Harvard's Endowment Cut Nvidia by a Third and Built a New AMD Stake

Harvard Management Company cut its Nvidia stake by 34.3% and opened a brand-new position in AMD worth $198.5 million, according to a third-quarter 13F filing submitted Friday.

The filing shows Harvard held 324,400 AMD shares as of September 30. At the end of June, the endowment's manager reported no AMD shares at all. The Nvidia position fell to 517,551 shares, leaving AMD at 5.0% of the reported 13F portfolio versus 3.0% for Nvidia.

The context matters here. Harvard's endowment stands at roughly $68 billion, and the 13F covers less than 6% of it — so this is a signal, not a strategy reveal. Harvard has not explained the trades, and the filing doesn't show when the manager bought or sold, or at what prices. Still, when one of the world's largest university endowments rotates between the two biggest AI chip designers, investors read it closely.

What did the swap do to Harvard's returns?

So far, the trade has worked out. Nvidia closed Friday at $162.85, which is 28.7% below its September 30 close. The drop followed a week in which Nvidia's board reportedly weighed removing longtime CEO Jensen Huang.

AMD held up better. It closed at $534.06, down 12.7% from its own September 30 close. By 24/7 Wall St.'s calculation, Harvard's AMD shares would be worth about $173.2 million at Friday's prices, and its remaining Nvidia shares about $84.3 million — assuming the endowment still holds them. The new mix has held up better than the old one would have.

Takeaway: the rotation cushioned Harvard from the worst of Nvidia's slide.

Why Taiwan Semiconductor shows this isn't an AI exit

Harvard didn't step away from AI hardware. It reshuffled inside the category. The clearest evidence is Taiwan Semiconductor: Harvard raised its TSM ADR position by 85%, to a stake worth $616.5 million.

That's the mechanism worth understanding. Taiwan Semiconductor manufactures chips for both Nvidia and AMD, so a bigger TSM position means Harvard keeps exposure to AI silicon whichever designer wins share. The pattern looks like a hedge inside the AI supply chain, not a retreat from it.

The deeper cuts went to cloud providers, not chipmakers. Harvard trimmed Microsoft by 41%, leaving that stake at $118.7 million — now smaller than the AMD position. Alphabet Class A shares took the biggest cut of all, down 62%.

Takeaway: Harvard sold software-and-cloud exposure and doubled down on the chip layer of the AI stack.

Why does AMD gain more from each new customer?

The numbers behind the trade are stark. AMD's data center sales rose 97% last quarter, to $8.546 billion, as Anthropic, Meta, Microsoft, OpenAI and Oracle brought their first Helios racks online. Management expects 2027 revenue to grow more than 80%.

Nvidia is still far bigger. It recorded $89.023 billion of data center revenue in its latest quarter at a 75.0% non-GAAP gross margin, and it expects fiscal 2028 growth of approximately 70%.

But the size gap is exactly the point. AMD is valued at $875.90 billion against Nvidia's $3.93 trillion. When you're a quarter of Nvidia's size, every new hyperscaler customer moves your growth rate much more than it moves Nvidia's. That arithmetic is what attracts investors to the smaller player.

Takeaway: AMD's percentage growth benefits from new AI customers more than Nvidia's does, simply because it starts smaller.

What did Lisa Su say about demand?

AMD's CEO gave investors the quote they wanted: "demand is higher than supply." Su's comment underpins the growth case — data center is expected to account for more than 80% of AMD's 2027 revenue.

The risk sits in the valuation, not the order book. AMD trades at a trailing price-to-earnings ratio of 202x, compared with 33 for Nvidia. Free cash flow came in at just $554 million last quarter, as inventory climbed to $11.463 billion. A company can have surging demand and still be priced for perfection, and AMD's multiple leaves little room for error.

Takeaway: AMD's demand story is real, but investors are paying an extreme price for it.

What will Nvidia's earnings tell us?

The next test of the AI selloff arrives soon. Nvidia is expected to report after the close on Tuesday, November 17, although the company has not confirmed the date. It guided to $108.0 billion in revenue with gross margins near 74.0%.

The demand side looks like the smaller concern. In August, Nvidia's CFO called the guidance "a supply-constrained outlook" — meaning the company expects to sell everything it can make, not that customers are pulling back.

That frames the earnings call simply. If Nvidia hits its guidance and keeps the Vera Rubin roadmap on schedule, the recent selloff likely went too far. If either slips, Harvard's decision to trim will look prescient.

Takeaway: Tuesday's report, not Harvard's filing, is what will actually move the market.

How much should anyone read into a 13F?

Not too much, by the source's own assessment. AMD's own results carry more weight for investors than Harvard's filing does. The 13F shows holdings only at a single point in time — September 30 — and covers a small slice of the endowment. Harvard has offered no explanation for any of the trades.

The filing's real value is as a data point in a broader debate: how to hold AI exposure when one chip designer trades at 33x earnings and the other at 202x. The source points to two alternative routes for investors who find AMD too expensive. Taiwan Semiconductor earns revenue from both chip designers no matter which one gains share. And the power, cooling and networking suppliers behind the AI data center buildout offer a way to play the same demand without paying a chip multiple at all.

Harvard hasn't commented on its reasoning, and the next 13F won't arrive until mid-February. Until then, the market's attention shifts to Nvidia's report and AMD's ability to keep converting supply into revenue at the pace its 202x multiple demands.