Cathie Wood ARK Invest AI Stocks: Why She Dumped $110 Million of AMD and Loaded Up on Nvidia Instead

Cathie Wood’s methods have an almost contrarian quality. Most managers would have stayed put and relished the ride when Advanced Micro Devices surged 184% in just one year and momentarily surpassed a trillion-dollar market valuation. Wood was sold. On September 29th, nearly $110 million was made in a single session. She then reversed course and invested about $81 million in Nvidia, a stock that had only increased by 22% during that time. It appears to an outsider that leaving the party at the height of the music is a conscious decision.

That transaction encapsulates a key aspect of ARK Invest’s current perspective on AI. Finding areas where value has changed or where the next phase of growth may be subtly emerging is more important than chasing momentum. AMD is pricing in years of perfect execution with a trailing price-to-earnings ratio close to 229x. In contrast, Nvidia’s valuation is at its lowest point since early 2015, at about 17 times forward earnings. The math is easy to understand, at least when written down.

Scale and margin have always been key factors in the Nvidia case. With gross margins close to 75%, its data center revenue reached $89 billion last year, up 117% from the previous year. Despite operating at a 56% margin on a much smaller base, AMD’s data center segment is truly impressive, up 107% to $6.72 billion. Both businesses are developing actual AI ventures. Simply put, they are constructing them from quite different angles. Wood seems to think that the gap is more important than the growth figures.

It’s important to remember that ARK’s September 29th session involved more than just switching from AMD to Nvidia. In the same day’s trading, the company also acquired roughly $27 million in Broadcom, $21.7 million in CoreWeave, and $17.3 million in Tesla. Tempus AI and 10x Genomics were among the sales positions. The pattern, which is distributed among four ETFs, resembles a general rebalancing rather than a single conviction call. It is more difficult to read too much into a single trade in isolation because ETF weight limits can compel such actions regardless of what a manager truly believes.

However, the larger argument ARK is constructing around AI infrastructure seems purposeful. According to the firm, global spending on AI will increase from $500 billion in 2025 to approximately $1.5 trillion by 2030. Whether you think it’s optimistic or not, that projection influences the reasoning behind nearly all of ARK’s positions. Customer deployments of Nvidia’s upcoming Vera Rubin platform are now underway. Over two million Nvidia GPUs will be deployed by AWS in 2027 and 2028. It appears that ARK believes the infrastructure wave is still developing rather than peaking.

The name that most likely receives the least attention but provides the most intriguing insight into the direction of ARK’s thinking is Cerebras Systems. ARK added about $26 million worth of Cerebras stock to two ETFs on August 26. The relatively new company Cerebras was founded on a wafer-scale AI chip, which is essentially a single, massive processor rather than a collection of smaller ones. Although the position is still modest—roughly 2.8% of the ARK Innovation ETF—the direction is important.

Cerebras is an example of the push toward specialized hardware competitors that could eventually carve out significant space in the AI compute market, as Wood has long advocated for going beyond the obvious hyperscaler plays. Although it’s genuinely unclear if that actually occurs, it makes sense to diversify away from single-stock concentration risk.

Observing the accumulation of ARK’s AI trades in recent months gives the impression that Wood is moving away from the hypergrowth-at-any-price strategy the company gained notoriety for during the 2020 and 2021 bull run and toward something more thoughtful. To date, the flagship ARK Innovation ETF has not outperformed the S&P 500; instead, it has been roughly tracking it.

The majority of ARK’s devoted supporters did not sign up for that particular profile. However, it’s possible that the more measured approach represents a sincere assessment of the cost of concentration risk as well as a wager that patient positioning will still be profitable given the amount of runway left in the AI infrastructure story. The question that no one in this market can honestly answer yet is whether or not that patience is rewarded.