Leopold Aschenbrenner Bought $8.5B of Puts. What Would It Pay to Sell Them to Him?

What this is: a thought experiment run on real filings and real option quotes, for entertainment and education. Infotainment. Not a trade idea, not a recommendation, not a forecast. Every number below carries the date it came from and the sources are itemised at the bottom. Please read the dates. The position data was four months old on the day this published, and the option quotes went stale the same afternoon.

---

If you sell options for income, and a fair number of people here do it quietly between meetings, you have probably never thought of yourself as anybody's counterparty. You pick a strike, you collect the credit, the cash sits in the account. It feels like a solitary activity.

It isn't. Somebody is always on the other side, and occasionally you get to see who.

Leopold Aschenbrenner's last public filing reports $8.46 billion of notional put exposure against chipmakers. SMH, NVDA, AVGO, AMD, MU, TSM, ASML, INTC. Sixty-two per cent of everything he disclosed. He is in his mid-twenties, he left OpenAI to start the fund in 2024, and for two years now he has been the most closely watched AI investor in the market.

Every one of those puts had a seller.

Which is a slightly uncomfortable thought if you run the wheel, because the seller of a put on Nvidia is, structurally, doing what you do. Not the same order, obviously. The same chair, though, facing the same way, paid the same way.

So we got curious. What would it actually pay to sit on each side of Aschenbrenner's book with a million dollars of cash collateral? We ran the numbers properly. The premium turned out to be the least interesting part.

What the filing actually says

Start with a correction, because most of the coverage has this backwards. He was long semiconductors through 2024 and 2025. He is not long them now.

We read the Q1 2026 13F information table off SEC EDGAR ourselves — accession 0002045724-26-000008, holdings as of 2026-03-31, filed 2026-05-18 — rather than take anybody's summary for it. Just as well: several of the write-ups sitting at the top of the search results describe a put book as long exposure. All 42 lines, sorted:

Sleeve	Reported value	Share	What's in it
Puts on the chip complex	$8.46B	62%	SMH, NVDA, ORCL, AVGO, AMD, MU, TSM, ASML, INTC
Long common	$3.86B	28%	Bloom Energy, CoreWeave, IREN, Core Scientific, Applied Digital, miners
Long calls	$1.36B	10%	MU, SNDK, TSM, CRWV, BE

Short the chips, long everything that keeps the chips running. His reasoning is public: in his June 2024 essays he argued the real constraint on AI is not silicon but electricity, and the line people quote is "Where do I find 10GW?" Chips, he reckoned, were still a modest share of leading-edge foundry output with room to grow. You can build wafer capacity. Try building a substation.

A word of caution on that $8.46B before anyone repeats it at lunch. A 13F reports options at the notional value of the underlying, not what was paid for them, so the actual money committed is a good deal smaller than the headline. And MU shows up as puts, calls and common; TSM as puts and calls. That is somebody building structures, not somebody slamming the sell button. The version of this story where he "shorted Nvidia with $8.5 billion" is really a filing convention being mistaken for a declaration of war.

The option market takes a different view

This is the part we did not expect.

We took his names, split them into the sleeve he is short and the sleeve he owns, and pulled live chains for a 0.10-delta put about a month out. Tradier broker data, close of 2026-07-29.

Sleeve	Median IV30	Median annualised CSP yield	Median quoted spread
Names he holds puts on	68%	19.5%	33%
Names he is long	122%	44.8%	62%

Roughly double the implied volatility to underwrite the things he owns, versus the things he is betting against.

Sit with that for a moment, because it says something about him rather than about the stocks. He bought his insurance in the cheap aisle. NVDA's 0.10-delta put was pricing off 49% vol; Nebius off 156%. He is short the crowded, well-behaved leg of the trade and long the volatile, cash-hungry leg, and he chose to buy protection on the side where protection happens to be inexpensive.

Perfectly sensible if you run a hedge fund. Rather less useful if you sell premium, since selling a put is the opposite side of buying one. The seller's version of "buy cheap convexity on the chips" is simply to leave those puts alone, which does not pay anything and makes for a short article.

What a million dollars would collect

Three books, then. All near 0.10 delta, 22 to 36 days out, collateral fully committed, every strike a real quote from 2026-07-29.

Book A sells puts on the nine names he owns puts on. Loosely, his side.

Book B sells puts on the power and neocloud names he holds long. You believe the man and you take the fat premium.

Book C ignores the story entirely and takes only what passes a tradability filter — open interest of at least 100 contracts, quoted spread inside 35% of mid. Five names survive: BE, NBIS, NVDA, ORCL, TSM.

On the quoted mid price, it is not close:

Book	Credit @ mid	% of $1M	Annualised
A, fade the chip shorts	$16,411	1.64%	~19.1%
B, ride the neocloud longs	$32,976	3.30%	~43.9%
C, liquidity-gated	$24,216	2.42%	~30.8%

Double the money for agreeing with him. Which is roughly where most people would stop reading and start clicking.

Then you try to get filled

Mid price is a polite fiction on a contract nobody trades. So we gave back half the quoted bid-ask, which is generous when the open interest on the thing is two:

Book	Mid	Realistic fill	% of $1M	Annualised	Keeps
A, fade the chip shorts	$16,411	$11,762	1.18%	~13.7%	72%
B, ride the neocloud longs	$32,976	$20,802	2.08%	~27.7%	63%
C, liquidity-gated	$24,216	$20,939	2.09%	~26.7%	86%

Book B and Book C end up $137 apart.

The extra money was never really there. It was sitting in the gap between the bid and the ask, and the moment you cross that gap to get a fill, it belongs to whoever took the other side of you. Book B needed 1,378 contracts across 14 names to arrive at the same place Book C reached with 82 across 5. On a $0.65 commission that is $896 of costs against $53. It also only gets there by selling 674 contracts of a five-cent option with an open interest of 2, which is less a trade than an act of optimism.

That the spread is wide and the premium is fat is not really two facts about these names. Both come from the same place, which is that nobody quoting them is very sure what they are worth.

Where the three books stop differing

We put all three through the same conservative whole-book crash — market down 25%, each name scaled by its own volatility, implied vol rising into the fall. This is our model rather than a broker's margin figure, and it is deliberately unkind.

Book	Names	Modelled loss @ −25%	% of NAV
A, fade the chip shorts	8	$626,361	62.6%
B, ride the neocloud longs	14	$642,066	64.2%
C, liquidity-gated	5	$608,530	60.9%

Three books, three different arguments, and the results land within 3.3 percentage points of one another.

Fourteen tickers bought nothing that five did not already have. The chips, the neoclouds, the miners, the fuel cells, the memory — pick any of them and you have made the same wager on AI capital spending continuing. Spreading a million dollars across more of them adds contracts and commission and a comforting sense of having done something. Diversification wants exposures that are genuinely unrelated, not simply more of them.

The same arithmetic, as it happens, explains his own bad month. Between 2026-03-31 and 2026-07-29 the short leg fell about 25% while the long leg fell about 50%: SMH −24.6% off its 2026 high, Nebius −48.3%, Bloom Energy −52.7%, CoreWeave −55.9%, IREN −56.8%. The hedge had been sized on notional and parked on the calmer side. On 2026-07-30 the Financial Times reported he was raising fresh capital, with some investors offered the chance to buy assets out of the portfolio. A book holding $8.5 billion of notional puts on chips still had to make that call. The protection was real enough; it just was not sized where the losses turned up.

What this is really about

Nobody should be picking sides of a hedge fund's book off a blog post. What the exercise does show is why three numbers have to be read together, and why any one of them on its own will mislead you.

The credit, first, and whether it survives the spread. A mid you cannot fill is a screenshot, not income.

The assignment exposure — chance of finishing in the money multiplied by the strike, added up across the whole book. Not delta, which is a different number entirely. Book C collected 15.0 cents of realistic premium per dollar of expected assignment; Book A managed 10.0.

And the whole-book crash loss, because per-trade max loss tells you very little about a portfolio where every line moves on the same news.

Read any single one of those and Book B wins comfortably. Read all three and it is the weakest of the three: same crash exposure, same realistic income, seventeen times the paperwork.

The risk, plainly

A cash-secured put is a promise to buy the shares at your strike. The premium softens a small fall and does nothing whatsoever for a large one.

The richest liquid line in the whole exercise was the Nebius 2026-08-28 $90 put at $4.00 a share, quoted 2026-07-29:

• Credit: $400 a contract
• Breakeven: $86.00, some 42% below the $148.22 close
• Max loss: $8,600 a contract if NBIS goes to zero
• Risk-neutral chance of assignment: ~18.9%

A 42% cushion sounds like a lot. Nebius fell 48% from its high in the space of a few weeks, and IREN fell 57%. In a name that moves like that, 42% is about one bad month, and the reason the premium is generous is precisely that everybody quoting it knows so.

Get assigned and you own a leveraged, capital-hungry business whose funding depends on the capital markets staying friendly, bought at a price that felt like a bargain and might not have been. The bear case does not need much imagination: AI capital spending slows a little, the financing window for neoclouds narrows, the equity in the most indebted names starts trading like an option, and the whole book goes down together regardless of how many tickers are in it.

Set the realistic figures against what your fixed deposit or a T-bill is paying at the moment, then decide whether the gap is worth that. It is a reasonable thing to conclude either way. It is not a reasonable thing to conclude without looking.

Nobody in this story knows what happens next, Aschenbrenner very much included.

Where all this came from

Primary documents, read directly. We avoided the aggregators on purpose: several of the highest-ranking pieces on this fund describe a put book as long exposure, and one carries a position at "40% of the portfolio", a figure that does not reconcile against any denominator we could construct.

What	Source	As of / dated
42 positions, put/call/common split	SEC 13F information table, accession 0002045724-26-000008	Holdings 2026-03-31, filed 2026-05-18
Nebius stake, 12,410,060 shares (5.63%)	SEC Schedule 13G	Filed 2026-05-27
SharonAI stake, 19.99% (capped)	SEC Schedule 13G / Form 3 / Form 4	2026-06-29 / 06-22 / 06-30
Option chains, IV, greeks, spreads, open interest	Tradier broker API	2026-07-29 close
Share prices and drawdowns	yfinance daily closes	2026-07-29
VIX 19.56, the crash model's beta denominator	yfinance	2026-07-29
AUM >$20bn, Anthropic ~20%, return figures	Hedgeweek, press-reported, not filed	2026-06-09
Capital raising after the rout	Financial Times via Bloomberg, press-reported	2026-07-30
"Where do I find 10GW?" and the power-constraint thesis	situational-awareness.ai, Aschenbrenner's own essays	June 2024
Assignment probability, greeks, crash stress	levelbox's own engines, our model, not a broker's	Run 2026-07-30

What we could not see. A 13F captures US-listed longs and options and nothing else, so it misses Anthropic — reportedly his largest holding — along with SK Hynix, MatX and Fluidstack. Call it two-thirds of the book, and not the important third. His Q2 2026 filing was not due until around 2026-08-14, meaning the freshest position data available here predates the July selloff completely. Four of the fourteen names in Book B returned a degenerate skew-corrected assignment probability, so we used the uncorrected N(−d₂) for those and are telling you rather than quietly patching it. The fill assumption, the commission rate and the crash model are assumptions, flagged wherever they are used.

---

Published 2026-07-30. Infotainment and education, not investment advice. levelbox.ai does not make recommendations or predictions, and this article makes none. The three books are illustrative constructions built to demonstrate a method; nobody traded them. Position data is a historical snapshot from SEC filings dated 2026-03-31 and June 2026. Option quotes are one broker's marks at the 2026-07-29 close and were stale within minutes of capture. Crash and assignment figures are levelbox's own conservative model estimates, not a broker's margin calculation and not a prediction. Options carry substantial risk of loss, including losses well beyond the premium received. Please do your own work.