Held unchanged, the disclosed stocks and warrants lost 21.7% from 30 September to 18 November, against 4.3% for the MSCI World price index in dollars. The chip and memory shares lost 35.4%. That is a measurement of the disclosed basket, not the fund's performance.
Copio's assessment is that the concentration raises the hurdle for compounded growth, and the filing does not establish effective crash protection. In an illustrative put scenario, the option's intrinsic value offsets just 5.9% of the basket's loss before its cost. The non-chip holdings lost less, but became more closely linked to chips during the sell-off.
What is in the fund?
The filing lists 17 lines at 30 September 2026: 15 stock and ADS positions, warrants on Grab and one put option, with a cover total of $10,309,080,051. Values are in whole dollars, as the SEC's Form 13F guidance has required since January 2023. Every share line divided by its share count matches the 30 September closing price, except Cerebras.
The September line items put the stocks and warrants at $9.80bn and the shares underlying the put at $511.6m. We use the line items to calculate weights, rather than the cover total, which differs slightly. The June filing reported $9.83bn across 20 lines, all shares or warrants.
30 June to 30 September 2026, share of share value: Nvidia 19.2% to 23.3%; Meta 7.8% to 12.6%; Cerebras 16.2% to 10.3%; Tesla new at 6.9%; Qualcomm, Applied Materials, KLA and Lam Research 9.0% to zero. Source: 13F-HR information table, 30 September 2026 and 30 June 2026; every line from December 2025 to September 2026 (CSV).
The three largest holdings make up 46.2% of share value and the ten largest 89.8%. The effective number of positions, which counts how many equal-sized holdings would give the same concentration, is 9.0, against 10.4 in June and 6.4 in March. Chip and memory companies (Nvidia, Cerebras, TSMC, Arm, Micron and Sandisk) are 48.3%; cloud and internet platforms (Meta, Amazon, Microsoft and Snowflake) 31.7%.
What changed since 30 June
The snapshots show fewer chip-equipment positions and the addition of Tesla, but they do not show when or how the positions changed.
- No longer reported. Qualcomm, Applied Materials, KLA and Lam Research, worth $882.5m in June, plus Alphabet and Synopsys.
- Added. Nvidia shares rose 6.3% to 10,006,069, Meta 25.4% to 1,698,703, Uber 19.5% and Micron 24.2%.
- New. Tesla, 1,915,440 shares worth $679.6m; Sandisk, 25,170 shares; and the semiconductor ETF put.
- Reduced. Cerebras 17.6%, CoreWeave 21.4%, SpaceX 40.5% and Microsoft 6.7%.
The fall in the chip and memory share, from 60.4% to 48.3%, was partly price. Cerebras's value per share in the filings fell from $221.00 to $169.76, and Altimeter also held fewer shares. A Schedule 13G/A filed the same day puts the 5,953,504 Cerebras shares at 3.7% of the class, and reports 360,097 more held by Gerstner personally.
The put
The SEC's Form 13F guidance explains that most fields on an option line describe the underlying security (question 44). So the $511.6m line is 840,000 shares of the VanEck Semiconductor ETF at their 30 September price of $609, not what the put cost. A put gives its buyer the right to sell at a set strike price; it has intrinsic value below that strike while it remains in force.
The filing gives no strike, expiry or premium. The underlying value equals 10.8% of the $4.73bn in chip and memory shares, and 5.2% of all share and warrant value. Those are ratios of reported dollars, not estimates of how much loss the put would offset. The June filing listed no options.
What the filing cannot show
The guidance excludes short positions (question 41), and a 13F shows no cash, borrowing or private holdings. Altimeter's latest annual Form ADV, filed on 31 March 2026, reports $18.75bn of regulatory assets under management in 36 accounts. The two figures have different dates and definitions, but the comparison shows the 13F covers only part of what Altimeter manages.
Why does Gerstner use this structure?
Altimeter describes itself as a long-term technology investor across private and public markets; Gerstner's own September talk adds an AI revenue thesis and an explicit view on position size. None of the sources explains the individual trades, the Tesla stake or the put.
On its website, Altimeter says it invests in both public and private markets and seeks to work with companies over the long term as they go public. Cerebras and SpaceX first appear in the June 2026 filing; the filings do not explain when those investments began or how long Altimeter plans to hold them.
In a market talk published by All-In on 17 September 2026 (automated transcript), Gerstner described his positioning as moderate. He linked taking more or less risk to AI lab revenue and oil prices, and warned against heavy leverage. Those remarks explain his thinking about sizing, not Altimeter's financing or any specific trade.
The filings and fund documents we found contain no periodic fund returns, so there is no matched geometric return, Sharpe ratio, Sortino ratio or frontier position for Altimeter versus MSCI World. The chart instead compares a fixed basket of the disclosed stocks and warrants with the dollar price index over identical dates, excluding dividends on both sides.
The changes in 13F value are not returns either. The reported share value went from $6.66bn at the end of 2025 to $5.70bn in March, $9.83bn in June and $9.80bn in September. Those changes mix price moves with changes in what is reported, which can include purchases, sales, or securities newly becoming reportable. They are not a time series of portfolio performance.
Hypothetical, 30 September to 18 November 2026: stocks and warrants -21.7% ($9.84bn to $7.71bn at closing prices); chip and memory shares -35.4%; MSCI World price index -4.3%. Source: 13F-HR for 30 September 2026, Yahoo Finance daily closes and the MSCI World price index in USD; daily path (CSV), price inputs (CSV).
Nvidia did most of the damage. Its shares fell 33.9%, taking 7.9 points off the basket; Cerebras fell 47.0% and took 5.1 points. CoreWeave, Meta, Arm and TSMC took 1.4 to 1.9 points each, while Tesla (+8.1%) and Amazon (+7.5%) added 0.6 and 0.5.
The Nvidia fall came as the business kept growing. Nvidia's board accepted Jensen Huang's resignation as chief executive on 11 November, according to its Form 8-K, and named its finance chief interim CEO. On 17 November it reported quarterly revenue of $115.8bn, up 103% from a year earlier. The shares closed at $173.73 on 11 November and $150.96 on 18 November, 13.1% lower.
The September basket did better in the sell-off than the June basket would have at fixed quantities. Marked over the same dates from 30 September, the June basket would have lost 24.2%, against 21.7% for the September basket. That comparison isolates the two snapshots, not the effect of any trades the manager actually executed.
The filing arrived 47 days after the date it describes. By 16 November, the day it was filed, the September shares were already 20.5% lower, against 2.6% for MSCI World. Over the third quarter the shares had an annualised daily volatility of 33.4%, against 9.9% for the index; we do not report Sharpe or Sortino ratios for a seven-week basket.
Copio's reading
Is it optimal for geometric growth?
The disclosed book is not an established geometric-growth optimum. The test is expected log wealth after costs over a decade or more, comparing the book with a broad world index or a less concentrated equity book under the same redemption, liquidity and financing constraints. The filing does not disclose those terms or the expected returns needed for that comparison.
First, is the concentration paid for? In a one-risky-asset diffusion model with cash, an unconstrained Kelly fraction equals expected excess return divided by variance; a fraction of one therefore requires the two to be equal. At the book's third-quarter volatility of 33.4%, measured retrospectively at September share counts, that is 11.2 points a year above cash. For full investment to represent half of that model's Kelly fraction, it would take 22.3 points.
In that same model, the gap between arithmetic and geometric growth is about 5.6 points a year at full exposure. That is a sizing illustration, not a forecast calibrated to jumps or a comparison with MSCI World's expected return. Nvidia's 103% revenue growth is not the expected return on its shares at the price paid. Our Kelly criterion guide shows the same arithmetic for a whole equity allocation.
Second, does the put improve whole-book growth after its cost? In a scenario where the put was struck at the 30 September price of $609 and still outstanding, its intrinsic value on 18 November would have been about $126m, or 5.9% of the basket's $2.13bn loss. That figure excludes any time value and premium, and does not measure what the actual option earned.
Third, do fees, financing and flows change the result? The snapshot cannot show whether the book was financed with borrowing or how fees and redemptions affect wealth through repeated falls. A matched after-cost return history could show whether the concentration has paid so far; establishing a growth optimum would also require forward assumptions and feasible alternatives under the same constraints.
Is it crash aware and protected?
The put and Gerstner's sizing remarks show crash awareness, but the disclosed position does not establish meaningful crash protection. Held unchanged, the September basket was down 8.7% by 16 October, 19.0% by 6 November and 21.7% by 18 November. Recovering from that level needs a 27.6% gain; our recovery maths article explains why losses need larger gains to repair.
The option references a semiconductor ETF, not the basket itself; even with a known strike, the ETF and the holdings need not fall by the same amount. For a household, compare losses under a joint sell-off with spending that cannot wait, and count the cost of any hedge in ordinary years as well as its payoff in a crash. Our study of put hedging budgets shows that trade-off.
Is it diversified without sacrificing growth?
The holdings looked loosely linked before the sell-off and moved much more together during it. A correlation of 1 means two daily return series move in lockstep and 0 means no linear relation. From 1 July to 30 September 2026, the ten largest holdings had pairwise correlations from -0.27 to 0.66, averaging 0.19.
1 July to 30 September 2026, 64 daily returns: pairs -0.27 to 0.66, average 0.19; TSMC with Arm 0.66, Nvidia with TSMC 0.52; Meta with Nvidia 0.07. Source: Yahoo Finance daily closes; correlation matrices for both windows (CSV).
From 30 September to 18 November the range was -0.07 to 0.82 and the average 0.40. The clearest change is in how the non-chip holdings moved with the semiconductor ETF that the put refers to.
Correlation with the semiconductor ETF, July-September to October-November 2026: Snowflake 0.12 to 0.73; Meta 0.02 to 0.46; Tesla 0.42 to 0.84; Amazon 0.06 to 0.40. Source: Yahoo Finance daily closes; price inputs (CSV). These are observed correlations from short samples of 64 and 35 daily returns, not a forecast of the next sell-off.
The non-chip basket fell 8.6%, against 35.4% for the chip and memory shares, with gains in Tesla and Amazon helping to soften the loss. Together with the stronger correlations, the two losses show some diversification within technology, not a reliably independent source of return. To assess a household portfolio, count holdings by the spending or financing cycle that drives them, not just by ticker.
Our falling-together study shows the same rise in correlations across asset classes in bad months. Copio does not grade this portfolio in its checker, because a book of single stocks, warrants and an option with undisclosed terms cannot be represented by an ETF replica.
Our review of Druckenmiller's Q3 13F offers another comparison of a disclosed share basket, and our Aschenbrenner follow-up examines a book run through calls. Our analysis of the November AI sell-off shows what the same weeks did to index investors.
FAQ
What did Brad Gerstner's Altimeter hold at 30 September 2026?
Its 13F, filed on 16 November 2026, lists 17 lines with a cover total of $10.31bn: 15 stock and ADS positions, Grab warrants and a put on the VanEck Semiconductor ETF. The shares and warrants are worth $9.80bn. Nvidia is 23.3% of that, Meta 12.6% and Cerebras 10.3%, followed by Tesla, Snowflake, TSMC, Amazon, Microsoft and Uber at 6.2% to 6.9% each.
Did Altimeter sell its chip stocks in the third quarter?
Some of them. Qualcomm, Applied Materials, KLA and Lam Research, worth 9.0% of the June filing, no longer appeared, nor did Alphabet and Synopsys. The disclosed Nvidia and Micron share counts rose by 6.3% and 24.2%, and Sandisk appeared in the filing. Chip and memory companies fell from 60.4% to 48.3% of share value, partly because Cerebras's price fell.
What is the semiconductor ETF put in Altimeter's 13F?
A put option on 840,000 shares of the VanEck Semiconductor ETF, reported at those shares' $511.6m value on 30 September, as the SEC's rules require. A put has intrinsic value when the ETF price is below its strike. The filing gives no strike, expiry or premium, so its cost and payoff are unknown; its underlying value equals 10.8% of the chip and memory shares in the filing.
How did Altimeter's disclosed holdings do in the November 2026 chip sell-off?
Held at fixed quantities, the disclosed stocks and Grab warrants fell 21.7% from 30 September to 18 November, while the MSCI World price index fell 4.3%. The chip and memory shares fell 35.4% and Nvidia alone 33.9%. This is a hypothetical basket, not Altimeter's return: the fund may have traded since 30 September, and Altimeter publishes no return series.
Can an investor copy Altimeter's 13F?
Not usefully. The filing arrived 47 days after the date it describes, and by then the disclosed shares were already 20.5% lower. It leaves out the put's terms, any short positions, cash and private holdings. Altimeter reported $18.75bn of regulatory assets in March 2026, against $9.80bn of shares in this filing, though the two figures have different dates and definitions.