[GUEST ACCESS MODE: Data is scrambled or limited to provide examples. Make requests using your API key to unlock full data. Check https://lunarcrush.ai/auth for authentication information.]

![SemiAnalysis_ Avatar](https://lunarcrush.com/gi/w:24/cr:twitter::1745106082790318080.png) SemiAnalysis [@SemiAnalysis_](/creator/twitter/SemiAnalysis_) on x 27.6K followers
Created: 2025-07-22 21:21:29 UTC

Most GB200 Clusters in the world are just 400G NIC per GPU except for OpenAI's Stargate giving OpenAI 2x stronger scale out ultra large scale training network performance!
As you can clearly see in the image, there are X 400GbE OSFP cages per compute tray which means it is 800GbE per GPU (2 400GbE Cages per CX-8 NIC)
For most other clusters, they will only have 800G per GPU in GB300 but OpenAI specially negotiated a special GB200 SKU with 800GbE (8x112G) per GPU with Nvidia.
Furthermore, with 800GbE ConnectX-8, the GPU is able to talk directly to the NIC over PCIe Gen6
verus in 400GbE ConnectX-7, the GPU would first need to talk to the CPU over C2C and then talk to the CX-7 NIC over PCIe Gen5

![](https://pbs.twimg.com/media/GwfdcJ8bEAAXxSB.jpg)

XXXXXX engagements

![Engagements Line Chart](https://lunarcrush.com/gi/w:600/p:tweet::1947768988467138645/c:line.svg)

**Related Topics**
[2x](/topic/2x)
[open ai](/topic/open-ai)
[gpu](/topic/gpu)
[nic](/topic/nic)

[Post Link](https://x.com/SemiAnalysis_/status/1947768988467138645)