US Networking & Communications Equipment: Bernstein initiates US networking with a positive AI-infrastructure view

Overview

This sector initiation contends that AI infrastructure is creating a structural upgrade cycle for networking and optics. Bernstein sees bandwidth constraints, accelerating switch speeds, constrained laser supply and data-center expansion as drivers of above-consensus growth, while differentiating sharply among individual stocks.

Core views

Bernstein’s central thesis is that AI infrastructure remains in a higher-for-longer investment cycle. The report notes that demand continues to exceed supply, consensus expects data-center capital expenditure to reach $1.3T in 2027E versus about $259B in 2024, and chiplet-adjusted accelerator volumes are expected to double over 2025-28E. Its medium-term framework assumes a roughly 15% ROI hurdle for hyperscaler AI investment and estimates that end-user AI spending would need to reach about $730B in 2027E, around three times current ARR estimates. Bernstein considers this demanding but plausible given reported rapid AI adoption. Longer term, it identifies continued improvement in model capability as the key leading indicator, while acknowledging that cost deflation and an eventual capacity digestion cycle remain material risks. The report argues that networking has become a structural AI beneficiary because compute performance has improved much faster than bandwidth. It describes a “bandwidth wall”: GPU FLOPs increased 17x from H100 to VR300 while SerDes speeds only doubled, leaving many workloads bandwidth constrained. Networking is estimated to be the third-largest AI data-center spending category after compute and memory/storage. Operators are responding with more networking and memory content, a shift to 1.6T switches and higher specifications. Although like-for-like equipment prices usually decline, the report expects higher-speed mix to dominate deflation, producing a 28% switch ASP increase in 2027 before growth moderates. Bernstein expects switch revenue to outpace accelerator volumes and sees Arista and Celestica gaining share as proprietary Nvidia networking gives way to open Ethernet standards in both scale-out and potentially UALink-based scale-up. Optics are the second major structural pillar. As data rates rise, copper’s viable transmission distance falls, making the transition to optical increasingly unavoidable. Bernstein forecasts transceiver-industry revenue to grow at a 40% CAGR from $57B in 2026 to roughly $300B in 2030. Laser-module supply is expected to remain a critical bottleneck through 2028, allowing Lumentum and Coherent to retain a meaningful share of industry economics as faster switches demand more expensive transceivers. The report believes concerns over co-packaged-optics yields are overstated for these companies because persistent CPO manufacturing difficulties could shift demand toward near-packaged optics rather than eliminate optical demand. It nevertheless flags that wider adoption of CW+SiPh lasers could reduce the InP bottleneck and pricing power. Beyond core AI switching, Bernstein identifies scale-across networking as an additional opportunity created by land and power constraints. When training clusters cannot be built on a single campus, operators connect multiple sites as one cluster. Management teams have indicated an $8-10B opportunity; Bernstein estimates around $8B of new-build spending and about $9B of retrofit potential, while its modeled global scale-across TAM is $8.9B annually. It stresses that this is mainly a training use case rather than an inference or last-mile latency opportunity. Traditional enterprise networking is also recovering cyclically, with spending up 16% from 2024 to 2025 versus a 7% CAGR over 2020-25, while stagnant service-provider spending is becoming a smaller drag. These inputs support Bernstein’s 18% industry-growth forecast for 2026-28, comprising 33% cloud and AI growth, 9% enterprise growth and 4% service-provider growth. The firm’s preferred stock is Celestica. It sees approximately $25B of possible incremental revenue from Google TPU, OpenAI Jalapeno, AMD Helios and 1.6T-switching programs, including roughly $8B from TPU, $11.5B-$12B from Jalapeno and $5B-$6B from Helios and broader high-performance switching. After execution haircuts, Bernstein models 2027 Connectivity & Cloud Solutions revenue up about $20B to $36B and company revenue of $40.0B with $23.79 EPS, versus consensus at $35.5B and $19.69. It rates Celestica Outperform with a $520 target, based on about 22x its $24 FY2027 EPS estimate, and identifies the October 27 analyst day as a high-conviction catalyst. Key risks are execution on new rack designs, TPUv8 ramp risk, Broadcom Tomahawk 6 availability, stronger ODM/EMS competition and weaker AI infrastructure investment. Arista is rated Outperform with a $250 target, based on 35x Bernstein’s $7.14 2028 EPS estimate. The report forecasts 2027 revenue of $18.6B versus $16.3B consensus, supported by its data-center-heavy mix and expected share gains as InfiniBand and NVLink are displaced by open Ethernet ecosystems. Bernstein estimates Arista’s AI back-end switching share rose from 3% in 2024 to 6% in 2025 and could reach about 9% in 2026E. Its valuation is described as full, but consistent with its history and offset by earnings upside; risks include Tomahawk supply limitations, white-box competition and an AI-capex slowdown. Bernstein rates Lumentum Outperform with a $1,220 target and Coherent Outperform with a $350 target. Both benefit from high-speed transceiver pricing and operating leverage, but Lumentum is preferred because it is closer to an optical-communications pure play, has greater laser-module exposure and should generate stronger incremental margins. Lumentum’s target applies 35x to Bernstein’s $35 FY2028 EPS estimate; Coherent’s target applies about 22x to its $16 FY2028 EPS estimate. Coherent’s vertical integration offers some geopolitical resilience but lowers incremental-margin potential; its industrial segment represents about 26% of revenue. A Chinese-transceiver import ban is considered unlikely by Bernstein because Chinese retaliation, potentially restricting indium exports, could damage the entire optical supply chain. Ciena is rated Outperform with a $440 target, based on 16x Bernstein’s $28 2028 EPS estimate. It is viewed as the scale-across leader with an estimated 70% share, but near-term revenue is constrained by ceramic packaging, pump lasers and memory. Bernstein expects supply conditions to improve beyond 2027 and models 2029 revenue of $14B versus $13.6B consensus. Its longer-duration appeal rests on an undemanding valuation and the possibility that scale-across demand exceeds current expectations. The principal risk is that easing land and power constraints reduce the need for multi-site training clusters, while execution failures by Ciena or suppliers could keep growth capped. Cisco and Corning are both rated Market-Perform. Cisco’s $110 target is based on about 20x Bernstein’s $5.60 2028 EPS estimate. The report sees tactical near-term upside from potentially conservative guidance, but argues that Cisco’s software pivot has coincided with decelerating security and observability growth—2% and 4% respectively in FY2026—and may have distracted it from the faster-growing AI-hardware opportunity. Corning’s $140 target, based on 25x Bernstein’s $5.65 2028 EPS estimate, recognizes strong AI optical-fiber content growth but judges its broader long-term framework aggressive. Optical Communications is modeled to grow strongly, but Glass Innovations and Automotive face weaker end markets; Bernstein models Glass Innovations at about 3.5% CAGR and Automotive at 4% CAGR. Corning’s capital intensity and full valuation limit the report’s enthusiasm despite AI-fiber upside.

Analysis framework

Bernstein builds an AI-infrastructure demand framework from capex, adoption, ROI and model-scaling indicators; translates this into networking unit economics, accelerator volumes, switch-port demand and optical supply-demand forecasts; then assesses company-specific share, program, supply-chain and margin implications. Valuation primarily uses P/E multiples relative to expected growth, checked against historical absolute and relative multiples as well as EV/FCF, EV/EBIT and EV/revenue comparisons.

Methodology notes

• Industry AnalysisSupply-demand framework

  AI infrastructure and optical supply-demand analysis

  The report assesses demand from AI capex, accelerator deployments and networking specifications against constrained laser, packaging, memory and switch-component supply.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  AI compute-to-networking-to-optics transmission

  It traces how accelerator growth and bandwidth bottlenecks increase demand for switches, transceivers, laser modules, optical transport and fiber.

• Valuation methodsP/E and PEG Valuation

  P/E multiples assessed against growth

  Price targets are mainly derived by applying stated P/E multiples to Bernstein EPS estimates, with attention to whether growth supports valuation premiums.

• Quantitative, Factor, and Portfolio TheoryBeta/alpha analysis

  Beta and WACC considerations in DCF

  The report considered DCF but judged recent AI-related beta to distort CAPM cost of equity and WACC, so it emphasized relative multiples instead.

Asset mapping & comparison

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

• Celestica (CLS)

  Top pick and AI rack, ASIC and high-performance-switching beneficiary

  Strengths

  Potential upside from Google TPU, OpenAI Jalapeno, AMD Helios and 1.6T switching; analyst day catalyst; valuation viewed as undemanding.

  Weaknesses

  Large new-program assumptions require significant execution.

  Comparison

  Bernstein expects differentiated growth relative to consensus and sees complex-program experience as difficult for ODM/EMS competitors to replicate.

  Risks

  New-rack manufacturing delays, TPUv8 execution, Tomahawk 6 availability, competition and AI-demand deceleration.

• Arista Networks (ANET)

  Structural AI-switching share gainer

  Strengths

  Data-center-heavy mix, expected Ethernet and UALink share gains, and earnings upside.

  Weaknesses

  Valuation is described as somewhat full.

  Comparison

  Expected to gain share as Nvidia proprietary ecosystems give way to open standards.

  Risks

  Switch-silicon supply constraints, white-box competition and AI-capex digestion.

• Lumentum (LITE)

  Preferred optical-module exposure

  Strengths

  Pure optical-communications exposure, laser-module economics and stronger expected operating leverage.

  Weaknesses

  More exposed to the optical cycle and supply-chain/geopolitical disruption.

  Comparison

  Preferred to Coherent for cleaner AI exposure and higher incremental-margin potential.

  Risks

  CW+SiPh substitution, indium-related retaliation risk and delayed optical adoption.

• Coherent (COHR)

  Vertically integrated optical-module beneficiary

  Strengths

  Benefits from optical demand and vertical integration provides some geopolitical resilience.

  Weaknesses

  Industrial segment and vertical integration reduce pure-play exposure and incremental margins.

  Comparison

  Bernstein likes both optical names but prefers Lumentum.

  Risks

  Pricing-power erosion, supply disruption and delayed optical adoption.

• Ciena (CIEN)

  Scale-across and optical-transport beneficiary

  Strengths

  Estimated 70% scale-across share, long-term growth potential and undemanding valuation.

  Weaknesses

  Near-term revenue constrained by ceramic packaging, pump lasers and memory.

  Comparison

  Seen as technology leader; rivals can gain orders when availability is constrained.

  Risks

  Easing land/power constraints, supply-chain execution issues and slower AI build-out.

• Cisco (CSCO)

  Covered networking incumbent

  Strengths

  Potential tactical upside from near-term guidance and some progress in Acacia and Silicon One.

  Weaknesses

  Lags peers in AI hardware share and software pure plays in cybersecurity; valuation viewed as fair.

  Comparison

  Lower AI-hardware exposure may cushion downside in an AI digestion but limits upside relative to preferred names.

  Risks

  Further software deceleration and share losses to cybersecurity pure plays.

• Corning (GLW)

  Optical-fiber beneficiary with diversified end markets

  Strengths

  AI optical content can drive rapid Optical Communications growth.

  Weaknesses

  Smartphone and automotive exposure drags growth; capital intensity limits upside.

  Comparison

  Less AI-concentrated than several peers, but its AI fiber upside is viewed as largely priced in.

  Risks

  AI-adoption deceleration, slower optical adoption and weak smartphone or automotive demand.

Key data

• Industry growth forecast18% growth in 2026-28Includes 33% cloud and AI growth, 9% enterprise growth and 4% service-provider growth.
• AI accelerator volumesDouble from 2025 to 2028EBernstein expects back-end switching revenue to outgrow accelerator volumes.
• Switch ASP growth28% in 2027Driven by the 1.6T speed transition before growth moderates.
• Transceiver industry revenue$57B in 2026 to approximately $300B in 2030Equivalent to a modeled 40% CAGR.
• Scale-across opportunity$8-10B management estimate; about $8B new-build and $9B retrofit opportunityLinked to data-center land and power constraints.
• Celestica 2027 forecast$40.0B revenue and $23.79 EPSVersus consensus of $35.5B revenue and $19.69 EPS.
• Arista 2027 revenue forecast$18.6BVersus $16.3B consensus.

Impact & implications

Bernstein believes the AI build-out is changing networking from a historically lower-growth hardware category into a higher-growth, higher-specification market. It favors companies with direct exposure to switching share gains, optical bottlenecks and scale-across demand, while remaining more cautious where AI exposure is diluted, supply limits near-term delivery, or valuation already reflects the opportunity.

Risks

• AI adoption could decelerate, reducing AI-infrastructure investment and triggering a capacity-digestion cycle.
• Greater use of CW+SiPh lasers could ease the InP bottleneck and weaken Lumentum and Coherent pricing power.
• Chinese retaliation against US transceiver restrictions could constrain indium or InP supply.
• Persistent CPO yield problems could delay optical adoption.
• Easing land and power constraints could reduce the scale-across opportunity for Ciena.
• Component availability, including ceramic packaging, pump lasers, memory and Broadcom Tomahawk supply, could limit revenue realization.
• More intense ODM/EMS or white-box competition could reduce win rates, share and margins.

What to watch

• AI data-center capex revisions in the October earnings cycle.
• Celestica’s October 27 analyst day and 2027 guidance.
• Evidence that model capability and AI adoption continue to scale.
• The 1.6T switch ramp and associated switch and transceiver ASP trends.
• Optical laser capacity and CPO/NPO manufacturing progress.
• Ciena’s ceramic packaging, pump-laser and memory supply availability beyond 2027.
• Meta Muse-related consumer-AI demand signals and possible hyperscaler capex revisions.
• March 2027 OFC and Nvidia GTC announcements.