S&P Says the AI Debt Is Bigger Than Disclosed. Korea Sold the Whole Rebound in a Day.
Tuesday, October 20, 2026  ·  Before Bell
Morning Note  ·  Equity Research

S&P Says the AI Debt Is Bigger Than Disclosed. Korea Sold the Whole Rebound in a Day.

October 20, 2026
Fund X
Steven Yuan
S&P Global Hyperscaler AI debt likely larger than filings show
AI capex 2025-30 Above $7tn across six names, negative FCF into 2027
Kospi 5,552.74, down 3.73%, erasing Monday's 1.76% bounce
SK Hynix Six-month low, below 1.2m won on Solidigm dispute
Brent $111.25, highest since May on a widening blockade
Housing starts 1.221m vs 1.305m expected, NAHB at a 2012 low
OAT-Bund spread 148bp, highest close since the euro crisis
3M Beat and raised, Amazon warrant on $6bn of optics
SpaceX First profitable quarter expected tonight on $12.9bn
Oct 28 hike odds 15% to 17%, December still near 90%

Tuesday, October 20, 2026, written at 9:15 a.m. Eastern, fifteen minutes before the open. Every future, yield, currency and commodity level in this note carries that stamp, and the European figures are intraday rather than settled. September housing starts have printed. The Federal Reserve is in its quiet period until October 28. Netflix and SpaceX both report after the close tonight, and Tesla, Meta and IBM follow on Wednesday, so the week’s information density peaks in the next thirty hours.

Top Call: The Bear Case Found a Third Leg, and It Is Opacity

S&P Global’s head of digital infrastructure told a London roundtable yesterday that it is safe to assume the largest technology companies carry more artificial intelligence debt than the agency can see. The mechanism he named is interconnected financing structures behind data centres and neocloud providers, where a hyperscaler supplies a guarantee or a long dated purchase commitment that never appears as debt in its own filings. S&P put combined capital spending across Amazon, Microsoft, Alphabet, Oracle, Meta and SpaceX above $7tn between 2025 and 2030, with negative free cash flow until at least next year, and it has changed no ratings.

That last clause is why this matters more than a downgrade would. A downgrade is a number you can trade against. A rating agency saying the leverage is probably larger than disclosed and declining to size it is an instruction to widen your error bars on every balance sheet in the complex. Credit told us in the first week of October that these assets cost more to finance than the plans assumed. DeepSeek told us on Monday that you may need fewer of them. Today adds the part that compounds both, which is that you cannot measure either exposure from the documents.

The tape behaved accordingly and the split is the cleanest in weeks. Dow futures are up 0.14% while Nasdaq 100 futures are down 0.50%, the Kospi gave back its entire Monday bounce and more, and 3M is up roughly 5% on a data centre supply agreement with Amazon. Capital has not left the theme. It has moved to the part of the theme where the revenue is contracted and the balance sheet is legible, and that distinction is now doing most of the work in this market.

Korea Sold the Rebound, and the Reason Was Governance

The Kospi closed down 3.73% at 5,552.74 after trading as low as 5,493.24, erasing Monday’s 1.76% gain and then some. The Kosdaq fell 3.33% to 781.97. SK Hynix dropped about 4.2% to a six-month low, breaking below 1,200,000 won for the first time since April and taking its decline since September 30 to roughly 31%. Foreign ownership of the shares has fallen to 48.85%, the lowest since December 2021.

Three things hit at once, and the ordering matters. The Financial Supervisory Service said a United States flotation of the Solidigm storage subsidiary could breach the rules protecting minority shareholders from dual listings, and a shareholder group has demanded management abandon the plan. Reports put the listing as early as 2027 at a valuation somewhere between $100bn and $150bn. Separately, CXMT said it has made progress on a denser 4F squared DRAM architecture with DDR5 server modules due by year end, and reported July to September sales up roughly sevenfold, although Citi argues volume production before 2028 is unlikely and the supply impact is therefore limited. Then S&P spoke.

Our read is that the governance item is the one with real content and the one the market is least equipped to price. A CXMT headline is a forecast about 2028 that can be argued with. A regulator questioning whether a company can monetise its most valuable subsidiary is a statement about who owns the upside in the asset you already hold. Korean memory has spent six weeks being treated as a pure demand proxy. It is now also being treated as a jurisdiction, and that is a different multiple.

The counterweight has not disappeared. The Philadelphia Semiconductor Index rose about 1.6% on Monday, ending eight consecutive declines and the longest such run since October 2008, with Nvidia up 1.4% and AMD up 1.8%. Those gains were not retraced overnight in the American names. CoreWeave is down 3.1% and Nebius 2.9% pre-market, Nvidia is marginally lower and Apple is off about 1%, which is soft rather than disorderly. A bounce that holds in the United States while Seoul takes another 4% is consistent with our rotation framing, and inconsistent with the view that the whole complex is being liquidated.

Housing Is the Part of the Economy That Already Broke

September housing starts fell 5.2% to a seasonally adjusted annual rate of 1.221m against a 1.305m consensus, with August revised up to 1.288m. Permits were essentially flat at 1.400m. The composition is worse than the headline: single family starts fell 4.6% to 859,000 and single family permits fell 4.0% to 845,000. Yesterday the NAHB builder sentiment index fell to 29 from 32, the lowest reading since June 2012.

This is what a 5.12% ten year yield does to the one sector that cannot hedge it. Builder sentiment at a fourteen year low is not a forecast of recession, since the rest of the economy has spent two years proving it can carry rates this high. It is evidence that the restriction is landing somewhere, and that the somewhere is narrow and rate sensitive rather than broad. For a committee that has spent the autumn arguing about whether one more hike is needed, a print like this is an argument for patience rather than for an easing cycle.

Treasuries are marginally better after Monday’s back up, with the ten year at 5.12% after reaching 5.14%, the thirty year near 5.46% and the two year 4.75%. October 28 hike odds sit between 15% and 17% and December is still priced near 90%. The high yield index is at 329 basis points, which tells you the stress this desk has been writing about for two weeks remains a sector event rather than a market one. That is a load bearing observation for the whole note, and it is the first thing we would abandon if it moved.

Energy: The Blockade Keeps Growing a New Edge

Brent reached $111 and change, the highest since May, and trades at $111.25. WTI set a fresh monthly high near $100.50 and sits at $100.60. The move is geopolitical and the escalation is directional rather than tactical.

Iran’s first vice president said the country has no choice but to control the Strait of Hormuz to ensure national security, which follows the IRGC’s October 17 declaration that the strait is closed to shipping of all nationalities. On Monday the IRGC added that the blockade is not yet complete and could be extended to Bab al-Mandeb and the Suez Canal. Kpler counts at least eighteen vessels attacked in the past two weeks. The IRGC says two tankers struck mines and burned south of the strait. Energy Secretary Wright puts Gulf crude flows back at roughly 80% of pre-war levels, and ANZ warns that recovery is now the thing at risk.

Two prices carry more information than the crude benchmark. Diesel is near $200 a barrel outright and European gas is around EUR 90 per megawatt hour, the highest since December 2022, with storage at 74% full against an 89% five year average going into winter. Our framing for a fortnight has been a 2027 crude surplus sitting on top of a present distillate shortage, and nothing this week has challenged it. What has changed is the breadth of the chokepoint risk. A Hormuz disruption is a crude and condensate problem. A Bab al-Mandeb and Suez disruption is a refined product and container freight problem, and it arrives in European and Asian energy bills rather than in the Brent curve. That is the escalation worth watching, and the gas price is already telling you it has partly happened.

Europe Priced France Again, and the Spread Is Now Historic

Monday closed with the Stoxx 600 down about 0.8% near 639.1, the DAX down 1.24% at 25,304, the FTSE 100 down 1.08% at 10,765 and the CAC 40 down 1.11% at 7,868. The French ten year rose about eight basis points to 4.94% and the spread over Germany reached 148 basis points, the highest close since the euro area debt crisis. Italy trades at 111 basis points over Germany, well inside France, which is the detail that should make people uncomfortable. The Stoxx 600 is down a further 0.76% this morning near 634.7 and the DAX is off about 0.65%.

The calendar is unforgiving from here. Moody’s reviews France on October 23, the ECB meets on October 29, and unions have called a national strike for November 5. Riot damage estimates have roughly doubled. Capital Economics now sees the spread reaching 200 basis points by the end of 2027. We wrote yesterday that European outperformance rests on a compositional argument about not owning the artificial intelligence complex and that the alternative carries its own repricing risk. One session later the alternative is trading at a euro crisis spread while the complex it was supposed to hedge is bouncing in the United States.

The United Kingdom supplied the morning’s genuine macro surprise. Employment rose 102,000 against a 5,000 forecast, unemployment fell to 4.7% against 4.9% expected, and average weekly earnings ran at 3.7%. That is a labour market reaccelerating into an inflation print due today, and it moves a November Bank of England hike from a debate to a base case. Elsewhere China left loan prime rates unchanged for a seventeenth month, and Japan’s twenty year auction drew the strongest demand on record with the yen at 158.13.

Earnings: Industrials Beat, Memory’s Customer Pays Up for Optics

A heavy morning, and the print that belongs in this note is not the biggest company on the list.

Company Consensus Actual Read
3M $2.29 adj EPS $2.50 adj, raised outlook Beat, plus an Amazon optics warrant
General Motors $3.63 adj EPS $3.70 adj on $50.0bn revenue Beat, $20bn buyback, EPS guide light
GE Aerospace $2.01 adj EPS $2.25 adj on $13.29bn Clean beat and raise
Steel Dynamics $5.41 EPS $5.26 on $6.509bn Miss, shares down 6%
AGNC $0.37 EPS Loss of $1.17, book value down 16.9% Duration is still expensive

3M earned $2.50 adjusted against $2.29 and raised full year guidance, and the shares are up roughly 5%. The reason is the attachment, not the quarter. 3M has issued Amazon a warrant for up to 3,162,015 shares at $164.89 over seven years, worth about $521m, with 85% of it vesting against cumulative payments for optical connectivity and other data centre products and full vesting at $6bn of sales. Our take is that this is the most useful single data point of the morning. A hyperscaler was willing to hand an industrial conglomerate equity in itself to lock in optical supply through 2033, in the same week a rating agency warned that hyperscaler obligations are larger than disclosed. Both things are true and they are the same thing. The spending is real, the contracts are long, and the structures through which they are financed are exactly what S&P says it cannot see. For the light complex specifically, a $6bn vesting threshold is the clearest public marker yet of what optical interconnect is worth to one buyer.

General Motors beat at $3.70 adjusted on $50.0bn of revenue against $48.76bn, with adjusted EBIT up about 21.6% to $4.1bn and an 8.2% margin. Electric vehicle realignment charges collapsed to $238m from $1,592m a year ago. The board authorised a new $20bn buyback. Full year adjusted EBIT was narrowed to $14.5bn to $15.5bn and the EPS range of $12.75 to $13.75 brackets a $13.41 consensus from below. This is a company that got profitable by building fewer electric vehicles and is returning the proceeds rather than reinvesting them, which is a rational answer to a 10% cost of capital and a reminder that most of the economy is not in the compute business. GE Aerospace beat on both lines and guided the year to $8.00 to $8.10. Steel Dynamics missed on earnings despite beating on revenue and is down more than 6%, which is the industrial demand signal of the morning and not a flattering one. AGNC lost $1.17 a share against a $0.37 estimate with tangible book down 16.9% to $7.13, which is simply what happens to levered mortgage exposure when the long end goes to 5.47%.

The regional banks were uniformly fine. Synchrony earned $2.68 against $2.40 with $1.27bn of provisions, Zions, Western Alliance and BOK Financial all beat, and W. R. Berkley beat on earnings while missing on revenue. Four weeks into this reporting season the pattern has not changed: nobody is losing money on credit, and the dispersion is entirely in balance sheet positioning.

For context on the season, FactSet now has third quarter blended growth at 29% year on year, trimmed from 29.4%. With roughly 9% of the index reported, 84% have beaten on earnings but the average surprise is only 2.6% against a five year average of 7.5%. Goldman sees growth decelerating to 18% in 2027. Beats are broad and thin, which is consistent with a market that is paying for certainty rather than for upside.

Tonight: SpaceX Is the First Honest Test of the Capex Question

SpaceX reports after the close with a webcast at 4:30 p.m. Eastern, and it is the most informative print of the week for the argument this desk has been making. Consensus is around $12.89bn of revenue and roughly $0.15 of adjusted earnings, which would be the first profitable quarter since the listing. Starlink is modelled near $4.9bn on about 14.3m subscribers, up from 12m at the end of June, with average revenue per user down from $85 in June 2025 to $66 in June 2026. Capital spending is forecast near $19.6bn against free cash flow of roughly negative $11.6bn, and analysts expect around $6.4bn of revenue attributable to artificial intelligence workloads. The shares closed at $146.03 on Friday and are down more than 10% this month.

Hold those numbers next to each other. A company expected to turn its first profit is also expected to spend about $19.6bn of capital in a quarter in which it books $12.9bn of revenue, and to fund the gap externally. That is the exact profile S&P was describing yesterday, in a name that sits on its list of six. Subscriber growth of nearly 20% in a quarter against a 22% decline in revenue per user over a year is the other tension, and it is the one that decides whether Starlink is a scaling network or a discounting one.

What we want from the call is narrow. Any language on how the 2027 capital programme is financed, any disclosure of commitments or guarantees associated with compute, and any revision to the Starship cadence after Flight 14 reached orbit on September 28. Flight 15 is scheduled before year end with a first attempt at catching the upper stage. Launch cadence is the only line in this business where the capital spending converts to capacity you can count.

Netflix also reports tonight, with consensus near $0.82 and $12.88bn against company guidance of $12.86bn, a 33.2% operating margin target against 28.2% a year ago, and roughly $700m to $800m of implied advertising revenue. The stock has missed earnings estimates in three of the last four quarters by an average of 7.1%, and options imply a move of 9% to 10%. It is a useful read on consumer pricing power and a poor read on anything else this market is arguing about. United Airlines, Capital One, Intuitive Surgical and Chubb also report tonight, and Tesla options imply about 6% into Wednesday.

Also On The Tape

The Record

Two open items and one small correction.

The Volta secondary test remains open and we are now near the end of its stated window. We said the loan should trade at or below its 95 cent issue price through its first week, and that a print above 96 would mean we had overstated how expensive this funding has become. There are still no reliable quotes. If none appear by Friday we will mark it unresolved rather than claim it, because a test you cannot observe is not a test you passed.

Wednesday’s test stands exactly as written. If the efficiency story has any purchase on actual plans, at least one of Meta, Tesla or IBM should soften capital expenditure language rather than raise it. We would add one line to it after this morning. S&P’s warning makes the disclosure around those plans as informative as the number itself, so we will also mark whether any of the three quantifies off balance sheet commitments or lease obligations tied to compute. If all three raise capital spending and none of them discloses the financing, the market will have learned that the opacity is a choice.

The correction is minor and we would rather log it than let it sit. Yesterday we wrote that the eight session semiconductor decline ended last week. It ended on Monday, the day we were writing, with a 1.6% gain. The point we were making about rotation was right and the chronology inside it was wrong.

Key Events Today