NVIDIA founder and CEO Jensen Huang made his third visit to mainland China this year at the Third China International Supply Chain Expo on July 15, 2025. This appearance follows NVIDIA's market capitalization briefly surpassing the $4 trillion threshold days earlier on July 9. Concurrently, NVIDIA announced via its official WeChat account that it had applied to resume sales of its H20 GPU in China, stating the U.S. government "has assured NVIDIA it will grant the license." The company expressed urgency to initiate deliveries promptly.

The H20 represents NVIDIA's previous specialized AI chip offering for China, which fell under U.S. export controls in April 2025. NVIDIA's announcement highlighted both its efforts to "strengthen AI infrastructure" in the United States and its desire to "explore AI's productivity potential" within China.

**The Compliance-Driven H20** The H20's development originated directly from updated U.S. Bureau of Industry and Security (BIS) export rules issued on October 17, 2023. These regulations imposed two quantifiable technical thresholds on advanced computing chips destined for China: "Total Processing Performance" (TPP), measuring raw computational power, and "Performance Density," the ratio of TPP to chip area, designed to circumvent restrictions via chiplet packaging. These thresholds barred NVIDIA's top-selling H100, A100, H800, and A800 data center GPUs from the Chinese market.

In response, NVIDIA pivoted its product strategy. By November 2023, reports confirmed the development of three China-specific chips – H20, L20, and L2 – meticulously engineered to comply with the BIS technical parameters. As the flagship model targeting AI training, the H20's specifications reveal its inherent "compliance logic." While sharing the Hopper architecture with the banned H100, the H20 features precise performance reductions.

For instance, the H20 delivers 148 TFLOPS of FP16/BF16 compute performance, starkly lower than the H100's 1979 TFLOPS at the same precision. Crucially, the H20's calculated TPP falls below the BIS 4800 TPP control limit, achieving compliance. Memory configuration stands as one of the few areas untouched; the H20 boasts 96GB of HBM3 memory, up from the H100's 80GB, alongside increased memory bandwidth of 4.0 TB/s versus 3.35 TB/s.

Interconnect capabilities present significant focus: the H20 maintains single-card NVLink bandwidth parity with the H100 at 900 GB/s, ensuring efficient intra-node GPU communication. However, a critical bottleneck emerges in cross-node cluster networking. U.S. restrictions on high-speed networking equipment prevent Chinese clients from accessing NVIDIA's Quantum-2 InfiniBand solution – capable of 400 GB/s (3.2 Tb/s) per node – essential for building DGX SuperPOD-scale clusters. This limitation drastically reduces communication efficiency for Chinese cloud providers requiring thousands of interconnected GPUs for large model training, escalating both difficulty and cost.

The H20's journey from conception to approval proved arduous. Originally slated for early 2024 launch, repeated delays ensued, widely attributed to extended U.S. government review. The July 15 announcement of an assurance marked the end of this prolonged wait.

**Market Challenges Amidst "Assurances"** The U.S. government's "assurance" of a license holds substantial commercial weight, enabling NVIDIA to initiate critical activities in China: negotiating sales contracts, securing purchase agreements, planning production capacity, and preparing supply chains for the H20.

Simultaneously, NVIDIA unveiled another facet of its multi-product strategy, announcing a "new, fully compatible" NVIDIA RTX PRO GPU positioned as the "ideal choice for building digital twin AI in smart factories and logistics." This pivot towards industrial AI applications aligns with China's focus on manufacturing upgrades and "new quality productive forces." Crucially, this domain faces relatively lower export control risks, potentially offering NVIDIA a more stable foothold. Markets like digital twins and industrial automation could become vital entry points.

Yet, neither the conditional return of the H20 nor the exploration of new paths like the RTX PRO can evade a transformed market reality: intense domestic competition. The window created by U.S. export controls accelerated China's indigenous AI chip industry.

Market share shifts provide the clearest evidence. A May 2025 TrendForce report projected the proportion of AI servers in China relying on imported chips (like NVIDIA and AMD) would drop from approximately 63% in 2024 to 42% in 2025. Conversely, domestic chip suppliers, bolstered by national AI chip policies, are expected to capture 40% of the market in 2025, nearly matching imports.

Further underscoring this trend, IDC's *China Semiannual Accelerated Computing Tracker (1H 2024)* reported China's accelerated server market reached $5 billion in the first half of 2024, surging 63% year-over-year. Within this high-growth segment, a pivotal change emerged: servers powered by non-GPU chips like NPUs and ASICs grew explosively at 182%, reaching nearly $700 million.

At the underlying accelerator chip level, the momentum is equally pronounced. IDC data shows over 900,000 accelerator chips shipped in China during 1H 2024, with domestic AI chip brands accounting for nearly 200,000 units – roughly 20% market share. Notably, local server manufacturers like Inspur, H3C, and Ningchang dominated, capturing over 70% of the market by revenue. China's server supply chain is rapidly integrating with domestic AI chips, forging a distinct ecosystem.

IDC China AI Infrastructure analyst Du Yunlong noted that due to prior large-scale procurements, China's AI server market remains in a "stock market" phase, providing a buffer for domestic technology upgrades. He emphasized that as demand for industry-specific large models intensifies, "building ecosystems compatible with China's computing facilities will become the primary focus."

A landmark commercial case vividly illustrates the domestic acceleration. In May 2024, China Mobile announced winners for its 2024-2025 New Intelligent Computing Center procurement project, involving 7,994 AI servers at a total value of ¥19.1 billion (tax excluded). The winning vendors – Kunlun, Sugon, Powerleader, Yangtze Computing, FusionServer, Baixin, and Huakun Zhenyu – share a critical commonality: all are hardware partners within Huawei's Ascend ecosystem. Huawei's Ascend architecture effectively monopolized the project's technical route.

Within the high-demand internet sector, leading cloud providers have also decisively embraced domestic AI infrastructure. Huawei Ascend confirmed its hardware platform now fully supports and optimizes Alibaba's entire "Tongyi Qianwen" model series, from the 7B-parameter Qwen-7B to the 72B-parameter Qwen-72B. Baidu Intelligent Cloud offers users the option to deploy its ERNIE models onto Ascend hardware via its "Qianfan" platform. Tencent Cloud integrates its "Hunyuan" model with Ascend servers for enterprise solutions. iFlyTek was among the earliest to adapt its "Spark" model to the Ascend platform.

Consequently, NVIDIA's H20 re-enters a market far less dependent and no longer a vacuum. The company must not only justify the performance compromises inherent in its compliant design but also contend with formidable domestic competitors wielding advantages in price, service, and ecosystem synergy. Furthermore, for Chinese clients deeply invested in adapting software and infrastructure to domestic alternatives, the decision to partially or fully reintegrate NVIDIA technology presents a complex and challenging commercial calculus.