China Mobile’s acquisition of 6,208 AI accelerator cards, 776 super-node systems, and a 2.06 billion yuan contract isn’t just another procurement deal—it’s the first major crack in China’s AI computing landscape, a domain where even NVIDIA CEO Jensen Huang holds no sway. A 2 billion yuan computing order, with not a single NVIDIA card in sight.

The CUDA Empire

To understand the significance of China Mobile’s move, we must first grasp what NVIDIA CUDA once symbolized.

CUDA is far more than a mere programming tool; it is the “lingua franca” of the AI world. For the past decade, nearly every global tech company training AI models has relied on CUDA at the code level. Every line of PyTorch code you write, every TensorFlow model you run, is underpinned by NVIDIA’s ecosystem. This ecological moat is more formidable than any hardware patent.

In 2023, NVIDIA’s data center business boasted a gross margin exceeding 70%, with its market value surpassing $1 trillion. This success isn’t solely built on GPU hardware performance but on the “taxation right” of the CUDA ecosystem: every AI company, every model trained, pays NVIDIA a toll.

One could argue that the world has long been under CUDA’s sway.

In this context, China Mobile’s procurement decision carries a hint of “rebellion.”

China Mobile’s purchase of 6,208 AI accelerator cards—equivalent to 776 super-node systems—with a procurement amount exceeding 2 billion yuan, explicitly specifies the adoption of Huawei’s CANN (Compute Architecture for Neural Networks) ecosystem in its technical specifications. The five winning bidders—Henan Kunlun, Changjiang Computing, Hua Kui Zhenyu, Baode Computer, and Huaqi Zhihui—all built their solutions around Huawei’s Ascend series chips.

It is reported that in the 2025 AI inference server procurement, CANN ecosystem devices already accounted for 70% of the total value. This super-node procurement represents a further leap, shifting from inference to training and from single-point to full-process coverage. Quantitative changes are triggering qualitative shifts.

A single procurement of domestic AI accelerator cards cannot shake CUDA’s ecological dominance overnight. However, the situation may change when the following three conditions are met simultaneously.

First, the chips must be viable. With high-end GPU exports to China continuously restricted, Huawei’s Ascend series chips are among the few mature options capable of providing large-scale AI computing power domestically. Notably, the Ascend 910 series and the latest 950PR products have been validated in training multiple leading large models. The “usable” threshold has been crossed for domestic chips.

Second, the software ecosystem must be ready. In August 2024, Huawei announced the full open-sourcing of CANN; by 2025, mainstream frameworks like PyTorch and TensorFlow had completed deep adaptation with Ascend, while Huawei’s self-developed MindSpore framework was rapidly iterating. This means the CANN ecosystem has largely resolved user adoption issues at the software level. As more users adopt it, the toolchain matures, fewer developers encounter issues, and confidence naturally grows.

Finally, there must be demand. China Mobile’s Jiutian large model matrix has deployed over 50 industry models, with nearly 200 million customers using AI-powered products. For a player of this scale, self-reliance and controllability have become an inevitable trend. These three conditions are now aligned. The tipping point for the domestic computing power boom is truly approaching!

The Tipping Point Arrives

The most striking aspect of China Mobile’s order is the complete absence of NVIDIA cards.

This isn’t due to China Mobile’s deliberate exclusion of NVIDIA but an inevitable outcome of geopolitical logic. Nevertheless, the result is clear: in the AI infrastructure landscape of China’s largest operator, CUDA’s penetration will gradually decline from near-ubiquity in the foreseeable future.

These 6,208 cards from China Mobile will be deployed between 2026 and 2027. The intelligent computing cluster formed by 776 super-nodes will become one of the largest AI training infrastructures within China’s operator system.

This is just the beginning. Operator procurements have strong demonstration effects. Once China’s largest mobile operator completes the localization of its AI infrastructure foundation, other industries and clients will follow suit at a much faster pace. Like a highway in China, once the first service area is built, the commercial ecosystem along the entire route will spring to life.

Of course, we must acknowledge the gaps. Challenges remain for the CANN ecosystem, such as the developer community still lagging behind CUDA in depth, the toolchain’s maturity requiring more scenario validation, and the performance gap with NVIDIA’s latest GPUs not closing in a year or two. However, these are issues of speed, not existence.

Moreover, the reshaping of the computing power landscape never happens overnight. It accumulates through bid after bid, order after order, and line of code after line of code, eventually converging into a perceptible change at a certain tipping point.