Less than 1%! NVIDIA's H200 returns to China — yet reveals a major underlying trend.

NVIDIA H200 has finally been sold to China.
On August 26, NVIDIA confirmed during its earnings call that H200 has been delivered to Chinese customers.
But there is another very telling number – data centre revenue from China accounts for less than 1%.
What does this mean?
In the past, when discussing China's AI computing power, the first question was: when can we get NVIDIA's most advanced GPUs?
But now, the question may have become:
"Does China's AI still need to rely so heavily on NVIDIA?"
This is what truly deserves attention in this news.

Because at the same time H200 was approved for entry into China, domestic substitution is undergoing a very important shift:
It is moving from "whether it works" to "whether the cost is worthwhile."

On August 26, Zhipu AI revealed during the GLM-5.3-Flash release that they had, for the first time, deployed domestic AI chips in large-scale traffic scenarios.
More importantly, Zhipu AI stated that through self-developed high-speed interconnects and cluster optimisation, the hardware efficiency and per-Token cost of domestic chips have reached levels comparable to mainstream NVIDIA GPUs.
Note this shift.
The real breakthrough for domestic chips is not about running a model – it is about running real business workloads.
Because "working" in the lab is completely different from handling hundreds of millions or billions of Tokens per day in production environments.

Tencent is also sending similar signals. In June this year, Tencent's management indicated they hope to secure more domestic computing power to support their cloud business in the second half of the year.
So what is happening now is:
NVIDIA is trying to re-enter China, while China's AI industry is working hard to "no longer be dependent on NVIDIA."
These two directions appear contradictory, but they are not actually contradictory.
Because what China truly needs to build is not a "domestic substitute for NVIDIA," but its own AI computing ecosystem.

01From Chip Specifications to Token Costs

From chips, to high-speed interconnects, to compilers, to model adaptation, to cluster scheduling – it all ultimately comes down to a very practical metric:
How much does each Token actually cost?
This means competition in AI chips is shifting from "who has better chip specifications" to "who can turn a chip into real productivity."

NVIDIA remains very strong.
Its latest quarterly data centre revenue reached approximately $89 billion, with total revenue of $96.2 billion – AI computing demand remains robust.
So this is not a story of "NVIDIA being replaced by domestic chips."
The real change is:
China's AI computing market is moving from single-supplier dependency toward multi-chip, multi-ecosystem, and multi-supply-chain dynamics.

02The New Logic of Enterprise AI Computing

The impact on enterprises is actually more significant than how many H200s were sold.
In the future, when building AI computing capabilities, enterprises will likely no longer simply ask:
"Do you have NVIDIA?"
Instead, they will ask:
"For my models, which chips deliver the lowest total cost, the most stable supply, and the lowest migration risk?"

This means AI computing is truly moving from "buying GPUs" into the era of "computing engineering."
And the fact that H200's revenue in China accounts for less than 1% precisely illustrates:

NVIDIA has reopened a door to the Chinese market, but China's AI industry is working hard to ensure this door is no longer the only one.