April 6, 2026 — In what may prove to be the most consequential AI hardware announcement of the year, Reuters confirmed on April 4 that DeepSeek’s upcoming V4 model will run entirely on Huawei’s Ascend 950PR chips — marking the first time a frontier-scale AI model has been built to operate on Chinese semiconductor infrastructure without relying on NVIDIA GPUs.
Why This Matters
Every major frontier model to date — GPT-5, Claude, Gemini — runs on NVIDIA hardware. DeepSeek V4 breaks that dependency. If the model delivers on its reported benchmarks, it demonstrates that Chinese-made chips can support cutting-edge AI, directly challenging the assumptions behind US export controls on advanced semiconductors.
According to Reuters, Chinese tech giants including Alibaba, ByteDance, and Tencent have already placed bulk orders for hundreds of thousands of Huawei Ascend chips in preparation for V4’s launch. Chip prices have surged 20% in recent weeks on the back of this demand alone.
What We Know About DeepSeek V4
DeepSeek V4 is a Mixture-of-Experts (MoE) model with approximately 1 trillion total parameters, activating roughly 37 billion per token. Key specs reported so far:
- Architecture: MoE with ~1T parameters, ~37B active per token
- Context window: 1 million tokens
- Modalities: Native text, image, and video generation
- Hardware: Huawei Ascend 950PR + Cambricon chips
- License: Expected open-source (MIT or Apache 2.0)
- Training cost: Reported at approximately $5.2 million
- Pricing: Estimated ~$0.28/1M input tokens — roughly 1/50th of Claude Opus pricing
A lighter variant, V4-Lite, has already been spotted in live testing on API nodes, with developers reporting a 30% inference speed increase and dramatically improved context recall (94% at 128K tokens, up from 45% on V3).
The Chip Story: Beyond the Model
DeepSeek reportedly denied NVIDIA early access to V4 while giving Chinese chipmakers an exclusive integration window. The company has spent months rewriting core code to work with Huawei’s CANN architecture instead of NVIDIA’s CUDA ecosystem.
This has three major implications:
- Proof of non-NVIDIA frontier AI. If V4 performs competitively, it demonstrates that US chip sanctions have not prevented China from building state-of-the-art models on domestic hardware.
- Shift in AI economics. Huawei chips at scale could push API prices down industry-wide. DeepSeek’s leaked V4 pricing (~$0.28/1M input tokens) would make it the cheapest frontier model by an order of magnitude.
- Ecosystem fragmentation. A successful non-CUDA AI stack could encourage other countries and companies to explore alternatives to the NVIDIA-dominated supply chain.
Expected Launch
DeepSeek V4 has been delayed twice — originally planned for mid-February, then March. Current reporting from Reuters and The Information points to a late April 2026 launch. The model is currently in its final “stress-test” phase.
Given the geopolitical timing — with a Trump-Xi meeting on the horizon — some analysts speculate the launch could be timed to demonstrate Chinese AI parity as a negotiating factor on chip export controls.
What This Means for AI Tool Users
For developers and businesses using AI tools, DeepSeek V4 could be a game-changer on price alone. If the model delivers near-Claude-Opus quality at 1/50th the cost, it would dramatically lower the barrier for AI-powered applications — from content generation to code assistance to data analysis.
The open-source release (if confirmed) would also enable on-premise deployment, giving organizations full control over their AI infrastructure without vendor lock-in.
We’ll be tracking this closely as the launch approaches. Check back for updated benchmarks and hands-on testing once V4 drops.
Sources: Reuters, The Information, FindSkill.ai, Huawei Central, TechWire Asia
Comments