DeepSeek 开源华为昇腾软件栈，试图降低对 Nvidia 的依赖

DeepSeek is exploring another way to challenge Nvidia: making competing hardware easier to use, even without matching its chip performance.

On September 30, DeepSeek open-sourced programming infrastructure for Huawei’s Ascend platform, including computing and communication libraries. The companies also highlighted work on a supernode built around 128 Ascend 950 processors, according to Reuters.

The challenge is captured by a striking estimate from months earlier. At a May investor meeting, DeepSeek founder Liang Wenfeng said training a model comparable in scale to the leading US systems could require roughly 50,000 Nvidia GB300s or 200,000 Huawei Ascend 950s. The Information reported that it had confirmed the authenticity of the meeting transcript. These were Liang’s estimates of the resources needed for a training effort, not a controlled benchmark proving equivalent performance between the two platforms.

The number point to an awkward trade-off. Why invest in making a hardware platform easier to use if it might require four times as many processors?

The answer begins with what DeepSeek can actually get its hands on and leads to a part of Nvidia’s business that is much harder to see than a chip specification.

Why DeepSeek Needs Another Way to Compute

For an AI lab, computing capacity is useful only when it is available in sufficient quantities, at a workable cost, and in a form its engineers can use.

US export restrictions have limited Chinese companies’ access to advanced Nvidia processors. Domestic chips offer another route, but supply remains a constraint. In his May remarks, Liang described computing resources as the principal gap between Chinese AI labs and their US counterparts. He also acknowledged that Huawei’s production capacity was limited.

Nor is switching to domestic hardware an automatic way to cut costs. In September, Reuters reported that Huawei had raised the indicated price of its Ascend 950DT accelerator card to more than 250,000 yuan, up 20% to 50% from earlier quotes, as a global shortage of high-bandwidth memory pushed up costs for Chinese chipmakers.

The price of a processor is only part of the calculation. What matters is how much useful work a system can deliver, how much electricity it consumes, how reliably it runs, and how much engineering it takes to keep it productive.

DeepSeek is trying to make a constrained source of computing capacity more usable. Even if Huawei’s hardware requires more processors for a given training job, it becomes a more meaningful option if the software makes those processors easier to program and deploy.

A supply of chips that developers struggle to use is not much of an alternative. That brings the software stack into the centre of the story.

Quality takes time, and we’ve done the heavy lifting for you. Each X.PIN deep-dive is built on exclusive industry access and weeks of investigation. Subscribe now to support our work. Don’t just follow the trend—PIN it!

TileLang and the Cost of Switching

Nvidia’s advantage extends beyond its GPUs. CUDA has become embedded in the software, optimized libraries, tools and development habits that AI teams rely on. Years of work have accumulated around it. Moving to a different processor can mean rewriting kernels, adapting workflows and then tuning the results all over again.

DeepSeek’s work with TileLang targets part of that problem.

TileLang is an open-source, high-level programming language designed to make it easier to write high-performance kernels for different hardware platforms. TileLang gives developers a common programming layer for expressing operations, while the compiler handles the platform-specific work needed to target supported hardware.

DeepSeek’s TileKernels repository added Huawei Ascend support on September 30.

The project says its kernels now have a second backend, allowing the same Python APIs to run on Nvidia GPUs and Huawei NPUs.

The practical scope matters. TileKernels includes optimized operations used in AI training and inference, such as mixture-of-experts routing and quantization. These are pieces of the software that determine how efficiently a model runs. Developers can reuse parts of that work rather than rebuild every operation for a new platform.

TileLang does not make every CUDA program run on Ascend unchanged.Different processors still need different optimizations, and a portable programming interface cannot replace the entire ecosystem of libraries and production tools that has accumulated around Nvidia.

Still, the cost of switching need not remain as high as it once was. If developers can reuse more of their code and optimization work, testing an alternative becomes a more realistic option. For a company that cannot count on getting all the Nvidia hardware it needs, that is a practical gain.

What 128 Chips Have to Do with It

A supernode built around 128 Ascend processors brings a less visible challenge into focus: coordinating them efficiently.

Imagine a factory with 128 machines. Buying the machines does not guarantee high output. Materials must reach each station on time, work needs to be coordinated, and a bottleneck in one part of the line can slow everything down.

An AI system faces the same challenge. Processors need to exchange data, share work and access memory efficiently. If they spend too much time waiting for information to arrive, a cluster can have enormous theoretical computing capacity without delivering the performance its owners paid for.

DeepSeek and Huawei have worked on both sides of that problem: computing operations and the communication that connects them. The software release includes libraries for each, while the 128-chip supernode brings the system-level challenge into focus.

An optimized kernel on one processor is only one piece of the puzzle. At cluster scale, developers also need to keep data moving and prevent communication overhead from eating into the gains from adding more hardware.

Without this coordination, adding chips can increase the bill faster than it increases useful output.

How Open Source Could Make This Bigger Than DeepSeek

Opening the software changes who can benefit from the engineering effort.

If every AI lab had to solve the same low-level problems independently, building on a new hardware platform would be expensive and slow. Shared tools and optimized kernels give other teams a starting point. They can start with existing code, test it, identify problems and contribute improvements.

That is one way a hardware platform can attract developers before it has an ecosystem comparable to Nvidia’s. The initial software may cover only particular operations and workloads, but each usable component reduces some of the work needed to build on the platform.

X.PIN is a reader-supported publication. Feel free to share this work if you enjoyed it!

Share

There is also a potential deployment path beyond frontier-model training. On September 4, Bloomberg reported that DeepSeek planned to deploy at least 160,000 Huawei Ascend 950DT accelerators at a data center in Inner Mongolia. According to people familiar with the plan, the chips were intended for running models—inference—rather than training DeepSeek’s largest models. The timing would depend on Huawei’s production capacity.

The reported order points to a potential large-scale inference use for Huawei accelerators. It does not establish that DeepSeek has moved its most demanding training workloads away from Nvidia.

Inference is a major computing workload in its own right. Every model put into practical use needs computing resources to respond to users. A platform that can serve substantial workloads reliably has a reason to attract more software work, even while it is still developing its capabilities for frontier training.

The question is whether that demand, combined with open-source software development, can sustain a broader ecosystem. Other teams need to find that Ascend works for workloads they actually care about, not just that a demo runs or a backend exists.

Why Nvidia Should Pay Attention

The implications extend beyond China. For Western AI companies, the relevant lesson is the strategic value of reducing dependence on a single hardware platform.

When software is deeply tied to one supplier, switching becomes expensive. That gives the supplier an advantage beyond raw performance: customers must account for migration costs as well as the hardware itself.

A competing platform could attract customers by running selected workloads reliably, at acceptable cost and with less software re-engineering, even if Nvidia remains ahead on other tasks. For companies managing large computing budgets, that possibility creates room to negotiate, diversify infrastructure and reduce exposure to a single supplier.

The strategic effect would be a less rigid hardware market: customers could shift selected workloads, diversify capacity and negotiate from a stronger position—even if Nvidia remains the preferred option for their most demanding jobs.

DeepSeek and Huawei are betting that software can make up some of the ground their hardware still has to cover. The key question is whether developers can put enough real work on Ascend to make the alternative worth maintaining and improving.

For Nvidia, the longer-term risk is a gradual erosion of CUDA’s lock-in as developers gain viable alternatives.

Your support is what keeps X.PIN growing. If you enjoyed this, please leave a like and subscribe. We read every comment, so let us know what you think.