Nvidia (and the US administration) are aggressively rethinking their China strategy. Restrictions on strategic AI chips have been relaxed. Nvidia now seems intent on applying an aggressive pricing policy to its H200 chips. The aim is to regain market share, of course, but also to slow down the advance of the Chinese AI chip industry.
Several Chinese media seem to have had access to Nvidia’s new strategy. They have estimated that the price of a cluster of eight H200 AI chips would be $200,000, similar to that of a comparable H20 configuration. The H200 AI chip is much more elaborate, and is estimated to deliver more than six times the performance. NVIDIA thus intends to make the H200 much more competitive on the Chinese market. Above all, with a very aggressive price tag, Chinese customers will no longer have to worry about local production. Here are a few technical comparisons of the two solutions:
| Details | NVIDIA H20 | NVIDIA H200 | Estimated gains (H200 → H20) |
|---|---|---|---|
| Architecture | Hopper (export-limited variant) | Hopper | – |
| Process Node | TSMC 4N | TSMC 4N | – |
| HBM Type | HBM3 | HBM3E | – |
| HBM Capacity | ~96 GB | 141 GB | ≈ 47% capacity |
| HBM Bandwidth | ~4.0 TB/s | 4.8 TB/s | ≈ 20% bandwidth |
| Intended Use Case | Inference-focused (restricted training) | Full training inference HPC | Functional uplift |
| Compute Throughput (FP8 / FP16) | Significantly reduced vs H100 (export-compliant) | Similar to H100-class, full Hopper capability | Substantial uplift |
| PCIe / SXM Form Factors | PCIe versions only (typically) | PCIe SXM | – |
| NVLink Support | Restricted / limited | Full NVLink | Major system-level uplift |
| Typical Deployment | China-compliant LLM inference | Global training & inference at scale | – |
While up until now there have been several rumors that the Chinese market was somewhat disdainful of this new version of Nvidia’s chips, an article in the Taiwan Economic Daily reports that Chinese behemoths involved in artificial intelligence, such as Alibaba, Tencent and ByteDance, are planning major investments following the granting of access to H200 GPUs. All in all, a budget of up to $31 billion has been earmarked for the infrastructure, essentially made up of adapted equipment supplied by NVIDIA and AMD.
Nvidia could make the first deliveries as early as February 2026.