a stock this big. ![On a horizontal baseline, the up arrow for good news is short while the down arrow for bad news is long, showing that the bigger the stock, the more asymmetric its returns.](/figures/giant-stock-asymmetry-en.svg) ## Last week: one man's exoskeleton The episode quotes Ben Thompson, and for me it is the sharpest read of the hour: NVIDIA's most important moat was the degree to which it was not a normal company; the founder organised it as an exoskeleton around himself, there is very little management, and he famously had 60 direct reports — the same reason there was no succession plan. That structure let a company this size still move like a startup, at the cost of nobody being able to check him; now that he's gone, the exoskeleton is more poised to collapse than you might expect. ![On the left, one center point sends out many direct lines with only a thin layer of management; on the right the center is gone and those lines dangle with nothing to connect to.](/figures/one-person-exoskeleton-en.svg) What follows is the schedule that gives that read its cash-flow statement. He testifies before Congress on 1 December, his lawyer the next day; the board's independent investigation has a law firm attached to it, and Josh notes that those processes never move quickly; regulators are already looking at his departure. On the day of recording he posted that if the board concludes his return is in the best interest of shareholders, employees, the company and America, he would come back. Josh couldn't believe it — under a week after resigning, with the congressional hearing still ahead. Josh had an analogy for the board's problem. The head of the household leaves, the siblings have to raise each other, and none of them listens to anyone; and the trouble with this governance structure is that no internal candidate carries enough weight to pick up the gauntlet. His read is that the only scenario where the board brings him back is one where the business and the stock deteriorate badly first. ## Earnings night: 11x earnings and a $150bn buyback That authorisation dates from September. Josh's comment was blunt: announcing a buyback with the stock at $220–240 and the market cap heading toward $6tn made little sense; with the stock cut to here and the multiple possibly the lowest in the company's history, it does. His framing: if you think this company survives, this is where you pull the trigger — everybody is selling, and a business generating this much cash shouldn't let its own share slide go by. ![A price line that rises and then falls, with the buyback authorization marked near the high and the genuinely cheap point sitting low, after a 36% drawdown.](/figures/buyback-authorized-high-usable-low-en.svg) What stayed with me is the half he admitted next. He owns the stock, he has ridden it through the 36% drawdown, he says it is unenjoyable, so he is hunting for a silver lining — and he can say out loud that this is what he's doing. He also says he might end up selling some, that he has no strong view right now: "if I felt strongly, I would say I feel strongly." Michael and Josh split on the low. Josh thinks the market hasn't processed it yet, with the accounting and the circular financing still to go under the microscope; Michael's version is that memories are short and this low could hold for a couple of months. The one thing they agree on: rallies get sold. ## Two years on: from 80–90% to 28% The episode showed a chart of announced frontier-lab compute agreements. Two years ago NVIDIA held 80 to 90%; in these announced deals it is 28% — Broadcom 23%, AMD's GPUs 18%. Broadcom builds custom silicon for several of NVIDIA's largest customers, and those customers are building their own tensor processing units and other purpose-built chips. ![Two horizontal bars: two years ago one company filled nearly the whole bar, while now the same bar is split into close to three equal parts plus a small slice for others.](/figures/compute-share-two-years-en.svg) Josh doesn't dispute that the CUDA software platform still locks in a great deal of GPU cluster demand; but once the conversation reaches price per unit of compute, the companies that spent the last two years working get a seat. He is explicit about what this episode means for a buyer of compute: the leading GPU supplier now has a leadership vacuum, an internal investigation, and a market speculating about a supply glut — the best position a large buyer could hope for. Anthropic's last filing before going public mentions standing up a model instance on AMD's Helios architecture in three days; a year ago nobody thought that was possible. The next round of GPU contracts gets priced across the table from a supplier that needs them to buy more than they need to buy. The other thread is more direct: this company's loudest voice in Washington argued for selling chips to China, and that voice is now gone or much diminished. As for whether something more complicated sits behind the timing, Josh ran through several versions of the speculation and said nobody knows which one is true. ## Where the money goes: a second source, and the memory it stopped hoarding Two mechanisms sit behind the list of beneficiaries, and I find them more useful than the list. The first is that the door to a second source is open. One other company sells a rack-scale AI system the way NVIDIA does, and it has shipping product; the figures the hosts quoted were data-centre revenue up 97% to $8.5bn last quarter and a $17bn guide at the midpoint for the next, which works out to 66% growth. The more structural reason is procurement: purchasing teams at the large cloud and software buyers now have to produce a plan for reducing supplier concentration. Josh's version is that a Pepsi always grows up next to the Coke, and nobody believed it would happen here. Along the same thread, the messier the mix of chips inside a data centre, the better the business of whoever sells the non-NVLink interconnect — the episode calls that company the Switzerland of connectivity. The second mechanism is memory. Apple's CFO said memory prices account for all of the quarter's sequential margin pressure, and the iPhone 18 Pro and Pro Max each went up $100 with memory costs named as the reason. Josh's inference: someone has been grabbing as much memory as possible to hold the position of only game in town, probably overbuying; if the incoming management is more disciplined about spend and that hoarding eases, the biggest beneficiary is the company that needs the same memory and can push the cost into its prices. Apple is up 30% this year while NVIDIA is down 16%; its customer is 2.5 billion devices paying cash, there are no financing shenanigans, and it has run its China business in accordance with international law. For money leaving the data-centre story, there aren't many large-cap places to switch into. ![One bar on each side of the zero line, up 30% this year on one side and down 16% on the other, with an arrow in the middle carrying memory cost from the falling side to the rising side.](/figures/memory-cost-two-sides-en.svg) ## Who is selling now: managers cut, retail takes the other side Two positioning charts, pointing opposite ways. Bank of America's fund manager survey shows the net overweight in tech falling by 25 percentage points in a single month, the biggest drop since 2006, with the bank's clients dumping tech over the same stretch. On the other side, the daily retail imbalance in NVIDIA is overwhelmingly buying, and Robinhood users bought a record dollar amount of single stocks last week — this one. Michael says he sides with the institutions, because that survey has been right more often than wrong. ![Two opposing forces on the same stock: on the institutional side the arrow points outward as they cut, on the retail side it points inward as they buy.](/figures/institutions-out-retail-in-en.svg) The source Josh assigns to the selling has nothing to do with earnings, and it's the passage I most want to keep. The multi-strategy model is renting a trader's edge: your edge has an expiration date — the thing you're expert in stops being tradable, or everyone else works out what you were doing, or the market stops caring about your sector. So these funds pay the best traders enormous amounts, and when the edge goes, so do they. These semiconductor stocks sit in those hands — people who can tell you anything about wafer size, process nodes, the Asian market — and then the call comes, three good years are over, and the book has to be cleared. Josh's point is that this process wrecks the stocks whether or not they grow earnings 20% next year; tell him the semiconductor ETF trades at a 22x median multiple next year and he'll believe you, say 12x and he'll believe that too, because the only thing it rests on is whether there are more sellers than buyers. ![On one base block representing next year's earnings, two bars grow to heights that differ by nearly double, labeled 12x and 22x.](/figures/same-earnings-two-multiples-en.svg) ## Why "it's cheap now" won't save you "This thing is down 40%, surely the bad news is out" — I've thought that sentence before buying something, and the Adam Parker work cited in the episode takes it apart cleanly: when a company misses, cheap stocks get punished harder than expensive ones, and by the widest margin in 25 years. Cheap stocks that fall a long way into a print have not de-risked themselves in advance. ![After an earnings miss, two stocks each drop a leg, and the cheap one that had already fallen a lot drops deeper; schematic.](/figures/cheap-punished-harder-on-miss-en.svg) That deserves a layer of thought. Cheapness worked as protection on the condition that the bad news was already in the price — and that the bad news in the price is the same event that later arrives. The cheapness in semis today is computed off forward earnings estimates, and the front half of those estimates is the 2027 and 2028 buildout plans of a handful of rented-compute cloud companies. Josh's read is that they won't shut down projects already underway; what they cut is the out years. When those get cut, the downward revisions bite back on the companies selling into them, and the share price reaction won't be linear. ![A before-and-after pair: before, the denominator is an earnings estimate propped up by build-out plans for the coming years; after those plans are cut, the denominator shrinks and the same price works out to a higher multiple.](/figures/cheap-depends-on-whose-estimate-en.svg) So when I look at something "cheap" now, I start by asking whose estimates the cheapness is computed from and where that person's money comes from. The episode also gives the other side: the same research says that after spikes in the forward earnings yield for semis, forecast net income growth has historically been 20% lower three months later — and Adam's own view is that it won't repeat this cycle, because the economy is holding up. Josh agrees, on the grounds that every earnings report says the same thing: demand isn't slowing, whoever is running the company. ## Why you keep looking at the brightest tile on your screen The closing stretch made me laugh. Michael put up a mystery chart with the price stripped off: one of the ten largest companies in the world, up from 500 to 600 since 1 October. Josh guessed Microsoft, then Eli Lilly, then Visa. It was Berkshire Hathaway. Michael said he'd hung his head looking at that chart earlier in the day — he owns it, and he wasn't excited when the breakout happened. Josh's explanation: people don't know what to look at first, so they open their screen and whatever is flashing brightest gets the attention, because there's no system. I take that as a note to myself. Everything in this episode — the founder, Congress, memory prices, interconnect chips, the manager survey — lives in the brightest tiles; and the name that went from 500 to 600 over the same stretch was on nobody's screen. The check is simple enough: list the time you spent reading this week, and see whether the names that time went to line up with the actual weights in your account. I did it once, and the name I read the most about was a middling position. ![Two columns of bars side by side: the name read about the most ranks only mid-pack by weight, while the largest weight got almost no reading time.](/figures/reading-time-vs-weight-en.svg) ## Further reading - The Compound and Friends, recorded on NVIDIA's earnings night, 18 November 2026 (Josh Brown and Michael Batnick) — the source for this piece - NVIDIA's quarterly report and call materials, and the buyback authorisation announcement - Ben Thompson's piece at Stratechery on the founder's departure - Adam Parker's (Trivariate Research) work on earnings punishment for cheap stocks and the forward earnings yield in semis - The Bank of America global fund manager survey, and the retail imbalance data cited in the episode - Apple's most recent earnings call and the CFO's remarks on memory costs and margins ## The one thing worth taking with you **If a seat can only be filled by one person, that seat is your largest risk — and how good that person is has nothing to do with it.** The episode plays this out: the same structure that let the company move like a startup is the structure that left it hollow the day he walked. Excellence hides the risk, because you never have to think about the alternative. Here's something I've tried that has nothing to do with investing. Think of the last time you handed something over and then took it back to do yourself — a file at work, or whoever deals with the older relatives at home. Write the reason you took it back as one sentence, then write down how the version they would have produced would have differed. I did this once, and what I found was that the gap was a standard of mine I'd never said out loud, where I'd assumed the gap was ability.