In April 2026, DeepSeek V4 β one of the most capable open-source LLMs β started running natively on Huawei Ascend chips. No Nvidia. No American GPUs. This is a signal that the Chinese AI stack is viable, and European enterprises finally have a real alternative.
What Just Happened?
DeepSeek, the Chinese AI lab that shocked the world with V3 in early 2025, has now fully optimized V4 for Huawei's Ascend 910 AI accelerators. The model runs on CANN (Compute Architecture for Neural Networks) β Huawei's alternative to Nvidia's CUDA.
The benchmarks are telling: inference performance on Ascend 910 is comparable to Nvidia A100 for large language models. At significantly lower cost. Without export restrictions. Without supply chain uncertainty.
"This isn't a theoretical proof-of-concept. DeepSeek V4 on Ascend is production-ready, deployed at scale in China, and now available for European enterprises."
Why This Matters for Czech & Slovak Enterprises
For years, European companies building AI infrastructure had one option: Nvidia. A100s, H100s, Blackwell β if you wanted performance, you paid Nvidia's prices and accepted their availability constraints.
But the landscape has changed:
- Sanctions & Supply: Export restrictions make high-end Nvidia GPUs increasingly difficult to source in Europe
- Pricing Power: With limited competition, Nvidia maintains premium pricing
- Lock-in Risk: CUDA ecosystem creates significant switching costs
- Data Sovereignty: Many enterprises need on-premise AI for compliance β cloud-only isn't an option
The Ascend Alternative
Huawei Ascend offers a genuine alternative:
- Available: No export restrictions, reliable supply chain
- Cost-effective: 30-40% lower TCO compared to equivalent Nvidia setups
- Capable: DeepSeek, Llama, Qwen β all running natively
- Sovereign: On-premise deployment, data stays in your data center
Real-World Use Cases
We're already seeing demand across sectors:
Banking & Finance: Internal document analysis, compliance checking, risk assessment β all on-premise, no data leaves the building.
Manufacturing: Technical documentation assistants, quality control AI, predictive maintenance models.
Telecommunications: Customer support automation, network optimization, fraud detection.
Public Sector: GDPR-compliant AI for citizen services, document processing, internal knowledge bases.
The Migration Path
Moving from Nvidia to Ascend isn't zero-effort β but it's far more manageable than most CTOs assume:
- CANN toolkit provides CUDA-compatible APIs for most operations
- PyTorch and MindSpore models transfer with minimal changes
- Inference workloads are easiest β training requires more planning
- Huawei's ecosystem is maturing rapidly with growing community support
"For inference-heavy AI workloads β chatbots, document analysis, code assistants β Ascend is ready now. DeepSeek proved it."
What's Next?
At INVEXTA Group, we're building the first specialized Ascend deployment practice in Central Europe. We've partnered with Huawei as a Official Huawei Partner, and we're working to bring Atlas 800 servers to our data center for testing and client demonstrations.
If you're evaluating AI infrastructure options, the question isn't whether Ascend is viable β DeepSeek answered that. The question is whether your enterprise is ready to diversify away from single-vendor dependency.
We're here to help you navigate that transition.