AI optical interconnect industry: AI optical interconnect demand is projected to expand, while technology transitions and component constraints shape the path.

Overview

This expert-call update examines the AI optical-interconnect chain, covering OCS adoption, competing OCS technologies and costs, 800G/1.6T optical-transceiver demand and pricing, and the development path for CPO, NPO and VCSEL solutions. The report’s central view is that larger AI clusters support expanding optical demand, although technology reliability, component availability and interface bottlenecks will influence adoption timing.

Core views

The expert sees OCS as an increasingly important architecture for large AI clusters because it can improve total cost of ownership, latency and failure rates. Google currently uses OCS for scale-out at the spine layer and for rack-to-rack scale-up connections in a 3D Torus architecture; other cloud service providers are also using OCS for rack-to-rack connections, with potential extension to scale-across networks. Google’s OCS demand was about 12,000-15,000 units in 2025, is estimated at 18,000-20,000 in 2026E, and could reach 35,000-40,000 in 2027F. Other CSPs remain at the sample stage in 2026E but may each require 1,000-2,000 units in 2027F, taking total 2027E demand to roughly 40,000-50,000 units. MEMS-based OCS is expected to hold 80%-85% market share in 2026F before declining to 60%-70% in 2027E as competing technologies develop. Over the longer term, the expert expects MEMS at 40%-50%, liquid crystal at 20%-30%, with piezoelectric beam steering and silicon-photonics OCS accounting for the remainder. MEMS offers short switching times and rapid assembly that may support mass production, but mechanical movement and 220V operation may reduce lifespan. Liquid-crystal OCS uses lower voltage and may have a longer lifespan, though reliability remains unresolved; Coherent is identified as a leading producer of its core liquid-crystal panels. Piezoelectric OCS, led by Huber+Suhner, is expected to make small validation shipments in 4Q26. Silicon-photonics OCS has relatively high insertion losses and requires more time to commercialize. The call also details OCS component economics. A 384-port MEMS OCS requires two collimator arrays, two MEMS chips, mirrors, a calibration system and assembly. Fiber arrays cost about USD2,500-3,000 each, lens arrays USD1,700-1,800, MEMS chips USD17,000-18,000 each, mirrors USD200, calibration USD5,000 and assembly about USD15,000, resulting in a total OCS price of roughly USD100,000-120,000. For a liquid-crystal 384-port OCS, the two-side configuration uses two 192-port modules, two liquid-crystal panels at USD6,000-8,000 each, four collimator arrays at USD3,000 each, and four deflection splitters plus crystal wedges totaling about USD14,000. For optical transceivers, North American 800G shipments in 2026 may reach 45m-50m units, with DR products representing about 70% and FR about 30%. DR is estimated to use silicon photonics and EML solutions in equal proportions, while FR is essentially all EML. The expert estimates 1.6T demand above 30m units in 2026, but actual shipments may be only 10m-15m because of material shortages. A shortage of 3nm DSPs for FR products could make DR about 70% of 1.6T volume in 2026; 1.6T DR is expected to comprise 70% silicon photonics and 30% EML. The report’s 2027 pricing estimates are about USD350 for 800G DR EML, USD300 for 800G DR silicon-photonics products and USD400 for 800G FR. For 1.6T, EML DR is estimated at USD900-950, FR above USD1,000 and silicon-photonics products at USD800-850, with annual price declines below 10%. The transition to 400G-per-lane faces technology constraints: while 400G InP laser chips have been released, SerDes remains limited to 200G and silicon modulators to 200G. The expert therefore sees thin-film lithium niobate as necessary for the modulator upgrade. NPO may begin small-volume shipments in 2027F and reach mass production in 2028F. CPO shipments are expected at 80,000-100,000 units in 2027F, mainly for scale-out networks. VCSELs, which use GaAs and are suited to 30-40m links, may be useful for scale-up networks because they are less temperature-sensitive; however, their use at 850nm or 1060nm would require custom optical fiber because prevailing laser infrastructure is mainly at 1310nm.

Analysis framework

The report summarizes an industry expert call, moving from AI-cluster network architecture and OCS demand forecasts to technology comparisons, component cost structures, transceiver volume and pricing estimates, and emerging CPO/NPO and VCSEL adoption paths. Its reasoning links AI-cluster scaling to interconnect demand while identifying component constraints and technology trade-offs that could affect timing and market share.

Methodology notes

• Industry AnalysisSupply-demand framework

  Expert-based demand, shipment and price outlook for OCS and high-speed optical transceivers.

  The report estimates end-market unit demand and compares it with component availability, using shortages and technology mix to explain potential gaps between demand and actual shipments.

• Competition & strategyValue chain analysis

  OCS bill-of-materials and supplier-role analysis.

  The report breaks an OCS into chips, optical arrays, calibration and assembly, and identifies companies associated with selected technology components or development paths.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Google (GOOGL US)

  Uses OCS for AI-network scale-out and rack-to-rack scale-up; its projected OCS demand is central to the market outlook.

  Strengths

  Current OCS deployment in spine-layer and 3D Torus architecture.

  Comparison

  Other CSPs are still at the sample stage in 2026E.

• NVIDIA (NVDA US)

  May use MEMS- or liquid-crystal-based OCS together with CPO/NPO for future scale-up interconnection.

  Strengths

  Potential adoption across multiple optical-interconnect approaches.

• Coherent (COHR US)

  Identified as a leading producer of liquid-crystal panels used in liquid-crystal-based OCS.

  Strengths

  Capability to produce the core liquid-crystal panel.

  Weaknesses

  Liquid-crystal OCS reliability remains an issue.

  Risks

  Technology adoption depends on resolving liquid-crystal reliability.

• Lumentum (LITE US)

  Developing MEMS-based OCS with Google, according to the expert.

  Strengths

  Participation in MEMS-based OCS development.

  Comparison

  MEMS is expected to lead OCS share near term but lose share over time to alternatives.

  Risks

  MEMS lifespan may be lower because of mechanical movement and high-voltage operation.

Key data

• Google OCS demand, 202512,000-15,000 unitsExpert estimate.
• Google OCS demand, 2026E18,000-20,000 unitsExpert estimate.
• Total OCS demand, 2027E40,000-50,000 unitsIncludes Google at about 35,000-40,000 units and other major CSPs at about 1,000-2,000 units each.
• MEMS OCS share, 2026F80%-85%Expected to decline to 60%-70% in 2027E.
• North American 800G shipments, 202645m-50m unitsDR is estimated at about 70% and FR at about 30%.
• 1.6T transceiver demand, 2026Over 30m unitsActual shipments may be 10m-15m because of material shortages.
• 2027 CPO shipments80,000-100,000 unitsExpected mainly in scale-out networks.

Impact & implications

The report links the expansion of AI clusters to greater demand for OCS and high-speed optical transceivers. Near-term adoption favors MEMS OCS, but its durability trade-off and the development of liquid-crystal, piezoelectric and silicon-photonics alternatives could change the technology mix. Transceiver growth may be constrained by materials and 3nm DSP supply, while CPO and NPO remain earlier-stage adoption opportunities.

Risks

• Material shortages may restrict 1.6T transceiver shipments to 10m-15m units in 2026 despite demand estimated above 30m units.
• A shortage of 3nm DSPs for FR products is expected to influence the 1.6T DR/FR mix.
• MEMS OCS may have a shorter lifespan because of mechanical movement and 220V operation.
• Liquid-crystal OCS reliability remains a development issue.
• Silicon-photonics OCS has relatively high insertion losses and requires further time for commercialization.
• SerDes and silicon-modulator limitations remain bottlenecks for 400G-per-lane upgrades.
• VCSEL deployment would require custom optical fiber for 850nm or 1060nm wavelengths.