Goldman Sachs: The future of AI must have light!

Innolight's explosive performance in the first quarter reflects a rising new trend in AI infrastructure construction—optical interconnect networks. Goldman Sachs' latest report significantly raised its target price from 791 yuan to 1187 yuan, and made this judgment: from now to 2027-2028, the addressable market for optical networks will expand from about $15 billion to about $154 billion, an increase of about ninefold.

Innolight's quarterly financial report released on Thursday showed that revenue grew nearly twofold year-on-year, net profit surged by 262%, share price jumped 4%, hitting another record high.

Goldman raised its target price by about 50% at the time of the results, maintained its buy rating, and made it clear in the report: it is optimistic about the optical communication network sector. The core logic lies in the evolution of data center architecture from horizontal to vertical, bringing higher bandwidth and more connection demands, significantly driving the expansion of the overall addressable market.

According to Chasing Wind Trading Desk, Goldman Sachs Asia Pacific tech analyst Allen Chang systematically reviewed the full-chain opportunities in the optical interconnect track—from optical modules, CPO, silicon photonics to optical circuit switching. The report’s core message is: optical interconnect is no longer a supporting facility for computing power expansion, but is becoming an independent, quantifiable investment main line.

TAM from $15 Billion to $154 Billion, Scale Up is the Real Driver

The previous market consensus regarding optical module demand mainly focused on the scale out direction—external interconnection between racks and data centers. The real driver behind the report's TAM estimate reaching $154 billion comes from scale up, that is, high-speed optical interconnects within racks and ultra-nodes.

Allen Chang provided specific specification comparisons: from the currently mass-produced GB300 NVL72 to the expected 2027-2028 rollout of Rubin Ultra NVL576, the dollar value of network interconnection per compute unit will jump from $315,000 to $9.4 million, a 29-fold increase. Behind this leap is the expansion of Nvidia GPU cluster scale from 72 to 576, with interconnection levels spreading from within racks to between racks, and the boundary of optical fiber replacing copper cables pushed to shorter distances.

Take Rubin Ultra NVL576 as an example: this is a super node made of 8 racks and 576 GPUs, needing second-level high-speed optical connections between racks, fully using CPO solutions. A single compute unit requires 324 optical engines, 162 external laser sources, 5184 optical fibers and MPO connectors, with the scale up material cost alone reaching about $800 million. Of the $154 billion total TAM, about 69% ($106 billion) is scale up, with CPO contributing about $91 billion at 29% penetration rate, accounting for about 59% of the total.

Addressing market concerns about different connection solutions competing and eating each other's market, the report believes that both scale up and scale out are expanding simultaneously, and different technologies are additive rather than zero-sum competitors.

CPO Reshapes Value Distribution, Pluggable Optical Module Usage Still Growing

CPO (Co-Packaged Optics) has clear technical advantages: moving the optical engine next to the chip, reducing the electrical signal transmission path from centimeters to millimeters, lowering latency and power consumption, and removing the need for DSP and retimers. Nvidia plans to start mass producing CPO switches in early 2026; Broadcom will deliver 102.4T Davisson CPO switches to customers in October 2025.

However, CPO has a structural defect: optical engines are closely tied to the switch ASIC, so if a failure occurs, it may require replacing the ASIC along with the optical engine. Maintenance costs and downtime risks are significantly higher than traditional pluggable optical modules. This feature means the two technologies do not substitute but complement each other based on use case—scenarios highly sensitive to bandwidth and power use CPO, while those needing flexible operations continue to use pluggable modules.

By 2028, CPO penetration rate in scale out is about 29%, but the absolute volume of pluggable optical modules is still growing. The reason is overall GPU computing power continues to expand, and demand for external interconnection only increases. From GB300 to Rubin Ultra NVL576, the scale out market value for pluggable optical modules is expected to expand tenfold.

Meanwhile, silicon photonics (SiPh) is rapidly replacing traditional EML (Electro-absorption Modulated Laser) solutions. At 800G specs, SiPh’s BOM (Bill of Materials) cost is 26% lower than EML, sales price is 15% lower, but gross margin is 9 percentage points higher (37% vs 28%); at 1.6T, BOM cost advantage rises to 32%, gross margin difference is 7 points (57% vs 50%). SiPh penetration in the data communication optical module market will rise from 6% in Q1 2024 to about 46% in Q4 2028. With product structure upgrades, optical module suppliers’ overall gross margin is expected to increase to the 48%-55% range.

InP Supply Tight, Geopolitical Risks Add Uncertainty

Laser supply is currently the most clear bottleneck in the industry chain. Both CW lasers and EML’s core raw material is InP (Indium Phosphide) substrate, and geopolitical risks add hard-to-quantify variables to the global supply chain.

Demand side pressure comes from three directions: sustained ramping of AI servers, speed upgrades from 800G to 1.6T and 3.2T, and additional laser demand brought by CPO optical engines. Supply side expansion plans are gradually underway—Lumentum plans to expand production by 40% between Q3 2025 and Q2 2026, Coherent promises to double capacity, domestic firms Yuanjie Tech and Yijing Tech continue to expand MOCVD lines, VPEC plans to increase InP MOCVD machines from 60 to 64 in the second half of 2026.

But capacity building takes time, tight laser supply is expected to continue until 2027, with possible balance in late 2028—the premise being CPO penetration progresses as expected, AI shifts from training to inference doesn’t further accelerate specification upgrades, and InP export controls don’t further tighten. All three premises are uncertain.

OCS: Cross-generation Compatibility Value Proven, Large-scale Deployment Still Watching

Optical Circuit Switching (OCS) is still at an early stage, but substantial backlog orders have emerged. Its core advantage is no need for optical-electric conversion; 800G, 1.6T, and 3.2T signals can pass through the same OCS without essential difference, so during fast iteration of AI clusters, there’s no need to replace the switch—this characteristic has practical commercial value in rapidly evolving computing infrastructure.

Real-world cases confirm this logic: Google TPU v7 super node interconnected 9,216 chips, using OCS solutions. Commercially, Lumentum disclosed in Feb 2026 that OCS backlog orders exceed $400 million; Coherent said over 10 customers are deploying OCS, covering 64×64 and 320×320 systems; Innolight plans to launch SiPh-based OCS products in 2027.

The current mainstream OCS uses MEMS solution, ASP ranges from $10,000 to $200,000, higher than traditional switches ($10,000 to $100,000). The ultimate scale of OCS penetration will depend on how much premium customers are willing to pay for "cross-generation compatibility", and whether silicon photonics can reduce insertion loss and cost to acceptable levels—the latter is still the main shortcoming of OCS.

Benefiting from the ramp-up of AI servers, continuous specification upgrades, and expanding use cases for optical connectivity, the growth trends for several links in the optical communications industry chain are expected to persist through 2028.

~~~~~~~~~~~~~~~~~~~~~~~~

The above exciting content comes from Chasing Wind Trading Desk.

For more detailed interpretations, including real-time analysis and frontline research, please join [Chasing Wind Trading Desk · Annual Membership]

Risk Disclosure and DisclaimerThe market contains risks, and investment should be cautious. This article does not constitute personal investment advice, nor does it consider individual users' unique investment goals, financial situations, or needs. Users should consider whether any opinions, views, or conclusions in this article fit their specific situation. Investments based on this content are at your own risk.