Wall Street giant Goldman Sachs has released its latest research report, "Optical Industry Survey: Ten Key Points on Technology, Demand, Supply, and Competition," revealing that tech leaders' unprecedented trillion-dollar-scale investments in AI computing infrastructure are simultaneously expanding the volume, speed, and scope of high-speed optical connections in data centers. Growth opportunities are extending from strong demand for high-speed optical module systems to lasers, fiber arrays, optical industry testing, and high-performance automated coupling equipment.
Goldman Sachs had previously raised its 2027 and 2028 demand forecasts for 800G and above optical modules by 39% and 36%, respectively, to 144 million and 171 million units. Following this latest round of research on leaders in China's data center optical product lines, analysts believe there is still room for upward revision. Meanwhile, DSP, laser, and PCB supply may continue to constrain 2027 shipments.
From an engineering perspective, the growth in optical interconnect demand stems from more frequent and intensive data exchanges between computing nodes: Mixture of Experts (MoE) model expert parallelism requires distributing and aggregating data across GPUs; Prefill-Decode Disaggregation requires transmitting KV Cache; and complex agent workflows increase the concurrency of model calls, tool execution, and data access. As clusters expand, network congestion causes expensive GPUs to wait for data, directly reducing the effective tasks delivered per dollar and per watt. The distance, loss, and power constraints of high-speed electrical connections are therefore driving broader optical connectivity, with NPO/CPO shortening electrical signal paths, EML, CW, and ELS providing light sources for optical communication, and OCS supporting workload-based connection reconfiguration.
As front-end AI application demand accelerates, the requirements for transmission bandwidth and efficiency in AI cluster collaboration are escalating. Goldman Sachs believes investment value in data center optical interconnect/communication depends on whether companies can secure key materials, complete high-speed product certification, and convert demand into deliverable capacity. The report emphasizes that while CPO commercialization takes time, testing and coupling equipment are already generating revenue, and profit clocks across the supply chain are not synchronized.
Regarding stock ratings under coverage, Goldman Sachs has assigned "Buy" ratings to Zhongji Innolight A/H shares, Robotechnik, LandMark Optoelectronics, and FOCI. The A-share target for Zhongji Innolight is based on a 35.6x P/E for 2027, while the H-share target assumes a 13% H/A premium. LandMark, Robotechnik, FOCI, and ASMPT are valued using more distant earnings discounted back. For Zhongji Innolight A-shares, Goldman Sachs set a target of RMB 2,645, implying a potential upside of 185.6% over the next 12 months; for H-shares, the target is HKD 3,267, implying a 179.2% upside. For Robotechnik, the latest target implies roughly a 26% upside. For Taiwan-based LandMark Optoelectronics (2455.TW), the target implies a 54% upside over the next 12 months.
The expansion of demand and price resilience form the basis of Goldman Sachs' bullish view on high-speed optical modules. Surveyed Chinese companies indicated global 800G and 1.6T module demand ranging from 130 million to 200 million units, with China's overall market expected to exceed 50 million units, dominated by 400G and 800G, while 1.6T is beginning to enter the market. However, these figures involve product scope differences and cannot be directly summed. On pricing, surveyed companies expect 800G module price declines to stay in single digits in 2027, with 1.6T prices maintained above USD 700 per unit, meaning shipment growth still translates into revenue growth.
The product iteration cycle for data communications has shortened to approximately one to two years, significantly faster than the roughly five-year cycle in the telecom market, making R&D, supply chain management, mass production, and cross-regional delivery capabilities key competitive barriers. Goldman Sachs noted that Zhongji Innolight secures supply through joint R&D, supplier investments, prepayments, and long-term agreements, demonstrating how leaders convert procurement capabilities into competitive advantages. Surveyed companies also see robust demand over the next two to five years, with some customers planning through 2030 and beyond. The roughly one-to-two-year payback period cited is an operational judgment by the surveyed companies and cannot be generalized as a guaranteed return for all AI projects.
Goldman Sachs indicated that laser and fiber array upgrades are simultaneously driving supply constraints and domestic supply capability improvements. Dongshan Precision reported that 200G EML lasers are in mass production, but output remains constrained by DSP supply tightness. The company has secured supply of 10 million DSPs for 2027, supporting subsequent laser and optical module shipment growth. Everbright Photonics expects 2027 EML capacity to be slightly higher than CW lasers, with EML still dominated by 100G and supplemented by 200G. LandMark Optoelectronics (VPEC) sees improved InP substrate supply and plans to increase MOCVD equipment from 62 units to 69 by Q2 2027, while planning new production bases. Its epitaxial wafer products are also upgrading from serving 70-100mW CW lasers to over 100mW for NPO and 300-400mW products for CPO, while expanding 100G and 200G EML.
Additionally, FOCI's FAU is upgrading from 40 channels serving 3.2T to 80 channels for 6.4T, with 100 channels still under development. August revenue grew 20% month-over-month, higher than July's 9%. These changes support a higher share of premium products, but cannot confirm that heat dissipation, yield, and certification issues for high-power lasers are resolved.
The early benefits of CPO are emerging in testing and automated equipment. Goldman Sachs forecasts CPO switch shipments of 10,000, 92,000, and 131,000 units for 2026-2028, while observing strong demand for 3.2T NPO engines from Chinese and overseas customers. Robotechnik's Q2 2026 revenue grew 172% quarter-over-quarter, 141% higher than Goldman Sachs forecast. The company and its subsidiary FiconTEC expect equipment shipments over the next year to exceed half of the cumulative shipments over the past 25 years, aiming to reduce wafer-level testing time by 50%-60% and quadruple chip-level testing speed next year.
Goldman Sachs analysts wrote that high-value photonic integrated circuits (PICs) must identify defects early, making double-sided wafer testing, Known Good Die (KGD) screening, and precision coupling critical for reducing scrap costs. Photon Technology's NightJar system identifies optical losses and defects early; Gallant Precision's dual-fiber array active alignment equipment handles both transmit and receive ends simultaneously, cutting coupling time by 50%; Huite's assembly systems cover optical modules, external light sources, OCS, NPO, and CPO applications; ASMPT's MEGA platform integrates high-precision mounting, dispensing, UV curing, and 3D inspection. Goldman Sachs analysts collectively favor a comprehensive upgrade of photonics manufacturing capabilities.
The stronger the AI agents become, the less data center networks and high-speed data transmission can afford congestion! Strong AI computing demand linked to the broader AI supply chain is already reflected in leaders' robust earnings and long-term capacity agreements. Nvidia's fiscal Q2 2027 revenue reached USD 96.2 billion, up 106% year-over-year, with data center revenue at USD 89 billion, up 117%. Anthropic's disclosed capacity arrangements include up to 5GW with Amazon, 5GW with Google and Broadcom, and over 300MW with more than 220,000 Nvidia GPUs via SpaceX. These agreements are at different delivery stages, supporting visibility for continued construction and expansion, but cannot all be counted as already operational capacity.
The market rebound also highlights the strengthening bullish logic for the entire AI computing chain driven by robust AI demand. Korea's KOSPI rebounded about 22% from its July 30 low by August 13, while the Philadelphia Semiconductor Index closed at 12,621 on August 17, up over 20% from its July 29 low, both entering technical bull markets. The recovery of memory heavyweights like Samsung and SK Hynix reinforced AI earnings expectations, though the KOSPI benchmark retreated in September, indicating notable expectation volatility during the AI computing rebound.
From a technical perspective, the strong optical interconnect growth in the AI inference era comes from "more concurrent workloads plus broader cross-node collaboration." OpenAI's recently launched frontier model Astra enhances capabilities in computer operations, software engineering, and complex workflows, making previously uneconomical tasks executable by agents. OpenAI's product head stating that demand is so unprecedented the company may suspend new Pro subscriptions is a significant signal of AI computing capacity pressure. Thus, the continued massive expansion of computing demand from frontier models will act as a new growth catalyst for data center optical interconnect product lines.
In MoE expert parallel systems, data must be distributed and aggregated across accelerators; Prefill-Decode Disaggregation requires KV Cache transmission; agent tool calls, database access, and shared storage increase system-level data traffic. When networks slow data exchange, GPUs cannot deliver effective work despite high peak computing power. The bandwidth, distance, and energy efficiency advantages of optical connectivity, plus NPO/CPO's ability to shorten high-speed electrical signal paths, therefore have strong and clear economic value. Demand growth does not require each task to consume more tokens: even if individual task efficiency improves, as long as new tasks and concurrency scale expand faster, total AI infrastructure and high-performance network equipment demand can still grow. However, port count, per-port speed, and optical adoption rates are the most direct variables for measuring optical module demand.
The Navier-Stokes research provides a concrete scale reference: OpenAI stated its stronger internal model organized approximately 10,000 concurrent agents, forming solutions in about 88 hours, after which Astra completed Lean formalization and verification in about 17 hours. All research attempts generated approximately 300 billion output tokens, with the Navier-Stokes portion around 130 billion, demonstrating the scale achievable in research-grade inference. Furthermore, Recursive Self-Improvement (RSI) opens additional demand space for research computing: AI participation in code writing, experiment design, data generation, evaluation, and training tool optimization may increase continuously running experiments and inference workloads. OpenAI has explicitly advanced research toward RSI.
OpenAI's GPT-6 Astra model and the RSI technical path focused on by AI leaders are expected to become two core drivers of exponential AI computing demand expansion. More powerful AI models, broader AI application tool usage, and next-generation AI training paths requiring stronger computing power are strengthening the case for sustained growth in AI infrastructure demand. Combined with Goldman Sachs' latest survey, the new growth curve has both technical and commercial basis. For investment, tracking high-speed module actual delivery, laser yields, testing equipment acceptance, and per-share cash flow is key: only companies that convert complexity and supply constraints into customer value can sustainably realize the profit premium of AI optical interconnect.
Comments