Looking ahead to the next year, a significant system dubbed GH200 is anticipated, integrating GPUs such as A100 and H100 and interconnected via Mellanox switches. Also check-Introduction to NVIDIA DGX H100. The sales dynamics of A100, H100, the rollout schedule of the new system DGX GH 200, and the overall sales landscape are poised for some adjustments. Presently, the ConnectX-7 network card operates at 400G, but an 800G network card, ConnectX-8, is slated for release next year. While the current usage of A100 and H100 includes a mix of 400G and 800G optical transceiver, future trends may lean towards 800G, prompting adjustments in demand forecasts and algorithms for these components.
Main Data Center Customer Demand Situation
Google operated within a 400G environment, utilizing an 8×50 electrical port converted into an 8×50 optical port, necessitating an 8-in-8-out DSP chip for data center architecture.
Amazon's 400G setup involves an 8×50 electrical port transformed into a 4×100 optical port using a DSP chip, indicating the utilization of different chip types between the two companies.
As for Microsoft and Meta, their data centers primarily relied on 100G infrastructure. While Meta initially considered transitioning to 200G or 400G, the current focus remains on the 200G range due to substantial demand, reportedly exceeding one million units annually.
Microsoft had plans to upgrade to 400G last year but postponed the firmware or upgrade implementation due to internal deliberations and a slowdown in market demand. There are speculations that Microsoft might proceed with the 400G upgrade in the latter half of this year, although reconsideration due to AI factors or a leap to 800G is also plausible.
Before NVIDIA's AI-related demands emerged, Google was at the forefront of numerical environments, swiftly advancing towards 800G, while Amazon settled at 400G, Meta at 200G, and Microsoft aimed for 400G but is yet to commence infrastructure development, representing the current landscape of existing data centers.
What About Marvell?
Marvell is a leading global semiconductor manufacturer providing a full range of broadband communications and storage solutions. Historically, Marvell commanded a substantial market share of at least 60-70% for PAM4 DSPs, covering single-channel speeds ranging from 50G to 100G, with the potential to reach 80% when assessed by shipment volumes.
Marvell's annual revenue is estimated to be around $6 billion, with DSP-related business accounting for approximately $1.4-1.5 billion. This encompasses modules catering to various speeds, including 800G, 400G, 200G, and 100G, as well as coherent DSPs tailored for applications necessitating long-distance transmission, such as inter-data center connectivity or telecom infrastructure. Additionally, there exists a customized module segment serving Microsoft, contributing to the overall revenue estimate in the range of $1.4-1.5 billion. If we narrow the focus to PAM4 DSP alone, the figure may hover around $700-800 million.
Forecast of the Situation in 2024
In the upcoming year, we anticipate that the revenue generated from 400G PAM4 DSPs may either stabilize or experience a slight decline. This is primarily due to the lack of significant growth in shipment volumes, compounded by Google's transition towards 800G demand, resulting in a deceleration of 400G demand. As a result, the main driver of growth will be the adoption of 800G optical transceivers. If this year's revenue stands at $200 million, we project it to reach $400 million next year.
Other segments within the DSP business are expected to see modest growth, likely ranging between 10-20%. Specifically, the customized module business for Microsoft and the Coherent DSP business may experience an average growth rate of approximately 20%, translating to an increase of $100-200 million.
Therefore, if this year's total revenue amounts to $1.4 billion, we anticipate a growth of $300-400 million next year, resulting in a total revenue range of $1.7-$1.8 billion.
