← Back to stories Industrial optical switch with connected rubber cables of different colors with stickers representing letters and numbers and plastic terminations
Photo by Brett Sayles on Pexels
钛媒体 2026-04-17

NPO, CPO and XPO — who will wield the discourse power over future data centers?

Power play at the optical layer

The fight over next‑generation data‑center optics has shifted from laboratories to procurement desks. Near‑Package Optics (NPO) is scaling now; Co‑Packaged Optics (CPO) is positioned as the “endgame” for ultra‑high density and low power; and a newly announced XPO — unveiled by Arista with more than 45 partners, it has been reported — has entered as a wildcard. Which technology will set the standards for the AI era? The answer depends as much on manufacturability, costs and maintainability as on raw performance.

Market forces and geopolitical context

Analysts see a structural boom: Yole Group forecasts the photonics‑packaging market swelling to about $14.4 billion by 2031, driven by AI data centers and HPC, while DataIntelo projects NPO growing rapidly through the late 2020s. AI workloads are pushing interconnects from 800G toward 1.6T and 3.2T, exposing the physical limits of pluggable optics and making integrated optics attractive. But scale hinges on advanced packaging and foundry capabilities — notably TSMC’s COUPE and other 2.5D/3D platforms — and here geopolitical frictions and trade controls matter. It has been reported that the lack of unified standards and cross‑vendor interoperability is already slowing CPO adoption, a dynamic that plays into regional strengths and supply‑chain resilience.

Industry moves: who’s doing what

Commercial momentum today favors NPO in which many Chinese suppliers lead. It has been reported that Google issued a large NPO order reportedly totaling 12 million modules, with InnoLight (中际旭创) and Xinyisheng (新易盛) capturing the bulk of that business. Domestic firms such as Huagong Technology (华工科技) and Guangxun Technology (光迅科技) have publicly demonstrated 3.2T NPO products and system validations; Haiguang Xinzhen (海光芯正) has shown approaches toward 6.4T engines. At the same time, global chip giants are betting on CPO: Nvidia and Broadcom are investing heavily in silicon‑photonic co‑packaging and advanced assembly. It has been reported that Meta presented reliability data at OFC that challenged earlier maintenance concerns about CPO, but high costs, thermal design and the need for liquid cooling remain practical hurdles.

Outlook: complementary routes, contested standards

NPO and CPO are unlikely to be pure rivals across all use cases. Short term, NPO is a practical bridge for large cloud operators seeking performance gains without remaking their supply chains. Mid to long term, CPO’s power and density advantages should make it the preferred choice inside the highest‑density, rack‑level AI stacks (Scale‑up), while NPO — and now XPO — may persist in Scale‑out and maintenance‑sensitive deployments. Who “wields discourse power” will be decided by who controls key building blocks (foundry packaging, silicon photonics IP, module assembly) and who sets interoperable standards — and that is as much a geopolitical contest as a technical one. Which ecosystem will win: the one driven by packagers and cloud buyers in China, or the one defined by Western chip and foundry ecosystems? The industry is racing to find out.

AI
View original source →