Nvidia H20 Returns, Huawei Scales AI Chips Amid US-China Tech Tensions

The landscape of US-China artificial intelligence relations saw significant developments this week, marked by Nvidia’s renewed access to the Chinese market, Huawei’s advancements in domestic AI hardware, and a substantial penalty for Cadence Design Systems due to export violations. These events underscore the ongoing competition and the intricate balance between technological advancement, national security, and economic interests.

Nvidia Resumes H20 Shipments to China

Nvidia has secured approval to resume shipping its H20 AI accelerators to Chinese cloud builders, a reversal of an earlier export ban. This decision comes after negotiations reportedly tied to rare-earth mineral trade, despite criticism from some US lawmakers who warn about the potential erosion of US technological leadership. The H20, while less powerful than Nvidia’s flagship H100 or Blackwell chips, remains the most advanced AI accelerator Nvidia is legally permitted to export to China, designed specifically to comply with US export controls.

Following the approval, Nvidia placed an order for 300,000 H20 AI chips with TSMC, adding to an existing stockpile of 600,000 to 700,000 units, indicating strong demand from Chinese customers. Chinese tech giants such as Tencent, ByteDance, and Alibaba had previously placed large orders for the H20 and are awaiting licensing approval from the Commerce Department, which Nvidia anticipates will be cleared soon. However, US officials have not yet formally confirmed these approvals. Nvidia CEO Jensen Huang has stated that future H20 production will depend on market demand, noting that restarting wafer fabrication could take up to nine months due to capacity cancellations during the export pause. This resumption of shipments is seen by the Trump administration as a strategic trade-off to ensure US access to Chinese rare-earth exports.

Huawei Scales Up Domestic AI Capabilities with CloudMatrix 384

In a clear demonstration of China’s push for domestic AI solutions, Huawei quietly unveiled its CloudMatrix 384 AI cluster at the World Artificial Intelligence Conference in Shanghai last week. This system, powered by 384 of Huawei’s proprietary Ascend 910C NPUs, is positioned as a domestic alternative to Nvidia’s Blackwell-based GB200 NVL72 racks. The CloudMatrix 384 utilizes a unique mesh UB mesh fabric, linking together a significantly higher number of accelerators than Nvidia’s comparable system.

On paper, the CloudMatrix 384 boasts impressive performance, yielding approximately 40% more FP16 throughput than an NVL72, despite each Ascend chip delivering only about one-third the per-chip performance of a B200 GPU. Huawei compensates for this individual chip performance gap by employing a larger number of processors and incorporating additional high-bandwidth memory for faster data movement. This design, however, leads to substantially higher power consumption, around 600 kW, which is roughly four times that of an NVL72, and a higher unit cost of approximately $8.2 million per 16-rack supernode, more than double Nvidia’s list price.

A significant challenge for Huawei remains software. While its MindSpore/PyTorch-compatible stack now supports ONNX and other popular LLM frameworks, it still lags behind Nvidia’s well-established CUDA ecosystem. Huawei is actively investing in “higher-level abstractions” to simplify code migration and is promoting “China-first workloads” like the DeepSeek-R1 LLM, which are optimized for Ascend hardware. Despite these efforts, the cost and complexity of migrating existing codebases from Nvidia GPUs continue to slow widespread adoption of Huawei systems. Huawei’s CloudMatrix appears to target large, state-backed AI datacenters in China, which may prioritize domestic supply chains and performance over energy efficiency or global interoperability.

Cadence Faces Penalty for Export Violations

As Chinese companies advance their domestic AI hardware, US regulators are demonstrating stricter enforcement of export rules. Cadence Design Systems has agreed to plead guilty and pay over $140 million in combined criminal and civil penalties for selling electronic-design-automation (EDA) software and hardware to China’s National University of Defense Technology (NUDT) through front companies between 2015 and 2021. NUDT has been on the US Commerce Department’s Entity List since February 2015, due to its supercomputers’ suspected support for nuclear explosive simulation and military activities.

The plea agreement includes a three-year probation period, forfeiture, and strict compliance reporting. Investigators revealed that Cadence subsidiaries completed at least 56 illegal exports and later transferred tools to Phytium Technology, another chipmaker linked to NUDT. This case highlights Washington’s dual approach: allowing some limited technology exports, such as Nvidia’s H20, while simultaneously enforcing strict penalties for violations to deter illicit transfers of sensitive technology. The US government emphasizes the importance of safeguarding American technological know-how from falling into the wrong hands, particularly concerning military applications.

Broader Implications for US-China AI Relations

These recent developments underscore the complex and evolving dynamics of US-China AI relations. The US aims to maintain its technological leadership through export controls, while China is aggressively pursuing self-sufficiency and domestic innovation in AI hardware and software. The re-entry of Nvidia’s H20 chips into the Chinese market, alongside Huawei’s powerful CloudMatrix system, illustrates the ongoing competition for market share and technological supremacy. Meanwhile, enforcement actions against companies like Cadence signal a continued commitment by the US to prevent the unauthorized transfer of sensitive technologies that could bolster China’s military capabilities. The “AI race” between the two nations is not merely economic but also has significant geopolitical and national security implications, influencing military strategies, intelligence operations, and global power dynamics.