ARK Invest’s largest transaction on Monday was not a purchase. Cathie Wood’s firm sold 60,458 shares of Deere & Co worth roughly $37.53 million, and on the other side of the ledger bought 122,422 Nvidia shares for about $26.6 million plus 39,020 shares of Broadcom across ARKK and ARKQ, around $16.69 million.
Read the two sides together and the trade is legible. Wood is funding semiconductors by selling a tractor company, which is a specific bet about where the money in artificial intelligence ends up, and it is not the bet most of her peers are making.
Down the Stack, Not Up It
The consensus AI trade of the last two years said value accrues to software. Build the application layer, charge subscription prices, run 80% gross margins, let somebody else own the depreciating metal. It is a comfortable thesis because it is how the last technology cycle actually paid.
ARK is positioned against it. Across last week the fund added SpaceX, CoreWeave, Cerebras, Nvidia, Circle and Coinbase, roughly $13.2 million of CoreWeave and $13.1 million of Cerebras among them, while trimming Roblox and Palantir. Monday extended the pattern: another $2 million of Palantir out through ARKQ, Broadcom and Nvidia in.
Sort those names and the list stops looking like a grab bag. Nvidia, Broadcom and Cerebras are silicon. CoreWeave is rented compute. SpaceX is launch capacity and orbital bandwidth. Circle and Coinbase are settlement rails. Every addition owns something physical or something with a regulatory moat around it. Everything being trimmed sits at the application layer, where the barrier to entry is a good team and an API key.
The implied claim is that in an AI economy the scarce input is capacity, not cleverness, and margin flows to whoever controls the scarce input. If that is right, the software layer gets competed down to something closer to a services business while the people selling chips and megawatts keep the surplus. It is a coherent position. It is also the opposite of what most growth managers have told their investors for two years.
The Deere Sale Is the Interesting Half
Selling Deere to buy Nvidia is not a rotation between two flavors of the same idea. It is a rotation out of the one holding with genuine pricing power in a rising-input-cost environment into the most crowded trade on the tape, and the timing sits awkwardly against the macro backdrop. Oil jumped about 5% on Monday on doubts about a Strait of Hormuz agreement, Brent reached $88 on Tuesday, and Wednesday’s US consumer price report is the event the entire rates market is waiting on.
An agricultural equipment maker with pricing power is precisely the kind of asset you want if that print runs hot. Wood sold it anyway, in size, to buy the assets that are most sensitive to a repricing of long-duration growth. That is either high conviction or a tell about how strictly the fund’s mandate constrains it, and it is worth noticing that ARK has been reducing Deere for several weeks rather than making a single call on Monday.
The Palantir trims deserve a flag too. Wood has been selling into a name where Michael Burry’s short position remains outstanding, which is not agreement between the two of them so much as a rare overlap in direction from opposite premises. Burry is betting on valuation math. Wood appears to be betting on where in the stack the money lands.
The SpaceX Position Is the One to Watch
The purchase that carries the most idiosyncratic risk is the smallest one on the tape. ARK added 20,318 SpaceX shares worth about $2.3 million alongside a $21 million block of Block on August 6, and has been building the position steadily.
SpaceX is the only major holding here whose price is not set by a market every second of the trading day, and ARK is buying it into a supply event we flagged last week: the lockup expiry that frees roughly 911 million shares. Buying ahead of a lockup release is a defensible call if you believe the float gets absorbed. It is a considerably less defensible one if it turns out the fund was the buyer providing the absorption.
None of that makes the thesis wrong. ARK has been early and correct on infrastructure before, and the demand data has not turned: TSMC’s July revenue rose 44.7% to a record NT$467.58 billion, and the company lifted its 2026 growth outlook above 40% in dollar terms.
What has changed is that the easy part of the trade is over. Buying compute in 2023 was contrarian. Buying it in August 2026, after Nvidia and Broadcom have already run, is a bet that the crowd is right and simply has not finished being right yet. Wood is funding that bet by selling the one holding that would have cushioned the fund if Wednesday’s inflation number lands badly.