(Source: Shutterstock)
The contest to build and control China’s proliferation of AI hardware intensified on three fronts this week. Nvidia secured the green light to resume shipping its pared‑down H20 accelerator, Huawei introduced a 384‑chip Ascend 910C cluster that it claims can stand beside Blackwell systems, and Cadence accepted a $140 million penalty for past export violations tied to Chinese military‑linked institutes. These approvals, Chinese domestic advances, and stricter enforcement highlight how Beijing and Washington keep shaping the same market from opposite directions, with China scaling its own solutions and the United States refining the export rules that decide what chips can leave its shores.
Nvidia Wins Approval To Resume H20 Sales to Chinese Cloud Builders
(Source: Rokas Tenys/Shutterstock)
Last week, Nvidia placed an order for 300,000 H20 AI chips with contract manufacturer TSMC—adding to an existing stockpile of 600,000 to 700,000 units—as unusually strong demand from Chinese customers prompted a shift in strategy beyond relying on prior inventory. This move follows a reversal by the Trump administration of the April export ban, allowing limited shipments of the China-specific H20 GPU to resume. The clearance came after negotiations tied to rare‑earth mineral trade, according to sources familiar with the talks, and drew further criticism from lawmakers.
Originally designed to comply with U.S. export controls, the H20 offers less computing power compared with flagship variants like the H100 or Blackwell chips, yet remains the most powerful AI accelerator Nvidia is legally permitted to ship into China. Chinese tech giants such as Tencent, ByteDance and Alibaba had lined up large orders for H20 before the ban and continue to await licensing approval from the Commerce Department—a process Nvidia expects will be cleared soon.
Despite assurances that licenses are forthcoming, U.S. officials have not formally confirmed approvals yet. Nvidia CEO Jensen Huang has emphasized that future H20 production hinges on market demand, signaling that restarting wafer fabrication could take up to nine months due to capacity cancellations during the export pause.
House China‑relations committee chair Rep. John Moolenaar has publicly criticized the resumption of H20 shipments, warning the chips for China could erode U.S. technology leadership. In typical contrast, the Trump administration framed the export pause reversal as a concession to support U.S. access to Chinese rare-earth exports—a strategic trade-off deemed necessary.
Huawei’s Ascend‑powered Supernode Shows China’s Push for Domestic AI, Despite Higher Costs
Huawei quietly unveiled its CloudMatrix 384 AI cluster at the World Artificial Intelligence Conference in Shanghai last week, positioning it as a domestic alternative to Nvidia’s Blackwell‑based GB200 NVL72 racks. Using a proprietary mesh UB mesh fabric, the system links together 384 Ascend 910C NPUs, which is five times the accelerator count of an NVL72. On paper, that scale yields about 40% more FP16 throughput than an NVL72 (≈ 250–300 PFLOPS vs 180 PFLOPS), though each Ascend chip delivers only one‑third the per‑chip performance of a B200 GPU.
Huawei’s CloudMatrix 384 supernode. (Source: Huawei)
This system may be more accessible to China than Nvidia’s products, but it will come with significant costs. Huawei compensates for its chip‑level performance gap by using a larger number of Ascend processors and additional high‑bandwidth memory located close to each chip for faster data movement. That design pushes power consumption to about 600 kW, roughly four times the draw of an NVL72, and raises the cost of each 16‑rack supernode to around $8.2 million, more than twice Nvidia’s list price.
An even bigger hurdle is software. Huawei’s MindSpore/PyTorch‑compatible stack now supports ONNX and other popular LLM frameworks, but it still lacks CUDA’s robust ecosystem, now almost a decade in the making, that many developers prefer for writing and tuning AI code.
Huawei is trying to close that advantage and make its platform more appealing on two fronts. First, it is investing in “higher‑level abstractions,” meaning programming tools that hide the gritty hardware details so a model written in PyTorch can run on Ascend chips without major rewrites. Second, it is showcasing “China‑first workloads” such as the DeepSeek‑R1 LLM, which are designed from the ground up for Chinese data and users and already perform well on Ascend hardware.
Even so, many companies have years of scripts and optimizations written specifically for Nvidia GPUs and moving that code to a new architecture still takes time, money, and skilled engineers. Those migration costs continue to slow widespread adoption of Huawei systems despite recent progress on the software side.
The bottom line: Nvidia’s go‑ahead to resume H20 shipments may satisfy cloud builders who simply need “good‑enough” AI accelerators, but Huawei’s CloudMatrix targets a different slice of the market. It could be popular for the large, state‑backed AI datacenters already existing all over China who are likely more willing to trade energy and floorspace for top performance and a dependable domestic supply chain.
US Regulators Tighten the Reins as Cadence Pleads Guilty on China Sales
While Chinese companies press ahead with their own AI hardware, a U.S. chip‑design software firm is discovering how costly it can be to break export rules.
(Source: Shutterstock)
Cadence Design Systems has agreed to plead guilty and pay more than $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. The plea includes a three‑year probation period, forfeiture, and strict compliance reporting.
NUDT has been on the U.S. Commerce Department’s Entity List, a roster of organizations subject to strict export licensing, since February 2015, as its supercomputers are thought to support nuclear explosive simulation and military simulation activities. Investigators say Cadence subsidiaries completed at least 56 illegal exports and later transferred tools to Phytium Technology, a chipmaker linked to the same university.
The case arrives as Washington balances selective openings, like allowing Nvidia’s trimmed‑down H20 back into China, with high‑profile enforcement actions to deter future violations. The $140 million fine does not appear to be much trouble for Cadence, as its share price rose 7.8% after it booked the charge, suggesting investors see the matter as financially contained.
Washington’s Mixed Signals Complicate the AI Race
The back‑and‑forth on export policy has made clear that Washington is trying to have it both ways—enforcing the rules when politically necessary, but easing them when trade leverage or domestic industry pressure outweighs strategic concerns. Reopening the door for Nvidia’s H20 shipments may help secure rare earth supplies and preserve short‑term revenues, but it also feeds China’s AI infrastructure at a time when U.S. officials are warning about the very capabilities those chips could support. Meanwhile, penalties like the one Cadence accepted signal that enforcement remains real, but perhaps uneven.
For U.S. tech companies, the message is simple: there’s money to be made, so long as you stay within the (shifting) boundaries. But for U.S. policymakers, the bigger question remains unanswered—whether short-term commercial wins are worth trading for the erosion of long-term technological dominance in one of the most strategically important sectors of this era.