In recent years, semiconductor companies have expanded their activities to include system development. This trend is not surprising since it provides critical infrastructure space, power, and performance optimization opportunities within datacenter deployments to keep pace with demanding workload requirements. Modern AI, including generative and agentic frameworks, will benefit from this development activity in many ways, especially given rising power consumption and throughput demands. Case in point: both NVIDIA and AMD are moving towards more fully integrated rack solutions, while hyperscalers and cloud service providers are designing both servers and switches to complement their own custom silicon designs.
As AI surges in importance, there is a need for shorter time-to-market. There is also a demonstrable need to manage overall development costs to improve margins by leveraging other core competencies of solution providers. Silicon system development addresses both of these challenges.
In this piece, I will highlight Marvell’s journey in system development and its proven ability to deliver optimized silicon to meet the needs of demanding applications and workloads.
How Marvell Optimizes Silicon Design to Optimize Device Design
Marvell continues to be at the forefront of enabling system development for infrastructure — especially AI infrastructure — within the datacenter, but in an unconventional way. Instead of starting with servers, it develops optimized companion systems that support compute infrastructure. The company is succeeding on many fronts, as demonstrated by its achievements in specialized processing and ecosystem collaboration.
With respect to specialized processing, Marvell is widely recognized for its ability to reduce power, space, and cost through its pursuit of optimized silicon design. Optimized silicon is then used to run new classes of devices that consume less power, take up less space, cost less to operate and buy, and over time reduce any performance gap with larger, more traditional systems.
Long-Distance Optical Connectivity: Marvell’s Product Strategy at Work
As an example of this, coherent DSPs were initially developed for subsea telecommunication cables that stretch across great expanses. Inphi, now part of Marvell, created these devices to support pluggable modules that can be incorporated into datacenter switches. The first versions that were produced more than a decade ago provided 100G of bandwidth for 40km connections, long enough for typical metro network architectures. By leveraging complementary technologies to steadily boost performance, pluggable coherent modules like Marvell’s COLORZ 800G can now deliver an astounding 800G over links that can extend for 1,000km and beyond.
This example also shows how product ecosystems can foster innovation, faster time-to-market, scale, and adoption. To this end, Marvell has a proven record of selling coherent chips along with complete modules. This hybrid marketing strategy has allowed an ecosystem for pluggables to grow and flourish. Customers benefit from having greater choice, and the supply chain benefits from resiliency achieved through diversity.
This ecosystem approach continues to accelerate innovation. A prime example of this comes from last year’s showcase of Coherent, Lumentum, and Marvell. The three companies were able to add probabilistic constellation shaping to modules running Marvell Orion DSPs, which enabled them to reach distances of 500km at full bandwidth and 2,500km at 400G. As a result, Marvell is the market leader in this space, which represents one of the fastest-growing segments in the semiconductor industry for module-based solutions.
Streamlining and Scaling Encryption in the Cloud
To turn to another example, hardware security modules are not widely understood, but they underpin an enormous number of mission-critical processes. HSMs manage encryption and authentication for credit card transactions, subscription services, and credential protocols. Traditionally, HSMs have been deployed as separate servers that are managed directly by customers within datacenters. This legacy approach to security has created not only complexity, but also OpEx and CapEx burdens.
To address these inefficiencies, Marvell has used the specialized OCTEON processors in its line of LiquidSecurity HSMs to shrink these devices to the size of a PCIe card. This allows them to be moved to densely-populated racks for the cloud. That way, instead of purchasing incremental compute hardware and dedicating operational management resources, customers can leverage the power of HSM services with more flexibility for however many transactions or users are required. This capability empowers cloud infrastructure to spin up new security services that will consume less power and space.
The current generation of LiquidSecurity HSMs is also capable of performing a stellar 1 million transactions and managing the same number of encryption keys. Rapid adoption points to its success, with eight out of ten hyperscalers now using LiquidSecurity HSMs for cloud workloads. Additionally, these modules are highly interoperable with platforms that include IBM HashiCorp and Cryptomathic, broadening their widespread market appeal. Marvell has also been working to certify LiquidSecurity for more services and use cases — such as cross-border transactions and agentic AI security across highly regulated markets — to scale its implementation potential.
System Development’s Long Tail
What Marvell — along with NVIDIA and AMD — is doing in system development is just the beginning. Given the insatiable needs of modern AI workloads, it is expected that this category will continue to grow rapidly. System development in the semiconductor industry could also open the door to new devices — including memory appliances, dedicated processing appliances, and scale-up fabrics.
In the modern AI era, the transition from discrete silicon solutions to systems is an interesting one to watch. One thing is certain: Marvell is trailblazing in the space through its strategic acquisitions, continuous engineering refinements, and a willingness to partner with others.
