The Illusion Behind China’s AI Compute Boom

In the past two years the story of domestic AI compute hardware has dazzled the market, with shipment volumes projected at 4 million cards in 2025 and a domestic share of 41%, rising to over 55% in Q1 2026 while Nvidia’s share in China falls to just above 40%. In government and financial sectors the penetration exceeds 70%, and inference workloads show a 62% share for Chinese chips.

However, field visits to data‑centers reveal a stark contrast: the proportion of truly large‑scale deployments of domestic compute hardware is under 10%. Many small‑ and medium‑size data‑centers even use consumer‑grade Nvidia 5090 GPUs for inference instead of the professional AI accelerator cards promoted by vendors.

Top‑tier domestic chips such as Huawei Ascend, Cambricon, and others match or surpass the hardware specifications of Nvidia’s H200 series, yet enterprises still favor overseas products because the total cost of ownership extends beyond the hardware price.

Cost comparison scenario: For a medium‑size enterprise needing 100 compute cards for AI inference, the Nvidia option costs 1.5–2 times more per unit but offers plug‑and‑play deployment within two to three weeks, a mature CUDA ecosystem covering most open‑source models and AI frameworks, and global technical support that allows a five‑person ops team to maintain the system. The domestic option reduces hardware cost by 30–40 % but incurs hidden costs: extensive model‑porting effort (months of work or inability to run models), actual performance often only 60–70 % of advertised specs, incompatibility between different Chinese chip brands requiring complete re‑engineering, and the need for a ten‑person adaptation team. Additionally, domestic chips have slower technical response and higher failure‑resolution costs.

The time cost is even more critical. A business that can launch with Nvidia in two to three weeks may face a delay of several months when switching to a domestic solution, losing market windows in the fast‑moving AI industry.

Root causes of the deployment gap:

Software stack lagging hardware: While Chinese AI chips boast impressive hardware parameters, their compilers, acceleration libraries, and adaptation frameworks are unevenly mature. Model migration often encounters missing operators, precision drift, and steep performance cliffs. In contrast, CUDA’s long‑standing ecosystem provides deep software support that cannot be quickly replicated.

Fragmented industry chain: Chip makers focus on single‑card performance, system integrators on hardware assembly, software teams on their own frameworks, and application vendors on business rollout. No entity coordinates end‑to‑end efficiency, leaving integration gaps that fall on the final user and causing severe industry‑internal friction.

Large‑scale cluster networking shortcomings: Training trillion‑parameter models requires massive multi‑card clusters. Domestic chips suffer from non‑standard interconnect protocols, insufficient intra‑cluster bandwidth, and immature power‑ and cooling solutions. This results in the “single‑card shines but cluster underperforms” situation, keeping Nvidia’s share above 85 % in large‑model training.

In summary, domestic AI compute has moved past the initial “do we have chips?” stage and now faces the practical challenges of usability, cost‑effectiveness, and ecosystem maturity. The key to real adoption lies in building a robust software stack, fostering coordinated industry collaboration, and strengthening engineering capabilities for large‑scale deployment.

Looking ahead, a full replacement of Nvidia will not happen overnight. Over the next two to three years Nvidia will remain the dominant choice for high‑end training in Chinese data‑centers, but continued progress in hardware and ecosystem will keep the domestic sector in a long‑term tug‑of‑war. The true turning point will be when enterprises, after comprehensive cost‑benefit analysis, find that domestic solutions deliver lower total cost, higher efficiency, and better long‑term returns. Achieving this will require joint ecosystem construction across chips, systems, software, applications, and operations.