China Doesn’t Need Nvidia Anymore: What DeepSeek V4 Really Proves

China just shipped an AI model that runs on its own chips and ranks just behind Google’s best system on public benchmarks. That model is DeepSeek V4, and it’s a direct challenge to the idea that US export controls could keep China years behind in artificial intelligence.

Whether every detail of its training story is true or not, one thing is clear: China is rapidly building an AI ecosystem that no longer depends on Nvidia. And that shift could matter more than who has the single best model.

What DeepSeek V4 Actually Is

DeepSeek V4 is a large language model (LLM) released by Chinese company DeepSeek. On open benchmarks, it reportedly:

• Beats every rival open model in math and coding tasks
• Trails only Google’s Gemini 3.1 Pro in general world knowledge
• Comes very close to OpenAI’s latest frontier models

The headline claim: V4 was trained and engineered to run inside mainland China, on chips designed by Huawei and Cambricon, rather than Nvidia’s high-end GPUs.

That alone is a big deal. US export controls were meant to keep China three years behind on advanced AI hardware. Instead, China is now shipping near-frontier models that can run at scale on its own silicon.

How China Closed the Gap Without Matching Nvidia

Nvidia’s latest Blackwell chips deliver around 20 petaflops of AI compute per chip. Huawei’s Ascend 950 is weaker on raw performance. On paper, the US still has the faster silicon.

So how did DeepSeek get close to US-level models anyway? The answer is engineering around the hardware limits, not trying to copy Nvidia directly.

1. Mixture-of-Experts (MoE) Design

DeepSeek V4 uses a mixture-of-experts architecture. Instead of activating the entire model for every request, it only turns on a small subset of parameters that are most relevant to the task.

Think of it like this: instead of powering an entire factory to run one machine, you only switch on the exact machines you need. The result is:

• Less compute used per query
• Lower costs to serve users
• Better scalability on weaker or fewer chips

2. Extreme Token Efficiency

DeepSeek also focused on “token efficiency” — how much useful work the model can do per token (the basic unit of text it reads and writes) and per floating point operation (FLOP).

By squeezing more value out of each unit of compute, they reduce the amount of hardware needed to reach a given level of performance. That’s critical when your chips are slower and harder to get.

3. Huawei’s Supernode Interconnect

The real bottleneck in training giant AI models is often not the chip itself, but the bandwidth between chips. When you train a model with hundreds of billions of parameters, those parameters have to communicate constantly across the entire cluster.

If that interconnect is slow, your cluster crawls, no matter how fast each chip is. Nvidia solved this with NVLink, a proprietary high-speed fabric that has become the gold standard in the US.

Huawei’s answer is its “supernode” interconnect. It’s not a copy of NVLink, but a homegrown replacement built entirely inside China, with no American IP in the signal path and no US licensing choke point.

In practice, that means China has built its own “highway” for AI training. Once you control your own highway, you care a lot less about who owns the toll booths elsewhere.

The Open-Source Play: Building Dependency, Not Just Models

One of the most important details: DeepSeek V4 is open source. Unlike most top American models, it can be downloaded, used, and modified freely.

Chinese open models are now leading the world in downloads on Hugging Face, with DeepSeek alone surpassing 75 million downloads since January 2025. That’s not just a vanity metric — it’s how you build a global user base and, more importantly, a global dependency.

When developers in Nigeria, Indonesia, Brazil, or the Gulf build their apps on top of a Chinese open model, they’re not just choosing a piece of software. They’re plugging into:

• Huawei’s hardware ecosystem
• China’s data formats and standards
• A regulatory and infrastructure environment outside Washington’s control

This is what strategists call a “sphere of influence.” In past centuries, Britain built one through shipping standards, and the US built one through the dollar. China is now trying to build one through affordable, open AI stacks.

If you’re curious how far that openness goes in practice, we’ve also broken down the hardware realities in a detailed look at whether you can actually run DeepSeek V4 locally.

Did DeepSeek Really Train on Huawei Chips?

This is where things get murky. A senior former US official told Reuters that DeepSeek’s latest model was actually trained on Nvidia Blackwell chips inside China — a likely violation of US export rules.

According to that account, thousands of Nvidia GPUs were allegedly dismantled in third countries and smuggled into China, then used to train V4. Nvidia has publicly called this “far-fetched” and says it hasn’t seen evidence of such a scheme.

So which story is true? Right now, we don’t know for sure:

• Version A: DeepSeek smuggled massive amounts of Nvidia hardware and secretly trained on it.
• Version B: DeepSeek really did train a near-frontier model primarily on Chinese silicon, and the smuggling story is political cover for a failed containment strategy.

Either way, the chip ban doesn’t look great. If Version A is true, the controls are being bypassed at scale. If Version B is true, China has already crossed the line the ban was meant to hold.

The Real Pivot: Training vs Inference

There’s a subtle but crucial distinction here: training vs inference.

• Training is the one-time, extremely expensive process of creating the model.
• Inference is the ongoing, day-to-day process of running the model to answer user queries.

Most of the long-term cost, electricity use, and hardware dependency lives in inference, not training. That’s where 90% of the real-world footprint sits.

Analysts note that DeepSeek V4 can run natively on Chinese chips at scale. Even if some training relied on Nvidia hardware, the model has been engineered so that its ongoing operation — the part that matters for an AI-powered economy — does not.

In other words, DeepSeek didn’t need to “win” training. It needed to win inference. And on inference, it now has a path that doesn’t require permission from Washington.

Is China Now Ahead in AI?

Not yet. According to the Stanford AI Index 2026, US companies still hold a slight lead at the absolute frontier. Nvidia, OpenAI, Anthropic, and Google remain at the top of the stack for now.

But the performance gap is closing fast. Chinese companies have effectively caught up on many practical benchmarks, and they’re doing it under the most aggressive semiconductor blockade in modern history.

The deeper shift isn’t about a single leaderboard. It’s about this:

• The US still has the best chips.
• China is building an AI economy that increasingly doesn’t need those chips.

That’s why we’re seeing a wave of free and low-cost Chinese AI tools — from language models like DeepSeek V4 to creative tools like the new wave of no-signup Chinese AI video generators. Together, they form an ecosystem that’s cheap, open, and very attractive to developers in emerging markets.

What This Means for the Future of the Chip War

Back in 1983, the US tried to slow the Soviet Union with a chip export ban that was supposed to buy a decade of advantage. It bought about 18 months. Under pressure, the Soviets stopped trying to copy US chips exactly and started building different ones.

China appears to have learned that lesson. Instead of chasing Nvidia spec-for-spec, it’s building:

• Different chips
• Different interconnects
• Different model architectures
• And a global open-source AI stack that locks in long-term dependence

The US still leads at the very top end of AI hardware and models. But this week showed something more important: China can now ship powerful, open-source AI that runs on its own standards, at its own prices, and is available to every developer on Earth.

The chip war was supposed to buy the US a decade. So far, it looks like it bought about three years — and pushed China to build an AI economy that no longer needs American chips to thrive.