The explosive first-quarter results of Zhongji Xuchuang highlight a rising new trend within AI infrastructure construction: optical interconnection networks. A recent Goldman Sachs report significantly raised its target price for the company from 791 yuan to 1187 yuan and presented a key projection: from now until 2027-2028, the total addressable market (TAM) for optical networks is expected to expand from approximately $15 billion to around $154 billion, representing a nine-fold increase.
The Q1 financial report released by Zhongji Xuchuang on Thursday revealed that revenue nearly tripled year-over-year, net profit surged by 262%, and its stock price rose 4%, again reaching a record high.
Coinciding with the earnings release, Goldman Sachs raised its target price by about 50%, maintaining a Buy rating. The report explicitly expressed optimism for the optical communication network sector, with the core rationale being the evolution of data center architecture from horizontal scaling to vertical scaling. This shift drives demand for higher bandwidth and more connections, significantly expanding the overall serviceable market.
According to analysis, Goldman Sachs Asia Pacific technology analyst Allen Chang systematically outlined the full-chain opportunities in the optical interconnect field, ranging from optical modules and CPO to silicon photonics and optical circuit switching. The report's central message is that optical interconnection is no longer just supplementary infrastructure for computing power expansion but is becoming an independent, quantifiable investment theme.
The market's previous consensus on optical module demand primarily focused on scale-out—interconnections between server racks and across data centers. The report identifies scale-up—high-speed optical interconnects within server racks and super-nodes—as the true driver pushing the TAM projection to the $154 billion level.
Allen Chang provided specific comparisons: from the currently mass-produced GB300 NVL72 to the projected Rubin Ultra NVL576 expected to ship in 2027-2028, the networking dollar value per compute unit is forecast to jump from $315,000 to $9.4 billion, a 29-fold increase. This leap is underpinned by the expansion of Nvidia GPU cluster sizes from 72 to 576 GPUs, with interconnectivity extending from within racks to between racks, pushing the boundary for replacing copper cables with fiber optics to even shorter distances.
Using the Rubin Ultra NVL576 as an example—a super-node comprising 8 racks and 576 GPUs—a second layer of high-speed optical connection is required between racks, entirely utilizing CPO solutions. A single compute unit would require 324 optical engines, 162 external laser sources, and 5,184 optical fibers with MPO connectors. The material cost for the scale-up portion alone reaches approximately $800 million. Within the total $154 billion TAM, about 69% (roughly $106 billion) is attributed to scale-up. Under a 29% penetration rate assumption for CPO, it contributes about $91 billion, accounting for approximately 59% of the total TAM.
Addressing market concerns about different connection solutions substituting for or cannibalizing each other, the report suggests that scale-up and scale-out directions are expanding simultaneously, with various technologies being additive rather than engaged in a zero-sum competition.
CPO's technical advantages are clear: moving the optical engine closer to the chip reduces electrical signal transmission paths from centimeters to millimeters, resulting in lower latency, reduced power consumption, and the elimination of DSPs and retimers. Nvidia plans to begin mass production of CPO switches in early 2026, while Broadcom already delivered its 102.4T Davisson CPO switch to customers in October 2025.
However, CPO has a structural drawback: the optical engine is highly integrated with the switch ASIC. In case of failure, both might need replacement, leading to significantly higher maintenance costs and downtime risks compared to traditional pluggable optical modules. This characteristic dictates that the two technologies are not substitutes but serve different scenarios based on needs—CPO for scenarios extremely sensitive to bandwidth and power, and pluggable solutions where operational flexibility is paramount.
By 2028, CPO penetration in scale-out is estimated at around 29%, yet the absolute volume of pluggable optical modules continues to grow. This is because total GPU computing power is persistently expanding, driving ever-increasing demand for external interconnects. From GB300 to Rubin Ultra NVL576, the value market for pluggable optical modules in scale-out is still projected to expand tenfold.
Simultaneously, Silicon Photonics (SiPh) is rapidly replacing traditional EML solutions. For 800G specifications, SiPh solutions have a 26% lower Bill of Materials cost and a 15% lower selling price than EML, yet achieve a 9 percentage point higher gross margin (37% vs. 28%). For 1.6T specifications, the BOM cost advantage widens to 32%, with a gross margin difference of 7 percentage points (57% vs. 50%). SiPh penetration in the datacom optical module market is expected to rise from 6% in Q1 2024 to approximately 46% by Q4 2028. As the product mix upgrades, the overall gross margin for optical module suppliers is forecast to increase to a range of 48% to 55%.
Laser source supply is one of the most evident bottlenecks in the current supply chain. The core raw material for both CW lasers and EMLs is Indium Phosphide (InP) wafers, and geopolitical risks add unquantifiable variables to the global supply chain.
Demand-side pressure comes from three simultaneous directions: continued volume growth of AI servers, speed specification upgrades from 800G to 1.6T and 3.2T, and new laser demand driven by CPO optical engines. On the supply side, capacity expansion plans are underway—Lumentum plans a 40% capacity increase between Q3 2025 and Q2 2026, Coherent has committed to doubling capacity, domestic players like Source Photonics and Eoptolink continue to expand MOCVD production lines, and VPEC plans to increase its InP MOCVD tools from 60 to 64 in the second half of 2026.
However, capacity building takes time. Tight supply conditions for laser sources are expected to persist until 2027, with a potential balance achieved only in the second half of 2028—contingent on CPO penetration progressing as expected, the shift from AI training to inference not accelerating specification upgrades further, and no further tightening of InP export controls. All three of these preconditions involve uncertainty.
Optical Circuit Switching (OCS) remains in early stages but is already seeing substantial order backlogs. Its core advantage lies in eliminating optical-electrical conversion; signals at 800G, 1.6T, and 3.2T can pass through the same OCS without fundamental differences, allowing AI clusters to upgrade without replacing switches—a feature with tangible commercial value in rapidly evolving compute infrastructure.
A real-world example validates this logic: Google's TPU v7 super-node interconnects 9,216 chips using an OCS solution. Commercially, Lumentum disclosed in February 2026 that its OCS backlog exceeded $400 million; Coherent stated that over 10 customers are advancing OCS deployments, covering 64x64 and 320x320 systems; and Innolight plans to launch a silicon photonics-based OCS product in 2027.
Current mainstream OCS technology uses MEMS solutions, with ASPs ranging from $10,000 to $200,000, higher than the $10,000 to $100,000 range for traditional switches. The ultimate penetration of OCS will depend on how much premium customers are willing to pay for "cross-generation compatibility" and whether silicon photonics can reduce insertion loss and costs to acceptable levels—the latter remains a key challenge for OCS technology currently.
Benefiting from the volume growth of AI servers, continuous specification upgrades, and the expansion of optical connectivity applications, growth momentum across multiple segments of the optical communication industry chain is expected to continue through 2028.