11 min read4 viewsinvesting

If That One Person Leaves, Does the Machine Still Turn? MacroMicro on the Post-Jensen Era

A substation inspection walkway at dusk receding into depth, a row of open circuit breakers in the foreground, a data hall far down the line with only a few windows lit

MacroMicro's After Meeting EP. 222, published 22 November 2026, spends a full episode on the ghost stories around Nvidia: Jensen Huang stepping down as CEO, rumours of Vera Rubin slipping, and AI circular financing named in a Fed report. Educational listening notes on which layer of the funding chain takes the strain and which numbers can settle the argument — no price targets, no stock recommendations.

  • Nvidia
  • Jensen Huang
  • AI circular financing
  • data centres
  • reading financials
Contents
  1. The line someone said in the studio
  2. Ghost story one: the numbers are strong, the margin slipped
  3. Jat answers the rumour by counting racks
  4. He’s gone, so do the signatures still hold?
  5. Pull the camera back: the machine going round in circles
  6. Is the engine stalling? Anthropic’s 47%
  7. The gear layer: 35% energised
  8. The bad news keeps coming — which one should I actually fear?
  9. The visionary left — do I need to revalue the company?
  10. Further reading
  11. One thing to take away

A substation inspection walkway at dusk receding into depth, a row of open circuit breakers in the foreground, a data hall far down the line with only a few windows lit

Toward evening my heart was uneasy, so I drove up to the ancient plain. The setting sun is beyond beauty, only it is close to dusk.

— Li Shangyin, “Climbing the Leyou Plain” (Tang dynasty, c. 9th century; translation mine)

MacroMicro’s After Meeting EP. 222, recorded on 19 November 2026 and published on 22 November, lines up the ghost stories around Nvidia one by one: stand-in host Vivi and researchers Dylan and Jat start from Jensen Huang’s resignation as CEO on 11 November, with CFO Colette Kress stepping in. Jat’s scorecard for fiscal 2027 Q3 shows revenue up 103% year on year, free cash flow of US$43.8 billion, and purchase commitments to suppliers raised to US$390 billion. Dylan puts the pressure squarely on the mid-tier neoclouds that borrow to build their own halls, where CoreWeave has energised only 35% of its contracted capacity. Their conclusion: it takes a large build-out delay, a sharp pullback by Nvidia and a hawkish surprise from the Fed all at once before this counts as systemic risk, and the turning point arrives in 2028 when the big cloud operators’ free cash flow goes positive. The sun is low on the machine, and the machine is still turning.

The line someone said in the studio

They taped on 19 November, a bad week to review. Taiwan’s index broke below 39,000 intraday on Thursday, its lowest since April; gold fell through US$3,800, WTI crude slipped toward US$80, the dollar index pushed past 103 toward 104, and the Taiwan dollar edged toward 33. India was the only major market in the green. Vivi’s description was that tech and Wall Street had been running a soap opera.

Someone on the episode mentioned putting on that leather jacket for the occasion.

Halfway through, Vivi said something I wrote down: she misses the era of Powell and Jensen, because now it is Warsh and there is no Jensen. The Fed changed hands and so did Nvidia — two names the market treated as reassurance, leaving within the same stretch of weeks. The hour that follows circles that one question: if that person is gone, does the machine still turn?

Ghost story one: the numbers are strong, the margin slipped

Jat lays out fiscal 2027 Q3. Revenue up 103% year on year, earnings per share of US$2.46, both above expectations; guidance for the coming quarter raised, with the cloud industry’s order backlog growing from US$2 trillion to US$2.5 trillion. Free cash flow hit a record US$43.8 billion, achieved while capital spending was also setting records.

The one line moving down is gross margin, for two reasons: Vera Rubin entering its production ramp, where yields and costs have not yet settled, and memory prices biting. Inventory days stayed high at 110, which Jat reads as stock being staged for the volume ahead.

Jat answers the rumour by counting racks

Alongside the results came a scare: warped boards, poor yields, real volume shipments of Vera Rubin pushed into 2027. Acting CEO Colette Kress told the call it had already shipped and was on plan. Rumour against official line — how do you call it?

Jat’s method is to count racks. Vera Rubin contributed close to US$20 billion of revenue in Q3. At US$5 million to US$7.8 million per VR200 NVL72 rack system, that works out to roughly 3,000 racks in the quarter. Compare the previous generation: GB200 shipped 1,000 to 1,500 racks in its first production quarter. This ramp starts at twice the scale.

Two vertical bars compared by height: the left one, the prior generation's first production quarter, is roughly 1,000 to 1,500 racks, while the right one, this generation's third quarter, is about 3,000 racks — visibly twice as tall.

He credits the design. Full liquid cooling removes fans, cabling and water-pipe assembly, and the steps that took heavy manual labour last generation are handled by robots this time. What I like about this passage is that it converts a rumour I could never verify into a division I can do myself — twenty billion divided by the price of one rack, and the answer speaks.

He’s gone, so do the signatures still hold?

Then comes the part that reads most like a ghost story. The commitments Nvidia disclosed on the call now total US$390 billion, US$100 billion more than the prior quarter, earmarked for memory, wafers and components from suppliers. On top of that sit US$100 billion of future commitments to buy cloud services and lease data centre capacity, and another US$100 billion backstopping OpenAI’s data centre in Ohio.

A single horizontal bar split into three segments by dollar amount: upstream purchase commitments take up about two-thirds at the far left, and the two following segments each take about one-sixth.

If those commitments shrink because the top job changed, the whole Taiwanese AI supply chain feels it. Jat’s judgement is that signed contracts do not lapse over a change of CEO; what is exposed is the batch of projects not yet negotiated.

His real worry lives further out. Nvidia’s position today traces back to a decision in 2006: pivot from graphics and gaming accelerators to general-purpose computing, then on to AI. There was no visible commercial return at the time, and Jensen Huang pushed resources into it anyway. Vera Rubin, Rubin Ultra and Feynman are already on the roadmap and unlikely to be disturbed by this transition. The question mark is the next one — whether a new management team will back a technology that takes five to ten years to pay off. Jat added one more thing: one person leaving is survivable; the engineers who followed him leaving too is what would loosen the research lead.

Pull the camera back: the machine going round in circles

Dylan takes the second segment, on the AI circular financing the Fed named in its financial stability report. In plain terms, the chip maker no longer just takes your money and ships. It lends to you, sometimes invests in you, and you spend that money back on its chips and compute, so the same dollar keeps circling the chain.

Three stacked boxes, from top to bottom chip maker, cloud provider and model company, with a thick line on the right running from the top layer down to the bottom one to represent money going out, and a line on the left running from the bottom layer back up to represent compute being bought, forming a closed loop.

MacroMicro split the machine into three layers early on: compute upstream is the fuel, the cloud operators in the middle are the gears, and the model companies at the end are the engine. Dylan names three weak points. First, the five largest cloud players will spend over US$800 billion in capex this year and more next year, pushing their free cash flow collectively negative for the first time and sending them to the bond market. Second, short-lived assets carrying long-dated debt: a GPU’s economic life is five to six years and its rent and residual value shrink once the next generation lands, yet the contracts written against it run ten to twenty years. Third, the supplier is also the financier, which ties Nvidia to the weaker operators downstream.

Two parallel horizontal timelines: the upper one, the asset's useful life, covers only the front portion, while the lower one, the debt contract, stretches more than twice as far, with the blank right-hand stretch marked as the gap.

Vivi pushed back on calling Nvidia the central bank of AI, and her version lands better: a central bank is the lender of last resort, arriving after the trouble starts, while Nvidia is out in front writing the cheques. Different position entirely.

Is the engine stalling? Anthropic’s 47%

The thing this machine fears most is an idling engine — monetisation at the far end failing to keep up. Anthropic’s listing documents show quarterly revenue growth falling from 144% in Q2 to 47% in Q3, and the market read it as demand cooling.

Dylan puts the slowdown down to not having enough compute to sell. The numbers he cites carry weight: net revenue retention of 900%, meaning existing customers are spending nine times what they spent a year ago; 92% of enterprise customers expecting to raise their token budgets over the next year; and Anthropic itself writing in its filing that it underestimated compute demand, which has constrained product development and deployment.

On the left, a sharply bending line falls from 144 down to 47, while on the right a very tall bar stands at 900, so the two sides point in opposite directions.

The most useful thing he dropped is an indicator: rental prices for last-generation GPUs are still climbing. If demand were softening, the oldest generation would crack first. It is still rising, which means even old cards are scarce. That is a thermometer harder to fake than guidance, and the prices are published.

The gear layer: 35% energised

The strain sits in the middle. The large cloud operators have gone free-cash-flow negative, but their core businesses still earn, with enterprise cloud revenue growing. The chip maker already has the cash. Caught in between are the neoclouds, which have no core business cushioning them and whose entire model is leveraged compute bought to be rented out.

Two pressures land on them at once. One is power. After the US midterms, anti-data-centre sentiment spread across Texas, Ohio and several other states, holding up grid connection and energisation. CoreWeave has energised just 35% of its contracted capacity, and customer prepayments and rents are tightly bound to on-time delivery — no power, no payment. Dylan makes the energisation rate the first indicator to check on operating cash flow.

A rectangle standing for contracted capacity is filled in only across about the left third, marking the energized share, while the larger empty stretch on the right is labeled as earning nothing.

The other pressure is money. Colette Kress drew a line under the US$118 billion of guarantees on the books: operators have to get land, power and halls online before Nvidia prioritises their allocation and the guarantee takes effect. Asked on the call how supply gets allocated, the answer was to the customers who can monetise immediately. Power and funding are therefore tied together — infrastructure that slips means no guarantee, private credit behind it pulls back, and financing cash flow dries up with it.

Dylan does not read either pressure as fatal. Most approvals do not affect projects already under construction, so 2027 supply looks much as it did before the midterms, and operators are moving toward states with better frameworks — SEPA counts roughly 36 states now offering workable rate-separation schemes. In a neutral case, about 8% of supply is affected by approvals through 2030.

As for the 16 November rule from the Commerce Department — any data centre anywhere renting more than 12 million TPP of compute to Chinese entities now needs a licence — Jat reads the impact as manageable. Taiwanese suppliers serve US cloud customers first; Goldman Sachs estimates roughly 10% of Nvidia’s non-US revenue is China-linked, and with compute this scarce, that gap gets filled by whoever is next in line. He notes that Anthropic and OpenAI are both still short of compute, so redirecting it could ease worries about the frontier labs’ cash flow instead.

The turning point lands in 2028. Dylan expects the large cloud operators’ free cash flow back in the black then and rising clearly through 2029, at which point they can take over the financing role and the chain returns to being funded from its own cash.

The bad news keeps coming — which one should I actually fear?

The headlines arrive in bursts: a resignation, a delay, an export rule, a moratorium. Every one is written like the end of the world, and after enough of them you go numb and then make an important decision on an unimportant story.

What this episode gave me is a sorting step: ask which layer the news hits. News that hits the engine — end demand — matters most, because the whole machine monetises there. News that hits the fuel layer is a cash-flow question, and Nvidia’s US$43.8 billion of free cash flow answers it. News that hits the gears depends on whether the operator has a core business underneath. The same sentence about an AI bubble means wildly different things depending on where it lands.

After sorting comes something falsifiable. Dylan’s three conditions — a large build-out delay, Nvidia pulling back hard, a hawkish Fed surprise — have to arrive together before this is systemic; one of them moving alone is localised strain. I copied three numbers into my own tracker: energisation rates, last-generation GPU rents, and Nvidia’s purchase commitments. Where those go next quarter tells me more than any headline.

Three overlapping circles: only the central area where all three meet is labeled systemic risk, while each circle alone is labeled local pressure.

The visionary left — do I need to revalue the company?

This is not only an Nvidia question. Somewhere in your holdings is a company whose story rests on one person — a founder, a chief technologist, the salesperson who won the customers back one at a time. The day they leave, the price drops first and you have to decide whether that is emotion or fact.

The episode’s method is to split what that person did into two piles. One pile has landed in contracts and institutions: signed purchase commitments, products on the roadmap, supply relationships already built. Those survive the handover. The other pile lives inside their judgement: backing a technology with no payoff for five to ten years, moving resources when nobody agrees. Nothing is written down, so a new occupant means starting the observation again.

A vertical line splits the frame into two groups: three solid boxes on the left stand for things already written into contracts, and two dashed boxes on the right stand for things that exist only inside one person's judgment.

In my experience the second pile does not show its colours within a quarter, so rather than guess, I turn it into something to track: what the new team picks the first time it makes a capital allocation decision, and whether core research staff start drifting away. Both are observable, and both take several quarters to answer — which is why they go in my holding notes rather than into a revaluation on the day the news breaks.

Further reading

  • MacroMicro After Meeting EP. 222, “Ghost Story Special: The Post-Jensen Era” (recorded 19 November 2026, published 22 November), with stand-in host Vivi and researchers Dylan and Jat
  • Nvidia’s fiscal 2027 Q3 results and earnings call, where acting CEO Colette Kress sets out the conditions attached to the guarantees in the Q&A
  • Anthropic’s listing filing, which carries the Q3 revenue figures, net revenue retention and the note about compute constraints
  • The Federal Reserve’s financial stability report, which names AI circular financing as an item to watch
  • SEPA’s tally of state-level rate-separation schemes, useful for tracking where data centre supply is moving

One thing to take away

One idea: whether a system can take a hit comes down to what happens when a person, a sum of money or a single line is removed — who picks it up, and how. Nvidia’s purchase commitments sit inside contracts, so they survive a change of CEO. Its next big bet lived inside one person’s judgement, so that one has to be watched again from scratch. Same company, two kinds of thing, wildly different resilience.

Here is something I tried that has nothing to do with investing. Take a sheet of paper and write down the things at work or at home that exactly one person knows how to do — who holds the tax password, who collects the kids on Friday, which technician is the only one who can tune that old machine. Next to each, add two lines: who covers it if that person is away for a week, and what they would need to know to cover it. The first time I did this, three of the answers were “nobody”, and those three got a backup afterwards. Nobody else has to see the sheet; writing it is what does the work.

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.