According to the latest trading disclosures, Cathie Wood's ARK Investment Management, via its ARK Innovation ETF (ARKK) and ARK Next Generation Internet ETF (ARKW), purchased 105,616 shares of the newly listed AI chipmaker Cerebras Systems (CBRS.US) on Thursday. Based on Wednesday's closing price, the combined position is valued at approximately $32.8 million.
The Silicon Valley-based AI chip company Cerebras Systems made its debut on the Nasdaq on May 14, staging a long-awaited tech IPO spectacle. The company priced its IPO at $185 per share, a significant increase from the initial range of $115 to $125. In the week leading up to the listing, due to overwhelming institutional demand—with subscription orders exceeding the available shares by more than 20 times—Cerebras twice raised its offering price and increased the size of the sale. Upon opening, the stock surged to $350, reaching an intraday high of $386, a gain of 108%, which triggered the Nasdaq's volatility circuit breaker. The stock ultimately closed at $311.07, marking a 68.2% increase from the IPO price. At the closing price, Cerebras's market capitalization was approximately $67 billion; on a fully diluted basis including options and warrants, the valuation reached about $86 billion.
The company sold 30 million shares in the offering, raising $5.55 billion. If underwriters exercise the over-allotment option for an additional 4.5 million shares, the total proceeds could reach $6.38 billion. This marks the largest global IPO since 2026 and the biggest tech IPO in the United States since Uber's listing in 2019. The offering attracted subscription orders more than 20 times the number of shares available.
This was not an isolated purchase. On the same day, ARK reduced its holdings in Taiwan Semiconductor Manufacturing Company (TSM) by 41,540 shares (approximately $16.6 million), Advanced Micro Devices (AMD) by 11,510 shares (approximately $5.13 million), and semiconductor test equipment maker Teradyne (TER) by 22,576 shares (approximately $8.2 million). Some analysts have characterized this portfolio adjustment as a "generational hardware shift"—taking profits from established semiconductor giants and pivoting to bet on emerging computing infrastructure, such as Cerebras with its wafer-scale engine (WSE) technology, which challenges the existing landscape.
For Cathie Wood, who consistently advocates for a "disruptive innovation" investment philosophy, this trade is a typical projection of her 2026 investment thesis. She stated earlier this year that a new major investment cycle combined with a trend toward deregulation would create significant tailwinds, with core themes including AI, robotics, computing power, and new energy. The substantial investment in Cerebras on its first trading day demonstrates a firm, long-term bullish bet on the AI infrastructure sector.
Cerebras has garnered significant attention primarily due to its unique technological approach. The company focuses on WSE (Wafer-Scale Engine) technology, fabricating a single AI processor from an entire 12-inch wafer, unlike traditional manufacturers that cut multiple individual chips from a wafer. Its third-generation product, the WSE-3, is built on a 5nm process, with a chip area of 46,225 square millimeters, integrating 900,000 computing cores and 4 trillion transistors, delivering AI computing power of 125 petaflops. Compared to traditional GPU clusters, the WSE-3's area is about 58 times that of NVIDIA's B200, with on-chip memory bandwidth 2,625 times greater, and inference speeds are claimed to be 10 to 20 times faster than mainstream GPU solutions.
CEO Andrew Feldman stated that the real turning point occurred in the first half of 2025: "At that time, AI models had become smart enough to be genuinely useful. Our business exploded last year." In his view, as AI transitions from the training era to the inference era, the WSE architecture, capable of generating tokens at lower cost and higher speed, will encounter structural opportunities.
However, technological advantage does not guarantee commercial success. Analysts point out that if a defect occurs in a wafer-scale chip, the entire wafer must be scrapped, not just a small section, presenting inherent constraints on production flexibility and yield. Although Cerebras claims to have developed a fault-tolerant architecture to bypass defective blocks, the stability of large-scale mass production remains to be verified. Furthermore, NVIDIA maintained an 86% market share in the AI accelerator market in 2025, a position that is difficult to challenge in the short term.
Behind this capital celebration, Cerebras's financial data harbors concerns masked by market exuberance. The prospectus reveals that Cerebras's revenue for 2025 was $510 million, a 76% year-over-year increase, and the company successfully "turned a profit"—reporting net income of $238 million, or earnings per share of $1.38, a significant improvement from a loss per share of $9.90 the previous year. At first glance, these numbers paint a picture of a high-growth, profitable AI chip company. However, a closer look at the accounting tells a more complex story.
In 2025, Cerebras recorded a GAAP net income of $237.8 million, but this included a one-time, non-cash accounting adjustment of $363.3 million—a paper gain resulting from the termination of a forward contract liability related to the UAE-based AI company G42. Excluding this item and adding back share-based compensation costs, the company's non-GAAP net loss for 2025 was $75.7 million, which was even larger than the $21.8 million loss in 2024. This indicates that Cerebras's "profitability" was essentially a result of accounting treatment, not the realization of operating profit. In an IPO where the valuation soared to nearly $70 billion, this financial detail warrants attention from every investor.
Another core risk Cerebras faces is extreme concentration in its customer base. In 2024, over 85% of Cerebras's revenue came from the UAE-based AI company G42. This high dependency directly triggered a national security review by the Committee on Foreign Investment in the United States (CFIUS), leading Cerebras to withdraw its IPO application in 2024. The updated prospectus shows that G42's revenue contribution decreased to 24% last year, but the Mohamed bin Zayed University of Artificial Intelligence in the UAE contributed 62% of revenue. These two Middle Eastern entities together still accounted for 86% of the company's 2025 revenue.
CEO Andrew Feldman responded to this by stating, "There are indeed some whale-sized customers in the market, which is one characteristic of this market." However, this is less a market characteristic and more a business model issue—a company positioning itself as a competitor to NVIDIA should ideally sell its chips to "everyone," not just a handful of sovereign-related clients.
To address this concern, Cerebras has made significant efforts over the past year to expand its customer base in the United States. In January 2026, the company signed a multi-year cooperation agreement with OpenAI worth over $20 billion in total, committing to deploy 750 megawatts of high-speed AI inference computing power for OpenAI, accompanied by a $1 billion loan and warrant arrangement. In March, Amazon Web Services (AWS) announced it would deploy Cerebras CS-3 chips in its data centers, making them available to developers through Amazon Bedrock. However, shifting dependency from G42 to OpenAI essentially substitutes "one form of reliance for another." According to Morgan Stanley forecasts, an estimated 90% of Cerebras's sales in the coming years are expected to come from OpenAI. The market can always defend high customer concentration by citing "large clients as quality clients," until the day those large clients stop placing orders.
With a market capitalization of $67 billion, Cerebras's price-to-sales ratio based on its 2025 revenue of $510 million is as high as 130 times. For a company that is still reporting losses on a non-GAAP basis, is this valuation reasonable? Market opinions are significantly divided.
A senior research analyst at Renaissance Capital noted that based on the $185 IPO price and revenue and EBITDA projections for 2028, the valuation might be reasonable; however, measured against the current trading price, the long-term valuation appears elevated. Another analyst stated bluntly, "The market is filled with fantasies about AI, but frenzy is often not a rational pricing tool."
Nevertheless, Cerebras's remaining performance obligations (RPO) stand at $24.6 billion, including the multi-year contract with OpenAI worth over $20 billion, which supports its expectations for rapid growth. If these contracts can be successfully delivered and converted into revenue, the current valuation is not entirely without fundamental support. However, the premise is that AI inference demand indeed explays as Feldman predicts—"increasing by a factor of one million"—and that Cerebras's wafer-scale chips can continuously maintain their technological advantage in this computing power race.
Comments