The daily disclosure file showed the shift before the market opened: ARK Invest funds had bought more Broadcom and Cerebras, and trimmed positions in AMD and CrowdStrike. The trades, disclosed in filings, extend a pattern that Cathie Wood’s firm has been building all year, a bet that the next phase of AI computing will be won by custom chips and specialized silicon rather than the general-purpose processors that have dominated so far.
The logic is visible in the names. Broadcom designs custom accelerators for some of the largest AI customers in the world, chips tailored to specific workloads rather than sold off the shelf, and its stock has been a beneficiary of the buildout of AI infrastructure that Nvidia’s earnings this week underscored. Cerebras makes a wafer-scale processor, a single enormous chip built to run AI models with less reliance on the networking that links thousands of smaller ones. Both companies represent the argument that AI compute is diversifying away from a single architecture.
The sells tell the same story from the other side. AMD has been Nvidia’s main rival in general-purpose AI accelerators, and its products have won design wins and market share, but its fortunes remain tied to the same off-the-shelf model that Wood’s thesis says will commoditize. CrowdStrike, a cybersecurity company, is a different kind of trim: its technology is not AI hardware at all, and its position in ARK’s funds has been shrinking as the firm concentrates on the compute chain.
The timing is pointed. Nvidia’s quarterly report, released Wednesday, showed revenue up 106 percent and a growth forecast that stunned the market, and it also sharpened the debate about whether Nvidia’s dominance can hold. Its largest customers are designing their own chips, and the custom-silicon market is growing faster than the accelerator market as a whole. Wood’s trades are an early vote in that debate, placed by a firm that has built its reputation on identifying technology shifts before the consensus does.
ARK’s history argues for caution in reading too much into any single day of trades. The firm’s funds are known for rapid portfolio turnover, and its positions swing with its research view of each company’s five-year potential. The consistent thread in this year’s buying, however, is a rotation toward companies whose revenue depends on AI inference at scale, the workload that matters most as models move from training to everyday use. Custom chips, which can be optimized for specific models and specific customers, are the companies ARK believes capture that growth.
The trades also reflect a view about the industry’s structure. Wood has argued publicly that the AI buildout will resemble earlier infrastructure cycles: enormous spending first, consolidation later, and value shifting to the companies that own the specialized layers. Broadcom’s custom-chip business and Cerebras’s wafer-scale architecture fit that template, while AMD’s general-purpose line does not. CrowdStrike’s trim is separate, reflecting ARK’s assessment of the cybersecurity market’s growth versus the rest of its portfolio.
The trades fit the pattern that built ARK’s reputation. Wood’s firm made its name betting on Tesla when electric vehicles were a niche, on genomics before the biotech boom, and on the platform companies of the internet economy, and it has built a research process organized around five-year time horizons rather than quarterly results. The AI compute chain has been ARK’s central investment theme for two years, and its funds have moved in and out of the chipmakers as its research has refined the thesis. The current positions, heavy on custom silicon and light on general-purpose accelerators, are the latest iteration of that process, and they carry the firm’s characteristic conviction in one direction.
The companies on the buy side have very different profiles. Broadcom’s custom accelerator business designs chips for a small number of very large customers, a model with high revenue visibility and high dependence on those customers’ spending plans. Cerebras, which listed its shares last year, has built its pitch around a single architectural bet: a chip the size of a dinner plate that eliminates the need to network thousands of smaller processors together. Both are expensive bets on the same idea, that AI workloads will diversify beyond the off-the-shelf GPU, and ARK’s purchases say it believes the idea is not just right, but right soon enough to matter within its five-year window.
The market has been testing the thesis. Broadcom and Cerebras have both risen strongly this year as investors have priced in the custom-chip boom, and their valuations now discount significant growth. AMD, meanwhile, has traded below its peak as questions have mounted about its competitive position. The trades suggest ARK sees the gap as still too narrow, and it has put its money on the side of the thesis it believes will be right in five years.
Whether the bet pays off depends on questions that Nvidia’s earnings made more urgent: how quickly hyperscalers adopt their own silicon, whether custom designs can match Nvidia’s software advantage, and whether inference workloads grow fast enough to support both. Wood’s answer, encoded in this week’s trades, is that the diversification is real and that the winners will be the companies building for it directly. The filings are a small signal, but for a firm whose public posture is built on bold calls, they are the clearest statement yet of where it thinks the AI computing market is headed.
