AInvest Option Flow Digest — 2026-09-04

Ten names, twenty legs, $221.07M of premium — and only one of them is a clean directional bet.

The rest are a calendar, a hedged risk reversal, a three-legged rates package, two rolls, an unwind, and a fully delta-hedged put. Each looks like a wager on where a price is going. None of them is. Today is a good day to learn how to tell the difference — because the three checks that separate them (open-interest history, the paired share block, the net delta) would have re-labelled four of these ten trades.

⭐ The one clean directional trade is also the day's best story. At 09:50 someone crossed a small TYRA call spread whose upper strike had prior open interest of exactly one contract. Later that morning the company announced a Phase 2 data readout for Wednesday, September 9 — nine days before the spread expires. The stock closed +15.08%.

⚠️ And the feed dropped three legs again. TYRA's second leg turned an outright call into a capped spread; LQD had a third leg that adds a short call to what looked like a put structure; and URA's missing leg turned a call purchase into a calendar roll.

⭐ The day's other standout is a date. Someone has been buying GE December $340 puts for a month — the line grew from 149 to 14,744 contracts over 21 sessions, and counting those sessions back lands on August 6, the day GE printed its all-time high of $388.84. They started hedging the session the stock topped. It worked: GE is −13.7% from that high. Today they took $15.5 million of that profit off the table.

⏳ Nothing is settled until tomorrow — and yesterday we learned that the hard way. We called fifteen legs on another name "proven opens" because each printed larger than the open interest on its strike. Open interest then fell 430 across 19,060 contracts. The rule is rewritten below, and it changed how every leg on today's board is labelled.

🔁 OI Review — Thursday's Predictions, Scored

The September 4 pre-market open-interest snapshot is in, and it settles all thirty-four legs from Thursday's board. Twelve of the twenty-seven legs where we published a lead branch landed on it. Fifteen did not — every single one of them PDD, whose $69.6 million ladder turns out to have created no open interest at all. Two more names moved off the fence in ways that change their story.

🔄 The inversions

1. PDD's fifteen-leg ladder created zero new open interest, and our "proven open" claim is refuted. We wrote it flatly: "Open-versus-close is already proven. These are opens" — because every leg printed larger than the open interest standing on its strike. Across 19,060 contracts on fourteen lines, open interest fell 430. Eight lines did not move by a single contract. We checked the three things that could make that a data artifact and it is none of them: six control strikes elsewhere on the PDD chain updated normally, the tape carries no cancellations and end-of-day volumes match the prints, and the January $130 line has now absorbed 7,382 / 2,258 / 2,552 / 4,822 contracts on four separate sessions since August 27 with open interest pinned at exactly 1,015 the entire time. Corrected PDD analysis

And the contrast we drew was backwards. That article set PDD against the near-identical SPCX parity-put program from September 1 — the one that created no net position — and argued PDD was "the mirror image; the sizing says this one is genuinely building something new." It was the same trade, not the mirror image.

2. NKE's $17.4 million "downside package" was an unwind, not a new bearish bet. We refused to call open or close on either leg, and the close branch won both. The $40 put line fell 126,956 → 39,059 (Δ −87,897, 99.9% capture) and the $45 call line 115,159 → 46,614 (Δ −68,545, 95.2%). Roughly 156,000 contracts of standing September-18 exposure came off the board. There is no new $231 million short-delta position — one was retired, ahead of a fortnight with nothing scheduled inside it. Corrected NKE analysis

3. BABA's $10.65M straddle mostly changed hands rather than getting put on. Branch 3 (transfer) on both legs: the $100 call moved 30,383 → 30,760 (Δ +377, 7.5%) and the $100 put 33,852 → 34,613 (Δ +761, 15.2%). Only ≈1,138 of 10,000 printed contracts are new. Corrected BABA analysis

4. SPCX split down the middle. The $245 long put opened in full — 1,000 → 2,700, a perfect 100.0% capture, beating the 40-75% band we published on the thinnest arithmetic floor of the day. But the $110 floor leg did not open: 9,750 → 9,439, a fall of 311 on a 1,700-lot print. The long position is real; the "floor" is better read as the price at which the package was financed than as a fresh short obligation. Corrected SPCX analysis

The lesson we are paying for this week — and it costs us a shortcut we have leaned on: size relative to prior open interest is a good prior, not a proof. It held on eleven legs and failed on fifteen, all on one flow type: a deep-in-the-money parity ladder crossed on the floor. The arithmetic floor assumes every print is either an opening or a closing trade. Those prints are neither. Where we refused to pick a branch, the snapshot did the work honestly. Where we used the word "proven" in advance, we were wrong.

🔀 Two forks the snapshot settled

• PCG was a ROLL, not a doubling. We published both readings and refused to pick. The October $20 line fell 315,665 → 71,868 (−243,797, 99.5%) while the November $20 line went 0 → 246,278 (100.5%). This was one 245,000-lot position moved out by a month for a 13-cent-per-contract net debit — not 490,000 contracts of new exposure. PCG
• MSFT opened cleanly, Branch 1. Size sat well below the standing 26,382 open interest so we published no lead branch; the line went 26,382 → 31,379 (+4,997, 99.9%), within three contracts of the full-open figure we predicted. The stock-substitution position is genuinely new, walking into the projected November 19 ex-dividend one day before expiry. MSFT

✅ Confirmations

• GLD's 175,000-lot payrolls vertical opened at ≈100% on both legs — $395 put 1,462 → 176,511, $392 put 2,694 → 177,764. Genuinely new, and direction is still unknowable and always will be: open interest is side-blind. GLD
• IREN's two-year short straddle is real — $120 put 2,744 → 9,962 (100.9%), $120 call 2,775 → 9,925 (100.0%). The frozen 2028 line genuinely woke up. IREN
• OKLO opened at 99.8% — 2,932 → 53,353, top of our 94-100% band. So the ≈1.99 million shares of delta on the short side are real, into a 16-17% short float. OKLO
• QQQ opened at 98.9% — 3,970 → 13,361, above our 70-100% band. The day-over-day reversal against Wednesday's put buying is genuine. QQQ
• SPY opened at 98.0% — 3,298 → 40,048. The hedged, non-directional structure is new and live into payrolls, CPI and the Fed. SPY
• NN opened at 135.1% — 211 → 10,244, meaning essentially the entire session's volume on that strike opened, not just the block. NN
• RARE opened at 111.3% — 1,343 → 6,909, the whole line, ahead of the September 19 PDUFA date. RARE

What changed, and the standing lesson

Thursday's board published as $306.02M net across thirteen tickers. Nothing about the money moved — every side and every price was right, so the net figure is unchanged. What moved is what those dollars bought: PDD's $54.01M created nothing at all, NKE's $17.38M retired ≈156,000 contracts instead of establishing $231M of new short delta, PCG's 490,000 contracts were one position moved out a month rather than two positions bought, and BABA's straddle mostly changed hands.

The score to remember: a headline premium tells you nothing about open versus close — and neither does size exceeding open interest, on the wrong kind of flow. A deep-in-the-money parity ladder crossed on the floor can print thousands of contracts a day for a fortnight without creating a single new position. Only the next-morning open-interest snapshot settles it.

📋 The Board at a Glance

Ticker	Premium (net)	Expiry tag	The trade	What it means	Catalyst
QQQ	$18.26M   (of $81.90M gross)	Monthly → EOM	A   $700 call calendar   — buy December, sell October	⚠️   Nets to −0.05 delta. A bet on WHEN, not where	Every mega-cap reports in the back window
SPY	$6.49M   (of $14.61M gross)	Quarterly   — Sep 18	32,460 $760/$745 put spreads,   6.5 : 1	The only clean directional bet today	Bought 2 hours after the jobs report
SMH	$3.25M credit   (of $26.73M gross)	Monthly   — Oct 16	A risk reversal   102.5% delta-hedged	Not directional   — a bet on skew, not price	ASML Oct 14, TSMC Oct 15 — the last 48 hours
EWY	$39.83M	LEAP   — Jun 2027	Three legs, 2:3:1 —   tranche two of an unwind	Profit-taking, not a new Korea bet	Up   +91.9% YTD ; SK hynix now has its own ADR
LQD	$1.09M   (of $5.79M gross)	Quarterly → LEAP	Three legs; a brand-new March-2027 $104 put line	Financed downside on credit   — pays on rates   or   spreads	At its   52-week low ; the strike is 20bp away
KWEB	$1.80M	LEAP   — Jan 2027	100,000 far-out-of-the-money calls	⚠️   The feed reports NO side. Direction unknown	US–China tariff truce expires Nov 10
URA	$0.82M   (of $1.53M gross)	Quarterly → Monthly	A $50 call   rolled out one month	Buys time   — moves the strike inside the cone	World Nuclear Symposium Sep 9–11
TYRA	$0.80M   (of $1.53M gross)	Quarterly   — Sep 18	A $30/$45 call spread,   3.7 : 1	Event-driven, defined risk	Phase 2 data September 9 — announced this morning
GE	$15.55M credit   (of $42.05M gross)	Monthly → LEAP	A put   rolled down and out , $340 Dec → $300 Jan	Takes profit and REDUCES protection	A month-long build that began on the all-time high
IGV	$5.31M	Monthly   — Dec 18	12,200 $100 puts,   101.9% share-hedged	Not a short   — a bet on volatility, not price	Oracle AND Adobe both report Sep 10
	$93.21M net		$221.07M gross

Net is the debit paid or credit collected per package. Gross double-counts every spread.

🔍 What's actually interesting today

TYRA: an $800,000 spread built five days before a readout announced that morning

At 09:50 ET someone crossed a $30/$45 call spread expiring September 18 — 2,500 contracts, $800,000 net, paying up to $2.95 million. The upper strike had prior open interest of one contract. It was purpose-built.

Later that day Tyra announced a conference call for Wednesday, September 9 at 08:00 ET to present initial Phase 2 SURF302 results for its oral FGFR3 drug. The spread expires nine days after that readout.

We cannot tell you whether the trade came before the headline or after it. The cross printed with the stock at $24.99; TYRA closed at $28.61, up 15.08% on ten times normal volume. The exact wire timestamp isn't something we can source, so we won't imply foreknowledge — but the structure fits the event precisely, and 22% of the float is short at nearly 13 days to cover.

The honest risk: a binary readout can take this to zero. The whole $800,000 is at stake on one morning's data.

The jobs report was a hawkish shock — and that changes how to read SPY

August payrolls came in at +162,000 against a consensus near +53,000 — roughly three times expectations — with June and July revised up a combined +55,000. July flipped from −23,000 to +21,000.

Watch what moved. Equities barely flinched (the S&P was −0.26% mid-morning). Yields jumped: the 2-year to 4.39%, its highest since January 2025, the 10-year to 4.78%, the 30-year to 5.24%. September hike odds went above 50%.

Two hours later, a 32,460-lot SPY put spread crossed. It needs −1.70% just to break even on a day the index fell a quarter of a percent. That is not a bet the economy is rolling over — it reads as a rates hedge into the September 11 CPI, the September 15–16 Fed meeting and triple-witching, all inside the expiry.

QQQ's calendar has a clean structural answer

Someone bought December $700 calls and sold October $700 calls — same strike, 10,000 each. The package nets to −0.05 delta. It is almost perfectly direction-neutral.

So what is it betting on? Timing. Look at what falls in each window. The short October leg expires thirteen days before the first mega-cap reports. Microsoft, Meta and Alphabet report October 29; Apple and Amazon October 30; NVIDIA in November. Add two Fed meetings, the December dot plot and December witching — all of it lands in the December expiry they bought.

They also paid 21.2% implied volatility for December and sold 19.1% for October — the reverse of the usual calendar. They wanted the back month specifically, and paid up for it.

EWY: the same position, being closed for the second day running

Three legs crossed for $39.8 million, all reported as buys. It looks like a large new Korea bet. The open-interest history says it is the opposite.

On the previous session, exactly 3,334 / 5,000 / 1,668 contracts traded on these same three lines — and open interest fell by exactly 3,333 / 5,000 / 1,668. A 100% close on all three. Today's sizes are precisely double those amounts, in the identical 2:3:1 ratio.

Why now: EWY is +91.9% year to date. The KOSPI peaked at 9,000 on June 18, then fell ≈10% on June 23 when MSCI declined to upgrade Korea to developed-market status. And a structural one worth knowing: SK hynix listed its own Nasdaq ADR on July 10 — so US investors no longer need EWY to own 21% of its weight.

(This read is built on open-interest history across sessions, not on comparing today's size to open interest. That distinction is exactly what we got wrong yesterday.)

GE: a month of protection bought on the exact day of the top — and today they cashed some in

At 12:35 a two-leg package crossed on the floor: sell 12,100 December-18 $340 puts at $23.80, buy 12,100 January-2027 $300 puts at $10.95. Net, $15.5 million came in.

The open-interest history is the whole story. That December $340 line went from 149 contracts to 14,744 over the last 21 sessions. Count the sessions back and you land on August 6 — the day GE printed its all-time high of $388.84. Someone began buying downside the session the stock topped, and GE is now −13.7% from that high and −11.1% over the trailing month, still +8.9% on the year.

But read what the roll actually does. The old put has a delta of −0.4803; the new one −0.2423. Same contract count, half the sensitivity — the roll hands back roughly 288,000 shares of downside while collecting $15.5 million. This is a hedger banking a profitable hedge and accepting thinner cover, not someone getting more bearish.

What it does not do is dodge earnings. Q3 results are confirmed for October 20 — inside both expiries. And a footnote worth having: not one of the 22 analysts covering GE has a target below either strike (average $404.90, low $347). The stock's problem is not the business — Q2 revenue was $13.3B, up 21%, with guidance raised, and the stock fell ≈5% the next day, the third straight beat-then-selloff. At 39.3× earnings, good news has been arriving pre-paid.

⚠️ The December leg gets no open/close call from us. 12,100 contracts printed against 14,744 standing — below the line, so the tape cannot prove whether it closed the old hedge or added to it. The January leg is different: 12,100 against 862, a 93% margin.

IGV: a $5.3 million put that is not a short at all

At 12:55 someone bought 12,200 December-18 $100 puts on the software ETF for $5.3 million — and paired it with 414,800 shares of stock at $104.90. We measured the package's delta independently: 406,870 shares. The block covers 101.9% of it.

A put plus the matching stock is not a bearish position. The price risk is cancelled by design. What is left is volatility and skew — this desk is long downside convexity in software and flat on where software goes.

⚠️ We should correct our own first instinct here. We assumed IGV was the classic "AI eats per-seat software" hedge. It isn't, by weight. Top five: PANW 10.43%, PLTR 10.24%, MSFT 9.31%, CRWD 7.46%, ORCL 5.78%. Cybersecurity is 21.3% of the fund and Palantir is an AI beneficiary. The AI-disruption names are only ≈17.9%.

The better-sourced driver is rates. On September 1, with the 10-year past 4.8% — the highest since October 2023 — IGV fell 3% while QQQ fell 0.9%, with CRWD −7% and PANW −6% despite a beat. And six days out sits a dated binary: Oracle and Adobe both report September 10 after the close — 9.36% of the fund on one evening. Oracle is the transmission line from the bond market into this ETF: −$23.7B of FY26 free cash flow, $55.7B of capex and $43B of debt raised.

Timing matters too: IGV set its 2026 high of $110.57 on August 31, two sessions before this put was bought, after a ≈47% rally off the April 10 closing low of $74.67 — itself the bottom of a ≈36.6% fall from the September 2025 high. Fund flows agree — ≈−$819.5M over the past month, having taken in $1.2B in February.

⚠️ Two honest caveats. A second 68,000-share print sits nearby; if it belongs to the package the hedge is 118.7%, not 101.9% — an over-hedge, and we cannot tell whether it belongs. And the $103-strike tranches from late August cannot be called "short exposure" unless we know they were hedged too. If they were, the right phrase is accumulating convexity, not shorting.

KWEB: the feed will not even tell us which way

100,000 January-2027 $35 calls crossed for $1.8 million — and the Buy/Sell column reads "N/A". The print landed exactly at the mid.

So there are two readings and we lead with neither. Bought, it is a cheap lottery ticket on a beaten-down index (KWEB is −23.5% YTD) ahead of the November 10 expiry of the US–China tariff truce. Sold, it is $1.8 million of premium collected against a large obligation. The strike is +34.9% away — outside even the longest implied-move cone we have, which is why it costs eighteen cents.

👥 What this means for you

🎰 If you trade for the big score. TYRA is the clearest shape you like — defined risk, a dated catalyst, 3.7:1. But be precise about what "defined risk" means: if Wednesday's data disappoints, the entire $800,000 premium is gone, not reduced. And notice the structure — they capped the payoff at $45 rather than buying calls outright, which cut the cost by nearly a quarter. The professionals bought the event, not the fantasy.

📈 If you swing trade. The levels worth writing down: SPY support $770 (a large negative-gamma shelf) with resistance at $772; QQQ $715 support / $720 resistance; LQD sitting on a very large negative-gamma shelf at $105, right at the money; IGV pinned between a very strong $105 resistance shelf and very strong $100 support — the exact strike that was bought; and GE with resistance at $340 (very strong) and support at $330 — again, the strike that was sold. And the dated calendar in front of everything: Oracle and Adobe September 10, CPI September 11, the Fed September 15–16, triple-witching September 18.

💰 If you sell premium. GE is today's cleanest example of the trade you are usually doing: $15.5 million collected by giving up a hedge that had already worked. Notice what it costs — half the downside sensitivity, into an earnings date that sits inside both expiries. Selling something that has paid off is not free; it is a decision to carry more risk from here. SMH is the other one. A risk reversal that looks like a bullish bet is 102.5% delta-hedged by a 702,000-share block — so the seller isn't taking a view on semiconductors at all, they're selling downside volatility and buying upside. And the window was chosen: NVIDIA doesn't report before the October expiry, though ASML and TSMC land in the final 48 hours. Picking the window is most of the job.

🌱 If you are new. Today's lesson is that a trade can look like one thing and be its exact opposite. EWY's $39.8 million of "buying" is a position being closed. SMH's bullish-looking risk reversal is direction-neutral. IGV's $5.3 million put purchase is not a short — the stock bought alongside it cancels the direction. QQQ's call calendar isn't a bet that QQQ rises. The way to tell is never the headline premium — it is the open-interest history, the paired share block, and the net delta. Those three checks would have re-labelled four of today's ten trades.

🧯 Risk control and patience

• Twenty of twenty legs were negotiated blocks — the fifth straight session with no lit prints anywhere on our board. No aggressor exists here; the buy/sell labels are allocation conventions.
• Four legs get no open/close call from us at all — SPY's $760 put, QQQ's October $700 call, URA's September $50 call and GE's December $340 put all printed at or below existing open interest.
• And we are not using the word "proven" today. Yesterday we called fifteen legs on another name "proven opens" because each printed larger than the open interest sitting on its strike. Open interest then fell. A wide margin makes an open the likely branch, not a settled fact — and we have rewritten the rule accordingly.
• Seven of today's ten structures are not directional at all — a calendar, a hedged risk reversal, a rates package, two rolls, an unwind and a fully share-hedged put. Headline premium describes the size of a structure, not the size of anyone's conviction.

Position sizes here run into the tens of millions. Yours does not have to. The most valuable habit in this business is being willing to do nothing on a day you do not fully understand what you are looking at.

⏳ Tomorrow morning, ≈06:30 ET

Leg	Prior OI	Size	Lead branch	What would refute it
SPY $745P	20,766	32,460	open   (40–100%)	a fall
QQQ Dec-31 $700C	2,198	10,000	open	a fall
SMH $535P / $615C	6,328 / 2,883	13,000	open	a fall
TYRA $30C / $45C	1,164 /   1	2,500	open	anything but ≈+2,500 on the $45
EWY $160C / $245C / $120P	8,243 / 11,212 / 5,575	6,667 / 10,000 / 3,333	CLOSE	any line RISING — our most falsifiable call
KWEB Jan-2027 $35C	28,706	100,000	open	a fall — ⚠️ direction still untestable
LQD $104P / $105P / $107C	0   / 398 / 1,153	20,000 / 10,000 / 10,000	open	a fall
URA Oct-16 $50C	2,287	9,052	open	a fall
GE Jan-2027 $300P	862	12,100	open   (93% margin)	a fall
IGV Dec-18 $100P	5,097	12,200	open   (58% margin)	a fall
SPY $760P · QQQ Oct $700C · URA Sep $50C · GE Dec $340P	all ≤ OI	—	NO CALL	—

EWY is the call to hold us to. We are saying three legs that each printed below standing open interest will see that interest fall — on the strength of the previous session's 100% capture at the same ratio. If any of the three rises, we are wrong.

This newsletter describes trades that appeared on the public options tape. It is not investment advice and nothing here is a recommendation to buy or sell. Options can lose their entire value. The tape cannot tell us who traded, why, or what else they hold — and we say so every time it matters.