Bernstein: Comprehensive Mapping of China's NPO and AI SuperPod Supply Chain
nashnova research
Bernstein maps out China's near-package optics (NPO) and AI SuperPod supply chain end to end, concluding that China's AI server market could resume growth in H2 2026 as domestic AI chip capacity ramps — a full-chain investment roadmap from optical engines to finished racks.
What is NPO, and why does it matter now?
NPO — near-package optics, mounting optical modules right next to the chip so data travels as light instead of electrical signals — is the core interconnect technology for next-generation AI servers.
Unlike traditional pluggable optical modules, NPO must be custom-designed for each GPU or NPU: optical engine, fiber array, PCB routing, and thermal packaging all co-engineered.
This means → NPO is not an off-the-shelf component but an entire supply chain that needs rebuilding. Whoever clears validation first wins the orders.
How far along is China's NPO supply chain?
China's NPO ecosystem has the key building blocks in place, but it is more fragmented than TSMC's integrated COUPE platform — design, foundry, and packaging sit in separate companies.
Current demonstrations use edge coupling + MZM modulators (Mach–Zehnder modulators — devices that convert electrical signals into light) + external lasers, targeting bandwidth of roughly 3.2T.
All segments have entered validation and sampling, with mass production targeted for H2 2027. In plain terms = the technology path is locked; now it is an engineering-validation sprint.
Who sits where in the chain?
Optical engines & packaging: Innolight and T&S Communications focus on overseas NPO opportunities. Innolight has sampled two offshore customers, plans H2 2027 mass production, and expects shipments to jump to tens of millions of units in 2028.
Domestic optical engines: Accelink and HC Photonics target China's NPO demand. On PIC design — photonic integrated circuits, chips that process signals with light — Lightelligence nearly tripled H1 2026 revenue year-on-year; Xphor grew 2025 revenue fivefold to RMB 461 million.
Foundry & lasers: Tower Semiconductor remains the main silicon-photonics foundry for many Chinese PIC designers; CanSemi is expanding domestic lines. Sourcephotonics, China's largest laser supplier, is developing 120/300/400 mW products for NPO.
This reflects a supply chain that, while fragmented, has companies pushing volume at every link — maturity is rising in parallel across the board.
Huawei's Atlas 960E — what does China's first NPO SuperPod signal?
Huawei launched the Atlas 960E SuperPod, scalable to 4,096 NPUs, with FP8 compute of 8 EFLOPS and up to 1 PB of high-bandwidth memory (HBM).
The system uses roughly 5,500 units of 7.2T optical engines, replacing about 48,000 ports of 800G pluggable modules and cutting power consumption by over 550 kW.
This means → NPO has left the lab. Huawei has turned it into a deliverable product, and core foundry partner SMIC (rated outperform) benefits directly.
How does the cost math work — Huawei vs. Nvidia?
Bernstein estimates Huawei's Ascend 910C rack costs roughly $815,000, or about $25,000 per accelerator.
Nvidia's GB200 rack runs above $45,000 per accelerator. In plain terms = for the same single accelerator, the Huawei solution costs a little over half of Nvidia's.
This reflects a domestic offering that is already compelling on price — though the final comparison still hinges on real-world compute efficiency and software-ecosystem maturity.
When does China's AI server market recover, and who benefits first?
Combined capex of Alibaba, Tencent, and Baidu grew roughly 70% in 2025; consensus expects another ~80% increase in 2026. ByteDance more than doubled 2025 capex and plans to invest about $30 billion in AI infrastructure.
Shipments in 2025–2026 face pressure from tighter U.S. restrictions on Nvidia GPU exports, but as domestic accelerators mature and advanced-node capacity expands, a recovery is expected in H2 2026.
First-in-line beneficiaries: interconnect components, switches, server ODMs (Huaqin and others), and ABF substrate makers (Unimicron's Kinsus, AT&S, and Zhen Ding's Liding together supply over half of China's ABF market). Luxshare Precision (rated outperform) is primarily positioned for overseas AI server components, with China expansion as upside.
This means → 2026 is the make-or-break validation year — whether domestic accelerator output ramps on schedule will determine if this entire supply-chain thesis plays out.
市场有风险，内容仅供研究参考，不构成投资建议。