Nokia’s account of discussions at the Open Compute Project Global Summit sets out why connecting AI chips is becoming an infrastructure challenge in its own right. The key point: there is no single optical design that wins everywhere, because power, reliability, repairability and serviceability matter alongside speed.
Nokia Watch analysis
What happened
Nokia says AI clusters are growing large enough to put greater pressure on the networks connecting their processors. It summarises the practical rule of thumb as “copper where it works; optics where distance, density, power or system architecture requires it”. The company also cites BNP Paribas’ estimate that networking equipment accounts for $15 of every $100 invested in AI infrastructure.
The options run from replaceable pluggable optics to near-packaged optics (NPO), which places optical engines close to a switch’s silicon, and co-packaged optics (CPO), which integrates them more tightly with the chip package. Moving optics closer can improve power efficiency, bandwidth density and signal integrity, but it also changes manufacturing, maintenance and failure-management requirements.
Nokia highlights Oracle and Meta’s reported 800G linear pluggable optics deployments, which it says showed power benefits and better link reliability than fully retimed optics. It also points to 1600G pluggables as the next speed milestone, with fresh power and signal-integrity challenges. These are examples and claims presented in Nokia’s account, not a universal verdict on the competing designs.
Why it matters
AI infrastructure is not just processors and electricity. Data has to move between chips quickly and reliably, and the choice of connection affects power use, how easily equipment can be replaced, and how widely a fault might spread. Those are operational decisions, not merely a contest to put the biggest number on a slide.
Nokia’s account also makes clear why “the next generation replaces the last” is too tidy a story. Pluggables, NPO and CPO have different strengths, and the right choice depends on the system and how it will be operated.
Our read
The useful takeaway is that AI networking is a menu of trade-offs, not a march towards one inevitable architecture. Nokia is a supplier with an interest in the field, but its description gives readers a concrete way to compare the approaches: ask what they save in power and space, and what they demand in reliability, repairability and supplier choice.
What to watch
- Whether more operators report results from 800G LPO deployments.
- How 1600G pluggables handle power consumption and signal integrity.
- Whether socketed NPO designs gain standardised interfaces and multi-vendor support.
- Which optical approaches prove easiest to maintain and isolate when something fails.
Discussion spark: For AI data centres, should operators prioritise the power and density gains of more tightly integrated optics, or the serviceability and supplier choice of replaceable pluggables?
Sources and evidence
- Source update (5 November 2026, 15:36 UTC)
Independent WittyWires tracker for public updates about Nokia. Not affiliated with or endorsed by Nokia; this is not an official account.