Thread by @wmhuo168

Nvidia’s China dream is dead.

Jensen Huang burned his bridges. Here’s how Chinese AI chipmakers are making him irrelevant.

NVIDIA's Next "China-Specific" Blackwell B30 AI Chip Rumored to See Significant Demand; Dominating The Inferencing Market share.google/Vc3PdnVLQyGCzv…

Jensen Huang is done in China.

His stripped-down B30 chip won’t save him. His Pentagon ties burned his bridges. And Chinese GPU makers are rising fast.

Nvidia is not irreplaceable. It’s obsolete. (1/15)

China has no illusion about Jensen's loyalties.
He went to Washington, got export bans.
He came to Beijing, got nothing.
Then he fled to Taiwan like it was a safe middle ground.
Beijing never forgot. (2/15)

Now he’s back with a neutered “Blackwell B30” chip.
No CoWoS.
No HBM.
No high bandwidth.
Just a low-performance placeholder trying to hold market share without violating U.S. export rules. (3/15)

China’s answer?
Build its own stack.
From silicon to software.
Backed by state capital, industrial policy, and market demand.
Nvidia’s monopoly is breaking. (4/15)

Here are the Chinese GPU makers rendering Nvidia obsolete.
Not in theory. In production.
Right now. (5/15)

🇨🇳 Biren Tech

BR100 chip rivaled A100

Cut off by TSMC, pivoting to SMIC

Focused on large model training

Deep state backing
Sanctioned? Yes. Crippled? No. (6/15)

🇨🇳 Moore Threads

Founded by Nvidia’s ex-China GM

General-purpose GPU for AI and gaming

Works with Baidu PaddlePaddle

Made on SMIC 7nm
They know CUDA. And how to kill it. (7/15)

🇨🇳 Innosilicon

Fantasy One GPU already in the wild

Also makes GDDR and LPDDR controllers

Building a vertically integrated ecosystem
Not flashy. Just effective. (8/15)

🇨🇳 Iluvatar CoreX

TI series chips for training and inference

Runs Transformer models at scale

Deployed in major Chinese data centers
They don’t talk much. They just execute. (9/15)

🇨🇳 Cambricon

Spun out of the Chinese Academy of Sciences

Focus on edge and cloud inference

Backed by deep research talent
Not a GPU in name. But a GPU in function. (10/15)

🇨🇳 Huawei Ascend

910B for training, 310B for inference

Powered by MindSpore, not CUDA

Runs Pangu models, vision models, and LLMs
You don’t need Nvidia when you control the full stack. (11/15)

🇨🇳 MetaX

Founded by ex-AMD engineer

Backed by Alibaba

Targeting general-purpose compute
Still early, but moving fast with local fabs and local software. (12/15)

The software problem is solved.
PaddlePaddle. MindSpore. OneFlow.
China doesn’t need CUDA anymore.
What it needs, it already has. (13/15)

The final nail: policy.
Beijing is banning foreign GPUs in state infrastructure.
Data centers. Research labs. Military applications.
If it’s not domestic, it’s not allowed. (14/15)

Jensen Huang isn’t bringing chips to China.
He’s bringing apologies.
But China doesn’t need either.
Nvidia is no longer a gatekeeper.
It’s a ghost. (15/15)