AnswersPublished17 sources
DeepSeek and Huawei Are Co-Designing AI Chips—but Nvidia Independence Is Unproven
DeepSeek is reportedly moving from buying Huawei Ascend chips to helping shape future designs, with pretraining on Ascend 950DT planned as early as late 2026. Open source Ascend software and a 128 chip supernode connect DeepSeek’s model needs to Huawei’s hardware, but chip supply and high end memory remain constraints.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- DeepSeek is reportedly moving from buying Huawei Ascend chips to helping shape future designs, with pretraining on Ascend 950DT planned as early as late 2026.
- Open source Ascend software and a 128 chip supernode connect DeepSeek’s model needs to Huawei’s hardware, but chip supply and high end memory remain constraints.
- Running or adapting a model on Ascend for inference is not the same as independently training a frontier model on it.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: How is DeepSeek’s collaboration with Huawei evolving beyond a conventional chip supplier–customer relationship into co-designing future AI c. Article summary: DeepSeek and Huawei are moving toward hardware–software co-design: rather than simply buying Ascend chips and adapting its models afterward, DeepSeek is reportedly influencing how future Huawei chips perform calculations. Topic tags: general, news, general web, user generated, academic. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks
DeepSeek and Huawei’s relationship appears to be moving beyond a conventional supplier–customer arrangement. According to secondary accounts of an October 19 report by The Information, Huawei engineers have been working with DeepSeek on future chip designs, using the AI lab’s feedback on how models calculate and move data between processors.8
11 DeepSeek is also reportedly planning to begin pretraining on Huawei’s Ascend 950DT as early as late 2026.
8
23
The strategic idea is to optimize the whole system—chips, programming tools and model workloads—together. That could make Huawei hardware more useful for DeepSeek and other Chinese AI developers. But the plans are prospective, and the available reporting does not establish that Ascend can already train frontier models as reliably or economically as Nvidia hardware.
From buying chips to shaping designs
In a conventional arrangement, a chipmaker builds a processor and its customer adapts software to run on it. The October reporting describes a closer feedback loop: DeepSeek shares workload requirements, while Huawei engineers incorporate that feedback into decisions about computation and processor-to-processor data transfer.8
13
That kind of collaboration could help Huawei tune future Ascend chips to real model-training demands. The report also says DeepSeek plans to use Ascend 950DT chips for pretraining and that Huawei intends to prioritize DeepSeek for much of the early supply of the next-generation 960DT.8
11 Those are reported plans, not evidence that large-scale training has already succeeded.
Software is a key part of the partnership
Hardware alone does not make a usable AI-computing platform. Developers also need programming tools and software libraries to run workloads efficiently. On September 30, DeepSeek announced open-source programming infrastructure for Huawei’s Ascend chips, including compute and communication libraries; Reuters reported that Huawei provided development support.1
2
The companies have also advanced a supernode built from 128 Ascend 950 chips.1 Together, the software and system work provide a practical basis for testing how models and multi-chip workloads perform on Huawei hardware. They are better understood as parts of a broader effort to build an alternative ecosystem than as a drop-in replacement for Nvidia’s CUDA platform.
1
2
Inference progress is not the same as training independence
A model can run on a chip to generate answers—known as inference—without having been trained on that chip. That distinction matters when assessing DeepSeek’s progress.
DeepSeek’s V4 was adapted for Huawei Ascend, and Huawei said its processors were used in part of V4-Flash’s training. But reporting also described DeepSeek’s training as still relying on Nvidia’s ecosystem. These developments indicate growing experience with domestic hardware, not proof that DeepSeek can move all demanding training workloads off Nvidia.
The separately reported plan to deploy at least 160,000 Ascend 950DT chips at an Inner Mongolia data center concerns inference, according to reporting on the project. That planned installation should not be treated as evidence of a frontier-model training cluster. The reported plan to pretrain on 950DT chips is a distinct and more direct test of Huawei hardware for training.8
23
Roadmap and supply will shape what happens next
Huawei has brought the Ascend 960DT launch forward to the first quarter of 2027, with the 960PR scheduled for the third quarter, according to The Straits Times.24 Earlier availability could give Huawei and DeepSeek another chance to apply workload feedback to the platform, but a launch date does not guarantee enough chips for large-scale deployment.
Supply is a separate constraint. One account of the 950DT rollout says high-end-memory shortages are limiting production and that filling DeepSeek’s reported order could take more than a year. If supply is tight, improvements to software and chip design may not translate quickly into enough installed capacity for training at scale.
What the partnership could—and cannot yet—show
Co-design could help DeepSeek make better use of China’s domestic accelerators by aligning chip behavior, programming tools and model workloads. That is strategically relevant as Chinese companies seek alternatives to Nvidia amid US export controls.1
2
The evidence so far points to a developing effort, not a completed substitute. The 950DT pretraining plan, future-chip design work and open-source software are meaningful steps to watch; whether they deliver frontier training at competitive speed, reliability and cost remains unproven. The October 19 claims are available here through secondary accounts of The Information report, rather than its full original article.8
13
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "DeepSeek and Huawei Are Co-Designing AI Chips—but Nvidia Independence Is Unproven"?
DeepSeek is reportedly moving from buying Huawei Ascend chips to helping shape future designs, with pretraining on Ascend 950DT planned as early as late 2026.
What are the key points to validate first?
DeepSeek is reportedly moving from buying Huawei Ascend chips to helping shape future designs, with pretraining on Ascend 950DT planned as early as late 2026. Open source Ascend software and a 128 chip supernode connect DeepSeek’s model needs to Huawei’s hardware, but chip supply and high end memory remain constraints.
What should I do next in practice?
Running or adapting a model on Ascend for inference is not the same as independently training a frontier model on it.