Nvidia reportedly ships H200 AI chips to China by mid-February

Nvidia plans to begin shipping its H200 AI chips to Chinese clients before the Lunar New Year in mid-February, according to three sources speaking to Reuters.

The US chipmaker aims to fulfill initial orders using existing stock, with shipment estimates ranging from 5,000 to 10,000 modules, or about 40,000 to 80,000 chips.

Nvidia also told clients it intends to add production capacity, with new orders opening in Q2 2026.

The timeline depends on approval from Chinese authorities, and the plan may change based on government decisions.

These would be the first H200 shipments to China after President Trump announced the US would permit such sales with a 25% fee, reversing a previous ban by the Biden administration.

Beijing is still reviewing whether to allow these imports.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

Chinese AI demand exceeds 40k–80k H200 chips, a symbolic rather than game-changing shipment

Huawei plans 400,000 Ascend AI accelerators in 2025 and over 1 million in 2026 1, while Nvidia shipped 3.76 million AI accelerators worldwide in 2023 2.
A 40k–80k H200 lot equals about 10–20% of Huawei’s 2025 Ascend output 1. This covers a small slice of local demand and will not slow homegrown chip work.
Alibaba and peers invest far less in AI infrastructure than U.S. hyperscalers (large-scale cloud platforms) 3, which limits near-term deployment plus integration of H200s.
H200 beats H20 by about 6x for training (building new models), but many firms already run millions of capable devices 4. Inference (running models) offers the broader rollout path.

Vendors can target Chinese data centers that need power and cooling upgrades for H200s

High Bandwidth Memory (HBM) 3e hits 4.8 TB/s on H200 5, which pushes liquid cooling plus stronger power delivery in dense racks.
Integrators plus cooling specialists can pitch retrofits to sites buying H200s, since Huawei’s Cloud Matrix 384 cluster (a large-scale AI training cluster) uses about 4x the power of comparable Nvidia systems 1.
Power capacity is often available in China 4. The constraint is accelerators, i.e., sites have enough electricity but not enough GPUs. That opens demand for efficient distribution and management gear.
Leasing firms can structure deals that absorb the 25% import fee, which lowers upfront cost while approvals remain pending.