"The next trillion-dollar company, ladies and gentlemen." NVIDIA CEO Jensen Huang described Marvell, a relatively unknown American semiconductor company to Taiwanese audiences, this way as he took the stage at Computex. This statement sent Marvell's stock price soaring by over 30%, marking the company's largest single-day gain in history.

Nvidia invests $2 billion in Marvell

Jensen Huang's statement carries a hint of endorsement, as Nvidia announced a $2 billion investment in Marvell two months prior. The collaboration covers technologies such as optics, silicon photonics, and NVLink Fusion. However, it also points to a new shift in AI infrastructure: as AI factories grow larger, the focus of competition is no longer solely on GPUs and HBMs themselves, but is also turning to "connectivity"—that is, how to connect massive amounts of computing resources at high speed, stably, and with low power consumption.

"When you break down a computational problem into many parts and distribute them across the entire data center (when using agentic AI), connectivity becomes absolutely essential," explained Jensen Huang. "That's why Matt does it so well, and that's why Marvell is so crucial."

Marvell CEO Matt Murphy described Marvell as having strategic partnerships with computing and memory companies, stating, "In a way, we're the Switzerland of this industry." In other words, Marvell aims not to be the most prominent GPU supplier in the AI factory, but rather a "neutral" entity providing high-speed data movement technology between different chips, systems, and cloud service providers.

Who is Marvell?

Marvell didn't start out at the center stage of the AI data center arena. Murphy recalled that when he took over as CEO in 2016, Marvell's annual revenue was approximately $2.3 billion, with data center business accounting for less than 10%, while consumer products accounted for over 60%. He even joked on stage that one of Marvell's key design projects at the time was the first Wi-Fi-enabled "Barbie's Neverland."

But Murphy predicted that the next wave of semiconductor growth would come from data platform companies like Google, Amazon, Microsoft, and Meta. These companies needed technologies not only for computing but also for moving, storing, processing, and protecting massive amounts of data. Therefore, Marvell began to reposition itself as a "data infrastructure" company.

Over the past decade, Marvell has shifted its product portfolio towards data centers, custom chips, switches, and high-speed connectivity through acquisitions and internal R&D. Murphy stated that Marvell invested approximately $22.5 billion in acquisitions, approximately $18 billion in internal R&D, and sold approximately $4.5 billion in assets, totaling approximately $36 billion in building its data infrastructure platform.

This transformation is also reflected in its revenue structure. Murphy pointed out that Marvell's data center business accounted for over 75% of its revenue in the most recent quarter, and the company looks significantly different from what it did 10 years ago. He also stated that according to Wall Street's consensus forecast for this year, Marvell's revenue is expected to grow to $11.4 billion, approximately 2.5 times that of the past five years. For Marvell, the AI boom is the result of its transformation bets over the past decade.

Why do AI factories need Marvell?

In his speech, Murphy posed the question: What determines the performance of AI infrastructure? Many would first think of GPUs, XPUs, process nodes, or memory bandwidth, but he believes these are not the complete answer. Because no matter how fast a single processor or how large the memory, it cannot independently support today's AI workloads. AI systems require tens of thousands of processors to operate together, and in the future, this may even lead to millions of processors cooperating.

This is also the core reason why Marvell is entering the AI factory. As AI infrastructure expands from a single chip to entire racks, data centers, and even cross-data center operations, the ability of processors to exchange data at high speed will directly affect system performance. "Ladies and gentlemen, we have already seen amazing breakthroughs in accelerated computing and the rise of HBM to address the challenges brought by AI. But the next major wave of innovation and scaling will come from the underlying connectivity technologies of these systems," Murphy stated. He added that the industry has faced bottlenecks in computing power and memory in the past, and now the bottleneck is shifting to connectivity.

Huang Renxun echoed this statement in the dialogue. He pointed out that useful AI has arrived, and the next generation of AI computing models is proxy AI. "These AI agents have a specific computing platform, which is a fragmented, distributed computing model." The computing will be distributed across the entire data center. When computing is broken down and distributed, Marvell's expertise in "connectivity" becomes a key technology.

What does "Switzerland in industry" mean? Why does it want to be a "neutral country"?

Marvell describes itself as "the Switzerland of the industry," not because it doesn't compete, but because its core strength lies in connectivity, and therefore it doesn't want to be tied to a single computing company or cloud service provider.

AI data centers contain GPUs, CPUs, memory, switches, and optical connections. These components may come from different companies, but ultimately they must be connected to form a truly functional AI factory. Marvell targets this layer: it provides high-speed input/output (I/O), SerDes (sequencer/deserializer), switches, optics, and silicon photonics technologies, enabling data to move rapidly between chips, servers, racks, and data centers.

In other words, Marvell's value isn't in betting on a single computing chip, but in providing the data movement capabilities needed by different computing platforms. The more cross-platform it can collaborate, the larger the market for this connectivity capability will be; being seen as serving only a single platform would limit its role.

Marvell's "neutrality" is more accurately described as a connectivity layer provider that doesn't tie itself to a single platform. It can collaborate with Nvidia, or serve the needs of cloud service providers for their self-developed chips, because what it wants to sell is not a specific computing platform, but the high-speed connectivity capabilities that enable different chips and systems to connect to AI data centers.

Why are optical interconnects and CPOs becoming the next focus?

Marvell is betting heavily on connectivity technology, with optical interconnects and CPO (Common Packaging Optics) being key areas of focus. Mo Fei pointed out that copper wires are simple and inexpensive, and the industry will use them for as long as possible; however, copper wires have physical limitations in bandwidth and distance. The higher the signal bandwidth, the shorter the transmission distance. He gave an example: a 100Gbps system can use approximately 5 meters of copper wire per channel, and about 2.5 meters per channel for 200Gbps. Given that the rack height is approximately 2 meters, plus internal wiring, 2.5 meters is close to the limit. When speeds reach 400Gbps per channel, copper wires will struggle to completely connect the entire rack.

This is why CPO is being discussed. CPO brings optical connections into the package, closer to the switching or computing chips, reducing the transmission distance of high-speed signals over copper wires, further reducing power consumption and increasing connection density.

Huang Renxun also stated in the discussion that the correct strategy is not to immediately replace all copper wires with optics, but rather to "use copper wires as much as possible, and only use optics where absolutely necessary." He believes that data centers will continue to heavily utilize copper wires and optical connections over the next 5 to 10 years; "the convergence of copper and light will continue for a long time."

Mo Fei also extended this connectivity technology transformation to a longer-term vision of data center architecture. He stated that future data centers may evolve into a data infrastructure "without distance limitations," where computing, memory, networking, and photonics technologies will operate as a unified system. When computing and memory resources can be managed uniformly, and infrastructure can be dynamically configured according to workload, AI system design will have the opportunity to shift from being "limited by interconnect distance" to "configuring resources according to model requirements."

Besides Huang Renxun, why is Wu Tianyu from ASE Technology Group also endorsing Marvell?

Another noteworthy aspect of the presentation was the participation of ASE Technology Holding Co., Ltd. COO and CEO Tien-Yu Wu in support of Marvell. In his speech, Mo Fei stated that customers need reliable, manufacturable, and scalable solutions, not just presentations, demonstrations, or proof-of-concept, and that Taiwan's supply chain can meet this demand.

Wu Tien-Yu added that ASE's decision to partner with Marvell wasn't just a bet on Marvell's success, but also a search for a partner who could foresee the needs of next-generation architectures and technologies. For Taiwan's manufacturing industry, much infrastructure and capital expenditure must be invested years in advance, and once production capacity is established, it's crucial to ensure its future market usability.

He also pointed out that Taiwan's semiconductor ecosystem has accumulated decades of experience, from PCs, wireless communications, mobile computing, and data centers to HPC, achieving economies of scale, cluster efficiency, and a strong talent pool. These conditions give Taiwan a unique and difficult-to-replicate role in the next-generation AI infrastructure supply chain.

Currently, Marvell's collaborations in Taiwan primarily focus on two areas: advanced processes and advanced packaging. TSMC supports Marvell's advancement of chip platforms from 5nm, 3nm to 2nm, for applications in data centers, network connectivity, and customized chips; ASE, on the other hand, focuses on advanced packaging and optoelectronic integration, assisting Marvell in meeting the needs of high-speed interconnection in AI data centers, CPO, and system integration under silicon photonics architecture.

Source: Digital Times

View more at EASELINK