Why the US Approved NVIDIA H200 Exports to China

Table of Contents

Media report that the US government has approved NVIDIA to export its high-end H200 AI accelerator to China—but with a major catch: NVIDIA must hand 25% of the revenue from each sale back to the US government as a “technology usage fee.”

On the surface it looks like a relaxation. In reality, the move is a carefully calibrated geopolitical and technological strategy—designed to protect US industry, maintain a controlled technology gap, and slow the momentum of China’s rapidly advancing AI hardware ecosystem.

🔍 Understanding the Real Role of the H200
#

Despite being powerful, the H200 is not NVIDIA’s leading-edge chip.

Key characteristics include:

Built on Hopper architecture (2023 release)
Equipped with HBM3e, offering 141GB memory and 4.8TB/s bandwidth
Provides 1.8× capacity and 1.4× bandwidth of the H100

However, it is significantly outclassed by NVIDIA’s next-generation Blackwell architecture, including:

B200, GB200, and related accelerators
Up to 4× the computational performance
Fully prohibited from export to China

In short, the H200 is a “previous-gen but still extremely capable” accelerator. Allowing its export helps China meet immediate compute needs while carefully preserving a full generational gap between what China receives and what US firms use.

This controlled lag—allowing advancement but preventing parity—is the core strategy shaping US tech export policy.

🇺🇸 The US Strategy: A Triple-Dividend Calculation
#

The approval of the H200 is not a concession—it’s a strategic maneuver optimized for US interests. Three key objectives drive the decision.

1. Protecting US Corporate Interests
#

After the 2024 export restrictions tightened, NVIDIA’s Chinese market—which previously generated over $50B annually—collapsed from 95% share to near zero. Jensen Huang publicly admitted in 2025 that NVIDIA had been “forced out of an irreplaceable market.”

Allowing the H200 reopens a major revenue channel and prevents long-term damage to NVIDIA’s global competitiveness.

2. Direct US Government Revenue
#

The mandated 25% revenue share operates as a built-in national profit mechanism.

H200 unit price: ~$40,000
US government cut per chip: ~$10,000

If NVIDIA sells $5B worth of H200s per quarter in China, the US gains $1.25B directly.

3. Slowing China’s Domestic AI Chip Ecosystem
#

From 2023 to 2025, China made rapid gains in homegrown AI chips due to NVIDIA’s absence:

Huawei Ascend 910B/C mass deployment
Cambricon Siyuan 590 scaling in cloud clusters
Biren BR100 regaining momentum

By reintroducing the H200—with its mature CUDA ecosystem—the US hopes to lure some Chinese customers back, slowing the maturation of fully indigenous alternatives.

🇨🇳 China’s Dual-Edged Opportunity
#

The H200’s entry into the Chinese market is a mixed blessing—beneficial but strategically risky.

Short-Term Advantages
#

Alleviates compute shortages for large-scale model training
Cuts training cycles (months → weeks)
Reduces R&D costs and improves time-to-market
Helps startups access high-end compute without high initial investment

Long-Term Risks
#

The 25% fee will be passed on, raising compute costs
Heavy reliance on imported accelerators could slow domestic R&D
Supply interruptions risk future instability

Yet China’s domestic chip momentum significantly reduces these risks:

Ascend 910C entered large-scale deployments in 2025
Cambricon’s cloud chip revenue surged
Domestic accelerator penetration jumped to 30%, projected 55% by 2027

Strong policy support, security review mechanisms, and “de-Americanization” supply chain mandates create a structural moat for Chinese chipmakers. The H200 may influence the market, but it cannot reverse the broader trend of indigenous substitution.