When DeepSeek dropped V4 back in April, the AI world paid attention. Here was an open-source model that matched the performance of Claude-Opus-4.6, GPT-5.4, and Gemini-3.1 — at a fraction of the cost. But the real story wasn’t the benchmarks. It was the hardware underneath.
DeepSeek V4 was built to run on Huawei chips. Not NVIDIA. Not AMD. Huawei’s Ascend series — the same chips the U.S. government spent years trying to keep out of China’s hands.
And last week, a Huawei-led team proved they could do more than just run inference on those chips. They used 1,000 Ascend 910C processors to post-train DeepSeek’s 1.6-trillion-parameter V4-Pro model from scratch. Training, not just running. That’s a fundamentally different — and harder — computational task.
What Actually Happened
Post-training is where a base language model gets refined through supervised fine-tuning and reinforcement learning. It’s the step that turns a raw text predictor into something that follows instructions, reasons through problems, and produces useful output. It requires enormous compute — and traditionally, that compute has come from NVIDIA’s GPUs.
According to reporting by the South China Morning Post, the team used a cluster of 1,000 Ascend 910C chips to complete the post-training phase. GIGAZINE confirmed that this marked a significant step forward in China’s AI independence from NVIDIA hardware.
The Ascend 910C isn’t a secret prototype. It’s Huawei’s current flagship AI chip — designed, fabricated, and deployed entirely within China’s semiconductor ecosystem. And it just handled one of the most computationally demanding tasks in modern AI development.
Why This Matters More Than You Think
There’s a common misconception that China can only run AI models on domestic chips, not train them. Inference — the process of generating output from a trained model — is computationally lighter than training. Running DeepSeek V4 on Ascend chips was impressive, but it didn’t silence the skeptics who argued that Huawei’s hardware couldn’t handle the heavy lifting of model development.
This changes that narrative. Training a 1.6-trillion-parameter model requires moving massive amounts of data through billions of matrix operations, coordinating across thousands of processors, and maintaining numerical stability across weeks of computation. If the Ascend 910C can handle that workload, it means China’s AI ecosystem has reached a level of self-sufficiency that export controls were specifically designed to prevent.
Jensen Huang warned earlier this year that Huawei chips being used for DeepSeek models would be “horrible” for the United States. He wasn’t being dramatic.
The Numbers Tell the Story
Consider the economics. DeepSeek V4-Pro costs $1.74 per million input tokens and $3.48 per million output tokens. Compare that to OpenAI’s GPT-5.4 or Anthropic’s Claude-Opus-4.6, which charge significantly more for comparable performance. MIT Technology Review noted that V4 exceeds other open-source models like Alibaba’s Qwen-3.5 and Z.ai’s GLM-5.1 across coding, math, and STEM benchmarks.
Then in May, DeepSeek slashed V4-Pro pricing by 75% — a permanent cut, not a promotional one. That’s not the move of a company worried about compute costs. That’s a company that’s found a cheaper, domestic supply chain.
Meanwhile, the Stanford AI Index 2026 reported that China has “nearly erased” the U.S. AI lead. The flow of AI talent from China to America has slowed. Chinese AI papers are being cited more frequently. And the gap in frontier model performance? Shrinking fast.
What This Means for NVIDIA
NVIDIA’s dominance in AI computing isn’t just about having the best chips. When I looked at NVIDIA’s roadmap at Computex, the focus was entirely on next-gen consumer and data center GPUs. But the real competition isn’t coming from AMD or Intel — it’s coming from a Chinese company that wasn’t even supposed to be in the race. It’s about the ecosystem — CUDA, libraries, tooling, and the decade-long head start in making GPUs the default platform for AI workloads. That moat is real.
But moats dry up when countries build their own water supply. China isn’t trying to compete with NVIDIA on NVIDIA’s terms. It’s building an alternative stack — Ascend chips, CANN (Compute Architecture for Neural Networks) software, and models like DeepSeek V4 that are designed from the ground up to run on domestic hardware.
The U.S. export controls that were supposed to slow China’s AI progress may have actually accelerated this process. When you cut off access to someone’s supply chain, you don’t just limit what they can do today — you force them to build their own. And once they do, you’ve lost your edge permanently.
As I covered in my analysis of Google’s $85 billion AI capital raise, the AI arms race is fundamentally about compute access. If China can train frontier models on domestic chips at competitive cost, the entire economic logic of AI export controls collapses.
The Filipino Developer Perspective
Here’s where this gets personal for developers in the Philippines and across Southeast Asia. DeepSeek V4 is open-source. That means any developer, anywhere, can download it, run it, fine-tune it for local languages, and build applications on top of it — without paying OpenAI or Anthropic a cent.
For Filipino developers working on Tagalog NLP, government digitization projects, or startup MVPs, having access to a frontier-class model that runs efficiently on affordable hardware is a genuine game-changer. The regulatory landscape around AI is still taking shape, but the tooling is already here.
And if Huawei’s Ascend chips continue to improve — which the May breakthrough in chip design suggests they will — we might see affordable AI training hardware that doesn’t require a Silicon Valley budget. That opens doors for universities, small companies, and government agencies across the region.
Bottom Line
Huawei training DeepSeek’s largest model on its own chips isn’t just a technical milestone. It’s a geopolitical inflection point. The assumption that China needs Western hardware to build frontier AI is no longer true. The export controls that were supposed to maintain a multi-year lead? They just got a timeline check — and the answer isn’t favorable.
For NVIDIA, the immediate revenue impact is limited. Chinese customers were already largely cut off from their latest chips. But the long-term strategic impact is significant: if Ascend chips prove reliable for training at scale, other countries under similar sanctions — or simply looking for alternatives — now have a viable option.
The AI race just got more interesting. And for the first time in a while, it’s not a two-player game between American labs. It’s a three-way race between U.S. ecosystems, Chinese self-sufficiency, and everyone else trying to figure out which stack to bet on.