DeepSeek Open-Sources Full Ascend Toolkit: Closing the Software Gap in China's AI Compute Ecosystem

DeepSeek Open-Sources Ascend Components — A Quiet but Critical Move

On September 30, DeepSeek^Ad announced the open-source release of a complete set of AI infrastructure components for Huawei's Ascend platform. More importantly, these tools map one-to-one with their NVIDIA counterparts. For most users, this feels highly technical with no immediate impact — but in the context of China's domestic AI supply chain, it's a genuinely significant piece of the puzzle.

For a long time, the industry's understanding of AI chips has focused almost entirely on hardware performance, overlooking an equally critical layer: the software toolchain. NVIDIA's dominance in the AI compute market isn't just about fast chips — it's about CUDA, a software ecosystem so mature it has become the de facto standard. Easy-to-use tools, comprehensive documentation, a massive community, and near-zero friction for developers: that's the real moat.

What Was Open-Sourced — and Why It Addresses the "Software Gap"

In the past, getting large models to run on Ascend chips required developers to rewrite substantial amounts of code for adaptation — a costly process. Think of it this way: the same software running on different hardware essentially requires "rebuilding from scratch." This migration overhead has been an invisible barrier that prevented domestic AI compute from scaling in practice.

The components DeepSeek open-sourced this time offer fairly comprehensive coverage, including:

• Operator programming language: the building blocks of low-level computation
• Matrix operations: core operations for model training and inference
• Multi-card communication: essential for large-scale cluster coordination
• Long-context processing: optimizations for today's long-context workloads

The real highlight isn't the breadth — it's the consistent interface design. Code that enterprise developers previously ran on NVIDIA servers can now be migrated to Ascend simply by switching the backend, without large-scale rewrites. This dramatically reduces migration effort, and marks the first time Ascend has meaningfully closed the gap with the CUDA ecosystem in terms of software usability.

Joint Optimization with Huawei, Targeting Large-Scale Clusters

Beyond the open-source toolchain, DeepSeek also worked with Huawei to jointly optimize large-scale cluster solutions for Ascend 950, with deep tuning applied to the two most compute-intensive scenarios: multi-card training and large model inference.

The significance here is clear: a single chip's performance is just the starting point. What actually determines whether a large model can be deployed in production is how efficiently hundreds or thousands of cards work together. Engineering details like multi-card communication, cluster scheduling, and memory management often represent the gap between "peak paper performance" and "real-world usable compute." Joint software-hardware optimization is precisely what bridges that gap.

Multi-card communication is one of the most overlooked yet most performance-critical bottlenecks in large-scale AI training. When hundreds or thousands of chips co-train a single large model, each card holds only a portion of the model and data, requiring frequent exchange of gradients and parameters between cards. If communication bandwidth is insufficient or protocol efficiency is low, even the fastest chips will spend most of their time waiting. NVIDIA specifically designed NVLink high-speed interconnects and the NCCL communication library to address this, forming a complete hardware-software solution. The joint optimization effort with Ascend is aimed at closing exactly this gap — ensuring that real-world usable compute in a cluster approaches theoretical peak performance, rather than remaining a number on a spec sheet.

What This Means for Everyday Users

Ordinary people won't interact with any of this low-level code directly — but the indirect effects are real.

As the software ecosystem matures, more companies will be able to adopt domestic AI chips. Domestic large model training and AI services will have more homegrown hardware options to choose from, reducing sole dependence on foreign compute. More options for enterprises means more negotiating leverage and improved supply chain security — and over the long term, this should help bring down the cost of AI services, which will eventually be reflected in the everyday AI products we all use.

Closing the Gap ≠ Catching Up Entirely — Ecosystem Is a Long Game

It's worth keeping a clear head about this: it does not mean NVIDIA's ecosystem can be fully replaced overnight. Building an ecosystem is a long process — sustained hardware performance improvements, full model compatibility, community cultivation, and continuous tooling refinement all take time.

This open-source release closes a critical gap in the software layer — it is not a declaration of full parity. The ecosystem moat CUDA has built over many years won't be filled by a single open-source release. But the direction is right: once the hardware is built, the software toolchain must follow — only then can domestic AI compute actually deliver on its hardware potential.

Since its release in 2006, the CUDA ecosystem has accumulated nearly two decades of developer community, third-party libraries (such as cuDNN and cuBLAS), debugging and profiling tools, and vast amounts of production-validated code. Today, the vast majority of mainstream AI frameworks globally (PyTorch, TensorFlow, etc.) default to CUDA as their backend, and virtually all publicly released model code in academia assumes a CUDA environment. This means that even with interface compatibility achieved, Ascend still faces a long-tail adaptation challenge: large amounts of code that depend on specific CUDA behavioral details, specialized libraries not yet ported, and decades of accumulated developer habits all need time to migrate. The software ecosystem's moat is fundamentally a network effect built from countless details and ingrained habits — a reality that must be honestly acknowledged even after this important open-source step.

Conclusion

This is a genuinely positive development for China's domestic AI supply chain. It won't immediately change the daily experience of ordinary users, but it is an important piece of a much larger puzzle. As hardware, software, and toolchains are progressively filled in, domestic large models will have a far more solid foundation to build on. The shift from CUDA's monopoly to a world where domestic AI compute begins establishing its own software ecosystem may be a small step — but it points toward a long-term trend worth watching.

Background: What Is an "Operator"?

An "operator" (算子) is the smallest functional unit in a deep learning framework that performs a specific mathematical operation — convolution, matrix multiplication, activation functions, and so on. Every hardware architecture (such as NVIDIA GPUs or Huawei Ascend NPUs) has its own instruction set and memory access patterns, and operators must be written specifically for each target hardware to run efficiently. This is why porting the same model to a different chip requires "rebuilding from scratch" — the model itself hasn't changed, but the underlying operators need to be reimplemented and re-tuned for the new hardware. The operator programming language and related components open-sourced this time effectively provide Ascend with a set of "foundational building materials" equivalent to what CUDA offers — meaning developers no longer have to start from zero, significantly lowering the technical barrier to adopting new hardware.