[IN-DEPTH ANALYSIS] Marvell Technology (MRVL): Betting Big on ASICs Amid Challenges

Key Takeaways

• ASIC Advantages: Marvell's AI ASICs deliver efficiency benefits but contend with constrained market share (10-15%) and concentrated demand.
• Competitive Challenges: Rivals like Nvidia and Broadcom outpace Marvell in development speed and ecosystem breadth.
• Reasonable Valuation: The stock appears fairly valued. Investors can wait for clearer signs of order recovery and competitive momentum before investing.

Marvell was co-founded in 1995 by Indonesian American Dr. Sehat Sutardja, his wife Ms. Weili Dai, and his brother Pantas Sutardja. The three founders have backgrounds in electronic engineering and software engineering, respectively, and have accumulated rich experience in the semiconductor industry. The brothers Sehat and Pantas are passionate about hardware design, but they feel that working for a large company limits their creativity. Weili Dai is good at turning technology into commercial success. With her management skills and execution capabilities, she helps the team turn innovative ideas into reality. Therefore, they form a perfect team, combining technology, management and execution capabilities to jointly promote the development of Marvell.

The company was originally established when the demand for computer hard drives skyrocketed in the 1990s. The founders foresaw that data would become more and more important, and storage was the foundation, but the read and write technology at the time was slow and unstable, and the high-end chips on the market were too expensive for small companies to use. Other large companies such as Intel were busy making CPUs, and no one focused on optimizing the niche field of hard drive chips. So, Marvell used a cheap and power-saving chip manufacturing method, CMOS technology, to make a faster and more stable read and write channel chip. Seagate became Marvell's first major customer. Marvell started from this point and later continued to expand its product line, as shown in the figure below.

Source: Marvell, TradingKey

Strong Revenue Growth

Marvell achieved revenue of $1817 million in the fourth quarter of 2025, up 27% year-over-year. This growth was driven by strong performance in the data center, where Data Center revenues reached $1366 million, or 75% of total revenues, up 78% year-over-year. This growth was driven by demand for AI infrastructure. Although Marvell's consumer and carrier infrastructure businesses experienced revenue declines, strong performance in the Data Center made up for these losses.

Source: Company Financials, TradingKey

Data Center: AI-related revenue accounts for more than 50%, among which 800G PAM4 DSP and 400G ZR interconnect products have strong demand, as shown in the figure below for the revenue and growth of each segment.

Source: Company Financials, TradingKey

It can be seen that although other segments have recovered, the driving force for Marvell's future growth is still in the Data Center. Next, we will focus on analyzing the product strength of the Data Center.

Marvell's products in the AI ​​era: computing power + transmission + storage

Marvell's current core positioning is a data infrastructure semiconductor expert. Its product line is closely centered around the needs of AI, cloud computing and 5G, serving a clear ecological chain: the upstream is chip manufacturing and technology providers, the midstream is Marvell itself as a designer, and the downstream is direct customers and final application scenarios.

1. Customized ASICs

Marvell's customized ASIC is a chip tailored for AI tasks. It uses a 5nm process technology and integrates high-bandwidth memory HBM3. The core of this design is to optimize matrix operations in AI models, especially tensor calculations in deep learning. Compared with general-purpose GPUs such as Nvidia's H100, Marvell's ASIC removes unnecessary functions such as graphics rendering and maximizes the computing efficiency per watt of electricity. The table below compares Marvell's Trainium with NVIDIA A100 in terms of energy efficiency, power consumption, training time, and inference latency, demonstrating its advantages in AI computing.

Source: Marvell, Nvidia, TradingKey

In the market competition, Marvell's ASIC mainly serves hyperscale cloud service providers. Such customers have their own AI ecosystem and do not want to rely entirely on Nvidia's GPU and CUDA software environment. Amazon AWS is Marvell's largest customer. Trainium and Inferentia designed by Marvell run through AWS's self-developed Neuron software stack, helping AWS reduce computing costs by 30% to 40%. However, Marvell's share of the AI ​​computing market is currently small, accounting for only 10% to 15%, especially in large model scenarios. Broadcom is another major competitor in the field of custom ASICs. Broadcom has a long-term accumulation in the ASIC field, and its TPU series in cooperation with Google has been iterated to the seventh generation. In summary, Marvell's disadvantage is that its customer concentration is too high. Marvell's ASIC mainly serves Amazon AWS and is highly dependent on AWS's needs. Resources are also mainly developed for AWS, making it difficult to develop new customers. Therefore, Amazon's bargaining power is higher than Marvell, and it lacks deep integration with optical module manufacturers and switch chip manufacturers, and its ecological support is relatively single. At the same time, it relies on TSMC for foundry and has a concentrated supply chain, which is greatly affected by tariffs. Chips imported from Taiwan to the United States may need to pay high tariffs, increasing costs, which may eventually be passed on to Marvell and affect its profit margin. The comparison with Broadcom's technology route and market strategy is shown in the figure below.

Source: Marvell, Broadcom, Tradingkey

Although self-developed ASIC is a general trend, it is full of uncertainty as to who will win in the end. Under the current market sentiment of AI risk off, these uncertainties are amplified, leading to a decline in the share prices of ASIC-related stocks, and betting on a certain ASIC company is risky. There are two current difficulties in the ASIC field:

First, it is industry related. At the current stage of AI development, the demand for computing power far exceeds the demand for efficiency. GPUs are more in line with this demand with their versatility and mature ecology. In contrast, ASICs are chips customized for specific tasks. Although they are cheaper and more efficient, they are less flexible. Once the design is finalized, the modification cost is high and the cycle is long, making it difficult to adapt to rapidly changing AI models and algorithms.

Nvidia's GPU update speed is accelerated, while the development cycle of ASIC usually takes 2-3 years, which is difficult to keep up with Nvidia's pace. GPUs are highly adaptable and can support various computing tasks and frameworks. Task switching only requires code modification. If the design direction of ASIC is wrong, it may fail and cannot be adjusted quickly.

The accumulation of AI computing power has not yet reached its limit. The focus at this stage is to run, not to run optimally. Hardware manufacturers are more focused on meeting diverse needs, while ASICs need to have clear and stable target scenarios. However, AI applications such as chatbots, smart assistants, and autonomous driving are still in the exploratory stage, with unclear needs and high risks of locking in specific scenarios, which makes it difficult for ASICs to play their specialized advantages.

In short, the time to invest in ASICs is not yet ripe. In the future, as computing power requirements reach physical limits or cost pressures increase, efficiency may become a focus, but the ASIC team needs to survive this stage first.

Second, more and more players are entering the ASIC field. Intel and Wipro are working together to accelerate ASIC design, and Nvidia supports MediaTek in developing ASICs. According to a report by Guangfa Securities, MediaTek has grabbed Microsoft's self-developed chip orders, replacing part of Marvell's share. At the same time, according to UBS and AWS order adjustments, the original plan to launch the 5nm enhanced version of Trainium in the second half of 2025 has now been changed to 3nm and postponed to the second half of 2026. Alchip may exclusively win the next-generation Trainium order, and Marvell faces greater competitive pressure. The following figure is Morgan Stanley's forecast of ASIC market share by 2027:

Source: Morgan Stanley, TradingKey

So, in the long run, Broadcom has more advantages in this field, because from the perspective of so many companies being able to jump directly into developing the latest generation of ASIC chips, the technical threshold for developing ASIC itself is not high. As long as the requirements of the software algorithm are clear, it is not complicated to design the corresponding hardware logic. But the real challenge lies in how to efficiently integrate ASIC with other system components, especially the transmission and communication links. This requires optimizing data transmission speed, reducing latency, ensuring compatibility, and balancing power consumption and cost. Broadcom has an advantage in this regard because it is not only good at ASIC design, but also has strong system integration capabilities, covering optical modules, switching chips and software stacks, and can provide end-to-end solutions with more competitive transmission efficiency and speed.

2. Optical DSP

Optical DSP is Marvell's core advantage product in the Data Center market segment. Marvell quickly entered the optical communication market through the acquisition of Inphi in 2021, avoiding the long cycle of research and development from scratch. Marvell's chip Orion is specially designed for pluggable modules and is suitable for use in cloud data centers and telecommunications networks. It has low power consumption. Orion is particularly suitable for AI clusters, because AI training requires many chips to work together to transmit massive amounts of data, such as data for model updates. If the old 400G module is used, the data volume will be stuck and the efficiency will be very low. Orion can greatly increase the transmission speed and help AI clusters expand from thousands of chips to tens of thousands of chips.

Marvell has also made new progress in optical interconnect technology. In January 2025, they launched a new architecture that directly installs optical interconnect technology into AI accelerators, making the connection speed between chips 100 times faster than traditional electrical signals, and the transmission distance can be increased from a few meters to hundreds of meters. This means that AI clusters can be larger and more efficient.

In the market competition, Marvell and Broadcom are the main competitors. Broadcom's DSP is more versatile and suitable for customers who need a complete solution. Customers need to purchase Broadcom's full range of products to achieve optimal performance. Marvell focuses on the high-end needs of AI and cloud data centers. AI cluster customers pay more attention to transmission speed, scalability and ecological compatibility.

Source: Marvell, Broadcom, TradingKey

Although Marvell's Orion consumes more power than Broadcom, it is fast, supports long distances and has good compatibility. Marvell's Orion is more flexible, and customers can freely choose optical modules and switching chips from other manufacturers. Therefore, it is more popular with major customers such as Google and Meta and has a larger market share.

3. Storage and network chips

Marvell's storage and network chips include the PCIe 5.0 SSD controller Bravera SC5 series and Teralynx switch chips, which are designed for AI infrastructure. The Bravera SC5 controller solves the AI ​​training storage bottleneck through high-speed reading, reducing the time for video generation models to load PB-level data sets from 2-3 minutes to 10-15 seconds, improving AI training efficiency by 20%, and ensuring data security to meet AWS and Meta requirements. Teralynx switch chips support distributed AI tasks with low latency and high throughput, such as Meta's advertising algorithm, reducing communication overhead by 50% and improving algorithm efficiency.

The performance of each company's products in the storage and network fields is not much different. Marvell's foothold lies in its precise high-end positioning and deep customization capabilities, rather than simple technological leadership or market share. Despite its low market share, Marvell has obtained stable orders through in-depth cooperation with AWS and Meta, and has responded quickly to customer needs with its flexible customization capabilities, but its success and failure also depend on large customers. In contrast, Phison and Silicon Motion occupy an important position in the consumer storage market, while Broadcom and NVIDIA lead in the network field by relying on full-stack integration and GPU ecological advantages. This reflects Marvell's shortcomings in market expansion and brand influence, and its high reliance on a few large customers. The technology iteration in the storage and network fields is rapid. If Marvell cannot continue to innovate, it may be further left behind by Broadcom and NVIDIA. Intensified market competition may also compress its profit margins, especially when its market share is low, and the scale effect is not enough to support long-term growth.

Financial Metrics

Profitability

Marvell achieved a positive GAAP net income of $200.2 million in the fourth quarter of 2025, mainly due to the strong growth of Data Center (Note: The abnormally high net profit in FY20Q4 was due to Marvell completing the sale of its Wi-Fi connectivity business in December 2019, which resulted in cash proceeds of $170 million. In addition, the company also received approximately $763 million in tax benefits through internal asset transfers, which had a significant positive impact on net income). Non-GAAP net income reached $531.4 million, reflecting the profitability of the company's core business after excluding one-time expenses. Previously, Marvell's GAAP net profit has been negative, mainly due to high non-operating expenses such as acquisition amortization and restructuring expenses. Marvell has strengthened its capabilities in ASIC and data centers through the acquisition of companies. Although these acquisitions have brought short-term financial burdens, they will help drive the company's growth in the data center market in the long run.

Source: Company Financials, TradingKey

Although Marvell's gross profit margin and operating profit margin have improved, they are still lower than the industry leader Broadcom. Broadcom's gross margin reached 78%, much higher than Marvell's 50%. The company still has room for improvement in cost control and profitability.

Source: Company Financials, TradingKey

R&D Investment

Marvell has maintained strong competitiveness in the semiconductor industry through continuous R&D investment and strategic layout. The company now is more focused and efficient in R&D investment in certain areas, rather than reducing overall R&D investment. In fact, Marvell's R&D expense rate (over 30%) has been higher than the industry average, mainly used to promote its technological innovation in data centers and AI.

Marvell's operating cash flow was $514 million in the fourth quarter of 2025, totaling $1.68 billion for the full year, and it is financially stable and can support R&D and shareholder returns. The decline in CapEx may be due to the company's optimization of production efficiency or adjustment of expansion plans, possibly due to the postponement of AWS projects.

Source: Company Financials, TradingKey

Valuation

Marvell's current forward P/E is 22x, based on consensus forecast of FY2026 EPS of $2.87, which is lower than the industry average of 25x. Taking into account the risk of order adjustment, competitive pressure and tariff, revenue growth may be lower than expected. Based on this, we lower the forecast EPS for FY2026 by 10%-20%, and the forward P/E rises to 24.5x-27.5x. Therefore, the current price is close to reasonable. The market has already priced in these risks. However, the market is still optimistic about the long-term growth potential of Marvell's optical DSP and AI ASIC. FY2026-2030 EPS is expected to grow by 20% annually. The future growth depends on order recovery and competitive dynamics.

Source: S&P Global, TradingKey