Jensen Huang says Nvidia will use co-packaged optical tech in two upcoming networking chips for its servers to cut energy usage, but GPU use will have to wait
Stephen Nellis / Reuters :
Context & Ripple Effects
Nvidia has framed the data center as an “AI factory,” a system-level approach discussed in an earlier AI-factory data-center strategy. This networking-chip decision extends that framing from compute and racks to the links that move data between them.
The move also fits Nvidia’s broader effort to deepen its rack-scale offering, alongside its prior work on custom server-rack designs and cloud services. It matters because networking power becomes a more visible constraint as server systems scale.
First-order effects
- Nvidia’s two upcoming server networking chips gain co-packaged optical technology, targeting lower energy use in the networking portion of its server platforms.
- GPU roadmaps are not receiving the same optical integration immediately, preserving a split between near-term network-chip deployment and later potential GPU adoption.
Second-order effects
- Server buyers evaluating Nvidia platforms will have another system-level efficiency attribute to weigh, rather than comparing GPUs only on compute performance.
- The staged rollout concentrates near-term implementation pressure on networking and optical-interconnect partners, while GPU-side component ecosystems have more time before a comparable transition.
Third-order effects
- If co-packaged optics proves viable in production networking silicon, data-center differentiation may increasingly shift toward power-efficient interconnects and rack-level design rather than standalone accelerators alone.
- The delayed GPU deployment underscores that broad optical integration remains a packaging and deployment transition, not yet a uniform architecture across the server.
The trend: AI infrastructure is evolving toward tightly integrated, rack-scale systems in which networking energy efficiency is becoming as strategically relevant as accelerator performance.