DeepSeek didn’t just release a new model. It released two—and both run entirely on Huawei’s Ascend AI chips, a move that rewrites the hardware dependency rules of the global AI race as of April 28, 2026.
Key Takeaways
- DeepSeek-V4 comes in two open versions—both designed for inference on Huawei Ascend chips, not NVIDIA
- The models are open, low-cost, and optimized for deployment without Western-made GPUs
- This marks the first time a major Chinese foundation model vendor has fully committed to domestic silicon at scale
- If performance holds, it could enable AI development in regions cut off from U.S. chip exports
- The move challenges the assumption that high-end AI requires NVIDIA’s CUDA ecosystem
China’s AI Play Just Got a New Chip
For years, the global AI arms race has had one constant: the need for NVIDIA’s high-end GPUs. Whether you were training a model in Mountain View or fine-tuning in Shenzhen, if you wanted speed and scale, you bought H100s—or tried to. But as of April 28, 2026, DeepSeek is betting that won’t always be true.
The company launched two versions of its DeepSeek-V4 model—both open, both optimized for inference on Huawei’s Ascend 910B AI chips. That’s not a footnote. It’s a strategic pivot. This isn’t just about software. It’s about detaching from U.S. semiconductor control, and doing it publicly, with documentation, weights, and benchmarks.
And they’re not hiding it. In the original report, DeepSeek stated plainly: these models are designed for deployment in environments where access to NVIDIA hardware is restricted or impossible. That includes large parts of China, of course—but also countries under U.S. export controls: Iran, Russia, Venezuela, and others.
That’s the subtext. The text is simpler: here’s a high-performance LLM. Here’s the chip it runs on. Here’s the code. Build on it.
The End of the CUDA Monoculture?
NVIDIA’s dominance isn’t just about raw compute. It’s about CUDA—the software layer that makes GPU programming possible at scale. Developers don’t just buy H100s; they buy into an ecosystem. Tools, libraries, documentation, community support. It’s not just faster chips. It’s fewer headaches.
DeepSeek-V4 challenges that. By fully optimizing for Ascend chips, they’re forcing a question no one in Silicon Valley wants to ask: can an alternative stack—MindSpore instead of PyTorch, CANN instead of CUDA—actually work at production scale?
Early benchmarks suggest yes, at least for inference. DeepSeek claims the V4 models achieve 92% of the throughput of equivalent models on A100s, but at a fraction of the cost and with no export restrictions. That’s not parity. But it’s close enough to matter.
What Huawei’s Ascend Stack Actually Delivers
- Ascend 910B chip: 256 teraFLOPS for FP16, 32GB HBM2e memory per chip
- CANN 7.0: Huawei’s compute architecture, now supporting dynamic batching and model parallelism
- MindSpore: Huawei’s open-source AI framework, now integrated with Hugging Face-style model hubs
- Latency for 70B model inference: 47ms per token in internal tests
- Power efficiency: 1.8x better than A100 in watt-per-token metrics, according to DeepSeek
These aren’t theoreticals. They’re numbers from the release. And while independent verification is still pending, the fact that DeepSeek is publishing them at all signals confidence.
Open Source as a Geopolitical Tool
Let’s be clear: DeepSeek isn’t a nonprofit. It’s not even pretending to be neutral. Open sourcing V4 on Huawei hardware isn’t altruism. It’s a play for influence.
By making the models freely available, they’re inviting developers in sanctioned or semi-sanctioned markets to build on a stack that doesn’t rely on U.S. components. That’s not just about cost. It’s about resilience.
Consider this: if you’re a fintech startup in Tehran or a logistics AI team in Caracas, your access to NVIDIA hardware is either nonexistent or routed through risky gray markets. Now, there’s an alternative. You can download DeepSeek-V4, deploy it on locally available Huawei servers, and start building.
That’s dangerous to the U.S. position. Not because the model is necessarily better. But because it creates an off-ramp from the Western AI ecosystem—one that’s documented, supported, and now, demonstrably functional.
The Cost Equation Changes Everything
DeepSeek isn’t claiming V4 outperforms GPT-4. It isn’t saying it’s the smartest model in the world. What it is saying is that it’s cheap and available.
The company reports inference costs of $0.0003 per 1,000 tokens on Ascend clusters—less than half the cost of running Llama 3 on A100s in public cloud setups. That’s not a minor difference. That’s the gap between prototype and production.
For startups, that’s survival. For governments, it’s scalability. And for Huawei, it’s validation. Every model run on Ascend hardware is a data point proving the stack works—and a step toward broader adoption.
Why Open Matters More Than Benchmark Scores
We obsess over MMLU, GPQA, and LiveBench scores. But in the real world, access matters more than accuracy.
A model that scores 80 on MMLU but can’t be deployed is useless. A model that scores 72 but runs on affordable, available hardware? That’s power.
And DeepSeek-V4 is fully open—weights, training data details, inference scripts, hardware tuning guides. That transparency means third parties can audit, optimize, and redistribute. It also means the model can evolve outside DeepSeek’s control.
That’s the real threat to the status quo: not that V4 is better, but that it’s escape-proof. Once it’s out, it’s everywhere.
What This Means For You
If you’re a developer outside the U.S. or EU, especially in a region with spotty access to NVIDIA hardware, DeepSeek-V4 is a lifeline. You can now build production-grade AI applications without begging for cloud credits or relying on sketchy hardware resellers. The tooling is there, the cost is low, and the performance is sufficient for most inference workloads. This is the first credible alternative stack that doesn’t require you to compromise on openness or performance.
If you’re building AI infrastructure, this is a wake-up call. The assumption that high-performance AI requires U.S.-made chips is eroding. You’ll need to evaluate Ascend-based deployments—not just for cost, but for compliance and risk diversification. And if you’re relying on cloud providers that haven’t integrated alternative hardware, you’re already behind.
Someone in Shenzhen just launched a model that runs on domestic chips, gave it away, and undercut the West on price. That’s not a warning. It’s a demonstration.
How long before others follow—not just in China, but in India, Brazil, or Indonesia—with their own Open Models on sovereign hardware?
Industry Reaction and Implications
The reaction from the industry has been mixed, with some experts praising DeepSeek’s move as a bold step towards reducing dependence on U.S. technology, while others have raised concerns about the potential risks of adopting a new, untested hardware stack. Companies like Alibaba and Tencent, which have significant investments in AI research and development, are likely to be watching DeepSeek’s progress closely, and may consider similar moves in the future.
Meanwhile, NVIDIA has downplayed the significance of DeepSeek’s announcement, pointing out that its own hardware and software stacks remain the gold standard for high-performance AI applications. However, the company may be forced to reassess its strategy in light of DeepSeek’s aggressive pricing and the growing demand for alternative hardware solutions.
Technical Dimensions and Challenges
From a technical perspective, the success of DeepSeek-V4 will depend on the ability of developers to optimize and fine-tune the model for specific use cases. This will require significant investments in software development, testing, and validation, as well as the creation of new tools and frameworks to support the Ascend hardware stack.
Additionally, there may be challenges related to ensuring the security and reliability of the model, particularly in sensitive applications such as finance, healthcare, and transportation. DeepSeek will need to work closely with its partners and customers to address these concerns and provide adequate support and maintenance for the model.
The Bigger Picture
The release of DeepSeek-V4 is a significant event in the global AI landscape, with far-reaching implications for the development and deployment of AI technologies. As the world becomes increasingly dependent on AI, the need for diverse, resilient, and secure hardware and software stacks will only continue to grow.
DeepSeek’s move is a clear signal that the AI industry is entering a new era of competition and innovation, one in which alternative hardware and software solutions will play an increasingly important role. As the industry continues to evolve, it will be interesting to see how other companies respond to DeepSeek’s challenge, and how the global AI ecosystem adapts to the changing landscape.
Sources: AI Business, South China Morning Post
