Thursday, October 15, 2026, written at 9:20 a.m. Eastern, fifty minutes after the September producer price and retail sales releases and ten minutes before the open. Every future, yield, currency and commodity level in this note carries that stamp. TSMC has reported. Anthropic showed third quarter numbers to investors. Chair Warsh and Beth Hammack both speak later today.
Top Call: The Question Changed From What AI Earns to Who Funds It
Oracle’s five year credit default swap went to a record 288 basis points overnight, up 26 on the session and up from 260 a week ago. CoreWeave’s went through 1,000. Its 2032 paper trades at 77 cents for a yield around 14.5%, a record for the name, against nearly $50bn of obligations. SoftBank widened 24 to 424. Amazon, Alphabet, Meta and Nvidia all set record wides of their own, far tighter in absolute terms but moving the same way on the same day.
That is the whole note. For six weeks the market has argued about demand, about whether the compute is sold and whether the chips are allocated. It has stopped arguing about that. TSMC just put a 68.7% profit increase and a fourth quarter guide roughly $2bn above consensus on the table and the ADR is down about 2% in the pre-market. ASML did the same thing yesterday and fell. The operating data is not in question. What is in question is the cost of capital for the entities that sit between the silicon and the cash flow, and that number has moved a long way in ten sessions.
The bond market said it more clearly than the equity market did. September producer prices rose 1.3% on the month and 6.3% on the year against forecasts of 0.5% and 5.3%, the largest monthly jump in about four years, and retail sales rose 0.8% against 0.3%. On that pair of prints the two year fell 7 basis points to 4.73%, the ten year fell 11 to 5.08% and the thirty year fell 10 to 5.43%. A curve does not rally into data like that because it has changed its mind about inflation. It rallies because somebody is buying Treasuries for a reason that has nothing to do with the inflation forecast.
So we are retiring the framing we used yesterday. The constraint on this market is no longer the discount rate applied to AI cash flows. It is the availability of the financing that builds the assets those cash flows come from. Those are different problems and they do not resolve the same way.
The Repricing Is Inside Equities, Not Against Them
The futures board is the most useful thing on the screen this morning. Dow futures are up 0.24% at 51,981. Nasdaq 100 futures are down 1.29% at 29,802. S&P futures split the difference at 7,743 and down 0.35%, and Russell 2000 futures are up at 2,752 after the cash index fell 1.93% yesterday into correction, more than 10% below its August peak.
That is a rotation, not a liquidation. Capital is leaving the levered end of the AI complex and going into Treasuries, gold at $4,155 and up 1.21%, and the parts of the equity market that do not need a syndicate to fund next year’s capital plan. Small caps rising on a day like this is the tell, because the same index fell 1.34% on Tuesday when the long end sold off. Small caps care about the front end, the front end just rallied 7 basis points, and nothing about a small cap balance sheet depends on whether a neocloud can roll its paper.
Yesterday’s close framed it. The S&P lost 1.11% to 7,718.94, its worst session since late July, the Nasdaq Composite lost 1.88% to 26,692.87 and the Dow was essentially flat at 51,621.48. The Philadelphia Semiconductor Index fell 4.3% for a sixth consecutive losing session. Underneath that, CoreWeave fell about 15%, Oracle 8.8%, Nebius 10% and Marvell 8.7%, while Nvidia and Broadcom lost close to 3%. The index numbers understate it. The dispersion is the story.
TSMC: A Record Quarter, a Raised Year, and a Lower Share Price
| Metric | Consensus | Actual |
|---|---|---|
| Revenue | US$45.6bn | US$46.72bn |
| EPS (per ADR) | US$4.45 | US$4.60 |
| Net income | n/a | NT$763.06bn, +68.7% y/y |
| Q4 revenue guide | US$50.5bn | US$52.3bn to US$54.3bn |
This is the sixth consecutive record profit quarter. Revenue rose 50.9% in local currency and 41.1% in dollars. The fourth quarter guide implies roughly 58% growth at the midpoint and the full year dollar revenue growth forecast went to about 44% from the previous more than 40%. Capital spending for 2026 stays at $60bn to $64bn, with management saying the 2026 to 2028 total will exceed the prior three years combined. C.C. Wei described demand as extremely robust and said he sees no sign of the AI trend slowing, with two nanometre capacity set to more than double in 2027.
The only soft line is gross margin at 66.4%, diluted by the initial N2 ramp, which is a known and temporary cost of the thing everyone wants them to be doing.
Our take is that this is a very good print and that the share reaction is still correct. A foundry that grows profit 69% and guides $2bn to $4bn above consensus and trades down is not being disbelieved. It is being told that its earnings were already in the price and that the marginal risk in the sector has moved somewhere its income statement does not reach. We wrote almost exactly this about ASML yesterday and the thesis improved overnight rather than weakened. Two capacity beats in two days, both sold. The capacity question is settled. The funding question is not.
Anthropic Showed Its Numbers and the Valuation Talk Fell by a Third
Anthropic told investors third quarter revenue reached $16.8bn, up about 45% from $11.6bn in the second quarter, against $4.7bn in the first. It said it was adjusted profitable for a second consecutive quarter, cash flow positive, with an adjusted gross margin near 85%. Those are genuinely strong numbers for a company at this stage and we want to be clear about that before the rest.
The rest is that investors who sat through the presentation now reportedly expect the listing to come below $1.5tn against the $2tn that had been discussed, and that the sequential growth rate fell from roughly 147% to roughly 45% in one quarter. Both can be true at once. A company can be executing well and still be worth meaningfully less than the number its own bankers were circulating a week ago, because the number was never anchored to anything except the previous quarter’s growth rate extended forward.
Hold this against the credit panel. The same six year compute agreement sits behind the Bitdeer lease that Volta’s term loan C collateralises. Yesterday we wrote that when the equity and credit marks on this asset disagree by thirteen times, the credit market is usually the one that has done the work. Today the equity expectation moved about 25% toward the credit market inside a single meeting. We are not claiming that as a forecast. We are noting that the gap is closing from one side only.
The Volta Test Is Unresolved, and That Is Itself Information
On Monday we pre-committed this desk to a specific test: Volta’s $5.012bn package would price at or wider than the 11.75% top of talk, or with a cut best efforts tranche, or the funding thesis was wrong. Commitments were due at 5 p.m. Eastern yesterday. As of this writing no final spread, no allocation and no clearing level has printed publicly.
We are not going to score that as a win. The honest statement is that the test has not resolved and we will carry it open.
What has changed around it matters more than the print will. The book was marketed at SOFR plus 625 to 650 with an original issue discount of 97 to 98, for 11.21% to 11.75%, with only $1.3bn of the $5.012bn underwritten and the $3.712bn GPU loan left to clear on a best efforts basis. Since that talk was set, CoreWeave’s own secondary paper has moved to a 14.5% yield and Oracle’s default insurance has set a record. Talk struck against a market that no longer exists is not aggressive pricing. It is stale pricing, and a book that goes quiet past its commitment deadline in a tape like this one is usually quiet for a reason.
Korea: We Were Wrong, and It Is Worth Saying Plainly
The Kospi closed down 7.57% at 5,554.22. It opened 5.59% lower, traded as much as 8.13% down, triggered a sell sidecar at 9:12 a.m. and a circuit breaker at 12:23 p.m., the tenth of the year and the first time since July 29 that both the Kospi and the Kosdaq halted on the same day. The Kosdaq fell 7.65% to 760.82. Samsung Electronics closed down 9.48% at 210,000 won. SK Hynix fell 11.06% to 1,262,000 won. Samsung Electro-Mechanics lost almost 17%. This is a seventh consecutive losing session and roughly 21% across the run.
On Monday we called the ChangXin reaction an overreaction in the near term. The index has fallen about 8% further since we wrote it and the market has halted twice. That call was wrong, and it was wrong for a reason worth naming rather than a reason worth excusing. We treated a supply side claim about memory architecture as a story about memory, and priced it as a terminal value question for two companies. It was never that. It was the first crack in a structure where Korean retail had financed semiconductor exposure with leveraged single stock products, where some holders are now reportedly down more than three quarters, and where the marginal seller is therefore mechanical rather than analytical. Mechanical sellers do not respond to a view about 4F squared cells.
Tokyo took the same hit through a different door. The Nikkei closed at 64,425.95, down 4.90% and 3,318.34 points, the third largest point decline in its history, with an intraday low of 64,092.74. It is now more than 10% below the 72,831.73 record set in June. Advantest fell almost 15%, SoftBank between 8% and 9%. The Topix fell only 2.4% to about 3,999, which is the cleanest available measure of how narrow this is. Hong Kong fell 1.47% to 23,955.59, the CSI 300 fell 2.00% to roughly 4,083, with sources still disagreeing on the exact level, and the Shanghai Composite fell to 3,707. Australia was flat.
Macro: The Data Was Hot and Nobody Traded It
| Release | Consensus | Actual | Prior |
|---|---|---|---|
| PPI m/m | +0.5% | +1.3% | +0.4% |
| PPI y/y | +5.3% | +6.3% | +5.4% |
| Retail sales m/m | +0.3% | +0.8% | +1.1% revised |
A 1.3% monthly producer price print is a large number by any standard and the biggest in roughly four years. Retail sales beat by half a point, although the August figure was revised lower and the series is not inflation adjusted, so real spending is softer than the headline suggests. Taken with yesterday’s 3.7% headline and 0.3% core CPI, this is an unambiguously inflationary set of releases.
The market’s answer was to buy bonds. We want to be careful not to over-read a single morning, so here is the narrower claim. October 27 and 28 hike pricing spiked above 50% intraday yesterday and settled back below 30%, with venues now quoting somewhere between 29% and 41%. December sits near 88% to 90%. One day volatility rose almost 32% while one year volatility rose 1.3%. That is a market repricing a window rather than a path. It thinks the next few weeks are dangerous. It does not think the destination moved.
Which is the right structure for what is actually happening. A credit event in AI infrastructure would be disinflationary and would argue for fewer hikes, not more, which is precisely why the curve can rally on a 6.3% producer price print without any contradiction. Warsh speaks later today with no forward guidance on offer, as has been his practice, and Hammack has already said this week that higher rates may be needed to stop an inflationary mindset taking hold. Neither of them will mention credit spreads. Both of them are now constrained by them.
Energy: Lower on a Sanctions Headline, Against a Ship on the Seabed
Brent is $104.04 and down 1.90%, WTI $94.81 and down 0.29%, after Brent settled at $106.06 and WTI at $95.09 yesterday, the latter a three week high.
The move down is a policy headline. Reuters reports Washington is weighing relief on sanctions for some shadow fleet tankers in order to bring record freight rates down. That is a supply of shipping story rather than a supply of oil story, and it tells you how much of the current price is logistics rather than barrels.
The move it is fighting is worse than it was yesterday. The GFS Pride, a Singapore registered container ship, sank in the Strait of Hormuz after being struck by an unidentified projectile and catching fire. All 21 crew were evacuated safely. A second Singapore registered vessel, the GFS Giselle, was also hit, with UKMTO logging two container ships struck on the day. The IRGC Navy claimed the sunken ship hit a naval mine while using what it called an unauthorised route and did not name the vessel. The maritime security firm Marisks disputes that account and describes the claim as part of a psychological operations campaign. Twenty six vessels have been struck this month and this is the third commercial ship sunk since the war began. Traffic is largely holding up under American naval protection.
Our read is unchanged and we will keep repeating it until the market prices it. Crude is a surplus story for 2027. Freight, insurance and distillate are a shortage story right now, and a sunken box ship in the strait raises the cost of the second without touching the first. Anyone reading $104 Brent as the all clear is looking at the only part of the energy complex that is behaving.
Europe: The Six Session Run Ended
The Stoxx 600 is around 635.0 and down 0.43%, the DAX 25,183 and down 0.46%, with the CAC near flat and the FTSE 100 the sole gainer at 10,778 and up around 0.47%. That follows a close at 637.76 yesterday, the highest since September 29.
We have argued for a week that European outperformance was compositional rather than analytical, that these indices simply hold banks, insurers, defence and luxury instead of memory and optical, and that the trade worked only while the problem stayed inside semiconductors. The problem left semiconductors overnight. Oracle is not a European index constituent, but the cost of capital for levered technology balance sheets is not a national question, and the first session in which Europe stops outperforming is the session in which that became visible. The FTSE holding up on energy and defence weight is consistent with the same explanation rather than a counterexample to it.
Also On The Tape
- Penske Automotive agreed to a take private by Penske Corp and Mitsui at $212 a share in cash, an enterprise value near $16.3bn
- Rocket Lab delayed the debut launch of Neutron into next year, the second slip on a vehicle carrying most of the equity story
- Ericsson reported nine month net profit of SEK 10.74bn, down 45.9%, on sales of SEK 161.5bn, down 3.5%
- TotalEnergies guided a third quarter refining margin of $33.2 a barrel against $13.5 in the prior quarter, the distillate squeeze showing up in a European income statement
- PNC beat at $5.09 against $4.89 with net income up 15%, on revenue of $6.56bn that slightly missed, while BNY grew revenue 12% to $5.68bn and beat
- Comtech missed badly, with quarterly sales of $66.2m against $111.7m expected
- Piper Sandler and Perella Weinberg have reportedly abandoned merger talks
- Cipher Mining and IREN fell with the neoclouds on the same borrowing cost concern
- Pre-market chips: Micron down about 4%, AMD down 2.5% to 3.5%, Nvidia down about 2%, TSMC ADR down 1.9%
- AAII sentiment: 38.7% bulls against 41.6% bears, the first bearish majority in several weeks
The Record
Yesterday we wrote that the market had raised its estimate of how long policy stays restrictive and had paid for it in the five to ten year part of the curve. That was a structural interpretation of a four hour move and it did not survive the session. The two year rose as much as 16 basis points and the ten year as much as 12 before both reversed, and this morning the entire curve sits below where it was before the CPI release. We captured a 9:15 a.m. snapshot and built a thesis on it. The lesson is narrow and we will apply it: an intraday move on a data day is not a repricing until it survives a close.
The thesis that did work is the one we have carried since October 7, that the binding constraint in this cycle is the cost and availability of credit rather than the demand for compute. Oracle at a record 288 basis points, CoreWeave through 1,000, and an equity tape that fell 1.1% on a day when its three largest reporting companies beat is that argument being paid out. We take credit for the frame and none at all for the timing, since we spent two of the last three sessions writing about inflation while this was building in spreads we were already quoting.
We were wrong on Korea and have said so above. We are retiring the near term overreaction call rather than defending it, and the specific error was mistaking a mechanical seller for an analytical one.
Two items stay open. Volta is unresolved and remains open on its original terms, at or wider than 11.75% or with a cut tranche. The second is new and we will mark it tomorrow: if this is a credit event rather than a growth scare, investment grade technology spreads should keep widening while the two year keeps falling. If the two year reverses higher and spreads come in together, then this was a positioning washout and we have over-read it.
Key Events Today
- Through the morning: TSMC call completed; Empire State and Philadelphia Fed surveys
- 10:30 a.m. Natural gas inventories
- Afternoon Chair Warsh speaks, with Beth Hammack at 4:30 p.m.
- After the close Alcoa, Interactive Brokers and J.B. Hunt report
- Ongoing IMF and World Bank annual meetings, Semicon West
- Watching Any public print on the Volta $5.012bn allocation