During the 2026 China International Fair for Trade in Services, Moore Threads Technology Co., Ltd. CFO Xue Yansong revealed that NVIDIA's share of China's AI chip market has plummeted from 95% to under 8%, nearing negligible levels. He noted that NVIDIA's gaming cards are also gradually exiting the domestic market, signaling the collapse of overseas chips' long-standing monopoly in China's computing power sector.

In Xue's view, the retreat of overseas brands presents a historic substitution opportunity for domestic computing power, with the industry transitioning from a "usable" phase to a "highly usable" stage. However, he acknowledged significant gaps remain in core areas such as supercluster engineering and computing cost control, which require sustained breakthroughs.

Xue disclosed that domestic AI accelerator cards now command over 60% of the China market share and continue growing rapidly, advancing steadily into high-end computing scenarios. Data projections indicate China's total computing power will reach 1.6 million PFLOPS by the end of 2025, solidifying its position as the world's second-largest market. Through continuous technological iteration, domestic computing power has achieved multiple hardcore breakthroughs, fully shedding its low-end label.

Xue detailed that Moore Threads Technology Co., Ltd., in collaboration with national laboratories, used a computing cluster exceeding 12,000 cards to complete pretraining of a scientific foundation model with 236 billion parameters. The company is now tackling trillion-parameter models, marking a landmark domestic advancement. Additionally, a Peking University team developed a world model based on Moore Threads Technology Co., Ltd.'s S5000 GPU cluster, which topped Stanford University's leaderboard for over 47 consecutive days, achieving China's first full-stack native world model training.

Despite these achievements, Xue candidly addressed the industry's shortcomings, emphasizing that computing competition has evolved beyond single-card performance into large-scale cluster engineering. "A 10,000-card cluster is now industry standard, with 100,000 cards as the next core target. Yet even when we complete a 100,000-card intelligent computing cluster with JD.com by 2027, we'll still lag the U.S.," he pointed out. The U.S. has already deployed 200,000-card H200 clusters, with industry leaders planning 500,000-card and million-card clusters, even defining supercomputing levels by 1GW power scale, far exceeding domestic capabilities.

Ecosystem and cost remain the other two core challenges constraining scaled domestic computing adoption. Facing NVIDIA's CUDA ecosystem moat built over two decades, Moore Threads Technology Co., Ltd. leverages its proprietary MUSA unified architecture to achieve full-stack compatibility across hardware, drivers, algorithms, and applications, closing the ecosystem gap. On costs, Xue stressed that "internet applications have near-zero marginal costs, but every AI invocation incurs real computing expenses. Continuously reducing comprehensive computing costs is the industry's core challenge."

Xue concluded: "Domestic computing power has moved past the 'can it work' phase and is accelerating toward 'does it work well.' It can now fully support model inference and small-to-medium model training substitution, but in frontier areas like high-end foundation model pretraining, scientific computing, embodied intelligence, and world models, we still need persistent catch-up. We look forward to collaboration between industry and finance to solidify the core foundation of domestic AI computing power."

During the same fair, Beijing's Shunyi District hosted a park-level financial services promotion event, bringing together banks, insurers, and financial leasing firms with local industry authorities, industrial parks, and key enterprises. The initiative targeted park development pain points and enterprise growth challenges, delivering specialized, scenario-based financial solutions and establishing a regular "government-park-enterprise-finance" connectivity platform.