Archived Edition 05Friday, Sept 25 2026Updated 9:30 AM ET

The driver changed, and I was still watching oil.

A five-year-high PMI took the 10-year to 5.11%. Oil fell anyway, through a hostile UN exchange. And Oracle declared force majeure on a Stargate campus.

Off-cadence · covers Sunday through this morning

§1 — The tape

Since Sunday

S&P 500
7,704.13
+0.70% week
−1.44% vs ATH
Nasdaq Comp
26,939.37
+1.57% week
Dow
51,349.98
−0.64% week
4th losing week
Russell 2000
2,835.57
−1.43% week
US 10-Year
5.11%
+11.0 bp
highest since 2007
US 2-Year
4.897%
+14.1 bp
highest since 2023
2s10s
21.3 bp
−3.1 bp
still flattening
Brent
$104.70
+1.16%
WTI
$93.12
−7.53%
spread blew out
Fed funds
3.75–4.00%
after Sept 16 hike

Thursday's session was a round trip — the S&P finished 1.90 points lower after two reversals. But the week underneath that flat surface was not quiet at all.

§2 — What actually happened

The economy printed a five-year high

S&P Global flash PMI · SeptemberActualExpected
Composite58.4—
Services58.756.0
Manufacturing57.0—

Composite activity at a five-year high, manufacturing at a 52-month high, services beating by 2.7 points. This is not an economy that needs help, and it landed one week after a Fed that had just told you 16 of 19 members want to hike again.

This morning added the same message from a different angle. August durable goods came in at 0.0% on the headline — but core capital goods orders, non-defence excluding aircraft, rose 1.6%. Headline flat, business investment accelerating. Strip out the lumpy aircraft line and companies are still spending.

The regime changed this week

For three weeks the story was oil → inflation → rates. This week it became growth → rates. The 10-year went to 5.11%, the highest since 2007, and it did so while crude was falling. Strong data, weak debt-auction demand and heavy government borrowing did the work that an energy shock was doing a fortnight ago. Same level, completely different cause — and a different set of things break.

§3 — The non-event

The UN went badly. Oil fell anyway.

Sunday's brief made the General Assembly the week's live variable. It delivered maximum hostility and minimum market impact. Trump threatened to "annihilate" Iran from the podium on Tuesday; Pezeshkian answered Wednesday that Tehran would not surrender and blamed Washington and Israel for global instability. No meeting, no channel, no de-escalation.

Brent rose 1.16% on the week. That is it. The supply premium that took crude to $108.75 ten days ago has continued draining on Saudi exports recovering and Aramco restoring pipeline flows, and two hostile speeches could not reverse it.

And look at the spread

Brent $104.70 against WTI $93.12 — a $11.58 spread, roughly double its normal range. WTI fell 7.53% on the week while Brent rose. That gap is the geopolitical premium isolating itself in the seaborne barrel: Hormuz risk prices into Brent, and US domestic crude simply doesn't care. If you were trading the Iran story through WTI, you were in the wrong contract.

§4 — The story of the week

Oracle declared force majeure on a Stargate campus

On Thursday Oracle sent a force majeure notice to Blue Owl Capital covering Project Jupiter, a 2.45-gigawatt Stargate data centre campus in New Mexico. The stock fell to $139.52 and is indicated another 5.5% lower pre-market this morning.

What went wrongDetail
Gas pipelineEnergy Transfer line delayed ~6 months, now targeting Feb 1 2027
PermitsRoute repeatedly denied by regulators, then rerouted
Air qualityFuel-cell permit pending, Nov 23 deadline
What the notice doesLets Oracle delay payments if the site misses its 2028 target
Oracle's position"Project Jupiter remains on our planned schedule"
Blue Owl's positionNotice "does not change the financial commitments"

Both parties are downplaying it and Oracle is not walking away from its role as anchor tenant. Take them at their word on the project. The signal is elsewhere.

Why this matters beyond ORCL

Every bear case on AI capex so far has been financial — too much debt, too little return, a 5% funding cost. This is the first big one that is physical. A pipeline that regulators would not permit. An air-quality approval with a November deadline. The constraint on the buildout turns out to be power delivery and paperwork, and you cannot solve either by raising more money. That is a much harder problem than the one the market has been pricing, and Oracle's $664B backlog is the number it lands on first.

§5 — Levels

The AI complex is now three different trades

These three names moved as one for a year. They have stopped.

52-WEEK HIGH NVDA −5.1% · $224.58 AVGO −29.4% · $349.71 ORCL −57.7% · $139.52 −60% −40% −20% 0

Distance below 52-week high, at the Sept 24 close. NVDA high $236.54 · AVGO $495.00 · ORCL $329.50.

Nvidia is 5% off its high and rose on the week. Broadcom is 29% off and fell. Oracle is 58% off and is still falling. Three names, one theme, and a fifty-point spread in drawdown. Whatever "the AI trade" meant, it does not describe a single position any more.

AVGO is back on the line

369.74shelf
366.78Sep 11 high — still untested
357.61Sep 18 close
350.24the floor — price is sitting on it
349.71last · Sep 24 close · −0.15% below
346.89Sep 24 low
337.90Sep 15 low

AVGO broke 350.24 on Sept 15, reclaimed it by Sept 18, and is now back within fifteen cents of it. A level being tested for the third time in ten sessions is a level doing real work. Added pressure this week: reports that China is scrutinising Broadcom's dominance in networking switches inside state-controlled data centres.

Pre-market Friday: AVGO $349.21 · NVDA $226.10 · ORCL $136.58.

§6 — Next week

Three catalysts in four days

Wed 30
August PCE · 8:30
The Fed's preferred gauge
This week
GDP revision
Nonfarm payrolls ahead

After a fortnight with nothing on the calendar, the pipeline refills fast. August PCE lands Wednesday Sept 30 — the first inflation print the Fed sees since it hiked, and the one that decides whether the 16-of-19 majority gets its second move. GDP and payrolls follow.

With composite PMI at 58.4 and core capital goods orders up 1.6%, the bar for a dovish surprise is now high. The risk has shifted: it is no longer that inflation is too hot, it is that growth is too strong for a Fed that has already decided it is behind.

10Y 5.11% 2s10s 21.3bp 2s10s 10bp AVGO 350.24 AVGO 366.78 Brent–WTI $11.58 SPX 7,800

§7 — Scorecard

Sunday's edition, graded

One clean hit, and a miss I have now made three editions running. The second one is the more useful of the two.

Hit
"You can't set an alarm for a diplomatic catalyst. Set the alert on crude, not on the headline." Exactly right, and it saved you from the obvious trade. The UNGA exchange was as hostile as it could have been — a threat to "annihilate" from one podium, defiance from the other — and Brent moved 1.16% on the week. Anyone positioned for headline risk got nothing. Made Sept 20 · resolved Sept 24
Partial
Scenario B — "nothing happens, the drift continues." Weighted 40% and closest to right on the tape: chips bid, cyclicals offered, Brent inside the $101–106 band. But the yield call inside it was wrong — B assumed the 10-year holds 4.95–5.05% and it broke out to 5.11%. Made Sept 20 · resolved Sept 24
Miss
Every scenario was built around oil, and oil was not the driver. A, B and C were all defined by where Brent went. The actual move came from a five-year-high PMI and a poorly-received debt auction — growth and fiscal supply, not energy. Third consecutive edition where the real driver was not in any branch. Made Sept 20 · resolved Sept 24
Naming the pattern, because three times is a pattern

Sept 13, Sept 16 and Sept 20 all made the same structural error: the scenarios were built around the previous week's driver. Oil explained early September, so every branch since has been an oil branch — right up until the market switched to trading growth data and fiscal supply instead.

The fix is a change to the format, not to the analysis. From the next edition, one branch is reserved for "the driver changes" — an explicit scenario for a variable that is not currently in the narrative. It will be wrong most weeks. It was right three times this month.

Free newsletter

Get the next one in your inbox