China NPO and AI SuperPod supply chain Report Interpretation

Overview

This China AI-hardware deep dive examines how near-packaged optics (NPO), domestic AI accelerators and superpod-scale systems could reshape China’s data-center supply chain. Bernstein’s central view is that local chip capacity and sustained cloud spending can revive AI-server demand from 2H26 while creating new content opportunities across networking, integration and advanced components.

Core views

Bernstein frames NPO as an emerging China AI-networking opportunity rather than a mature, standardized product market. China has the required elements—PIC designers, optical-engine packagers, fiber-array-unit suppliers, lasers, PCBs and system builders—but its ecosystem is more fragmented than TSMC’s integrated CoWoS-related optical platform. Domestic demonstrations generally use edge-coupled, Mach-Zehnder-modulator optical engines with external 100/200mW continuous-wave lasers and target roughly 3.2T bandwidth. Unlike pluggable transceivers, NPO must be tailored to each GPU, NPU or switch ASIC: electrical and SerDes interfaces, optical engine, fiber routing, sockets, PCB channels, and thermal and mechanical packaging all require joint design. Bernstein says suppliers are validating and sampling products, with volume production targeted for 2H27. The report highlights both technical advantages and commercialization constraints. Edge coupling can offer lower coupling loss, broader optical bandwidth and better thermal tolerance than grating coupling, but it entails a larger optical engine and may face chip-edge space, fiber-array density and assembly-yield limits as bandwidth rises beyond 3.2T. External laser-source architectures separate lasers from ASIC heat and retain replaceability, while Huawei’s embedded-laser approach can simplify routing and integration but raises thermal, reliability and repairability challenges. Bernstein identifies system integration, socket reliability, PCB routing and uneven port-level electrical performance as key hurdles; domestic ASIC performance must become more uniform as optical, electrical and packaging elements are jointly tuned. Huawei’s NPO deployments provide the clearest domestic superpod example. HiSilicon has demonstrated a 3.2T VCSEL optical engine for short reach and a 7.2T silicon-photonics NPO engine for the Atlas 960E SuperPoD. The system connects up to 4,096 NPUs, delivers 8 EFLOPS at FP8 and supports up to 1PB of HBM. Its 5,500 optical engines replace 48,000 800G optical modules; Bernstein cites more than 550kW of power savings, while a Huawei comparison indicates a two-thirds reduction in power use and a 90% lower failure rate against 800G transceivers. This architecture raises demand for advanced logic, with SMIC described as the core foundry partner, and expands the addressable market for optical, copper, cooling, power, switching and substrate suppliers. Bernstein expects China’s AI-server market to recover from 2H26 as domestic accelerators and advanced-node foundry capacity ramp. Chinese cloud and internet companies have entered a new AI-led investment cycle: aggregate 2025 capex for Alibaba, Tencent and Baidu rose about 70%, with consensus indicating another 80% increase in 2026. Alibaba indicated AI/cloud-related spending well above RMB380bn for 2025-27; Tencent’s 2026 capex is estimated near US$25bn, and ByteDance’s reported AI infrastructure plans are substantial but explicitly characterized as speculative. Server shipments declined in 2025-26 because tighter US restrictions constrained NVIDIA GPU availability, but Bernstein expects the maturing domestic accelerator ecosystem to reverse that trend. Superpods compensate for weaker single-chip domestic performance by linking more accelerators, improving parallel processing and utilization for trillion-parameter and mixture-of-experts workloads. The report compares Huawei’s domestic architecture with NVIDIA’s. Huawei CloudMatrix 384 integrates 384 accelerators across 12 racks, versus 72 GPUs in NVIDIA GB200 NVL72. Bernstein estimates an Ascend 910C rack with 32 accelerators costs about US$815,000, or roughly US$25,000 per accelerator, compared with more than US$45,000 for an NVIDIA GB200 rack. Huawei’s 17 September Atlas 960E announcement also underscores the scale-up direction, while Alibaba has introduced the Panjiu AL128 and Tencent and ByteDance are advancing deployments. Bernstein concludes that greater rack-scale integration shifts more value to domestic OEM/ODMs, which handle switching, non-chip procurement, power and liquid cooling alongside assembly. Within NPO, Bernstein differentiates supply-chain roles. Innolight is pursuing overseas NPO opportunities and has reportedly sent samples to two overseas customers, with mass products slated for 2H27; management expects only a few million shipments next year before potential expansion to tens of millions by 2028. TFC participates in fiber-array units and external-laser modules through a three-year NVIDIA NPO/CPO development partnership. Accelink and HG Genuine are focused on domestic optical-engine demand, while Lightelligence and Xphor are pursuing domestic PICs. External laser sources create demand for high-reliability CW lasers, precision packaging and polarization-maintaining fiber. Yuanjie’s 70mW and 100mW CW lasers are in mass production, with 120mW, 300mW and 400mW products under development; EugenLight’s UHP ELSFP module exceeds 23dBm, or over 200mW, per channel and targets mass production in 1Q27. For server integration and networking, Bernstein identifies Huaqin, Inspur/IEIT, xFusion and Luxshare as important suppliers. IEIT reported 1H26 revenue of RMB84bn, up 5% year on year, a 4.2-percentage-point gross-margin expansion and RMB3bn net profit, up 270% year on year, which the report attributes to domestic-accelerator platform mix and liquid-cooling scale. Huaqin is described as a leading supplier to Tencent and Alibaba with 30-40% share, a major non-Huawei superpod supplier, and a company targeting more than RMB10bn of superpod revenue in 2026, likely to more than double next year. Bernstein argues its deeper responsibility for design, integration and procurement can support margins despite a buy-sell model. Luxshare’s communication-segment revenue rose about 50% year on year in 1H26 to nearly 10% of group revenue, aided partly by domestic server assembly demand. The report also sees a favorable domestic setting for independent switch vendors because China’s accelerator ecosystem is more diverse than Huawei’s proprietary stack. H3C and Ruijie can benefit from switch trays for intra-pod scale-up, although qualification requires multi-year collaboration with chip vendors, cloud providers and ODMs. In substrates and PCBs, China’s ABF market was about US$4bn in 2025, or 25% of the global market, but remains dominated by overseas suppliers because of high technological and capital barriers. Qunce, AT&S China and Leading together account for more than half of domestic ABF supply. The shift from 800G to 1.6T optical transceivers also supports mSAP HDI PCB adoption, as finer routing requirements move from regular HDI at 400G toward mSAP HDI at 1.6T.

Analysis framework

Bernstein maps the NPO value chain, compares Chinese architectures with NVIDIA and TSMC-related approaches, and links technical design choices to component demand. It then combines cloud-capex trends, domestic-chip supply, server-cost comparisons, company operating data and supply-chain market shares to identify potential beneficiaries.

Methodology notes

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  NPO and AI-superpod supply-chain mapping

  The report traces how superpod and NPO adoption flows into demand for optical engines, lasers, fiber, switches, server integration, substrates, PCBs, cooling and power systems.

• Industry AnalysisSupply-demand framework

  Domestic AI-server recovery driven by local accelerator supply and cloud capex

  Bernstein links strong AI infrastructure spending with a recovery in server shipments as domestic chips and advanced-node foundry capacity expand.

• Other

  Rack bill-of-materials and cost-per-accelerator comparison

  The report compares Huawei Ascend 910C rack costs and system configuration with NVIDIA GB200 to explain the cost rationale for domestic superpods.

Asset mapping & comparison

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

• SMIC (688981.CH)

  Core foundry partner positioned to benefit from higher advanced-logic demand created by Huawei’s NPO superpod architecture.

  Strengths

  Bernstein rates SMIC Outperform with PT of RMB220/HKD120.

  Risks

  Dependence on domestic AI-chip and advanced-node capacity ramp.

• Hygon (688041.CH)

  Domestic AI accelerator supplier positioned for recovering local AI-server demand.

  Strengths

  Bernstein rates Hygon Outperform with PT of RMB450.

  Weaknesses

  Domestic accelerators lag leading overseas chips in single-chip performance.

  Comparison

  Domestic superpod architectures compensate by connecting more accelerators.

  Risks

  Execution of domestic chip supply ramp.

• Cambricon (688256.CH)

  Covered domestic AI-chip beneficiary of China’s localized AI-server ecosystem.

  Strengths

  Bernstein rates Cambricon Outperform with PT of RMB1,340.

  Risks

  Execution of domestic chip supply ramp.

• Luxshare (002475.CH)

  Supplier of AI-server components and assembly, with domestic demand supplementing its overseas AI infrastructure strategy.

  Strengths

  Communication revenue grew about 50% year on year in 1H26; Bernstein rates it Outperform with PT of RMB86.

  Comparison

  Focused more on overseas AI-server components than domestic peers.

  Risks

  Reliance on continued AI infrastructure demand.

• Unimicron (3037.TT)

  ABF substrate supplier with local subsidiary Qunce serving China’s expanding substrate market.

  Strengths

  Bernstein rates Unimicron Outperform with PT of NT$990.

  Weaknesses

  Advanced ABF manufacturing has high technical and capital barriers.

  Comparison

  Qunce, AT&S and Leading together supply more than half of China’s ABF market.

  Risks

  Capacity expansion and yield execution.

Key data

• China NPO volume-production target2H27Chinese suppliers are validating and sampling NPO products before volume production.
• Huawei Atlas 960E scale4,096 NPUs; 8 EFLOPS FP8; up to 1PB HBMBernstein’s cited specifications for Huawei’s NPO-enabled SuperPoD.
• Atlas 960E optical-engine substitution5,500 7.2T engines replacing 48,000 800G modulesAssociated with power savings of more than 550kW.
• Huawei Ascend 910C rack costc.US$815k per 32-accelerator rack; c.US$25k per acceleratorCompared with more than US$45k per accelerator for an NVIDIA GB200 rack.
• Alibaba, Tencent and Baidu capex growthc.70% in 2025; c.80% projected in 2026Aggregate AI-driven capex trend cited by Bernstein.
• China ABF substrate marketUS$4bn in 2025; 25% of global marketOver half of domestic supply is attributed to Qunce, AT&S and Leading.
• IEIT 1H26 performanceRMB84bn revenue; RMB3bn net profitRevenue rose 5% year on year, gross margin increased 4.2 percentage points, and net profit rose 270% year on year.
• Huaqin superpod revenue guidanceRMB10bn+ in 2026The report says this could more than double in the following year.

Impact & implications

Bernstein argues that China’s localized AI stack can expand value capture beyond accelerators into optical connectivity, switching, integrated server racks, cooling, power, ABF substrates and advanced PCBs. The report particularly emphasizes domestic suppliers able to co-develop systems with cloud providers and chip companies.

Risks

• NPO commercialization may be delayed by optical-engine attachment, socket reliability, PCB-routing, yield and system-integration challenges.
• Embedded-laser designs face thermal, lifetime reliability and repairability challenges.
• China AI-server recovery depends on domestic accelerator supply and advanced-node foundry-capacity ramp-up.
• ABF substrate expansion remains capital-intensive and technically difficult, while some domestic ABF businesses remain loss-making.

What to watch

• Progress from NPO validation and sampling toward planned 2H27 volume production.
• Domestic AI-accelerator availability and China advanced-node foundry-capacity expansion.
• Cloud and internet-company AI-capex execution, particularly at Alibaba, Tencent, Baidu and ByteDance.
• Huawei Atlas 960E deployment and the adoption pace of other domestic superpod architectures.
• Huaqin’s superpod revenue trajectory, independent-switch qualification progress and ABF capacity/yield development.