The more stunning AI capabilities become, the faster data must move! Marvell (MRVL.US) leverages 2nm optical interconnect to open new space for its AI infrastructure business.

Focusing on AI ASIC and data center optical interconnect chips, Marvell Technology, Inc. (MRVL.US) is actively showcasing its next-generation 2nm optical interconnect technology to global investors at the 2026 European Conference on Optical Communication (ECOC), drawing close attention to its stock performance. The showcase comes as large AI data center operators demand faster connections and more efficient data transmission among AI accelerators, CPU processors, HBM memory, and even data center NAND, while avoiding significant increases in connection and transmission power consumption. Based on the "Jevons Paradox," overall demand for AI computing resources will continue to expand in the future as inference-side computing costs decline. From the perspective of inference system architecture, more complex tasks often involve longer contexts, multi-turn model calls, tool execution, and result verification: Prefill needs to process inputs, Decode continuously generates outputs, and KV Cache occupies more memory as context and concurrency scale expand. Computing throughput, memory bandwidth, and capacity need to improve in coordination. The core inference for the AI computing industry is that GPUs and TPUs handle model computation, CPUs handle tool execution and task orchestration, and HBM, server DRAM, storage, high-performance network infrastructure, and data center optical interconnect components jointly support efficient data transport and state management; ultimately, a complete and increasingly large-scale AI computing server cluster will be needed to deliver sustainably operating services. The recent sustained strong stock performance of data center optical product suppliers such as Marvell and Lumentum also stems from market reports indicating that "AI chip leader" NVIDIA Corporation (NVDA.US) is adjusting the architecture of its next-generation AI rack computing system, namely that the next-generation AI computing cluster Rubin Ultra may reduce HBM configuration per rack and instead use an optical interconnect technology architecture to connect multiple AI racks. As HBM/DRAM memory systems and NAND storage chip products have seen continuously soaring prices in recent years, the focus of AI computing infrastructure investment appears to be shifting from GPU/TPU single-point performance and HBM capacity toward overall data center architecture efficiency, with high-speed optical interconnect potentially the most critical direction. This ECOC showcase further reinforces the growth path for Marvell's optical interconnect: the 2nm process upgrade and single-channel 400G optical connection target higher bandwidth density and lower energy per bit. The 2nm here mainly refers to the manufacturing process for electronic chips such as DSPs; at nominal rates, eight 400G channels can support 3.2T connections. Another 102.4Tbps co-packaged optics (CPO) platform using single-channel 200G silicon photonics technology shortens high-loss electrical channels by bringing optical connections closer to compute and switch chips. From the perspective of large data center construction, this helps carry more communication traffic within a given power supply, cooling, and rack space, improving the utilization of expensive accelerators. For Marvell, its commercial value will be reflected through customer design wins, high-speed interconnect product upgrades, and subsequent volume shipments, and is expected to become a new catalyst for upward revisions to revenue and profit expectations. OpenAI Astra's improved capabilities in software engineering, computer operation, and multi-step tasks, as well as suspected testing news about Gemini 4 Promarket reports indicate that Alphabet Inc. Class C has apparently quietly put its flagship model Gemini 4 Pro on the AI Arena for testing under the name "gemini-3.8-flash"are continuously strengthening market expectations for a new round of AI application expansion. From the underlying data center engineering logic, as coding, scientific research, and multimodal tasks shift on a large scale toward agent execution, and application penetration and concurrency scale expand, this will drive cross-chip collaboration, context cache access, and cross-rack data exchange in distributed inference, placing higher demands on network throughput, latency, and energy efficiency. Marvell's ECOC showcase of single-channel 400G optical interconnect technology based on the 2nm process, supporting network upgrades to 3.2T, is precisely aimed at helping AI clusters reduce data transmission bottlenecks and improve effective computing power by increasing bandwidth density and lowering power per bit. This is also the consistent logic behind Wall Street analysts' bullishness on Marvellthat is, application expansion brought by breakthroughs in AI capabilities is expected to simultaneously drive demand for custom ASICs and a new round of high-speed optical interconnect, enabling Marvell to gain business growth at both the computing and connectivity ends. Marvell steps up its AI data center high-speed connectivity layout with 2nm optical interconnect technologyECOC highlights at a glance As AI clusters require faster connections without data center power consumption growing at the same proportion, Marvell is pushing its optical communications portfolio toward more advanced process nodes. Marvell Technology, Inc. (MRVL) is showcasing its next-generation 2nm optical interconnect technology at the 2026 European Conference on Optical Communication (ECOC), drawing attention to its stock. The showcase comes as AI data centers require faster connections among accelerators, processors, and memory while avoiding significant increases in power consumption. ECOC is Europe's largest professional conference and exhibition for optical fiber communication and optical networking. The conference is currently being held in Malaga, Spain, attracting more than 340 exhibitors, with special sessions on AI infrastructure and advanced technology demonstrations. Marvell's stock rose more than 3% in Monday pre-market trading in the U.S. According to the company, Marvell's demonstrations include the first single-channel 400G optical PAM4 solution using a 2nm process, a 2nm 800G ZR/ZR+ pluggable module with security features, and 2nm 1.6T ZR and lightweight coherent technology. These products are designed to increase bandwidth density and reduce energy per bit as AI infrastructure scales. The company said its advanced optical and silicon chip designs leverage leading processes to achieve improvements in performance, bandwidth density, and power consumption, which are critical for efficiently scaling AI infrastructure. Faster AI chips can only deliver their value when Beijing Vastdata Technology can transmit between chips fast enough. This is why the AI data center optical interconnect/optical communications architecture led by Marvell Technology, Inc. is moving from 1.6T connections toward 3.2T and higher speeds, increasing demand for faster and more energy-efficient optical links. Marvell's single-channel 400G demonstration is intended to support this shift, while its 1.6T technology targets high-bandwidth connections within and between data center facilities. The company is also showcasing its 102.4 terabits per second co-packaged optics (CPO) platform. As traditional copper connections encounter distance and power limitations, this platform integrates optical connections more closely with AI accelerators and switch chips. At ECOC, Marvell plans to present 38 technology demonstrations covering 224G serializer/deserializer (SerDes), 1.6T optical connections, coherent interconnect, and optical memory interconnect networks, further advancing its overall layout to address the connectivity bottlenecks brought by scaling AI systems. In addition, Lumentum (LITE.US), Qualcomm (QCOM.US), and Corning Inc (GLW.US) will demonstrate a high-density die-to-die optical interconnect solution for AI scale-up systems at ECOC. The solution combines Qualcomm's interface technology, Lumentum's 1060nm vertical-cavity surface-emitting laser (VCSEL) optical platform, and Corning Inc's multimode fiber, aiming to provide higher-bandwidth, lower-power connections as AI systems push beyond the practical limits of electrical interconnect. However, on the Stocktwits platform, retail sentiment around Marvell's stock remains bearish. Year to date, the stock has risen more than 173%. By comparison, the iShares Semiconductor ETF (SOXX), which holds Marvell shares, has risen nearly 70% over the same period. The more capable AI agents become, the less the data center network and high-speed data transmission can afford to be blocked! The core logic behind Marvell's sharp stock rise this year is that custom AI chip volume growth and high-speed optical interconnect upgrades are simultaneously opening up revenue space. Cloud providers are expanding deployments of self-developed accelerators, bringing incremental growth to its custom application-specific integrated circuit (ASIC) business; the expansion of accelerator clusters in turn drives demand for optical communications digital signal processors (DSPs), high-speed interfaces, and switch chips. The financial results already reflect this trend: in the second quarter of fiscal 2027 ended August 1, 2026, data center revenue reached $2.1715 billion, up 46% year over year, accounting for 79% of total revenue. In September, Marvell management further provided guidance for fiscal 2027 and fiscal 2028 revenue of $12 billion and $18 billion, respectively, and expected custom business revenue to double in fiscal 2028. From an investment logic perspective, these growth expectations provide strong and specific grounds for the market to reassess its AI computing infrastructure business landscape. This ECOC showcase can be said to further reinforce this strong growth path of optical communications product upgrades and volume growth. The joint demonstration by Qualcomm, Lumentum, and Corning Inc also confirms, from another technology route, the trend of high-speed optical interconnect technology in data centers extending closer to core AI computing chips: the three companies combine D2D interfaces, 1060nm wavelength VCSELs, and high-density multimode fiber to extend die-to-die connections to tens of meters, targeting future CPO, near-packaged optics (NPO), and UCIe interfaces. This solution targets scale-up, that is, expanding tightly coupled accelerator compute domains. The investment implication behind this is that demand for optical connections is extending further from traditional network ports to near compute chips and advanced packaging; Marvell covers custom chips, optical DSPs, CPO, and optical memory interconnect, and therefore has the opportunity to participate in value creation across more system links. ASICs expand computing scale, while optical interconnect improves operating efficiency after scale expansion, and together they enhance its long-term growth space. The effective throughput of AI systems is jointly constrained by computing power, memory bandwidth, and cross-chip communication efficiency: distributed training requires synchronizing gradients and exchanging intermediate results, mixture-of-experts (MoE) models require routing data among different experts, and long-context inference increases the capacity and access requirements of KV Cache. Custom ASICs can optimize chip design around customers' operators, numerical precision, memory hierarchy, and data flow to improve energy efficiency for specific workloads; optical interconnect alleviates the distance, bandwidth density, and power constraints faced by high-speed data transmission, reducing the time accelerators spend waiting for data. Therefore, the expansion of global AI deployment increases both demand for computing chips and requirements for connectivity and memory systems. Marvell's unique advantage lies in participating simultaneously in "improving per-unit computing efficiency" and "improving cluster collaboration efficiency," helping customers obtain more effective computing power within limited power and capital budgets and reduce the processing cost per token. Wall Street financial giant Bank of America Corp maintains a $365 price target on Marvell, especially valuing its ability to expand revenue share per AI system through custom AI accelerators (AI ASICs) and accompanying optical interconnect chips built around massive AI agent inference demand. The highest price target is $400, from KeyBanc analyst John Vinh, who raised his target to $400 from $385 and reiterated an "Overweight" rating on August 28. The $400 price target implies about 63.77% potential upside. Analyst Vinh views Marvell as one of the main long-term beneficiaries of AI network expansion, and his team has previously emphasized that the growth opportunity in the data center optical interconnect business is becoming stronger and more sustainablethis bullish logic complements ASICs: cloud giants are increasing deployments of custom self-developed AI chips while also needing to expand the high-speed optical connection components that connect these chips, allowing Marvell to participate in capex growth at both the computing and interconnect ends.