AI rack upgrades enter the "battle for system efficiency"! As the OCP Summit approaches, Jefferies decodes 1.6T volume ramp, memory expansion, and high-voltage power delivery.

U.S. earnings season approaches, with the Nasdaq hitting a record high and Marvell management raising its revenue growth target, providing a positive catalyst for upward revisions to profit expectations across the global AI optical interconnect supply chain On October 6, the Nasdaq Composite rose 0.45% to close at a record high of 27,599.79; Marvell (MRVL.US), a leader in ASIC and AI data center optical interconnect chips, raised its fiscal 2028 revenue target from $18 billion to $20 billion, an increase of approximately 11.1%, with its stock rising 5.8% that day, both reflecting the market's positive expectations for expanding demand for AI data center accelerated computing chips, CPUs, HBM/DRAM/NAND memory chips, and data center optical interconnect components. Although the Nasdaq and the Philadelphia Semiconductor Index fell approximately 0.22% and 1.15% respectively on October 7, the previous round of tech stock rally had already pushed the upcoming U.S. earnings season and AI computing industry catalysts into the market spotlight. Wall Street firm Jefferies Financial Group Inc. (Jefferies) released a research report on October 7 titled "Summit Preview: Power Architecture Upgrades, Optical-in-Copper-out Interconnect Evolution, and the Battle for Scale-up Optical Interconnect Dominance," which, centered on the OCP Global Summit held October 12-15 in San Jose, outlined five interrelated investment themes: 800G to 1.6T optical interconnect upgrades, Ethernet, UALink, and NVLink competing for Scale-up interconnect dominance, near-package optics (NPO) and co-packaged optics (CPO) technology routes reshaping device value distribution in AI data centers, inference workloads driving memory architecture expansion, and 400V/800V DC power delivery reshaping rack power systems. Jefferies Financial Group Inc.'s research report most clearly favors MACOM and Lumentum as 1.6T plays, while also suggesting that the ALC and OmniConnect opportunities of Credo, a leader in dual high-speed transmission routes of copper cable and optical interconnect, have not yet received sufficient attention, and also highlighting Coherent's PhotonLink commercialization, Astera Labs' memory controllers and switch chips, Lattice's low-power controllers, and the growth space in power semiconductors. From an engineering and investment logic perspective, the investment opportunities that Jefferies Financial Group Inc. focuses on collectively point to one change: as AI computing cluster scale continues to expand at an accelerating pace, high-speed data movement and transmission, memory configuration, and power conversion capabilities within AI data centers are becoming important constraints determining whether GPUs can continuously deliver effective computing resources and whether capital expenditure can be converted into revenue, and related suppliers are therefore expected to capture higher per-rack value and profit contribution. AI rack upgrades are arguably entering a "battle for data transmission efficiency": as cutting-edge AI agents like Muse and Astra gain global popularity, AI computing cluster scale continues to expand, and the bandwidth, latency, and energy consumption of high-speed interconnects increasingly affect GPU utilization and overall computing output. On the eve of the OCP Summit, Jefferies Financial Group Inc. identified the 800G to 1.6T upgrade, NPO/CPO architecture evolution, and Scale-up interconnect competition as core areas to watch, and explicitly named MACOM and Lumentum as its most favored 1.6T beneficiaries, focusing on the former's opportunities in photodetectors, transimpedance amplifiers, and drivers, and the latter's positioning in high-speed lasers and external light sources. The key investment takeaway is that AI infrastructure expansion is elevating the value of critical optoelectronic components, and suppliers that can convert technological advantages into customer adoption, scaled delivery, and profit growth are poised to become important winners in the optical interconnect upgrade. Optical Interconnect Moves to the Investment Forefront: 1.6T Ramp and the Battle for "Connection Dominance" The investment value of the 1.6T upgrade is reflected in simultaneous changes in optoelectronic device demand, product mix, and supplier share. Jefferies Financial Group Inc. (Jefferies) equity strategist team named MACOM and Lumentum as its most favored 1.6T beneficiaries, with reasons respectively centered on key components at the receive end and transmit end. For MACOM, the focus is on the competitive position and revenue contribution of photodetectors in 800G and 1.6T deployments, as well as the demand for more channels of transimpedance amplifiers (TIAs) and drivers as module makers adopt NPO designs; Semtech's data center-themed event on October 15 may also provide new industry benchmarks for MACOM by disclosing the TIA and driver market size. For Lumentum, the most noteworthy near-term items to track are the 1.6T transmit-end product mix, pricing, capacity, and margins, as well as the demand structure for electro-absorption modulated lasers (EMLs) and continuous wave (CW) lasers. Jefferies Financial Group Inc. added that Broadcom Inc., an ASIC and optical chip leader larger than Marvell Technology, Inc., is at the center of observation in the optical module DSP chip share competition, and the report noted that early channel signals are favorable for it. The investment judgment formed by Jefferies Financial Group Inc. strategists is this data center high-speed optical interconnect rate upgrades can expand industry demand, but companies' profit elasticity also depends on device share, manufacturing capability, and product mix, and MACOM and Lumentum are precisely the two companies Jefferies believes have outstanding exposure to this upgrade. MACOM focuses on optoelectronic devices and high-speed analog chips for AI optical interconnect, including photodetectors that convert optical signals into current, transimpedance amplifiers (TIAs) that convert and amplify weak photocurrents into voltage signals, and high-speed driver chips that drive lasers or modulators, providing key transmit/receive components for 800G/1.6T optical modules; Jefferies Financial Group Inc. is particularly focused on its photodetector share and the multi-channel TIA and driver opportunities brought by NPO architecture. Lumentum focuses on high-performance light sources, optical transceiver modules, and optical circuit switching, including electro-absorption modulated lasers (EMLs), continuous wave (CW) lasers, external laser sources (ELS) for NPO/CPO, and optical circuit switching (OCS) systems that support dynamic reconfiguration of network optical paths; its investment highlights center on the 1.6T product ramp, laser product mix upgrades, and demand growth from new optical interconnect architectures. Regarding NPO and CPO, Jefferies Financial Group Inc. stated that NPO and CPO push optical devices closer to compute or switch chips, and value distribution will be redistributed along lasers, packaging, and complete assemblies. From an engineering perspective, shortening the transmission distance of high-speed electrical signals helps alleviate signal integrity and power consumption pressures of high-bandwidth connections; whether near-package optics (NPO) or co-packaged optics (CPO) is adopted will affect demand for external laser sources, drivers, optical fiber components, and package integration. Lumentum will discuss CPO external laser sources and wide-parallel VCSEL interconnects at the summit, and will jointly present an automated optical circuit switching (OCS) network with Oracle Corporation, the key to which is whether customer interest can be converted into deployments meeting actual operational requirements. Coherent's PhotonLink, launched in September, has engaged with more than ten customers in each of the CPO and NPO directions, with core customers and long-term agreements already in place for both architectures, and is expected to begin contributing revenue at scale starting in Q4 2026; it covers lasers, VCSEL arrays, silicon photonics, optical fiber components, and detectors, and whether customers purchase individual components or complete assemblies will directly affect revenue and profit contribution per deployment, with customer qualification, manufacturing yield, and ultra-high-power platform progress equally important. Meanwhile, Jefferies is watching whether ams OSRAM's micro-VCSEL can enter NVIDIA Corporation's future Scale-up architecture, including a potential NVL1152 Feynman opportunity; this route could change previous assumptions about the value of indium phosphide lasers. Scale-up competition determines how accelerators form larger coordinated computing systems, and also determines the share of switch chips, copper connections, and optical connections. Jefferies is closely watching the competition among Ethernet/SUE, UALink, and NVLink: NVIDIA Corporation is expanding ecosystem participation through NVLink Fusion and may disclose an NPO route for NVL576 Scale-up; AMD launched Helios in July and plans to begin delivering to customers including Microsoft Corporation by late 2026, with investment focus shifting further to deployment execution, UALink adoption progress, and participation opportunities for Astera Labs and Broadcom Inc. Marvell's announced 100G-per-channel UALink switch chip is expected to tape out within the year, and AMD's next-generation rack plans to combine optical and copper connections to achieve cross-rack Scale-up. The Jefferies strategist team stated that Credo offers another noteworthy technology path: it is expected to showcase for the first time ALC, which combines its own connectivity platform with microLED technology acquired from Hyperlume, targeting power consumption and reliability performance approaching active electrical cables (AECs) over distances of up to approximately 30 meters; the report judges that microLED is more likely aimed at Scale-out or even Scale-in, while a potential micro-VCSEL route may target Scale-up. In addition, whether NVIDIA Corporation is considering qualifying Credo's ZF Optics to address repeated link interruptions in new cloud service provider clusters using NVIDIA Corporation platforms is also a key point of observation. Jefferies emphasized that connection bandwidth, latency, and stability directly affect the coordination efficiency of expensive accelerators, and suppliers that can improve these metrics have the opportunity to capture value beyond simple port count growth. Memory and Power Move to the Interconnect Forefront: The Second Strong Profit Curve of AI Rack Upgrades The expansion of massive AI inference workloads is opening up growth space for memory controllers, switch chips, and tiered memory architectures. Jefferies noted that Marvell has emphasized that CXL and memory controllers can become important companion opportunities growing with XPU deployments, and Astera Labs' Leo product line provides a concrete landing point for this direction: Leo 2 E supports CPU memory expansion through CXL 3.2 and PCIe 6, while Leo 2 P supports cross-host memory pooling and sharing; both support DDR4 and DDR5, enabling customers to reuse existing memory and improve resource utilization without adding CPU sockets, and the new product series is currently sampling to hyperscale cloud customers. The GPU-oriented Leo X, meanwhile, connects dedicated memory tiers through PCIe and platform-specific protocols together with Scorpio switch chips, forming another product path distinct from CPU-side CXL expansion; depending on customer configuration, Astera Labs may capture revenue from both memory controllers and connectivity switch chips. The report also highlights opportunities for Amazon.com, Inc. and AMD's next-generation racks to adopt UALink, and whether Amazon.com, Inc. will allocate supply share between Astera Labs and Marvell, and once again emphasizes that Credo's OmniConnect opportunity is undervalued. From the underlying logic of inference systems, Jefferies noted that longer context and higher concurrency will expand cache and working set capacity requirements, and rationally combining high-bandwidth memory with substantial capacity expansion and memory pooling helps improve the utilization of expensive computing resources, also making memory connectivity and management chips a more important part of AI capital expenditure. High-density AI rack high-speed interconnect is simultaneously elevating the value of power conversion and system control, but the deployment sequence of 400V/800V determines the specific pace of revenue realization that is, optical interconnect mainly addresses the bandwidth, latency, and energy consumption issues of data transmission; high-voltage DC power delivery reduces current and line losses in high-power distribution, and through efficient power conversion meets chip power supply needs against the backdrop of high-speed transmission, with both jointly supporting the accelerated expansion of AI clusters; "deployment sequence determines the pace of revenue realization" mainly means that the deployment timing of different power supply solutions will affect when power semiconductor and related control device suppliers receive incremental revenue. In the view of Jefferies strategists. Jefferies expects that NVIDIA Corporation's October 12 keynote, as well as technical discussions around 800V DC, grid connection, liquid cooling, and MGX, will further clarify opportunities for Texas Instruments Incorporated, Analog Devices, Inc., and onsemi. The report noted that delays in the Kyber rack and the 800V transition timeline have led to pullbacks in related stocks; near term, the 400V solution is expected to be seen first in the OCP Diablo 400 rack, with the assumption that subsequent Oberon/Taycann racks will continue to increase power levels. High-voltage DC distribution can reduce current and cable losses at the same power level, while the chip side still requires high-efficiency step-down conversion; system architecture, conversion efficiency, solid-state transformers, and rack power distribution design will jointly determine demand for next-generation GaN and SiC devices. The Jefferies strategist team stated that Navitas Semiconductor has high business sensitivity to 800V adoption, and side-mounted power shelves may become a transitional path before full rack-level 800V deployment. At the same time, Lattice's low-power FPGAs can handle system management, platform security, hardware monitoring, and rack control, acting as continuously running controllers in AI servers, liquid cooling, and power supply systems, and edge AI systems and machine-to-machine intelligent coordination may also expand their demand for local programmable processing.