While ChatGPT’s wave swept the globe and domestic giants waged a “hundred-model war,” Huawei’s AI strategy seemed a bit “quiet.” It wasn’t until recently—with the full launch of Ascend AI cloud services and the release of Pangu 5.0—that a series of moves made people suddenly realize: that Huawei, which had already staged the “spare tire becomes the main tire” drama in chips and operating systems, is quietly completing a full-stack reconstruction from underlying hardware to top-level services in the deep waters of AI.
Background: why is Huawei AI’s “change” so critical?
As a technologist, I’ve been through more than one “earthquake” in the tech stack. From IOE (IBM, Oracle, EMC) to de-IOE, from virtual machines to containers—each upheaval in the underlying architecture meant a wholesale reshuffling of the upper-layer application ecosystem. Today’s competition in AI is, in essence, a new contest over the right to define “infrastructure.”
The OpenAI-NVIDIA alliance defined the mainstream paradigm of current AI compute: top algorithms + top GPUs. But there’s a fatal problem with this path: it’s built on a not-exactly-solid geopolitical and technological supply chain. For any economy or enterprise with long-term ambitions, betting the foundation of AI—the future skyscraper—entirely on someone else’s ground is a self-evident risk.
The importance of Huawei’s “change” lies not in its releasing a model with more parameters, but in its attempt to answer a fundamental question: when the most advanced compute is restricted, how do we build a sustainable, evolvable AI system? This is not a purely technical problem, but a complex systems engineering one involving chips, frameworks, ecosystems, and business. Huawei’s answer hides in the architectural details of this “big change.”
Technical breakdown: how were the “hard bones” of full-stack autonomy cracked?
Many people, when talking about Huawei AI, focus on Pangu’s impressive performance in weather forecasting and drug R&D. But that’s just the tip of the iceberg. What really supports these applications is the massive, self-contained “AI iceberg base” underwater. From bottom to top, we can break it into four layers:
Layer one: the Ascend compute foundation—the climb from “usable” to “easy to use”
Ascend chips are the starting point of it all. Unlike NVIDIA GPUs’ general-purpose computing architecture, Ascend adopts the “Da Vinci Architecture,” a heterogeneous computing architecture designed for the characteristics of AI computation.
|
Anyone who’s built enterprise systems knows that the hardest part of in-house chips isn’t a successful tape-out, but getting the upper software ecosystem to “be willing to use, and use well.” Huawei built a heterogeneous computing architecture called CANN (Compute Architecture for Neural Networks). You can think of it as an efficient “translator” and “dispatcher.”
- Ascend C, an Ascend compute language benchmarked against CUDA: Lets developers write operators using a programming model similar to CUDA C, lowering migration cost.
- Automatic differentiation and graph optimization: The compute graph emitted by frameworks (like MindSpore) undergoes deep optimization by CANN, including operator fusion, memory optimization, and pipeline parallelism, squeezing out hardware performance as much as possible.
Layer two: the MindSpore framework—not just a PyTorch alternative
If TensorFlow and PyTorch are “the Windows and macOS of the AI world,” then MindSpore’s ambition is to become “the HarmonyOS of the AI world.” Its core design philosophy is “natively adapted to large models and scientific computing” and “fusion of AI and HPC (high-performance computing).”
A key feature is the “unity of dynamic and static.” PyTorch’s dynamic graph is handy but hard to deploy; TensorFlow’s static graph deploys well but is hard to debug. MindSpore tries to unify them at the syntax level: one piece of code can be debugged in dynamic-graph mode (context.set_context(mode=context.PYNATIVE_MODE)) and switched with one command to a static graph for extreme performance optimization and deployment (context.set_context(mode=context.GRAPH_MODE)).
More noteworthy is its native support for “large-scale distributed training.” In Pangu’s training, it achieved:
- Auto-parallel: The system can automatically partition the model and data based on the compute-graph structure and cluster state, finding the optimal parallel strategy (a mix of data parallelism, model parallelism, pipeline parallelism, optimizer parallelism, etc.).
- Second-order optimizer support: Optimizers like LAMB, suited to very large batch sizes, are “first-class citizens” in MindSpore, with deep optimization.
Layer three: the Pangu large models—the “alchemy” of industry know-how
The cleverest part of the Pangu large models is that it avoids a head-on clash with GPT-4 in “general chitchat” and picks “industry reshaping” as its main battlefield. Its architecture is layered:
|
Take Pangu Weather as an example. Its breakthrough isn’t many parameters, but encoding physical laws into the AI model. Traditional numerical weather prediction (NWP) has to solve complex physical equations, running for hours on a supercomputer. The Pangu weather model discretizes the atmospheric dynamics equations and injects them as prior knowledge into the model architecture and loss function, teaching the AI to “think like a physicist,” thereby completing global weather forecasts in seconds—with accuracy surpassing traditional methods.
Layer four: Ascend AI cloud services—making compute as easy to get as water and electricity
This is the most commercially ambitious part of this “big change.” Huawei Cloud packages Ascend compute into cloud services, offering a “unified cloud training and inference” experience.
- ModelArts platform: Provides a full-process toolchain from data annotation, model training, and evaluation to deployment, with the underlying layer automatically scheduling Ascend clusters.
- “Train instead of recruit”: You can directly call the Pangu large-model API, or fine-tune on top of Pangu with your own data, paying for the resources actually used.
The essence of this layer is lowering the barrier to applying full-stack AI. Customers don’t need to understand Ascend chips, nor master MindSpore, nor even build a large AI team, to use the most advanced AI compute and model capabilities.
My cool-headed take: the “sweetness” and “burden” of full-stack autonomy
From an engineering standpoint, Huawei’s full-stack closed loop of “chip-framework-model-cloud service” is undeniably a feat. It demonstrates a technology roadmap for achieving self-reliance under extreme pressure. But as practitioners, we must keep our cool and see the other side of the coin.
1. The ecosystem’s “chicken-and-egg” dilemma remains severe.
NVIDIA’s CUDA ecosystem took over 15 years, accumulating millions of developers and forming an impregnable moat. Ascend and MindSpore are still in the “chaser” position. Although Huawei invests heavily to incentivize developers through the “Wotu Program,” true ecosystem prosperity requires a massive, spontaneous, commercially motivated choice. That takes time, and it also requires Huawei to show a unique, irreplaceable value beyond the NVIDIA-OpenAI combination. For now, that value point seems to be in “industry AI” and “extreme software-hardware co-optimization.”
2. The eternal game of performance vs. ease of use.
The advantage of full-stack optimization is a high performance ceiling and good power control (especially on the device and edge sides). But the cost is that it may sacrifice some flexibility and openness. When the academic world churns out new model architectures daily (like Mamba, MoE), can MindSpore adapt quickly and provide optimal implementations? This is a huge test of its team’s cutting-edge tracking and engineering capabilities. A framework that, in pursuit of extreme optimization, makes researchers feel it’s “hard to use” may hurt its penetration into the most innovation-active academic community.
3. The double-edged-sword effect of “spare tire culture.”
Huawei’s “extreme survival” assumption shaped its full-stack in-house strategy, which is undoubtedly a correct survival strategy. But in the commercial market, what customers need isn’t just “safety,” but “leading” and “cost-effectiveness.” If the overall cost-performance of Ascend cloud services (including performance, price, ease of use, and model richness) can’t significantly surpass NVIDIA-based cloud services, then it may mainly attract government and enterprise customers with a strong need for “self-controllability,” and struggle to become the first choice for the broad mass of internet companies and startups. How to go from “safe option” to “preferred option” is the core challenge of Huawei AI’s commercialization.
What this implies for product builders: how to find your ecological niche in an era defined by giants
Huawei AI’s path teaches a lesson to all technical product people: in fields where the foundational paradigm is defined by giants, a head-on clash is worse than a flanking breakout.
Look for “high-barrier, slow-changing” vertical domains: Don’t fixate only on “artificial general intelligence”—that dazzling but winner-takes-all track. Fields like meteorology, geology, pharma, and industrial manufacturing have deep professional know-how barriers, hard-to-obtain data, and relatively slow change. Digging deep here, using AI to solve pain points that have been unsolved in the industry for decades, makes it easier to build an advantage that’s hard to copy or shake. Pangu choosing industry large models follows this logic.
“Software-hardware co-design” is the ultimate weapon for differentiation: As algorithms become increasingly open-source and model architectures converge, software-hardware co-optimization becomes the new moat. The fusion of Apple’s M-series chips with macOS is the exemplar. For product managers, this means that when defining a product early on, you should think about whether you can do joint design at the chip, sensor, and algorithm levels—even just deeply customizing an FPGA or making good use of specific hardware instruction sets may bring a generational leap in experience.
Keep the complexity to yourself; give the simplicity to the user: Huawei’s Ascend AI cloud service approach is worth learning. However complex the underlying layer (in-house chips, in-house frameworks, cluster scheduling), the user-facing surface (cloud console, API, SDK) must be extremely simple. The core value of an AI product is not showing off how impressive the tech is, but letting users solve real problems faster, better, and cheaper. The essence of productization is encapsulating complexity and delivering certainty.
Building a “data flywheel” is more sustainable than chasing a “parameter flywheel”: The competition in large models once descended into a param arms race. But Huawei’s practice in industry AI shows that high-quality, high-value, industry-specific data is the scarcer resource. If a product can design a clever mechanism to, while serving users, legally and compliantly accumulate unique, well-annotated industry data continuously, it can form a data moat that gets smarter with use and is hard for latecomers to catch up with.
Conclusion
This “big change” in Huawei AI is far more than releasing a few new products. It’s a grand narrative and solid practice about technological-system autonomy and independent development paths in the AI era. It tells us that in the high frontier of technology, there are no shortcuts; you must start from the most fundamental physical and mathematical principles and build the whole edifice step by step.
This road is destined to be lonely and hard, requiring enormous strategic resolve and long-term investment. It may not be the most efficient path, but it may be the most solid and the most resilient against storms.
For each of us technologists, the story of Huawei AI is both an inspiration and a reminder: while chasing the trendiest algorithms and models, it doesn’t hurt to occasionally look down and check whether the “foundation” underfoot is solid. Because in the end, what can support great applications is always that complete, solid, controllable, and ever-evolving technology system. This, perhaps, is the engineer’s “long-termism.”
(The end)
评论
0 条评论