AnswersPublished23 sources
How Nvidia’s Optical Networking Plans Could Scale AI Clusters
Nvidia’s optical strategy pairs production co packaged optics switches with research prototypes and a future 1,152 GPU cluster design. At SEMICON West, Nvidia presented optics on interposer research and a DWDM demonstration reporting 256 Gb/s per fiber; neither has an announced commercialization timeline.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Nvidia’s optical strategy pairs production co packaged optics switches with research prototypes and a future 1,152 GPU cluster design.
- At SEMICON West, Nvidia presented optics on interposer research and a DWDM demonstration reporting 256 Gb/s per fiber; neither has an announced commercialization timeline.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: How do Nvidia’s optical-interconnect research, existing products, manufacturing expansion and future roadmap support larger AI clusters, inc. Article summary: Nvidia’s strategy combines optical-interconnect research, production networking hardware, and expanded manufacturing capacity to make larger AI systems practical. The key distinction is that commercial co-packaged-optics. Topic tags: general, general web, user generated, documentation. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks,
Nvidia’s optical roadmap has three distinct parts: research into closer, denser optical connections; co-packaged-optics (CPO) switches that the company says are in production; and supplier capacity targets intended to support larger systems. The distinction matters: a working prototype or manufacturing expansion does not mean every optical technology is ready for deployment.
Two research approaches to moving more data
At SEMICON West 2026, Nvidia Principal Research Scientist Nikola Nedovic presented “Optical Integration for the Next Generation of AI.” The event schedule lists his session on October 13, and the conference was held in San Francisco. 21
23
28
One focus was optics-on-interposer (OOI). The prototype places optical engines on the silicon interposer alongside compute chips, rather than routing the electrical connection through the organic package substrate. The aim is to shorten the electrical path and potentially improve bandwidth density and power efficiency. 17
21
A second focus was dense wavelength-division multiplexing (DWDM), which carries data on multiple wavelengths of light through one fiber. The reported demonstration used eight wavelengths at up to 32 Gb/s each, for an aggregate 256 Gb/s per fiber. That addresses how much data a fiber can carry; OOI addresses the placement of optical engines and the electrical path to them. 6
These demonstrations are not evidence of commercial readiness. Reporting says Nvidia has announced no commercialization timetable for OOI or DWDM, and describes the technologies as research-stage. 6
33
What is already in production: Spectrum-X Ethernet Photonics
Nvidia’s Spectrum-X Ethernet Photonics is a separate, commercial CPO switch platform. CPO places optics in the same package as the switch ASIC; Nvidia says the approach can reduce power consumption and increase reliability. 41
Nvidia announced the product as “now in production” in May 2026 and lists CoreWeave, Lambda, Meta, Microsoft and Oracle Cloud Infrastructure among its first adopters. 37
45 Production status is not the same as evidence of broad deployment: later reporting described limited shipments and said it was unclear whether recipients were evaluating the switches or using them commercially.
36
SPIL’s capacity target is not a shipment forecast
Nvidia says SPIL is targeting a 15-fold increase in optical-engine capacity for CPO by mid-2027, with another doubling of the expanded capacity by the end of 2028. 13
If both targets are achieved sequentially, capacity would be roughly 30 times the starting level. But Nvidia has not provided a baseline volume in the cited announcement, so the multiplier does not establish an absolute production count or predict how many switches will ship. 3
13
The capacity plan concerns optical engines for CPO. It does not announce when the OOI or DWDM research prototypes might become products. 6
13
The larger-cluster roadmap
Nvidia’s technical blog describes Kyber as a future rack design with 144 GPUs per rack and says the architecture can scale to a Kyber NVL1152 system connecting eight racks with direct optical interconnects. 32 The 1,152-GPU figure is the system’s stated configuration—not evidence that the research prototypes shown at SEMICON West will be used in it.
Taken together, the plans point toward more optical connectivity as Nvidia scales AI systems: CPO switches are the production step, supplier expansion is a capacity target, and OOI and DWDM remain research. Keeping those milestones separate is essential when assessing how soon the roadmap could support larger clusters.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "How Nvidia’s Optical Networking Plans Could Scale AI Clusters"?
Nvidia’s optical strategy pairs production co packaged optics switches with research prototypes and a future 1,152 GPU cluster design.
What are the key points to validate first?
Nvidia’s optical strategy pairs production co packaged optics switches with research prototypes and a future 1,152 GPU cluster design. At SEMICON West, Nvidia presented optics on interposer research and a DWDM demonstration reporting 256 Gb/s per fiber; neither has an announced commercialization timeline.
Sources
wccftech.comNVIDIA's Next-Gen CPO Technology Puts Optics Directly On The Interposer While DWDM Pushes Optical Speeds To 256 Gb/s, First Prototypes Already Made For Future Chips
nvidia.comNVIDIA at SEMICON West 2026 | October 13–15, San Francisco
semiconwest2026.eventscribe.netFull Schedule - SEMICON West 2026
technews.tw輝達發表矽中介層光學封裝原型晶片，布局 2028 年 Feynman 平台