What the Hell Happened?

Loading tweet...

view on X ↗

That was the first hour of Thursday's session. It got worse.

By the close the Nasdaq-100 was down 6.36%. In twenty years it has had six worse days. Three were Covid. Three were 2008. Liberation Day isn't on the list. Neither is the Monday after America lost its AAA.

Semis fell 12.25%, worse than any day of the dot-com crash. SanDisk lost 22%. Micron lost 17.5%. By Friday morning it had reached Seoul, where KOSPI hit its circuit breaker three minutes after the open.

The night before, the at-the-money SPY put cost $2.80. The options market was pricing a 0.76% day. By Thursday's close that put was worth $35.35.

The video and audio versions use an AI voice.

Everyone has a reason for it: data center moratoriums, AI capex, the Democrats. We don't think any of them explains the day.

Here's ours. After a big event passes, traders do the same thing: the risk is gone, so they sell options and buy stocks back. Funds that size their positions off volatility do it automatically. We think that's a large part of why the VIX closed at 16.6 on Wednesday. It wasn't a forecast. It was a lot of money positioned for nothing to happen, and when something did, that money all had to sell in the same session.

The Anthropic IPO is the next big date on the calendar, and we think the same setup is forming around it.

What happened

Tuesday was the midterms. Democrats took both chambers, and in Texas Gina Hinojosa beat Greg Abbott after calling for a data center moratorium. Wednesday the S&P rallied 1.6%, the VIX fell 20% to 16.61, and the one-day VIX halved. It was already splitting under the surface: that morning Microsoft was up 5.5% and Amazon 7% while CoreWeave fell about 4%.

Thursday the S&P fell 4.36%, the Nasdaq-100 6.36%, and the Philadelphia Semiconductor Index 12.25%. It went well past chips. CNBC's tally had Iren down 17%, Nebius 15% and CoreWeave 10%. Anything with data center exposure got sold.

SanDisk fell 22%. Micron fell 17.5%. If you read our SanDisk piece in August, that ordering won't surprise you: memory is the highest-beta way to own AI capex, and it gives it all back first. Nvidia fell 9%. Apple barely moved.

How bad was it? Depends what you're measuring, and most of the coverage got this a bit wrong.

For semis this was nearly a record. Only March 16, 2020 was worse. Thursday beat every day of the dot-com crash and every day of 2008.

For the S&P it wasn't close to a record. It ranks 43rd since 1970. But flip to "Everything" in that chart and look at the last column. The S&P's fall was 6.6 times its recent daily volatility, the largest multiple of anything on the list. The index had been so quiet that a medium-sized bad day looked like an extreme one.

That's the first clue.

The vote wasn't the surprise

Start with the obvious suspect. If the election had shocked anyone, that would explain Thursday. It didn't.

On Polymarket a Democratic sweep had climbed steadily for months, from about 44% in August to 74.5% by election day. Options priced the event about as well as you can: the VIX1D closed Tuesday at 23.22, an implied S&P move of 1.46% for Wednesday. The S&P moved 1.58%.

So the market got the event right. The mistake came next. It decided that once the vote was done, the risk was done, and all the premium came out in one session.

Click through those days. On Tuesday the front of the curve sits above everything, because the vote is the only thing being priced. On Wednesday it collapses and the curve slopes up. Saxo called the biggest one-session falls in the VIX and 9-day VIX in its data set. The one-day VIX closed at 12.08, 27% below the VIX. That priced Thursday at 0.76%.

The S&P did 4.36%. That's 5.7 times the implied move. Goldman's desk put it at about five times Wednesday's straddle, so we're in the same place.

One nuance worth being fair about. Skew barely moved on Wednesday. Cboe's skew index went from 150.0 to 150.9. So people kept paying for the crash tail. What got sold was the middle of the distribution: the ordinary, boring 2% day. Saxo's line was that the market "stopped paying for the vote but not for the chip debate." We'd go further. It stopped paying for anything that wasn't on the calendar.

The moratorium story doesn't hold up

The headline explanation for Thursday was data centers, and there's a real case for it. Democrats won Congress. Hinojosa won Texas, where Wells Fargo puts 27% of planned data center gigawatts. Dozens of local measures restricting data centers were on ballots in 10 states. Top Democrats called for a national moratorium after the sweep, and the Sanders and Ocasio-Cortez bill from March was back in circulation. Dan Ives said the buildout would slow and that Big Tech capex growth now likely peaks in 2027.

If that's what moved the market, you'd expect the people betting directly on moratoriums to care. They didn't.

Indiana's contract has traded $5.1M. It went from 18.5% to 18.0%. Texas is the sharper test: the candidate who wanted a moratorium won, and the December 2027 contract stayed at 21.5%. Click a state to see what it would take. The short version is that Yes needs a statewide law, and Abbott's August pause was already blocking new grid connections during a state audit.

There's a limit to this test, and we should say it. These contracts only pay on a statewide law. Local fights can make the buildout slower and pricier without ever resolving Yes. Barclays' Dominique Toublan called the pushback "bipartisan but local, making the buildout pricier rather than stalling it." So flat odds don't prove nothing changed. They do show that the thing people said happened on Thursday, a sudden jump in the odds of a statewide stop, didn't happen.

And semis were already leaving.

From October 1 to the vote the SOX fell 17% and Micron 22%, while the S&P was down about 1%. On Wednesday, the day of the relief rally, the SOX fell 3.1% and Micron 7% while the S&P rose 1.6%. It was breaking before anyone posted about data centers.

It wasn't an election trade on the way down, either. Daily changes in sweep odds and SOX returns had a correlation of 0.06 from August to the vote. The headline gave people a reason. The trade was already leaving.

The capex version has the same timing problem. Thursday's scariest line was Elon Musk saying SoftBank was defaulting on turbine orders, which nobody has confirmed. The serious version came from Capital.com's Kyle Rodda, who told CNBC the AI trade was already fragile because big tech earnings had fallen short of growth expectations. We agree with that, and it's our point. Fragile describes positioning. Apart from that unconfirmed claim, we can't find anything about capex that was known on Thursday and not on Wednesday, when the same stocks were already being sold into an index rally.

The cheapest thing in the market was the index

This is the part we think most people missed.

Everything was underpriced. But it wasn't evenly underpriced.

SPY missed by 5.5x. The single names missed by 3 to 4x. The thing that fell the least was the most mispriced.

That sounds backwards, so here's the mechanism. An index is a basket. If one stock falls 5% and another rises 5%, each had a wild day and the basket had none. The VIX prices the basket, so it only sees moves that happen together.

Wednesday is the cleanest example you'll get. Micron fell 7% and the SOX fell 3.1%, and the S&P rose 1.6%. If you were long chips it was an awful day. If you were looking at the VIX, it was a relief rally. From October 1 to the vote the SOX fell 17% while the S&P lost about 1%. Money leaving AI went into the rest of the index, so the basket barely noticed.

Traders call that low correlation, and the trade built on it is dispersion. For most of this year AI went one way and everything else went the other, so single-stock vol stayed high while index vol stayed low. We wrote about how much of this term structure was AI back in April.

Here's a toy version. Drag the slider:

Take a market that's 40% AI stocks running at 45% vol and 60% everything else at 14%. If the two halves barely move together, a correlation of 0.1, the index runs about 20.6% vol. Push that to 0.9 and it runs about 25.8%. No stock got any more volatile. They just started moving the same way, and the index went up by a quarter.

So a low VIX in a market like this isn't telling you the market is safe. It's telling you the market has a big short position in correlation, and it's been right for a while. Anyone holding the index at 16 vol was on the short side of that bet, whether they knew it or not.

Thursday everything fell together. The best hedge for an AI crash turned out to be the boring one: a Wednesday SPY put went up 12.6 times, and an Nvidia put went up 5.8 times on a stock that fell twice as much. Some of that is mechanical, because cheaper options have bigger multiples. The implied-vs-actual column is the honest number, and it says the same thing.

Who was actually selling

The obvious answer on Twitter was GEX and vol sellers. That's only half right, and the wrong half is interesting.

Dealers ended Wednesday estimated long gamma, about +$32.7B per 1%. In that state they buy dips, so you'd expect them to slow a selloff. They never got the chance. Wednesday's close of 7,689 sat only about 30 points above the 7,659 flip. The S&P opened near 7,678, just above it, broke it within about two minutes, and was under Thursday's 7,634 flip by about 09:38. That cuts against the tidy version of this story, where dealers absorb the first leg and then give up. On these estimates they were adding to the fall within minutes of the open. GEX doesn't explain why the market gapped lower, but it explains more of the day than the Wednesday number suggests. And dispersion trades didn't blow up: Risk.net reports some gamma-weighted versions made almost 7%, because single-stock vol rose more than index vol.

What the tape looks like to us: discretionary longs started it. Nomura's Charlie McElligott called it "shooting the generals," funds protecting 2026 gains in the year's best names. He wrote that the unwind had been "wildly acute" for more than a month and was nearing "round trip" levels in some trades, which fits a market that had been leaving since October. The perp fell 4.3% in the first 90 minutes, with no dealer cushion under it. The mechanical sellers were in there too, though as you'll see, we can't time them.

Two of those belong next to GEX, and one of them comes with a problem.

Leveraged ETFs have to sell after a down day. That's how they work: a 3x fund holds a fixed multiple of its assets, so when it loses money it's over-exposed and has to cut. It's short gamma by design, and it was sold to retail. SpotGamma's estimate, posted at 12:44 ET, was about $42B into the close, $19B of that in semis. Nvidia's share alone was about $4.4B, roughly 20% of its 20-day average volume. For just SOXL, SOXS and TQQQ, SpotGamma's figures add up to about $29B, and our own arithmetic below gets about $32B. Both are estimates, not reported trades. None of it shows up in a GEX chart built on options open interest. The SOXL perp on Hyperliquid fell 38% in a day.

Here's the problem. We expected this selling to show up as heavier volume at the close, and it didn't. The last 30 minutes were busy, but the whole day was busier, so the close was a smaller share of the day's volume than normal. The size estimates hold up. The timing doesn't: we can't show this landed at the close, and some of it may have traded earlier.

Vol-control funds are the cleanest example of the thesis. They hold more equity when realized vol is low and less when it's high. As an illustration, a fund targeting 10% vol runs about 125% exposure when the market is realizing 8%. If realized jumps to 20%, it wants 50%. So weeks of calm literally built the position, and a vol spike is literally the sell order. Bloomberg's estimate was about $61B over this episode. Calm wasn't a description of the market. It was on their books.

CTAs, for what it's worth, were already short. Citadel Securities had them at -3.3 standard deviations before Thursday. Not them.

A caution on the numbers: estimates of dealer gamma ranged from about -$7B to -$95B per 1% depending on who and when. Every one of those models guesses who's long what. Use the direction. Don't quote the number like it's a fact.

The day after gets priced like any other day

We went in believing the day after a known event is the least protected day of the month. We tested it, and that's too strong.

VIX1D has about three and a half years of history, so we ran all 29 Fed decisions since it started, plus the last two elections. After a Fed day the next session was priced at a median 0.95%, above the 0.80% of a normal day, and it moved 0.58%. It beat its price 10 times out of 29, or 34%, against 26% for all sessions. Wednesday's 12.08 wasn't cheap either. It sat in the middle of the last three years and above most of the prior 60 sessions.

So selling the day after has been a good trade, and that's the point. It works often enough that it gets crowded. The event premium goes out in one session, VIX1D fell 48% on Wednesday, and what's left is an ordinary price for an ordinary day. Nothing is left over for the second-order risk, the thing that follows from the event and has no date.

Thursday moved 5.7 times its price. That's the largest miss in all 889 sessions we have. The next largest is 4.0 times, on June 5, 2026. Two elections and one blowup is not a law. It's a reminder of what an ordinary price buys you.

The trade flipped just as fast the other way. By Thursday's close VIX1D was 36.8, an implied 2.32% move for Friday. The S&P closed up 0.3%, about a seventh of that. If you bought protection at Thursday's close you paid the top. Options are cheapest when you don't want them, and that isn't a coincidence.

Our call: Anthropic is the same setup

There's another date on the calendar.

Anthropic filed its S-1 on October 26, for what would be the largest IPO ever, with a roadshow planned for the week of November 9 and a valuation target of $1.5T or more. The filing showed $16.3B of first-half revenue, $9B of compute spending, and five customers making up about 25% of revenue through September. The filing lifted the whole chain: Benzinga called it a $570 billion order book for chipmakers, and the SOX rose about 3% that day. Ten days later that same chain had its second-worst day since 1994.

This time you can watch it being priced, in public, around the clock.

The ANT perp on Hyperliquid, a synthetic bet on the expected share price, fell about 20% over Thursday and Friday, from 117 to around 94. It lost another 12.6% on Friday while the SOX bounced 4%. Polymarket's odds of a listing by Nov. 30 fell from 82% at Wednesday's close to about 51% early Friday, then came all the way back to about 81% by Friday night. We found no report that explains the rebound. The Block saw Kalshi's odds of a listing before Dec. 1 fall from about 88% to 60% during the drop. The most likely valuation bucket dropped from $1.50T to $1.75T down to $1.25T to $1.50T. All of that happened while Reuters was reporting investors asking for a lower price, and Bloomberg had buyers talking about $1.2T to $1.3T. Nobody needed a banker to tell them where it was clearing.

So here's the call. The date may move, and the price certainly has. But whenever this prices, everyone is going to hedge into it, and everyone is going to take those hedges off after. If it goes fine, the market will treat that as the all-clear, the way it treated the vote. Vol comes in, the front of the curve collapses, and the day after will be priced like any other day.

We think the risk is the day after, not the day of.

We could be wrong. If vol stays bid after pricing, or the AI trade just rallies through it, this call was bad and we'll say so. The bull case is real too: TrendForce still has Q4 DRAM contract prices up 10 to 15%, hyperscaler capex plans haven't changed, and the SOX is on its lowest forward multiple since 2019. With CTAs this short, a squeeze wouldn't be strange. Friday's chip rally was a preview.

Who bought it here

We checked our own tape. Almost nobody bought the hedge here either, and the numbers are small, because we're small.

Between Wednesday's close and Thursday's open, three traders bought puts on our stock and S&P options, about $55,000 of notional, nearly all of it Microsoft and Apple. A normal weekday in October saw about $452,000 of puts bought. At Thursday's close those puts were worth about 3.8 times what they cost, on intrinsic value alone.

Most of that wasn't collected. The largest buyer sold within half an hour of the open at about 1.35 times cost. Held to the close, the same puts were worth about 4 times.

Then the flow turned the other way. During Thursday's session customers bought $526 of put notional and sold $67,000, and they bought $161,000 of calls. On the second-worst day for semis since 1994, the tape here sold fear and bought the dip.

It's a tiny sample, but it's the same trade at a smaller size. The hedge was cheap the night before, and hardly anyone wanted it.

Why we care

Back to Seoul. When KOSPI stopped for 20 minutes on Friday, with SK Hynix leading it down, the SK Hynix perp on Hyperliquid kept trading. It went up about 4% during the halt.

These names trade around the clock now. Hyperliquid's equity perps gave you a price on Friday morning when Korea couldn't. We wrote about what that looks like in the S&P in September. One twenty-minute window isn't proof of anything, but a halt doesn't stop people wanting out. It just moves where they go.

Perps give you price. They don't give you convexity. A perp short on Wednesday night would have made money. A Wednesday SPY put made 12 times its cost, with a known maximum loss and no liquidation price. The lesson from this week is that the time you need a hedge isn't the time you know you need one, and it might not be during New York hours.

That's the gap we're building for: options on these names, on the same rails as the perps, margined together, open when the rest of the market isn't. It's early and the books are thin next to the listed market. But Wednesday night was a pretty good demonstration of why it should exist.

What to watch on Monday

Everything above is a claim about how the market treats a date. Monday is the first live test of it, and the place to watch is the ANT perp.

Here's where it sits. It closed Friday's UTC session at 94.29, down 19.7% from Wednesday's 117.41. Friday's low was 93.06. Realized vol over the past week ran about 101% annualized, up from 72% the week before. Open interest is about 610,000 contracts, roughly $58M, and funding is close to flat, so nobody is paying up to hold either side.

The perp trades through the weekend, so there will be an answer before New York opens. Three things matter.

93. That's Friday's low. If it breaks with the roadshow about to start, the market is saying the $1.2T to $1.3T talk was the ceiling.

105 to 108. Thursday's low was 104.89 and its close was 107.86. Getting back above that says Friday was a flush, not a repricing.

The option curve. Our ANT options are expected to start trading Monday, so we have a stake in this and you should read it that way. The perp tells you the price. Options tell you what people will pay for the path. Compare what the market charges for the listing with what it charges for the days after it.

If ANT options price a big listing day and a calm day after, that's Wednesday night again. That's what we'll be watching on Monday.