At the Ascend Artificial Intelligence Ecosystem Conference held at the Shenzhen Convention and Exhibition Center, Engineer Zhang paced back and forth at the registration desk with his reservation code—his automotive company has been waiting for three months to obtain a testing quota for the Ascend 910B chip. “It’s not that we don’t want to buy off-the-shelf products; it’s that we simply can’t get them,” he told reporters with a wry smile. “There’s a saying circulating in the industry now: getting an order for Ascend chips is harder than securing financing from a certain new energy vehicle company.”
This “chip shortage” that occurred in the summer of 2024 does not involve NVIDIA’s H100 but rather Huawei’s Ascend. From internet giants to traditional manufacturing enterprises, from AI startups to research institutions, a battle for Ascend computing power is reshaping the competitive landscape of China’s artificial intelligence industry.
1. “Desperate for Chips”: Who is Competing for Ascend?
“We just signed the procurement agreement for the next batch of Ascend 310P chips, but the delivery has already been scheduled for Q2 next year,” lamented the CTO of a leading autonomous driving company in an industry group. This company originally relied on overseas computing platforms to train models but was forced to switch to domestic solutions due to supply chain fluctuations earlier this year. After testing, they found that the computing density and energy efficiency of the Ascend 310P exceeded expectations—offering 16 TOPS (INT8) of computing power, supporting real-time processing of eight camera feeds, while consuming 30% less power than similar products.
Such stories are unfolding every day:
-
Large Model Vendors: A startup focused on industry large models used an Ascend 910B cluster to compress the training cycle of a model with hundreds of billions of parameters from 45 days to 28 days, reducing costs by 40%;
-
Smart Manufacturing Enterprises: A leading industrial quality inspection company in the Yangtze River Delta developed a defect detection solution based on the Ascend AI hardware and software platform, reducing the false detection rate on production lines from 2.1% to 0.3%, saving over ten million annually per production line;
-
Research Institutions: An AI laboratory at a top domestic university leveraged the elastic computing services provided by Ascend to shorten GPU usage time from a week of waiting to “immediate use upon application.”
According to the latest data from the Huawei Ascend community, over 5,000 enterprises have developed AI applications based on Ascend, covering ten key industries including manufacturing, healthcare, transportation, and energy. The most direct signal of the “chip shortage” is that the chip delivery cycle for Ascend’s computing industry ecosystem partners has extended from 8-12 weeks last year to 20-24 weeks now, with some popular models even requiring “prepayment to queue.”
2. Why is it “Difficult to Obtain a Chip”? The Hard Core Strength of Ascend
Behind the market’s vote is a dual breakthrough in the technical hard power and ecological soft power of Ascend chips.
1. Computing Power Foundation: Redefining the “Chinese Standard” for AI Chips
Taking the Ascend 910B as an example, this flagship chip, which competes with NVIDIA’s H100, uses advanced 7nm process technology, achieving 384 TFLOPS of FP16 computing power and 48 TFLOPS of full precision (FP64) computing power—this means it can easily handle “computing-intensive tasks” such as large model training and scientific computing. More importantly, the liquid cooling design of the Ascend 910B supports high power consumption of 800W, combined with Huawei’s self-developed CANN heterogeneous computing architecture, improving computing power utilization by over 30% compared to traditional solutions.
“We have tested that for the same ResNet-50 training task, the unit computing cost of the Ascend 910B cluster is 25% lower than that of overseas solutions,” admitted a technical leader from a cloud service provider. “More importantly, Huawei has not ‘locked’ the computing power like some vendors; the interfaces open to partners are flexible enough.”
2. Ecological Breakthrough: From “Chip” to “Full Stack” Closed Loop Capability
If the chip is the “heart” of computing power, then the ecosystem is the “vascular network.” The brilliance of Ascend lies in its early establishment of a full-stack system of “hardware + software + toolchain + industry applications”:
Hardware Layer: From the Atlas series modules, boards to servers, providing diverse computing power carriers;
Software Layer: The MindSpore open-source framework supports automatic parallelism and mixed precision training, allowing developers to “migrate to Ascend without changing a line of code”;
Toolchain: The MindX developer platform integrates over 300 pre-trained models and more than 200 industry SDKs, significantly lowering the barriers to AI application development;
Industry Layer: Collaborating with partners like ChinaSoft International and Dongfang Guoxin to launch large models for industries such as finance, electricity, and mining, effectively addressing scene pain points.
This “chip-framework-application” collaborative optimization makes the stickiness of the Ascend ecosystem far exceed the simple hardware sales model. As one developer who joined the Ascend community said: “Previously, when using overseas chips, we spent 60% of our energy solving adaptation issues; now with Ascend, 60% of our energy can be directly used to optimize business models.”
3. Behind the “Chip Shortage”: A Breakthrough for AI Sovereignty
The “difficult to obtain” status of Ascend is essentially a collective breakthrough for domestic AI computing power.
In recent years, China’s AI industry has developed rapidly, but core computing power has long relied on overseas sources. The “supply chain disruption” crisis in 2022 made many enterprises realize that without a self-controllable computing foundation, AI innovation becomes a “tree without roots.” The emergence of Ascend has precisely filled this critical gap—it not only provides computing power that matches international top levels but also reduces the migration costs for the entire industry through an open ecosystem.
“Currently, 80% of the companies consulting us about Ascend come with a clear demand for ‘domestic alternatives,'” revealed the sales director of a partner in the Ascend ecosystem. “However, after collaborating, they value the long-term benefits brought by Ascend more: such as lower computing costs, more autonomous technology control, and the opportunity to participate in defining industry standards.”
This shift is reshaping the underlying logic of China’s AI industry. As more and more enterprises develop applications based on Ascend, and as more developers contribute code around MindSpore, China’s “computing power sovereignty” is transitioning from concept to reality.
Conclusion: A Small Chip, Crucial for the Future
Standing in the exhibition hall of the Ascend Computing Industry Ecosystem Conference, watching the real-time computing power data on the big screen—over 100,000 Ascend chips have been deployed nationwide, supporting over 200 billion AI inference requests daily. These pulsating numbers are not only a market choice but also a strategic move for a country in the AI era.
While the “difficulty in obtaining chips” will eventually ease, the changes brought by Ascend will not stop. It proves that Chinese technology companies can not only “produce” high-end chips but also “utilize them well” and “thrive in the ecosystem.” Perhaps this is the most precious gift that this “chip fever” leaves for China’s AI industry: once the seeds of independent innovation sprout, they will eventually grow into a forest that supports the future.
Source:Behind the Scarcity of Huawei Ascend Chips: A Breakthrough Battle for Domestic AI Computing Power Please delete if infringing