Is Huawei New AI Chip Is Bad for Nvidia? New US Export Licensing Is the Real Issue (Part 2)
Apparently, on news of this new Huawei chip in WSJ, Nvidia’s stock decline -2% yesterday. That’s strange – not new news and the real issue is new US export license requirements.
Nvidia’s problem is not Huawei, but US restrictions on selling GPU to China. Nvidia announced writing off US$5.5bn on 9 April 2025 (for H20). AMD estimates US$800m write off (for MI308).
The best investment conclusion is that the China-US decoupling is now complete. Nvidia and AMD will not sell any AI chips to China. Same for HBM memory (SK Hynix, Micron, Samsung).
What is Huawei new AI processor Ascend 910C or D and Ascend 920? What is the performance?
What is this chip? How do we know the performance?
Huawei is 1) a private company 2) target of US sanctions since 2019 3) engaged in a bunch of Chinese Government, public security and defense business. As a result, transparency is very low. Huawei actually has reasons to not communicate on its semiconductors. The result is that it is hard to confirm information. Regarding this new chip:
Most reports call it 910C. Most reports think its made by SMIC at 7nm (also called 6nm or N-2), for ex TrendForce,
The WSJ thinks it’s called 910D,
SemiAnalysis and RAND Corp thinks 910C is actually made of two 910B manufactured by TSMC and stitched together (Apple and Nvidia also do this),
How do we know the performance of this chip? Apparently, there was a DeepSeek evaluation summarized in a weixin post here which is now gone
This says “this content violates the Internet User Public Account Information Service Management Regulations". That’s the regulations of CAC:
You’ll find some reports like this one and this one that mention that the evaluation of Ascend 910C was done by DeepSeek by a certain Mr. Yuchen JIN. Mr. JIN says he is the co-founder and CTO of Hyperbolic Labs, no association with DeepSeek on his bio. He has a very busy X page which shows this 1 single tweet:
A few sources have announced an Ascend 920 chip. The first source seems to be Digitimes and SCMP based on someone posting something on weixin which was also deleted.
As per sources above, Ascend 920 was supposedly announced in a Huawei event here but it was not the kind of event to unveil a new chip – it was a data center opening. It was not announced in Huawei’s Cloud conference (you can watch the presentations videos).
So that’s all folks. It took me 2 days to thread the needle and backtrack where all the “information” comes from. A lot of posts and articles referencing each other, a very clear lack of primary sources, some people are affirmative that the chips was made by TSMC, others admit that they just don’t know.
Ascend 910C made by SMIC at 7nm 2nd Gen
The most reasonable assumption is that Ascend 910C is made by SMIC at 7nm 2nd Gen, in my opinion. What’s important here? Is it a good chip?
Let’s use the little bit of data published by western websites on the chip. The “best” Chinese source is here and you’ll see there is very little on performance / specs. Let’s look at the efficiency of the chip: how much calculations do you get per Watt of electricity? Here is a comparison of Nvidia’s data center GPUs and Ascend 910C.
So you can see that these claims are in the ballpark:
Semianalysis Ascend 910C cluster uses 4.1x the power of a GB200 NVL72 (I calculate 4.25x)
Yucheng JIN: Inference performance on Huawei 910C achieves 60% of the H100's performance
TrendForce: Huawei’s Ascend 910C Reaches 60% of NVIDIA H100’s Inference Power
Is it bad? No, it’s pretty good given the limitations that Huawei is facing: only access to 7nm, without EUV, lower density, lower yields. Not bad.
CloudMatrix 384
Huawei does not sell its AI chip. Huawei builds big super computers and sells Cloud services (AI services) that run in its data centers. You could say its similar to AWS designing its own chips (Graviton, Trainium, Inferentia) and selling Cloud services, not chips. That’s what CloudMatrix 384 is, an AI super computer.
How much computing capacity, etc is available here (news.cn) and here (semianalysis).
I don’t doubt that comments and analysis of CloudMatrix 384 are correct. Again, Huawei rents Cloud capacity and doesn’t sell chips.
Impact on Nvidia and AMD
is it bad for Nvidia, AMD (GPU) and SK Hynix, Micron, Samsung (HBM Memory)? Yes obviously – but they’re already under export licensing, i.e. no export to China
Huawei new chip hit the news after both Nvidia and AMD announced possible write off on the latest US export license requirements to China.
On April 9, 2025, the U.S. government, or USG, informed NVIDIA Corporation that the USG requires a license for export to China (including Hong Kong and Macau) and D:5 countries, or to companies headquartered or with an ultimate parent therein, of the Company’s H20 integrated circuits and any other circuits achieving the H20’s memory bandwidth, interconnect bandwidth, or combination thereof. The USG indicated that the license requirement addresses the risk that the covered products may be used in, or diverted to, a supercomputer in China. On April 14, 2025, the USG informed the Company that the license requirement will be in effect for the indefinite future.
The Company’s first quarter of fiscal year 2026 ends on April 27, 2025. First quarter results are expected to include up to approximately $5.5 billion of charges associated with H20 products for inventory, purchase commitments, and related reserves.
On April 15, 2025, Advanced Micro Devices completed its initial assessment of a new license requirement implemented by the United States government for the export of certain semiconductor products to China (including Hong Kong and Macau) and D:5 countries, or to companies headquartered or with an ultimate parent in such countries (the “Export Control”). The Export Control applies to the Company’s MI308 products. The Company expects to apply for licenses but there is no assurance that licenses will be granted. The Company expects that the Export Control may result in charges of up to approximately $800 million in inventory, purchase commitments and related reserves.
These restrictions are older, 2 Dec 2024.
The best investment conclusion is that the China-US decoupling is now complete. Nvidia and AMD will not sell any AI chips to China. Same for HBM memory (SK Hynix, Micron, Samsung).