Nvidia plans to introduce a new AI chip for the Chinese market, priced between US$6,500 and US$8,000.

This is significantly lower than its recently restricted H20 model, which ranges from US$10,000 to US$12,000.

The new chip, based on the Blackwell architecture, features weaker specifications and simpler manufacturing requirements.

It will utilize conventional GDDR7 memory instead of high bandwidth memory and will not use advanced packaging technology from Taiwan Semiconductor Manufacturing Co.

Mass production is expected to begin as early as June, though the final name of the product is not yet confirmed.

This development marks Nvidia’s third GPU tailored for China in response to United States export restrictions.

CEO Jensen Huang said the older Hopper architecture used in the H20 cannot be modified to meet current US regulations.

In the last financial year, China represented 13% of Nvidia’s sales, but US export curbs have significantly impacted its market share, which has declined from 95% before 2022 to 50% now.

These restrictions have forced Nvidia to write off US$5.5 billion in inventory and forgo US$15 billion in sales.

The new chip is designed to comply with US regulations, which limit memory bandwidth to approximately 1.7-1.8 terabytes per second. This is lower than the H20’s bandwidth of 4 terabytes per second, aligning with the new export limits.

.source-ref{font-size:0.85em;color:#666;display:block;margin-top:1em;}a.ask-tia-citation-link:hover{color:#11628d !important;background:#e9f6f5 !important;border-color:#11628d !important;text-decoration:none !important;}@media only screen and (min-width:768px){a.ask-tia-citation-link{font-size:11px !important;}}

🔗 Source: Reuters

🧠 Food for thought

1️⃣ US export controls have created an unintended AI chip arms race

Export restrictions on advanced AI chips have triggered a transformation in China’s semiconductor landscape, inadvertently accelerating domestic innovation.

Nvidia CEO Jensen Huang recently acknowledged this reality, stating that US export controls have spurred Chinese development and caused Nvidia’s market share in China to drop from 95% to just 50% 1.

This pattern of restriction followed by adaptation has forced Nvidia to develop specialized variants of its chips three times now, with the Blackwell-based design representing their latest attempt to maintain access to what Huang describes as a “$50 billion data center market” 2.

The restrictions created significant economic consequences for Nvidia, including a $5.5 billion inventory write-off and approximately $15 billion in lost sales opportunities from the H20 ban alone.

These developments illustrate how export controls often produce complex market consequences that diverge from their intended security objectives.

2️⃣ Huawei’s rise demonstrates China’s technological resilience

As Nvidia’s market position weakens in China, Huawei has emerged as a key competitor with its Ascend 910B chip filling the void created by US restrictions.

Huawei began positioning itself as a serious AI chip competitor even before the latest export controls, with its Ascend 910 chip claiming performance metrics that rival Nvidia’s offerings: 256 TeraFLOPS for FP16 operations compared to Nvidia’s comparable models 3.

The company has developed a comprehensive AI stack including hardware, software frameworks like MindSpore, and application enablement layers, creating an ecosystem that parallels Nvidia’s integrated approach 4.

This technological advancement occurred despite US sanctions, illustrating how trade restrictions can sometimes strengthen a nation’s determination to achieve technological self-sufficiency.

Huawei’s success demonstrates how competitive alternatives can emerge when market leaders face regulatory barriers, creating new industry dynamics that reshape global supply chains.

3️⃣ Memory specifications have become the new battleground in AI chip design

The technical specifications of Nvidia’s new China-focused chip reveal how memory bandwidth has become a critical factor in navigating export restrictions.

US regulations now specifically target memory bandwidth capabilities, capping allowable performance at approximately 1.7-1.8 terabytes per second, below the 4 terabytes per second that the now-restricted H20 model could achieve 5.

To comply with these limitations while maximizing performance, Nvidia’s new chip will use conventional GDDR7 memory instead of more advanced high bandwidth memory (HBM) and avoid TSMC’s advanced packaging technology.

This technical workaround exemplifies how chip designers must now balance regulatory compliance with performance optimization, creating distinct product categories for different markets.

The memory bandwidth restriction specifically targets AI workloads, which require extensive data processing capabilities, showing how regulators are becoming increasingly sophisticated in targeting the technical specifications most relevant to AI development.

Recent Nvidia developments

……