AI Rack Upgrades Enter a "System Efficiency Battle"! With the OCP Summit Approaching, Jefferies Decodes 1.6T Scaling, Memory Expansion, and High-Voltage Power Supply

According to Zhihui Finance APP, as the US stock earnings season approaches, the Nasdaq's new high and Marvell's management raising its performance growth target have provided positive catalysts for upward revisions in the global AI optical interconnect industry chain's profit expectations. On October 6th, the Nasdaq Composite Index rose by 0.45% to a new closing high of 27,599.79; ASIC and AI data center optical interconnect chip leader Marvell (MRVL.US) raised its FY2028 revenue target from $18 billion to $20 billion, a roughly 11.1% increase, with its stock price rising 5.8% that day. These moves reflect the market's optimistic expectations for the expanding demand for AI data center accelerated computing chips, CPUs, HBM/DRAM/NAND storage chips, and data center optical interconnect components.

Although the Nasdaq and the Philadelphia Semiconductor Index fell by about 0.22% and 1.15% respectively on October 7th, the recent sharp rebound in tech stocks had already made the upcoming US stock earnings season and the AI computing power industry the market's focus.

On October 7th, Wall Street financial institution Jefferies released the research report "Summit Preview: Power Architecture Upgrades, Optical-Copper Interconnect Evolution, and the Battle for Data Center Scale-up Optical Interconnect Leadership," outlining five interlinked investment themes centered on the OCP Global Summit in San Jose from October 12-15: upgrading from 800G to 1.6T optical interconnects, competition for scale-up interconnect leadership among Ethernet, UALink, and NVLink, the impact of AI data center Near Package Optics (NPO) and Co-Packaged Optics (CPO) on device value distribution, inference workload driving expansion of memory architectures, and 400V/800V DC power supply reshaping rack power systems.

The Jefferies research report identifies MACOM and Lumentum as the clearest beneficiaries of the 1.6T upgrade, and believes that the opportunities for both copper and optical dual high-speed transmission leader Credo’s ALC and OmniConnect have yet to be fully recognized. The report also highlights Coherent's PhotonLink commercialization, Astera Labs' memory controllers and switching chips, Lattice's low-power controllers, and the growth space for power semiconductors.

From an engineering and investment logic perspective, Jefferies focuses on how these investment opportunities indicate a major shift: With the continued and accelerating expansion of AI computing clusters, the speed of internal data transfer in AI data centers, memory configuration, and power conversion capabilities are becoming key constraints for sustained GPU performance and the conversion of capital expenditures into revenue. Relevant suppliers may therefore gain higher per-rack value and profit contributions.

The upgrade of AI racks is entering a "data transfer efficiency battle": As cutting-edge AI agents such as Muse and Astra sweep the globe and AI compute clusters continue to expand, the bandwidth, latency, and power consumption of high-speed interconnects are increasingly affecting GPU utilization and overall compute output. Ahead of the OCP Summit, Jefferies identified the 800G–1.6T upgrade, NPO/CPO architecture evolution, and the competition for scale-up interconnects as key areas of focus, explicitly naming MACOM and Lumentum as its top 1.6T beneficiaries, emphasizing the former's opportunities in optical detectors, transimpedance amplifiers, and drivers, and the latter’s position in high-speed lasers and external optical sources. The main investment insight is that the expansion of AI infrastructure is boosting the value of critical optoelectronic devices, and suppliers who can translate technical advantages into customer adoption, large-scale delivery, and profit growth are likely to be major winners in the optical interconnect upgrade.

Optical Interconnects Move to the Investment Forefront: 1.6T Ramp-Up and the Battle for "Interconnect Leadership"

The investment value of the 1.6T upgrade is reflected in the synchronized changes in demand for optoelectronic devices, product structure, and supplier market share. The Jefferies equity strategy team lists MACOM and Lumentum as their preferred 1.6T beneficiaries, based on their key roles in the receiver and transmitter domains, respectively. For MACOM, the focus is on the competitive position and revenue contribution of optical detectors in 800G and 1.6T deployments, as well as the demand for more channels of transimpedance amplifiers (TIAs) and drivers after module vendors adopt NPO designs. Semtech’s data center event on October 15th may provide new industry benchmarks for MACOM by revealing the market size for TIAs and drivers.

For Lumentum, the most noteworthy aspects are its 1.6T transmitter product share, pricing, capacity, and margins, as well as the demand structure for electro-absorption modulated lasers (EML) and continuous wave lasers (CW). Jefferies adds that industry leader Broadcom, with larger ASIC and optical chip scale than Marvell, is at the center of competition for optical module DSP chip market share, and early channel signals are favorable to it. Jefferies’ strategists thus conclude that—while data center high-speed optical interconnect upgrades can expand industry demand, profitability also depends on device share, manufacturing capacity, and product portfolio; MACOM and Lumentum have the most exposure to this upgrade, in Jefferies’ view.

MACOM focuses on optoelectronic devices and high-speed analog chips for AI optical interconnects, including optical detectors converting optical signals to electrical current, TIAs amplifying weak photocurrents into voltage signals, and high-speed driver chips for lasers or modulators, providing the key transceiver components for 800G/1.6T optical modules. Jefferies is particularly interested in its optical detector market share and multi-channel TIA and driver opportunities brought by NPO architectures. Lumentum, meanwhile, concentrates on high-performance light sources, optical transceiver modules, and optical circuit switches, including EMLs, CW lasers, external laser sources (ELS) for NPO/CPO, and Optical Circuit Switch (OCS) systems enabling dynamic optical path reconfiguration for networks. Its investment highlights are the volume ramp-up of 1.6T products, upgrades in laser product structure, and demand growth from new optical interconnect architectures.

In terms of NPO and CPO, Jefferies notes that NPO and CPO drive optical components closer to compute or switch chips, and value distribution will shift along lasers, packaging, and complete modules. From an engineering perspective, shortening high-speed electrical signal transmission distances helps mitigate signal integrity and power consumption issues of high-bandwidth links; the adoption of NPO or CPO architectures affects demand for external laser sources, drivers, optical fiber components, and package integration. At the summit, Lumentum will discuss external laser sources for CPO, wide-parallel VCSEL interconnects, and, together with Oracle, introduce automated OCS networks. The key here is whether customer interest can translate into deployments that meet practical operational requirements.

Coherent's PhotonLink, launched in September, is already in talks with more than 10 customers in both CPO and NPO directions, with core customers and long-term agreements secured in both. Revenue growth is expected to kick in from Q4 2026. Spanning lasers, VCSEL arrays, silicon photonics, optical fiber components, and detectors, whether customers buy individual components or complete modules will directly affect revenue and profit per deployment; customer qualification, manufacturing yield, and ultra-high-power platform progress are also vital. Meanwhile, Jefferies monitors whether ams OSRAM’s micro-VCSEL may be adopted in Nvidia’s future Scale-up architecture, including the potential NVL1152 Feynman opportunity, which could change previous assumptions about the value of indium phosphide lasers.

Scale-up competition determines how accelerators are integrated into larger synergistic computing systems, as well as the market shares for switching chips, copper, and optical interconnects. Jefferies focuses on the competition among Ethernet/SUE, UALink, and NVLink: Nvidia is expanding ecosystem participation via NVLink Fusion and may reveal NPO plans for NVL576 Scale-up; AMD launched Helios in July and plans to deliver to customers like Microsoft by the end of 2026. Investment attention will turn to deployment execution, UALink adoption progress, and participation opportunities for Astera Labs and Broadcom. Marvell announced 100G-per-channel UALink switching chips expected to tape out within the year, and AMD’s next-generation racks plan to combine optical and copper connections for cross-rack Scale-up.

Jefferies’ strategists indicate that Credo provides another noteworthy technical path: it will debut the ALC, integrating its own interconnect platform with acquired Hyperlume’s microLED technology, aiming for active cable (AEC)-like power consumption and reliability at distances up to ~30 meters. The report suggests that microLED is more likely for Scale-out and even Scale-in expansion, whereas micro-VCSEL could be a path for Scale-up. Additionally, whether Nvidia will certify Credo’s ZF Optics to solve repeated link drop issues in new cloud service provider clusters based on Nvidia platforms is a key observation point. Jefferies emphasizes that bandwidth, latency, and stability directly affect the synergistic efficiency of expensive accelerators, and vendors who improve these metrics could realize value beyond mere port count growth.

Memory and Power Move to the Interconnect Forefront: The Second Strong Profit Curve of AI Rack Upgrades

Explosive growth in AI inference workloads is opening up new opportunities for memory controllers, switching chips, and hierarchical memory architectures. Jefferies points out that Marvell has emphasized CXL and memory controllers as important accompanying opportunities as XPU deployments grow, with Astera Labs’ Leo product line providing practical approaches in this direction: Leo 2 E supports CPU memory expansion via CXL 3.2 and PCIe 6, while Leo 2 P supports cross-host memory pooling and sharing. Both support DDR4 and DDR5, allowing customers to reuse existing memory and improve resource utilization without adding CPU sockets. The new product line is currently being sampled to hyperscale cloud customers.

The Leo X, for GPUs, connects to dedicated memory layers via PCIe and platform-specific protocols, forming another product path distinct from CPU-side CXL expansion and, depending on customer configurations, Astera Labs may realize revenue from both memory controllers and switching chips. The report also tracks Amazon and AMD's next-gen racks using UALink, and whether Amazon will split supply between Astera Labs and Marvell, reiterating that Credo’s OmniConnect opportunity is underestimated. From the underlying logic of inference systems, Jefferies points out that longer context and higher concurrency will increase the demand for cache and working set capacity; effectively combining high-bandwidth and high-capacity memory pools helps fully utilize expensive computing resources, making memory interconnect and management chips a more critical part of AI capital expenditures.

High-density AI rack high-speed interconnects are simultaneously increasing the value of power conversion and system control, but the order in which 400V/800V systems roll out determines the specific revenue recognition pace. That is to say, optical interconnects mainly address bandwidth, latency, and energy consumption for data transfer; high-voltage DC power reduces current and line losses in high-power distribution, and supports chips under high-speed transmission scenarios through efficient power conversion, jointly underpinning rapid AI cluster expansion; "deployment order determines revenue timing," mainly means deployment timing will affect when power semiconductor and relevant controller suppliers see incremental revenue.

In Jefferies strategists’ view, Nvidia’s October 12th keynote and technical discussions about 800V DC, grid-connectivity, liquid cooling, and MGX will further clarify opportunities for Texas Instruments, Analog Devices, and onsemi. The report notes that delays in Kyber rack and 800V transition timelines have driven related stocks down. It's expected that OCP Diablo 400 racks will lead with the 400V solution, with subsequent Oberon/Taycann racks increasing power levels further. High-voltage DC distribution can reduce current and cable loss for the same power, but chips still require highly efficient step-down conversion; system architecture, conversion efficiency, solid-state transformers, and rack power distribution design will jointly determine the demand for next-generation GaN and SiC devices.

Jefferies strategists state that Navitas Semiconductor is highly sensitive to 800V adoption, and auxiliary power cabinets could serve as transition paths before comprehensive rack-level 800V deployment. At the same time, Lattice’s low-power FPGA can perform system management, platform security, hardware monitoring, and rack control functions, acting as persistent controllers for AI servers, liquid cooling, and power systems. Edge AI systems and machine-to-machine intelligent collaboration may also broaden its local programmable processing demand.