TL;DR: US export controls have delayed China’s top-tier AI frontier development, but they have also accelerated Beijing's transition toward domestic silicon self-reliance. While the Biden administration pursued a "freeze-in-place" strategy to limit Chinese access to advanced hardware, the Trump administration's policy shift in 2026 permits sales of lower-tier chips like the NVIDIA H20 to prevent Huawei from dominating China's domestic market.

Scott Bessent, a key economic adviser, argues that China's bid for artificial intelligence supremacy using hardware like NVIDIA GPUs is America's greatest macroeconomic and security threat. To address this threat, US policymakers are debating two distinct economic strategies: a "freeze-in-place" containment model versus a "sliding-scale" market engagement strategy. See our Full Guide on how these policies reshape global supply chains. The decision by the Trump administration to grant NVIDIA licenses for its H20 graphics processing unit (GPU) is a departure from previous restrictions. This move aims to prevent Chinese national champion Huawei from monopolizing its domestic AI market. Bessent views the technological lead as a primary lever of geopolitical power, meaning whoever leads in AI infrastructure controls the next era of global finance and national defense. This 2026 policy pivot highlights a fundamental division in Washington on how to manage this risk.

How effective are US chip export controls in slowing China's AI development?

US chip export controls have delayed China's training of frontier-class large language models by forcing domestic labs to rely on inefficient, clustered legacy hardware. Biden's October 2022 restrictions aimed to cap Chinese node access at 14-nanometer logic and 18-nanometer DRAM. This forced Chinese firms to cluster thousands of lower-performance chips, increasing power consumption and latency. ByteDance, Tencent, and Alibaba face hardware constraints that make training models comparable to OpenAI’s GPT-4 or Claude 3.5 Sonnet highly expensive.

However, the containment strategy has suffered from enforcement gaps. ASML and Tokyo Electron continued to service advanced lithography equipment already inside China, enabling Semiconductor Manufacturing International Corporation (SMIC) to manufacture 7-nanometer chips for Huawei’s Ascend 910B series. SMIC relies on deep ultraviolet (DUV) lithography machines from ASML. Using DUV for 7-nanometer and 5-nanometer processes requires multi-patterning, which drastically reduces wafer yields to below 40% and increases production costs. This low yield prevents Huawei from scaling its Ascend 910B and 910C production to match the massive volumes required by Chinese hyperscalers.

Furthermore, proxy purchasing networks bypassed direct restrictions. The discovery that Chinese chip design firms Sophgo and PowerAir were front companies to secure advanced logic dies from Taiwan Semiconductor Manufacturing Company (TSMC) for Huawei reveals the limits of unilateral export bans. TSMC halted shipments to Sophgo after discovering the design matched Huawei's Ascend processor, illustrating the constant struggle of US export enforcement. These loopholes allowed Huawei to build an estimated 200,000 to 700,000 Ascend GPUs. Consequently, export controls have hindered China's top-tier AI capabilities but have not frozen them entirely.

The shift to a sliding-scale policy targets Huawei's market dominance

The US transition from a "freeze-in-place" ban to a "sliding-scale" licensing model aims to starve Chinese domestic chipmakers of revenue by allowing US companies to sell competitive mid-tier hardware. Under the previous "freeze-in-place" paradigm, the US set static performance thresholds for GPUs. This policy inadvertently created a protected domestic market for Huawei's Ascend AI series. Because Chinese cloud providers could not buy NVIDIA's flagship H100 or H800 chips, they purchased Huawei's Ascend 910B.

The new Trump administration policy reverses this approach. Following meetings between NVIDIA CEO Jensen Huang and President Donald Trump in July, the White House agreed to resume licensing applications for the NVIDIA H20 GPU. By allowing NVIDIA to sell the H20 and the RTX Pro GPU in China, Washington intends to undercut Huawei's sales. The H20 is a down-specced version of the H100, designed specifically to comply with the 4,800 total processing power (TPP) and performance density limits set by the US Department of Commerce.

While its raw computing performance is roughly 15% of an H100, NVIDIA optimized its high-bandwidth memory (HBM) capacity to make it useful for large language model inference. This software-hardware integration is NVIDIA’s strongest defense. Chinese developers are deeply integrated into NVIDIA's proprietary CUDA ecosystem. Switching to Huawei’s MindSpore or Ascend Community software requires rewriting millions of lines of code. The Trump administration's "sliding-scale" strategy leverages this software lock-in to keep China's tech ecosystem dependent on US standards. By late 2026, the success of this sliding-scale experiment will become clear as Huawei tries to deploy its upgraded Ascend series.

Will the NVIDIA H20 sales slow down China's domestic silicon self-reliance?

Re-introducing the NVIDIA H20 to the Chinese market will not halt Beijing's push for technology self-reliance, as Chinese firms now view US hardware as an unreliable long-term bet. Beijing's state-directed industrial policy treats dependency on US technology as a critical vulnerability. The Chinese government instructs domestic tech firms to source up to 40% of their AI chips locally. This procurement target aims to shield Chinese infrastructure from future policy shifts.

Even if the H20 provides superior performance per dollar, [Chinese cloud providers](/articles/chinese-tech-giants-are-building-the-ai-tools-that-could-replace- hollywood-studios/) like Baidu and Alibaba will continue to fund domestic alternatives. They must hedge against the risk that a future US administration might revoke NVIDIA's licenses. The H20's restricted memory bandwidth and interconnect speeds also limit its utility for training massive frontier models, making it a temporary patch rather than a permanent solution for Chinese buyers.

Furthermore, the Chinese Ministry of Industry and Information Technology (MIIT) continues to push mandates for state-owned enterprises and government agencies to replace foreign hardware with domestic alternatives. While private giants like Tencent and Alibaba might buy the H20 to run immediate cloud workloads, they will maintain parallel development pipelines using Huawei's Ascend chips. This dual-track strategy ensures they can survive a sudden complete ban on US technology. Consequently, this transactional policy change will likely fail to alter Beijing's long-term sovereign AI trajectory.

Key Takeaways

  • Hedge hardware dependencies: Global enterprises must expect continued volatility in US export policy, making a multi-cloud strategy with diverse hardware architectures essential for operations in East Asia.
  • Monitor software lock-in: NVIDIA’s CUDA ecosystem is the primary defense against domestic Chinese alternatives; businesses should evaluate how easily their AI workloads can migrate between CUDA and open-source frameworks.
  • Anticipate localized supply chains: Expect Beijing to accelerate its internal substitution mandates regardless of US licensing concessions, ensuring that Chinese cloud providers will maintain dual-track silicon strategies through 2026.