Jensen Huang’s China Visit: Could It Revive GPU Prospects? Inside Nvidia’s DGX H200 Cluster Design

In January 2026 the United States formally approved export of Nvidia's DGX H200 to China, but imposed per‑unit approval, a safety review, and a 25% sales‑royalty; as of mid‑May no units have been delivered and no revenue has been recorded.

Jensen Huang was added to the Chinese delegation at the last minute, invited directly by President Trump, and met Chinese officials to accelerate the approval process and clarify compliance details; Trump’s presence may also ease congressional pressure on the export policy.

The author believes Huang’s visit significantly raises the short‑term probability of H200 deployment, potentially speeding approvals and loosening terms, yet domestic alternatives and US political constraints limit a return to the pre‑ban large‑scale exposure.

DGX H200 cluster wiring overview : Stable operation relies on two core data‑transfer networks—compute and storage—both built on optical connections using 400 Gbps InfiniBand/RoCE to meet AI large‑model training bandwidth demands.

Compute network ports : Each DGX H200 rear panel provides four OSFP ports supporting NDR 400 InfiniBand or 400 G Ethernet via dual‑port OSFP transceivers.

Storage network ports : Two NDR 400 InfiniBand or 400 G Ethernet ports are available, supporting 400 G/200 G optical modules for high‑speed data exchange with storage devices.

Compute network topology : An optimized full fat‑tree (leaf‑spine) architecture is used. A scalable unit (SU) comprises 32 DGX H200 servers (256 GPUs), eight leaf switches, and four spine switches. Within an SU, a node reaches any of the other 31 nodes in one hop; inter‑SU communication traverses the spine layer in two hops, leveraging low‑latency optical links.

Optical link design for the compute network includes:

2×400 G‑DR4 long‑reach links (≤500 m) using 2×MPO APC (Base‑8) to 2×MPO APC (Base‑8) cables for leaf‑to‑spine connections.

2×400 G‑SR4/VR4 short‑reach links (≤50 m) using 2×MPO APC (Base‑8) to 4×MPO APC (Base‑4) cables for leaf‑to‑spine connections.

400 G‑SR4/VR4 short‑reach node‑to‑leaf links (≤50 m) using MPO APC (Base‑8) to MPO APC (Base‑8) cables.

For a four‑SU configuration (32 nodes per SU, one node removed for UFM), the core compute network requires roughly 2 040 × 400 Gbps optical cables; the detailed cable breakdown is shown in the accompanying table image.

Storage network design mirrors the compute network’s optical technology, establishing point‑to‑point high‑speed links between storage devices and DGX H200 nodes. Each node’s storage port uses dedicated 400 G/200 G optical modules.

Optical link design for the storage network includes:

400 G‑SR4/VR4 to 2×200 G SR2/VR2 short‑reach links (≤50 m) using MPO APC (Base‑8) to 2×MPO APC (Base‑4) cables.

400 G‑SR4/VR4 short‑reach node‑to‑leaf links (≤50 m) using MPO APC (Base‑8) to MPO APC (Base‑8) cables.

In the same four‑SU scenario, the storage network’s core optical link count is presented in the second table image, allowing flexible scaling with storage capacity.

Source: iCONEC智连 (Nvidia).