After Lao Huang shouted sell the news, MRVL fell a lot - now may be the time to get to know this company again Last week, Lao Huang named MRVL in a high-profile manner at Computex, saying that it would become the next trillion dollar company, and then skyrocketed by 30% in a single day. Then this week, following the market's decline, many people chased after the high and were trapped. Someone asked me in the comment section what I think, so I reorganized this company and explained it clearly to friends who are not yet familiar with MRVL. What exactly does MRVL do If Nvidia is responsible for building cars, then MRVL is responsible for repairing highways. The more cars there are, the greater the demand for roads. If the future AI data center expands from 100000 GPUs to 1 million GPUs, it will benefit not only car sellers but also road builders. The specific business is divided into three parts: The first area is optical communication, producing DSP chips and optical components, with the core customer being AI data centers. This is currently the fastest growing engine, and the company expects the growth rate of its optical communication business to exceed 70% this year, compared to only 30% six months ago. The second part is custom chip ASIC, helping cloud vendors such as Google and Amazon to make custom AI processors. This market has been hyped for two years, and now it is finally starting to truly increase volume. The company expects to double its growth next year and maintain a scale of over 4 billion US dollars the following year. The third block is the switch, responsible for traffic scheduling within and between data centers. This market has hardly been priced in yet, which is the biggest potential surprise. Why is the position of MRVL becoming increasingly important The main bottlenecks of early AI were computing power and storage, whether the GPU was fast enough and whether there were enough HBMs. But now we are running inference models and hybrid expert models, which require the transmission of large amounts of data across clusters, and the communication bandwidth between GPUs has become a new bottleneck. The larger the center is built, the more the railway carrying data cannot be interrupted. The connection layer where MRVL is located is an indispensable part of AI infrastructure. financial data The company expects a total revenue of approximately $11.5 billion for the fiscal year 2027, a year-on-year increase of 40%. In the fiscal year 2028, it surged to 16.5 billion US dollars, an increase of 45%, which is an additional 1.5 billion US dollars higher than the previous expectation. More importantly, the cost structure. The revenue growth rate is 45%, but the operating expenses growth rate is only 15% to 20%. The gap widened in the middle is all converted into profits. The target operating profit margin is 38% to 40%, and the management says that it will reach the upper limit of this range in the 2028 fiscal year. In financial terms, this is called the release of operational leverage, where revenue skyrockets but costs do not follow suit. This is what institutions truly want to see. The management also directly advanced 1 billion US dollars to TSMC's lock production capacity, and the shipment schedule was moved up from the originally expected fourth quarter to the third quarter. Voting with cash is much more trustworthy than telling stories with your mouth. Cooperation with NVIDIA should be taken seriously This time, MRVL and Nvidia have reached a deep cooperation on three levels: AI wireless access network, NVLink high-speed interconnection integration, and joint development at the silicon photonics level. But what is more worth considering is why Nvidia is willing to do so. Cloud vendors want to use self-developed chips to reduce costs, but they also need compatibility with the Nvidia ecosystem. Nvidia is not convenient to directly provide highly customized solutions for customers, so MRVL will handle this matter. The current role of MRVL is to serve as a bridge connecting major technology giants, not just a component supplier. The impact of this identity change on valuation is structural, not simply adjusting the price to earnings ratio. There are still a few items that have not been priced in the market The switch business has almost not been included in the forecast for the 2028 fiscal year, which is purely a potential excess return. The guidance for optical communication still has room for further improvement, as stated by the management themselves. In addition, there were three obvious evasions at the French speaking meeting: there was no direct answer to whether the new front-line major customers would exclusively supply, the issue of the customized ARM CPU market was bypassed, and specific project details such as Microsoft and Amazon refused to be disclosed. There is a high probability that there is a bigger strategic plan behind these evasions that has not yet been revealed. But it should be noted that things that are not spoken out can only be used as clues, not as events that have already occurred to add to the inventory. Risks need to be clearly stated The data center business accounts for 76% of the total revenue, highly concentrated in a few North American cloud vendors. Any major client that reduces capital expenditures will experience significant fluctuations in revenue. The forecast for fiscal year 2028 is based on the assumption that cloud vendors' capital expenditures will continue to grow by over 30%. If this assumption is not met, there will be significant financial pressure. Operating expenses are still expanding this year and will not see significant convergence until the 2028 fiscal year, putting pressure on short-term profits. There is also the issue of equity dilution, as mergers and acquisitions and cooperation with Nvidia have led to an increase in the number of shares to 915 million. Although there are $200 million repurchases every quarter, it is necessary to continue to observe whether it is enough to offset. How do you view this price now MRVL has recently followed the market's decline and has since retraced significantly from its high position after Old Huang's call for orders. If the market continues to decline, as mentioned earlier, there is an opportunity to fill the gap around 220, which is a serious entry opportunity worth taking seriously. The current debate in the market is about the degree of MRVL redemption and the timing of repricing. This company is transitioning from a custom chip supplier to a full suite provider of AI interconnect layers. If this transformation is fully priced by the market, the valuation logic will be completely different. But the best strategy right now is to wait. Waiting for a gap, waiting for better prices, building positions in batches according to the plan, or if the performance is stronger, 250 may not necessarily fall below. For those who have not yet entered the market, DCA can be divided into batches, and the minimum drop standard is to reach a gap of 220.