LIVE
CHINA AI

DeepSeek open-sources Huawei Ascend kernels and plans massive Inner Mongolia AI supercluster on Ascend 950DT chips

In what analysts call "Jensen’s nightmare," DeepSeek has ported its core TileLang, DeepGEMM, and FlashMLA libraries to Huawei silicon while embedding Huawei engineers to co-design next-generation Ascend interconnects.

UpdateNext Staff
Editorial team
5 min read
1 view
Huawei Ascend 950 AI accelerator chip architecture
DeepSeek and Huawei are co-designing software kernels and optical interconnects around Huawei’s Ascend 950 series accelerators. (Photo: Huawei Central)

For years, Nvidia’s deepest moat in China was not just raw transistor density—it was CUDA, the proprietary software layer that made switching away from American GPUs painfully slow for frontier model researchers. Now, DeepSeek and Huawei are dismantling that barrier from the inside out.

According to semiconductor research from SemiAnalysis and reports in the South China Morning Post, DeepSeek has open-sourced full Huawei Ascend support across its entire low-level kernel stack—including TileLang (its custom GPU/NPU kernel language), DeepGEMM, FlashMLA (Multi-head Latent Attention), and DeepEP (expert-parallel communication libraries).

Embedded Huawei engineers and the Inner Mongolia Ascend 950DT megacluster

Rather than treating Huawei as a passive hardware vendor, DeepSeek has embedded Huawei HiSilicon and CANN software engineers directly inside its infrastructure teams. Operational telemetry from DeepSeek’s Mixture-of-Experts workloads is being fed directly into the silicon layout of Huawei’s Ascend 950PR/DT accelerators and upcoming Ascend 960 architecture.

Surpassing Nvidia in mainland data-center shipments

Backed by foundry partner SMIC’s multi-patterning DUV capacity, Huawei’s Ascend accelerator shipments have surged across mainland cloud providers in late 2026. DeepSeek is now preparing to anchor a flagship mega-scale AI data center in Inner Mongolia utilizing Huawei Ascend 950DT clusters to train and serve future generations of its reasoning models.

✓ WE KNOW
  • DeepSeek has open-sourced Huawei Ascend-compatible versions of its low-level programming language TileLang, matrix multiplication library DeepGEMM, FlashMLA, and MoE communication stack DeepEP.
  • Huawei engineers have embedded directly with DeepSeek’s systems team in Hangzhou to shape memory bandwidth and all-to-all interconnect topologies for Ascend 950DT and upcoming Ascend 960 chips.
  • Supply-chain filings indicate DeepSeek plans to anchor a massive new AI training data center in Inner Mongolia around Huawei Ascend 950DT supernodes.
? STILL UNCLEAR
    ADVERTISEMENT
    Get the daily AI briefingSubscribe
    “The symbiosis between DeepSeek’s kernel engineers and Huawei’s Ascend silicon team is narrowing the software ecosystem gap with Nvidia faster than Washington export planners anticipated.”
    — SemiAnalysis / South China Morning Post Hardware Report, November 2026
    SOURCES:South China Morning Post — Jensen’s nightmare: Huawei and DeepSeek narrow AI chip ecosystem gap with Nvidia ↗Dataconomy — DeepSeek and Huawei team up on AI chip design to challenge Nvidia ↗Huawei Central — Huawei Ascend chips surge in China AI market share ↗
    > how_we_reported_this
    • ✓Sources: written from primary sources and official statements, linked above.
    • ✓Fact-check: names, dates and claims checked against the original sources.
    • ✓AI use: researched with AI tools; written and reviewed by our editors. Our AI policy
    • ✓Independence: UpdateNext has no financial relationship with the companies covered.
    1 view
    > react:
    UpdateNext Staff
    The UpdateNext newsroom covers AI and tech from primary sources, every day. Every story is checked against official statements.
    Write for us →
    Comments
    Comments are moderated · Sign in to comment
    > read_next
    ENTERPRISE AI SHOWDOWN
    Alibaba launches Qwen Office 2.0 with autonomous 'Team Agents' across DingTalk as China’s AI office user base tops 140 million
    CLOUD TOKEN WARS
    China’s Model-as-a-Service (MaaS) market surges past 106 billion yuan as ByteDance Volcano Engine and Alibaba Cloud battle on tokens
    HUAWEI FULL-STACK AI
    Huawei debuts Xiaoyi Work AI agent on HarmonyOS PC and scales Ascend SuperNode clusters for DeepSeek V4-Flash workloads
    What's next in AI.

    The daily briefing on AI news, tools and breakthroughs.