NVIDIA is unwilling to give up the Chinese market! It aims to launch a "China-specific" AI chip again

According to the latest media reports, the global "AI chip giant" NVIDIA (NVDA.US) has notified its most important clients in the Chinese market, including ByteDance, Alibaba, and Tencent, that the company is revising its AI chip design architecture to comply with the latest export restrictions from the U.S. government while insisting on continuing to supply AI chips to Chinese companies.

During a recent high-profile visit to the Chinese market, NVIDIA CEO Jensen Huang revealed this latest special AI chip plan for China to clients, as reported by The Information.

This visit occurred after the U.S. government informed NVIDIA that it would require strict government approval to sell AI chips to clients in the Chinese market—specifically, the U.S. government's AI restriction list for AI chips has expanded to include the previously performance-restricted H20. The performance of H20 has been significantly reduced compared to H100/H200, making it an exclusive custom version of the AI chip that NVIDIA is permitted to export to the Chinese market. However, the latest restrictions from the Trump administration effectively prohibit the sales path of H20 to Chinese clients.

As a result, NVIDIA expects to incur up to $5.5 billion in additional costs in its upcoming quarterly financial report, a situation disclosed in its Form 8-K filing, leading to a nearly 7% drop in its stock price thereafter.

Despite the escalating tensions between the U.S. and China, NVIDIA continues to commit to developing new AI chips for clients in the Chinese market that comply with regulatory restrictions, highlighting the importance of this Eastern market to the performance of the semiconductor giant headquartered in Santa Clara, California.

The Information reported that Chinese tech giants like ByteDance, Alibaba, and Tencent have ordered over $16 billion worth of H20 AI chips in the first three months of this year, and it remains unclear how the latest U.S. government ban will affect these AI chip orders.

The report also stated that NVIDIA has informed some clients in the Chinese market that the company must obtain regulatory approval from the U.S. Department of Commerce before launching any new versions of AI chips in the Chinese market.

For the fiscal year 2025 ending January 26, NVIDIA achieved sales of up to $17.11 billion in the Chinese market, accounting for approximately 13% of the semiconductor giant's total revenue of $130.5 billion.

Will the upcoming "special version AI chips" for the Chinese market take the ASIC route instead of the general GPU route?

Some semiconductor industry analysts indicated after The Information's latest report that NVIDIA may shift its AI chip technology route from general GPUs to AI ASICs specifically designed for AI training/inference to launch a version of AI chips tailored for the Chinese market that complies with U.S. government export bans.

These analysts noted that the specific architecture of GPUs means that NVIDIA cannot launch AI chips that comply with U.S. export restrictions in the short term without significantly reducing performance, but a substantial reduction in performance may make NVIDIA's AI chips less cost-effective compared to domestic AI chips. However, other analysts suggested that NVIDIA's strategy for AI chips in the Chinese market may focus on "making quick and moderate downgrades on the AI GPU architecture to avoid regulatory red lines"—for example, reducing NVLink interconnect rates, cutting bandwidth, or thresholding tensor computing power Continue to assess in the medium to long term whether to launch dedicated AI ASICs for AI inference.

AI ASICs are also known in the industry as "customized AI chips," "dedicated AI chips," or "AI application-specific integrated circuits." Unlike traditional general-purpose processors (such as CPUs and GPUs), AI ASICs are deeply customized for specific AI tasks (such as deep learning, artificial intelligence inference, and training systems) to execute efficiently, aiming to enhance artificial intelligence computing efficiency, reduce power consumption, and improve performance, especially demonstrating significant energy efficiency advantages when executing large-scale AI parallel computing. For example, the TPU (Tensor Processing Unit) developed by Google in collaboration with Broadcom is a typical AI ASIC, primarily used for deep learning inference and training, optimizing key computational operations such as matrix multiplication to enhance AI computing performance. Broadcom and Marvell Technology are currently the leaders in the AI ASIC field.

The low-cost paradigm of DeepSeek indicates that AI inference can be fully optimized through algorithm engineering to reduce inference costs, allowing large models to be deployed more conveniently and cheaply, which also means that the advantages of AI ASICs at the inference end will become even more substantial in the future. Although NVIDIA's general-purpose AI GPUs are powerful, their power consumption, enterprise purchase costs, and computing power rental costs are much more pressured in large-scale inference scenarios. Microsoft, Amazon, Google, and Meta are all collaborating with Broadcom or developing their own AI ASIC chips with Marvell for massive inference computing power deployment. For example, the TPU (Tensor Processing Unit) developed by Google in collaboration with Broadcom is a typical AI ASIC.

Looking ahead to the future of computing power, NVIDIA's AI GPUs may focus more on ultra-large-scale frontier exploratory training, rapidly changing multimodal or new structure rapid experimentation, as well as general computing power for HPC, graphics rendering, and visual analytics. AI ASICs, on the other hand, focus on extreme optimization of specific deep learning operators/data flows, excelling in stable structure inference, high throughput, and high energy efficiency. In the long run, both will coexist harmoniously, with the AI ASIC market share expected to expand significantly in the medium term. NVIDIA's general-purpose GPUs will focus on complex and variable scenarios and cutting-edge research, while ASICs will focus on high-frequency stability, large-scale AI inference loads, and some mature and stable fixed training processes