The Nvidia Era Is Over — At Least in China

DeepSeek’s next flagship model, V4, will run exclusively on Huawei Ascend chips. Not as a backup. Not as a proof of concept. As the entire inference stack.

That’s not a press release talking point. That’s a tectonic shift.

For years, China’s AI labs quietly depended on Nvidia silicon — H100s, A100s, whatever they could get their hands on through official channels or creative workarounds. That dependency is now ending, and it’s ending fast.

The Numbers Tell the Story

DeepSeek V4 is a beast: roughly one trillion parameters using a Mixture of Experts (MoE) architecture, with 37 billion parameters active per inference call. It’s efficient by design — MoE lets you scale capacity without scaling compute linearly.

The hardware underneath? Huawei’s Ascend 910C processors, and the newer Ascend 950PR chips that Alibaba, ByteDance, and Tencent have ordered by the hundreds of thousands.

Individually, the Ascend 910C delivers about 60% of the H100’s raw performance. On paper, that looks like a gap. In practice, it’s a different calculation entirely.

Cluster Economics Beat Spec Sheets

A single chip benchmark doesn’t tell you much about running trillion-parameter models. What matters is cluster-level throughput, interconnect bandwidth, power efficiency, and — critically — cost.

Huawei’s Ascend chips are domestically produced. No export license anxiety. No supply chain chokepoints in the Taiwan Strait or a Commerce Department office in Washington. The chips show up when you order them, and you can order as many as you want.

When your alternative is paying a premium for smuggled hardware or begging for export waivers, 60% per-chip performance with 100% availability starts looking like the better deal.

The Export Control Paradox

Washington’s chip export controls were supposed to slow China’s AI progress. Instead, they may have done something far more consequential: they forced China to build its own hardware stack from scratch.

Before the restrictions, Chinese labs had no urgent reason to abandon Nvidia. The chips were better, the software ecosystem (CUDA) was mature, and switching costs were enormous. Export controls removed the option of staying comfortable.

Now Huawei has a captive market of the world’s most demanding AI customers, a government willing to subsidize the transition, and a clear mandate to close the performance gap. Every generation of Ascend chips will get closer to parity. And every generation will make the return to Nvidia less likely.

Two AI Worlds Are Forming

This isn’t just about chips. It’s about the emergence of two parallel AI ecosystems with different hardware, different software stacks, different supply chains, and increasingly different incentives.

The US ecosystem runs on Nvidia GPUs, CUDA, and a network of Taiwanese and South Korean fabs. The Chinese ecosystem is converging on Huawei Ascend, domestic software frameworks, and mainland fabrication.

DeepSeek V4 running on Huawei silicon isn’t a symbolic gesture. It’s proof that the second ecosystem works. That trillion-parameter models can train and serve on non-Nvidia hardware at scale.

What Happens Next

The short-term impact is straightforward: DeepSeek ships V4 on Huawei chips, proves it works, and other Chinese labs follow. The Ascend order books fill up. Nvidia loses a market it once dominated.

The long-term impact is harder to predict but potentially larger. Two competing hardware ecosystems mean two competing innovation paths. Competition could accelerate progress on both sides — or it could fragment the global AI research community in ways that slow everyone down.

One thing is clear: the assumption that cutting off chip exports would maintain American AI dominance is being tested in real time. And the early results aren’t what Washington expected.