China's government has quietly reversed its hardline stance on Nvidia chips, permitting four of its largest tech companies to purchase the company's H200 processors after an 18-month ban. In early July 2026, officials from China's Ministry of Commerce summoned representatives from ByteDance, Alibaba, Tencent, and DeepSeek to announce they could now submit applications to buy Nvidia's most advanced export-legal chips. The reversal contradicts nearly two years of public messaging about achieving complete AI chip independence through domestic alternatives like Huawei's Ascend processors.

What Changed in China's AI Chip Strategy?

The policy shift emerged without fanfare. No press releases announced the decision; news leaked through industry publications before Reuters confirmed it. The terms were restrictive: total purchases would be capped below 200,000 units across all four companies combined, with firms required to justify each purchase and pay Nvidia upfront. The H200, while capable, was already one generation behind Nvidia's H100 and two generations behind its newer Blackwell B200 chips available globally. Yet even this limited access represented a dramatic reversal from Beijing's April 2025 decision to ban even lower-tier H20 sales, which had gone beyond U.S. export restrictions.

The dominant narrative about China's semiconductor future had seemed ironclad. Analysts, policymakers, and even Chinese tech executives believed the country would achieve full AI chip independence within three to five years. The evidence appeared compelling: ByteDance had committed $5.6 billion to Huawei Ascend 950PR chips, Alibaba Cloud and Tencent had placed massive orders, and Huawei was targeting production of 750,000 Ascend 950 chips for 2026. DeepSeek had optimized its V4 model specifically for Huawei silicon, achieving immediate compatibility across Ascend's entire product line. Domestic chipmakers like Cambricon had posted their first profit after nine years of losses, with revenue surging 453 percent year-over-year.

Why Is China's Domestic Chip Push Falling Short?

The reversal stemmed from three mounting pressures that the triumphalist domestic-chip narrative had systematically underplayed: supply constraints, performance gaps, and the arithmetic of scale.

• Supply Shortfalls: Huawei's 750,000-unit production target for 2026 was already understood within the industry to be optimistic. By June 2026, supply chain reports indicated actual output was tracking closer to 400,000 to 450,000 units, enough to satisfy perhaps one-third to one-half of China's actual demand across all major cloud providers and AI labs.
• Performance Limitations: Huawei's Ascend 950PR delivered approximately 2.8 times the FP4 performance of Nvidia's H20, but still lagged on key metrics like memory bandwidth, which is critical for large model inference. DeepSeek's own engineers were reportedly frustrated by memory bandwidth limitations during V4 training, forcing the company to design the model with smaller context windows and more aggressive quantization than originally planned.
• Economic Inefficiency: Building a domestic chip industry from scratch is capital-intensive and inherently less efficient than purchasing from a mature global supplier. Every yuan spent subsidizing yield improvements and every engineer-hour spent porting code represented opportunity costs that China's AI labs could no longer ignore.

The math was brutal. ByteDance alone, serving over 100 million daily active users through its Doubao assistant and processing trillions of tokens monthly through its Volcano Engine cloud platform, likely needed 200,000 to 300,000 high-performance inference chips just to maintain current service levels. Add Alibaba's Qwen cloud services, Tencent's Hunyuan deployment, and the training clusters required for next-generation models, and China's domestic production was covering a fraction of actual demand.

How Are Chinese Tech Giants Responding to the New Policy?

The policy reversal signals that China's largest AI companies are pursuing a pragmatic hybrid strategy rather than full decoupling. Rather than betting entirely on domestic silicon, they are now positioning themselves to use both Huawei Ascend chips for certain workloads and Nvidia H200 processors where performance or compatibility demands it. This approach acknowledges a hard reality: the transition to fully domestic AI infrastructure cannot happen as quickly as Beijing had publicly committed.

The reversal also reflects shifting calculations about the long-term cost of independence. While Huawei's Ascend ecosystem has matured considerably, with a functional CUDA compatibility layer and CANN software stack, these tools still require manual optimization work that Nvidia's mature ecosystem does not. For companies racing to deploy cutting-edge models and maintain competitive advantage, the friction costs of domestic-only silicon had become prohibitive.

Nvidia's CFO Colette Kress had warned in May 2026 that China's AI chip market was becoming "structurally inaccessible," a statement that most observers had accepted as the final word on the company's prospects in the region. The July policy shift proved that assessment premature. While Nvidia's access remains tightly controlled and far below pre-2024 levels, the company has regained a foothold in the world's second-largest AI market at a moment when Chinese tech giants are racing to scale their models and services.

The quiet reversal exposes the tension between Beijing's sovereignty ambitions and the practical demands of competing in global AI markets. China's tech giants cannot afford to fall significantly behind their U.S. and European counterparts in model capability and inference performance. Achieving that performance requires access to the best available hardware, even if that hardware comes from a geopolitical rival. Beijing's decision to permit limited H200 purchases suggests the government has accepted this trade-off, at least for now.

" }