SpaceX Wants $40 Billion of Debt to Buy Nvidia Chips. The Ten Year Is at 5.35%.

Wednesday, October 7, 2026, written at 8:45 a.m. Eastern, forty five minutes before the open. Every future, yield, currency and commodity level in this note carries that stamp and can move before the bell. Where we compare to Tuesday we say whether we mean Tuesday’s close or Tuesday’s morning. Mainland China returns tomorrow after the National Day break, so the Southbound Connect is still shut in Hong Kong today.

Top Call: The Buildout Moved to the Bond Market on the Day Its Cost Hit a 24-Year High

SpaceX is in talks with banks and investors to raise about $40bn to buy chips from Nvidia. The structure reported is roughly $10bn of bank loans and $30bn of investment grade debt, with Apollo Global Management expected to lead and Pimco among a small group of lenders in discussions. The transaction is expected to close in 2027 and the talks are preliminary, so it may not happen at all. If it does it would be among the largest debt financings ever arranged for the artificial intelligence buildout.

Within the same twenty four hours, speaking at the Forbes Global CEO Conference in Singapore, Ray Dalio called artificial intelligence a classic bubble approaching the point where it bursts, and named the trigger precisely: the amount of debt being taken out to fund it, and rising interest rates. He singled out Korea and Taiwan as showing early signs of overinvestment. This morning the ten year is at 5.35%, up eight basis points and a fresh twenty four year high, and the thirty year is at 5.70%.

We do not need Dalio to be right to take the point. The important thing is structural rather than directional. For three years this buildout has been funded out of the operating cash flow of five or six companies that generate more of it than anyone in history, and a capital expenditure programme paid for in cash is indifferent to the long end. A $30bn investment grade tranche is not indifferent to anything. It is a duration product, priced off exactly the curve that has been making new highs every other session since late September, and once the marginal dollar of accelerator purchasing is borrowed rather than earned, the buildout becomes rate sensitive for the first time. That is a different asset class than the one the index has been pricing, and the market started finding out this morning: all four contracts are lower, the Russell weakest at 0.86%, volatility up five percent, and SpaceX itself down 1.71% to $168.98 after rising almost 8% on Monday. The equity is being asked to pay for the debt.

One detail deserves more attention than it is getting. Apollo was one of six financial institutions that partnered with Nvidia in August to build compute financing platforms aimed at mobilising more than $500bn of third party capital. The firm arranging a customer’s borrowing to buy Nvidia chips is also in business with Nvidia to finance purchases of Nvidia chips. None of that is improper and none of it is new, since the telecom equipment market ran the same structure in 1999. What it changes is the information content of the order book. When the buyer’s lender is a partner of the seller, a backlog stops being an independent read on demand and becomes partly a read on credit appetite, and those two things diverge at different points in the cycle.

Rates: 5.35% and 5.70%, and the Reprieve Lasted One Session

The ten year is 5.35%, up eight basis points from Tuesday’s 5.27% close. The thirty year is 5.70%, the highest since 2002. The two year is 4.83%, up three, which puts it back where it closed on Monday. Two year to ten year is near 52 basis points and the widest of this cycle.

The proximate cause is oil. Yesterday we wrote that Saudi Arabia’s official selling price cut was a reprieve rather than a resolution and that it was worth exactly as much as the next tanker. It was worth one session. Brent is back over $101 from under $98 and the long end gave back everything it had taken, plus five basis points.

The shape of this curve has not changed in a week and it is worth restating because it keeps being reported as confusion. The front end prices a committee that holds in October, with hike odds between 16% and 21% and December close to fully priced. The long end prices the inflation that committee has decided not to fight, plus the term premium required to hold thirty year duration against this deficit, this supply calendar and a hundred dollar barrel. Those are not contradictory positions. They are the same position held by two different buyers with different holding periods.

Today’s calendar is unusually dense for a Wednesday. The minutes of the September 15 and 16 meeting arrive at 2:00 p.m., the meeting that lifted the funds rate to 3.75% to 4.00% on a twelve to nothing vote and carried a Summary of Economic Projections. We said on Monday that minutes three weeks stale against four subsequent data releases are a weak instrument, and we still think so, with one qualification: because September carried the dot plot, the minutes should show how the committee argued about individual 2027 paths rather than just the October decision, and that is the part worth reading.

The $42bn ten year auction at 1:00 p.m. is the better event. Tuesday’s three year cleared without incident, but three year duration into a priced front end is an easy test. Selling ten year paper at a yield that has set new twenty four year highs on three of the last five sessions is the first real question of whether the long end has found a clearing price or is still looking for one. A tail here would tell us more about the next month than the minutes will. Consumer credit follows at 3:00 p.m., with consensus near $15bn against $18.06bn.

Gold is down 1.71% at $4,115.30 and bitcoin is down 3.39% near $83,371, its first large move after four flat sessions. Front end yields are higher, the dollar is firmer between 101.6 and 102.0, and both metals and crypto are lower, so the carry framework with the dollar term we added on Monday passes on both legs this morning.

Energy: The EIA Put Our Thesis in Its Own Model

Brent is just over $101 and pushing toward $102, WTI near $89.97. The cause is not a demand signal. UKMTO has logged nine attacks in the Strait of Hormuz already this month, in six days, which is half of what it reported for all of September across the Strait and the Persian Gulf combined. Five of those came on Monday alone: four tanker strikes, one with an engine room fire the crew was still fighting, and an Islamic Revolutionary Guard Corps hail eleven nautical miles north of Khasab ordering an inbound tanker to reverse course or be hit. Tanker rates are at $1.3m a day, forty three times January levels, and roughly forty percent of flows now bypass the Strait entirely.

That last figure is the one to hold. A market that has rerouted forty percent of its volume around a chokepoint has absorbed the disruption structurally, which is why Brent has traded a $95 to $105 band rather than going parabolic. The adjustment is not free, it is being paid in freight, insurance and the war risk premium that Saudi Arabia tried to absorb into its own margin on Monday, and this morning the market handed that premium straight back.

The EIA’s October Short Term Energy Outlook is the week’s most underweighted document. The agency raised its 2026 Brent forecast to $96.32 from $91.01 and its 2027 forecast to $83.74 from $73.74, citing falling global inventories and tight diesel markets. Global stocks fell an average 1.9m barrels a day in the third quarter and the agency expects another 0.7m a day in the fourth. The 2027 revision is the larger of the two and it is the one that matters, because raising next year by ten dollars is a statement that the disruption has stopped being a spike and started being a level. The agency also named tight diesel in its own reasoning, which is the official forecasting apparatus writing the distillate argument into its model after we have spent seven weeks making it from outside.

The product data continues to say the same thing. Distillate inventories are 105.2m barrels and down nearly 15% year on year, with East Coast stocks at 19.3m barrels, down about a third and a record low for the series, and the EIA projects a fall toward 93m barrels by month end. The diesel crack sits near $106. API reported crude down 2.09m barrels and gasoline down 1.37m for the week ending October 2. OPEC+ held the November quota at 31.01m barrels a day on Sunday as expected.

Equities: A Record With the Russell Down, and Then a Lower Open

Tuesday was the S&P 500’s first record close since August 13 at 7,818.93, up 0.58%. The Nasdaq Composite closed at a record 27,599.79, up 0.45%. The Dow added 253.38 points or 0.49% to 51,521.28. The Russell 2000 fell 0.57% to 2,830.97.

Hold those last two sentences together. On the session the broad index printed its first record in nearly two months, small caps fell. On Monday we narrowed our breadth claim to the middle of the large cap index on the grounds that the Russell had beaten equal weight. Tuesday is the clean counterexample to that narrowing, and we are restoring the broader version: this is a market where a handful of very large capacity-constrained names carry the tape and the rest of the listed universe is being discounted at 5.35%.

This morning the futures agree. S&P contracts are at 7,842.00 and down 0.41%, Dow futures 51,400 and down 416 points or 0.80%, Nasdaq 100 futures 31,263.00 and down 0.70%, Russell futures 2,823.70 and down 0.86%. The VIX is 15.77 and up 5.1%. The ordering is instructive: the weakest contract is the one with the most leverage and the least pricing power, and the second weakest is the one with the most industrial cyclicality. That is a rate move being traded, not an artificial intelligence move.

Nvidia closed near $238.90 for a market value around $5.78 trillion. Roughly $248 a share still buys the first $6 trillion market capitalisation in history.

Asia: Korea Sold the Best Quarter in Its History

Seoul is the session’s most interesting chart and it is not close. The Kospi fell 137.49 points or 1.98% to 6,803.90. Samsung Electronics fell 1.29% and SK Hynix fell nearly 3% for a second consecutive decline. Foreigners sold a net 2.6087 trillion won and institutions another 672.7bn won, with SK Hynix alone accounting for 1.3633 trillion won of the foreign selling.

That happened the day before Samsung reports preliminary third quarter results. Analysts expect operating profit near 108.1 trillion won, roughly a ninefold increase and, if it clears 100 trillion, the first such quarter in South Korean corporate history. A market that sells its two largest companies hardest into the single best earnings print the country has ever produced is not expressing a view on semiconductors. It is expressing a view on the won and on global duration, and it is the same foreign selling that ran for five straight months through September. Dalio named Korea and Taiwan from Singapore yesterday as early signs of overinvestment, and whatever one thinks of the thesis, the tape he was describing is visible.

Tokyo closed at 70,150.00, down 0.85%, giving back part of Monday’s first ever close above 70,000. Hong Kong fell about 0.62% to roughly 23,968 with Tencent down 1.8% and Xiaomi down 1.3%, as renewed oil strength revived rate concerns and the Southbound Connect stayed shut. Mainland exchanges and the Connect both return tomorrow, which means the CSI 300 reopens having missed payrolls, both ISM reports and the entire repricing of the long end.

Europe: The Continent Sells, London Does Not

Germany is down about 0.9% at the open from Tuesday’s close of 25,449, which was itself up 0.77% on the day. The CAC is off about 0.23% and the FTSE 100 is essentially flat at 10,534.42. The divergence is mechanical rather than mysterious: the London index is heavily weighted to energy and the continent is not, so a morning that is bad for duration and good for crude splits them. It is still worth noting, because it means the European selling is concentrated in exactly the industrial and export names that a rising global long end and a hundred dollar barrel hit from both sides.

Corporate: The Nuclear Deal Was Four Times the Size We Reported

Alphabet and Constellation Energy confirmed terms, and they are much larger than the figures we carried yesterday. Constellation will spend more than $4.3bn to uprate and modernise eleven reactors at six facilities in Illinois, New Jersey and Pennsylvania, adding about 890 megawatts under a twenty year power purchase agreement with the first capacity expected in 2028. Alongside it sits a separate fifteen year supply agreement for a further 2,700 megawatts from existing plants inside PJM. Constellation closed up 12.2%, Vistra rose about 8% and Talen Energy about 7%.

We reported this yesterday as roughly $1bn for 890 megawatts with the stock up about 4%. The structural reading we gave was right and the magnitude was wrong by a factor of four, which matters, because the magnitude is what a reader trades. The corrected numbers strengthen rather than weaken the argument we made: roughly 3.6 gigawatts of firm existing baseload has been removed from the merchant pool on fifteen and twenty year terms by a single buyer, and the $4.3bn is being spent on uprating licensed reactors rather than building new ones. Every other load behind the same interconnection now competes for a smaller pool. This remains a distributional question before it is a market one, and three utilities repricing between 7% and 12% in a session says the market is only starting to work out who the counterparties are.

AMD said it is actively pursuing collaboration with Samsung in both memory and contract manufacturing, after Lisa Su’s Taipei comments that demand will stay extremely high, that AMD will substantially raise supply in 2027 and that it could still use more. That is the second consecutive day a leading designer has publicly asked competitors for allocation. Designers do not do this when the constraint is demand.

HubSpot is cutting 660 jobs, about 7% of its workforce, with $65m to $75m of restructuring charges and full year guidance unchanged, and the chief executive’s letter attributes it to customers adopting artificial intelligence and to changed expectations of what a software provider should be. Raymond James cut the stock to Market Perform on near term demand. This is the most honest datapoint of the week on artificial intelligence and labour, because it is a software vendor cutting its own headcount while telling you that its customers’ adoption is changing what they will pay it. Both sides of the substitution appear in one filing, and the guidance reaffirmation is the part that will be tested next quarter rather than this one.

Apple has begun dismissing engineering programme managers in hardware engineering, including roughly half a dozen directors, in the first visible restructuring under new chief executive John Ternus. Disney cut close to 300 roles. McDonald’s has been sued over an artificial intelligence pricing tool used by its American franchisees, which is the first consumer facing legal test of algorithmic pricing we have seen at that scale. Zscaler’s chief executive used his platform to warn customers not to trust artificial intelligence agents on security grounds, which from a security vendor is simultaneously a product pitch and a real disclosure.

In ratings, Melius upgraded Microsoft to Buy with a $665 target and Morgan Stanley took Wells Fargo to Overweight at $102, while BNP Paribas cut GlobalFoundries to Neutral and took the target to $51 from $80. That last call is the two speed semiconductor market in one line: leading edge logic and high bandwidth memory are sold out through 2027 while a mature node foundry just lost 36% of its target price in a single note.

Elsewhere, Mattel rose 3.81% on takeover speculation with Ariel Investments pressing for alternatives, Caribou Biosciences halted two CAR-T programmes and cut staff, and the deal tape from earlier in the week stands with C.H. Robinson taking RXO at $5.8bn, Kroger offering $1.65bn for Giant Eagle, and Clayton Dubilier & Rice with McKesson agreeing $32.05 a share for Option Care Health.

Earnings: Constellation Brands Beat Everything and Fell 5%

Constellation Brands reported fiscal second quarter net sales of $2.63bn against a $2.54bn estimate, up 17%, with comparable earnings of $3.74 against roughly $3.56. Net income rose 21% to $565.8m for GAAP diluted earnings of $3.32. Shipment volumes rose 15.4% and depletions 10.2%, led by Mi CAMPO at about 51% and Kim Crawford at about 11%. Full year comparable earnings guidance was reaffirmed at $11.20 to $11.90. The stock fell close to 5%.

Our take: look at where the margin came from. The company attributes the improvement to recoveries of American tariffs and to savings across marketing and other overhead from optimisation and restructuring. It did not come from price. This is the second consumer staples print in two days with the same shape, after Lamb Weston beat by sixteen cents on 1% sales growth entirely on cost savings. Two companies, two sessions, one mechanism: the consumer is not paying more and the companies are spending less. Reaffirming rather than raising full year guidance after a quarter that beat on every line is the tell, and a 5% decline on those numbers says the market read the reaffirmation as the forecast and the beat as the noise. Set against an ISM services prices paid index at 74.0, this is what the input cost impulse looks like when it lands on a company that cannot pass it through.

Tonight brings Applied Digital and Levi Strauss after the close. Applied Digital is the one that matters and it is the reason the top call has a live test today. Consensus sits at a loss near $0.30 on revenue somewhere between $116m and $134m, it has just secured access to a potential gigawatt in Finland and activated another 75 megawatts at Polaris Forge 1, and it funds its buildout with borrowed money. On the morning a private company asks the bond market for $40bn to buy accelerators, the cost of capital disclosure from a leveraged data centre developer is the cleanest read available on what that market is actually charging.

Key Events Today and Ahead

Thursday is the week. Samsung’s preliminary number and TSMC’s September revenue land within hours of each other, and together they settle a question Foxconn opened last week when it reported third quarter revenue up 47% with artificial intelligence servers past half of total for the first time while profitability slipped. If TSMC decelerates against August’s 53.3% while the assemblers post records, the complex is converting inventory rather than adding capacity. If both accelerate alongside a hundred trillion won quarter in Korea, the constraint is exactly where Lisa Su and Sanjay Mehrotra keep saying it is, and the only open question becomes who is paying for the next tranche and at what spread.

Accountability

The Saudi reprieve lasted one session and we will not claim more than the mechanism. We wrote yesterday that the official selling price cut was worth exactly as much as the next tanker. Brent is back over $101 within twenty four hours and the tanker arrived as nine Hormuz attacks in six days. We had a coin flip on direction and we are not claiming it. What we will claim is the reasoning, which was that a producer discount does not touch a shortage that lives in freight, insurance and refined product.

We underpriced the Google nuclear deal by a factor of four and that is the week’s worst error so far. We reported $1bn or more for 890 megawatts with Constellation up about 4%. The confirmed terms are $4.3bn of uprate capital across eleven reactors, a separate fifteen year agreement for 2,700 megawatts, and a 12.2% move with Vistra up 8% and Talen up 7%. Being right about contract structure while being wrong about size by four times is not a good trade for anyone reading us, because size is the part that gets positioned.

The breadth claim goes back to its broad form. On Monday we narrowed it to the middle of the large cap index because the Russell had beaten equal weight on Friday. Tuesday the S&P set its first record since August and the Russell fell 0.57%, and this morning the Russell contract is the weakest of the four. The narrowing did not survive two sessions and we are withdrawing it rather than defending it.

The carry framework passes on both legs and we are holding the split open anyway. Yesterday we separated gold from bitcoin after four sessions in which bitcoin ignored a softer dollar and a lower front end. This morning yields are higher, the dollar is firmer, gold is down 1.71% and bitcoin is down 3.39%, so the amended framework works on both. One session of reconvergence does not undo four of separation, so the split stays on the board as an open question rather than being quietly closed because it became inconvenient.

The distillate thesis is now inside the EIA’s model. The October outlook raised 2027 Brent by ten dollars and named tight diesel in its reasoning. Seven weeks on the structure, three of them with the wrong trigger named, and the confirmation has now arrived from a producer’s price list on Monday and a government agency’s forecast on Tuesday.

The crude discipline rule pays a fourth round trip in four sessions. Monday over $101, Tuesday under $98, Wednesday back over $101. Carrying no directional view on the barrel while holding a view on the product crack has now been paid through four complete reversals in less than a week, and the reason it keeps paying is that the barrel is being priced by naval incidents while the crack is being priced by refinery capacity. Only one of those is forecastable.