🎉 Giảm 20% cho đơn hàng đầu tiên!✦🚚 Miễn phí vận chuyển cho đơn trên 500.000₫✦⚡ Flash Sale — Số lượng có hạn!✦🎁 Quà tặng kèm cho đơn từ 1.000.000₫✦🎉 Giảm 20% cho đơn hàng đầu tiên!✦🚚 Miễn phí vận chuyển cho đơn trên 500.000₫✦⚡ Flash Sale — Số lượng có hạn!✦🎁 Quà tặng kèm cho đơn từ 1.000.000₫✦
DeepSeek trains V3 on Nvidia cards but no longer uses the Nvidia ecosystem: TileLang, Huawei 950, and the race to escape CUDA
Quay lại Blog

DeepSeek trains V3 on Nvidia cards but no longer uses the Nvidia ecosystem: TileLang, Huawei 950, and the race to escape CUDA

16 tháng 11, 20268 phút đọcVũ Khôi
English
#English
Vũ Khôi

Vũ Khôi

Tác giả tại Wotech Electronics Blog

Liang Wenfeng and the DeepSeek logo

DeepSeek trains V3 on Nvidia cards but no longer uses the Nvidia ecosystem: TileLang, Huawei 950, and the race to escape CUDA

Liang Wenfeng, the man behind DeepSeek, has just shared notable remarks about how the company is reducing its dependence on Nvidia — and the surprise is this: DeepSeek still uses Nvidia GPUs to train V3, but is finding ways to leave the software ecosystem long considered the US chipmaker's greatest advantage. According to a post by the page Hỏi Dân IT Official on 15 November 2026, the key is a high-level compiler called TileLang, along with the ambition to build independent computing capacity on Huawei hardware. The figures and assessments in this article follow the post and the evaluations cited in it, and have not been independently verified.

Quick answer: According to the post, DeepSeek still uses Nvidia GPUs to train V3 but no longer uses the Nvidia ecosystem: the company developed the high-level compiler TileLang, built its compute components on top of it, and actively controls the software layer between the AI model and the hardware, instead of letting the entire system depend on CUDA. As a result, when moving to Huawei GPUs, DeepSeek can reuse most of the work already done instead of rebuilding everything from scratch. According to the shared assessment, the Huawei 950 supernode can handle tasks equivalent to GB200 and GB300 systems with equivalent latency; a 50% higher cost is still acceptable, 100% higher can still be considered, and even a 200% gap is not too great a barrier if the goal is independent computing capacity — in exchange, Huawei must use around 4 cards to replace one Nvidia card and is about two years behind its rival in technology. DeepSeek has also published AI components for Huawei Ascend, including TileLang, DeepGEMM, DeepEP, TileKernels, and FlashMLA.

CUDA is Nvidia's parallel computing software platform: millions of developers, numerous AI frameworks, and countless compute libraries have been developed and optimized on it, creating a barrier that makes switching to another chip far more expensive than simply finding a GPU with equivalent specifications. TileLang, according to the post, is a high-level compiler developed by DeepSeek — the software layer between the AI model and the hardware, allowing control over computation without depending on a single chip vendor's ecosystem.

How does DeepSeek train V3 without "using the Nvidia ecosystem"?

According to the post, DeepSeek's approach is quite interesting. The company developed a high-level compiler called TileLang, then built its compute components on top of this platform. Instead of letting the entire system depend on CUDA, it actively controls the software layer between the AI model and the hardware.

The benefit is that when moving to Huawei GPUs, DeepSeek can reuse most of the work already done, rather than rebuilding everything from scratch. In other words, DeepSeek is solving the problem at its root: reducing dependence on the software platform first, then expanding hardware choices.

The quoted remarks attached to the post put it directly: when training V3, the company uses Nvidia cards but no longer uses the Nvidia ecosystem; instead, it wrote the high-level compiler TileLang first, then completed everything else based on the TileLang ecosystem — so it is almost no longer dependent on Nvidia's ecosystem, and only needs to repeat the process on Huawei cards.

How does the Huawei 950 supernode compare with GB200 and GB300, according to the cited assessment?

The most eye-catching part of the post is the set of figures about the Huawei 950 and Nvidia's GB200 and GB300:

Aspect According to the shared assessment
Task capability The Huawei 950 supernode can handle tasks equivalent to GB200 and GB300 systems
Latency Assessed as equivalent
50% higher cost Still acceptable
100% higher cost Still considerable
200% higher cost Not too great a barrier if the goal is independent computing capacity
The price to pay Around 4 Huawei cards must be used to replace one Nvidia card
Technology gap Huawei is about two years behind its rival, according to the cited assessment

The quoted remarks attached to the post also state a larger ambition: to completely reverse the previous prejudice that the domestic card ecosystem is very poor, with the assessment that the timing and favorable conditions are basically in place and only time is lacking — and that within a year, many people will realize the problem has basically been solved, with production capacity being what remains. The speaker expressed optimism about domestic computing capacity, along with the remark that on this point, Nvidia is "digging its own grave". This is a statement quoted in the post, expressing DeepSeek's side's view, and has not been independently verified.

Quoted remarks on TileLang and the Huawei 950 supernode attached to the original post

Why is CUDA Nvidia's greatest advantage?

According to the post, for many years Nvidia built its advantage not only with powerful GPUs but also with CUDA. Millions of developers, numerous AI frameworks, and countless compute libraries have been developed and optimized on this platform.

A business wanting to switch to another chip must account for the costs of rewriting software, optimizing performance, testing, and training its team. Those barriers make replacing Nvidia far harder than finding a GPU with equivalent specifications.

The most notable point in the DeepSeek story, according to the post, is not that Huawei has surpassed Nvidia, but that one of the biggest barriers protecting Nvidia's position is being challenged from the software layer itself. Once software is no longer locked tightly to one GPU architecture, AI hardware competition will enter a very different phase: a rival does not necessarily have to build a stronger GPU immediately — it can start by making software less dependent on Nvidia GPUs.

What tools has DeepSeek published for Huawei Ascend?

Notably, according to the post, DeepSeek has also published AI components for Huawei Ascend, showing the company is investing in the ability to optimize computation across multiple hardware architectures, rather than focusing only on model development:

  • TileLang — the high-level compiler, the foundation for the compute components.
  • DeepGEMM — a high-performance general matrix multiplication (GEMM) library.
  • DeepEP — a communication library for Mixture-of-Experts models.
  • TileKernels — compute kernels built on TileLang.
  • FlashMLA — an efficient implementation of the MLA (Multi-head Latent Attention) mechanism.

What does this mean for developers and businesses?

According to the post, for developers, compilers, kernel optimization, and high-performance computing systems may become increasingly important areas in the AI race — because competition is shifting down to the software layer between the model and the hardware.

For businesses, having more hardware choices means being more proactive about supply, costs, and infrastructure investment strategy. When access to advanced chips is restricted, the problem is no longer simply buying the strongest GPU at the cheapest price; more important is having enough computing capacity to keep developing large models without depending entirely on a single vendor.

The post closes with an open question: if Huawei GPUs cost twice as much but still get the job done, would companies be willing to switch to reduce dependence on Nvidia?

Frequently asked questions

1. What hardware does DeepSeek train V3 on? According to the post, DeepSeek still uses Nvidia GPUs to train V3, but no longer uses Nvidia's software ecosystem.

2. What is TileLang? According to the post, TileLang is a high-level compiler developed by DeepSeek; compute components are built on top of it, letting the company control the software layer between the AI model and the hardware instead of depending on CUDA.

3. How does the Huawei 950 supernode compare with GB200/GB300? According to the assessment shared in the post: equivalent task capability and equivalent latency; in exchange, around 4 Huawei cards are needed to replace one Nvidia card, and Huawei is about two years behind in technology.

4. How much higher a cost is still acceptable? According to the post: 50% higher is still acceptable, 100% higher can still be considered, and even a 200% gap is not too great a barrier if the goal is independent computing capacity.

5. Why is CUDA the biggest barrier protecting Nvidia? According to the post: millions of developers, AI frameworks, and libraries have been developed on CUDA; switching to another chip means the costs of rewriting software, optimizing performance, testing, and training teams.

This article was rewritten from a post by the page Hỏi Dân IT Official on 15 November 2026. The figures, assessments, and quoted remarks in this article follow the original post and the evaluations cited in it, and have not been independently verified.