News Huawei reportedly acquired two million Ascend 910 AI chips from TSMC last year through shell companies

[Admin]

According to CSIS, Huawei has millions of AI chips from SMIC despite yield problems.

Huawei reportedly acquired two million Ascend 910 AI chips from TSMC last year through shell companies : Read more

 

[Pierce2623]

If TSMC doesn’t want your business, it’s pretty slimy to sneak if in like that. I used to claim that Huawei shouldn’t be treated as a direct proxy of the CCP, but they’ve clearly proven me wrong with this behavior.

 

[Marlin1975]

I don't get how this works? TSMC had to be in on this at least a little for it to work.

Company A can't buy custom designed chips but company B comes along and wants to make the exact same chips that would be protected by company A. So TSMC does not see an issue and just makes them?
Even if they thought they were licensed they would pick the phone up and make a few calls and run it past their legal office.

Just does not pass the sniff test.

 

[TCA_ChinChin]

If TSMC doesn’t want your business, it’s pretty slimy to sneak if in like that. I used to claim that Huawei shouldn’t be treated as a direct proxy of the CCP, but they’ve clearly proven me wrong with this behavior.

I don't think TSMC doesn't want their business, but that they have no choice but to comply with US sanctions against Huawei. If TSMC could take Huawei's money without upsetting the US, they would gladly do so.

 

[phead128]

Greg Allen is an MBA think-tanker with CSIS. He is one of the biggest proponents of export controls on China, which has failed spectacularly. His target audience is the dinosaurs in Congress who couldn't tell you the difference between an CPU and GPU. I would take what Greg Allen he says with a grain of salt.

 

[George³]

I don't understand. Huawei acquired chips that were their own development and probably ordered for TSMC to burn in wafer before the embargo on this lithography process took effect. Where is the problem?

 

[phead128]

I don't understand. Huawei acquired chips that were their own development and probably ordered for TSMC to burn in wafer before the embargo on this lithography process took effect. Where is the problem?

The question is how an MBA think-tanker who has "industry sources" knows more about sanction violations than even Bureau of Industry and Security (BIS) and Department of Commerce (DOC) and even the CIA/NSA/FBI ??

 

[AkroZ]

I don't understand. Huawei acquired chips that were their own development and probably ordered for TSMC to burn in wafer before the embargo on this lithography process took effect. Where is the problem?

Ascend 910C begin production in 2024, the embargo was instaured in 2019.
TSMC use many IP from USA, US law said they can restrain products worldwide when they use at least 25% US technologies.
To make a Ascend 910C with TSMC, globally around 40-50% american technologies are used.
Huawei invest heavily to reduce the dependencies on US technologies, but they are nowere near to produce "legally" Ascend 910C.

 

[bit_user]

I don't get how this works? TSMC had to be in on this at least a little for it to work.

Company A can't buy custom designed chips but company B comes along and wants to make the exact same chips that would be protected by company A. So TSMC does not see an issue and just makes them?

How would TSMC know what chip they're making? They're not in the business of reverse-engineering their customers' IP. By the time they get the design, it's not much more than a set of masks. Sure, you can figure out certain incredibly basic things about it, but probably nothing that's a dead giveaway. Given how many chips they make for how many different customers, many of which are not sold to the general public, I don't find this too surprising.

The main way you probably figure out what you're making is if you can require information about their downstream use, and then you actually do your own independent checks to see if the story is plausible and if that's where they ended up going. That costs money and takes personnel.

 

[tooltalk]

Greg Allen is an MBA think-tanker with CSIS. He is one of the biggest proponents of export controls on China, which has failed spectacularly. His target audience is the dinosaurs in Congress who couldn't tell you the difference between an CPU and GPU. I would take what Greg Allen he says with a grain of salt.

What's not working in so-called "targeted sanction" against China vs. country-wide sanction that many CCP shills oppose.

 

[Pierce2623]

I don't get how this works? TSMC had to be in on this at least a little for it to work.

Company A can't buy custom designed chips but company B comes along and wants to make the exact same chips that would be protected by company A. So TSMC does not see an issue and just makes them?
Even if they thought they were licensed they would pick the phone up and make a few calls and run it past their legal office.

Just does not pass the sniff test.

I don’t think it’s TSMC’s responsibility to examine every design to make sure they don’t I fring on others’ IP. That’s the responsibility of the people designing the chips. Small Asian companies are getting matrix math chips fabbed like crazy right now and TSMC would lose a ton of money skipping on small batch orders paying 2-3x the per-wafer price of their big customers. They don’t heavily examine anything coming in for n7 these days anyways….

 

[nivagh]

How would TSMC know what chip they're making? They're not in the business of reverse-engineering their customers' IP. By the time they get the design, it's not much more than a set of masks. Sure, you can figure out certain incredibly basic things about it, but probably nothing that's a dead giveaway. Given how many chips they make for how many different customers, many of which are not sold to the general public, I don't find this too surprising.

The main way you probably figure out what you're making is if you can require information about their downstream use, and then you actually do your own independent checks to see if the story is plausible and if that's where they ended up going. That costs money and takes personnel.

Indeed NOT reverse engineering its customers' designs, is of critical importance to being trustworthy as an outsourced fab.

 

[George³]

Ascend 910C begin production in 2024, the embargo was instaured in 2019.
TSMC use many IP from USA, US law said they can restrain products worldwide when they use at least 25% US technologies.
To make a Ascend 910C with TSMC, globally around 40-50% american technologies are used.
Huawei invest heavily to reduce the dependencies on US technologies, but they are nowere near to produce "legally" Ascend 910C.

The Ascend 910 is launched in 2019. I think that developing phase is earlier. The deals for production for 910 and future versions I think that is before embargo and it's are for years ahead.

 

[Pierce2623]

Indeed NOT reverse engineering its customers' designs, is of critical importance to being trustworthy as an outsourced fab. reverse engineering is so

The Ascend 910 is launched in 2019. I think that developing phase is earlier. The deals for production for 910 and future versions I think that is before embargo and it's are for years ahead.

You don’t seem to realize that the 910c is a newer model with higher compute performance and higher bandwidth compared to the 910 or 910B. It’s the equivalent of a 3090, 4090 and 5090 (not in numbers or performance but as in they’re different generations of products at the same tier)

 

[phead128]

What's not working in so-called "targeted sanction" against China vs. country-wide sanction that many CCP shills oppose.

Greg Allen and James Lewis of CSIS literally posted articles saying how export controls are ineffective and easily bypassed.

Gina Raimondo famously said it's a "fools errand" to try stop technological development in China using sanctions.

Doesn't stop these idiots from trying. But at what cost?

 

[bit_user]

Greg Allen and James Lewis of CSIS literally posted articles saying how export controls are ineffective and easily bypassed.

You're leaving out some key details, not sure why.

The Bureau of Industry and Security (BIS) said Tuesday that it’s struggling under a massively increased workload thanks to an outdated and almost entirely manual system.
...
The process used to collect licensing data was designed in 2006 and relies heavily on manual processes. BIS says this method consumes a considerable amount of its resources and includes a heightened risk of human error. While the bureau is working to improve its internal systems and processes, there’s still a heavy need for human interaction. Furthermore, the agency noted that its foundational systems were never designed to communicate seamlessly with one another."

https://www.tomshardware.com/tech-i...nforce-restrictions-on-chinese-tech-companies
​

More recently:

"According to the Senate report, the BIS only has 11 export control officers to cover the entire globe. This would make it nearly physically impossible for the agency to conduct end-use checks and physically verify even just a tiny fraction of the companies that order American chips. Aside from that, the Department of Commerce severely lacks the required numbers of subject matter experts in China and Chinese speakers, meaning it would have great difficulty in monitoring who gets U.S. semiconductors.

The committee suggests that the Commerce Department should get more funds to expand the BIS, allowing it to deploy more personnel to help enforce export controls. It also recommended that penalties for companies that violate the sanctions must pay larger fines and that their export controls must be checked by a third party, not just done in-house by the companies themselves."

https://www.tomshardware.com/tech-i...st-rely-on-voluntary-compliance-by-chipmakers
​

Okay, so:

• The agency is woefully understaffed, compared to the scale of the task.
• This is made worse by the processes they rely on being very manual and labor-intensive.
• The manual processes are error-prone.
• There's poor communication between systems & agencies.
• Oh, and the penalties for violators amount to basically a slap on the wrist.

It's a wonder that it works at all! This is not how you set yourself up for success.

Before decrying the effort as futile, maybe it would be a good idea to have some penalties with real teeth, modernize the enforcement mechanisms, and adjust staffing and contract out the work so you can meet it at the proper scale.

Gina Raimondo famously said it's a "fools errand" to try stop technological development in China using sanctions.

I think she was just making an excuse, in an attempt to deflect blame for the poor success rate. She is not a subject matter expert. She has a PhD in Sociology, not an MBA or any particularly relevant professional experience, and was overall head of the Department of Commerce (which does many other things, besides this). I don't give a lot of weight to her statement.

 

[Pierce2623]

Greg Allen and James Lewis of CSIS literally posted articles saying how export controls are ineffective and easily bypassed.

Gina Raimondo famously said it's a "fools errand" to try stop technological development in China using sanctions.

Doesn't stop these idiots from trying. But at what cost?

There is minimal cost. China would be working just as hard on fabs and spending just as much money on it either way. The US government will continue to throw out hurdles on that path to slow them down. Limiting their access to GPU compute has actually been a massive hurdle to running the simulations involved in designing new process nodes.

 

[phead128]

There is minimal cost.

Tell that to the $545 billion in lost revenue by US semiconductor companies due to export controls.

Summary of BIS Licensing Data
In total, from 2018-2023, the data analysis described above shows that BIS and its interagency partners reviewed 3,934 license applications that involved a PRC-based Entity Listed party. Of this total, 2,641 licenses totaling approximately $335 billion were approved, and 1,293 licenses valued at $545 billion were either denied, revoked, or RWA
Bureau of Industry and Security (BIS)

China would be working just as hard on fabs and spending just as much money on it either way.

The problem is Chinese semiconductor companies never gave domestic options as chance because foreign semiconductor companies were much more advanced, cheaper, reliable. Now the Chinese semiconductor equipment and suppliers is thriving thanks to the US sanctions, in ways not previously possible despite many decades of state-led investment and top-down orders.

The US government will continue to throw out hurdles on that path to slow them down.

As Gina Raimondo has said, it's merely a "speed bump", not enduring slow down, at the cost of $545 billion lost revenue for US semiconductor companies, and China still has been able to achieve 7nm nodes and D1z HBM3 which is more than sufficient for advanced GPU inference compute for LLM models.

Limiting their access to GPU compute has actually been a massive hurdle to running the simulations involved in designing new process nodes.

SMIC's CEO has said they have designs solidified for 5 and 3nm node has been completed in 2020 (source), it's a matter of implementing them using EUV lithography, which is the chokepoint, not designing new nodes. You don't need GPU compute to design new process nodes. Even sanctioned Huawei can design 3nm or 2nm nodes a long time ago, it's EUV that's the limiting step.

 

[bit_user]

Tell that to the $545 billion in lost revenue by US semiconductor companies due to export controls.

That's what the industry claims. I don't believe it. Those numbers must be wildly inflated. They were probably selling those licenses at like 1/10th the price, but then turned around and said they're "worth" full price. I'll bet that figure dwarfs the total, cumulative revenue of those companies, in the year prior.

The problem is Chinese semiconductor companies never gave domestic options as chance because foreign semiconductor companies were much more advanced, cheaper, reliable. Now the Chinese semiconductor equipment and suppliers is thriving thanks to the US sanctions, in ways not previously possible despite many decades of state-led investment and top-down orders.

As Gina Raimondo has said, it's merely a "speed bump", not enduring slow down, at the cost of $545 billion lost revenue for US semiconductor companies, and China still has been able to achieve 7nm nodes and D1z HBM3 which is more than sufficient for advanced GPU inference compute for LLM models.

The path China was taking was first to achieve parity with foreign competition, at any cost and by any means. SMIC wanted to do whatever it took to match TSMC. Then, backfill their supply chain with more and more indigenous solutions, over time.

The new strategy of blocking their access is forcing them to build out that indigenous supply chain first. The end state would be the same, but meanwhile their progress on lithography has virtually halted and their chip designers are greatly hampered. This way does force them to become more self-sufficient, sooner. However, it's also the best way to slow their progress.

SMIC's CEO has said they have designs solidified for 5 and 3nm node has been completed in 2020

All based on ASML and Western technology. They were close to reaching parity, yet experts now believe it has put China about 10 years behind:

https://www.tomshardware.com/tech-i...ouncil-believes-china-is-10-years-behind-tsmc
​

The key question is how best to use this current gap, because it does seem like only a matter of time before it's gone.

You don't need GPU compute to design new process nodes.

Turns out that it's surprisingly relevant there:

https://www.anandtech.com/show/21309/nvidias-culitho-gains-support-from-tsmc-and-synopsys
​

But the GPU sanctions are about AI, not semiconductor fabrication.

 

[phead128]

That's what the industry claims. I don't believe it. Those numbers must be wildly inflated. They were probably selling those licenses at like 1/10th the price, but then turned around and said they're "worth" full price. I'll bet that figure dwarfs the total, cumulative revenue of those companies, in the year prior.

Most of these US semiconductor companies are publicly traded, so they are not allowed to "lie" or "inflate" about their revenue decline or they get in serious violation with SEC.

Also, this figure is for a 5 year period (2018-2023), or about $108 billion per year in lost revenue, which is not unreasonable since China imports $385 billion in semiconductor every year, or 25% of the pie.

The new strategy of blocking their access is forcing them to build out that indigenous supply chain first. The end state would be the same, but meanwhile their progress on lithography has virtually halted and their chip designers are greatly hampered.

Tell that to the EUV lithography machine that China is building.


China's EUV breakthrough: Huawei, SMIC reportedly advancing LDP lithography, eye 3Q25 trial, 2026 rollout
China’s Homegrown EUV Machines Rumored for Q3 Trial Production, Spelling Trouble for ASML

Again, China is making rapid progress towards an EUV, really thanks to all the US sanctions.

If China achieves an EUV, it's really game over.

 

[bit_user]

Most of these US semiconductor companies are publicly traded, so they are not allowed to "lie" or "inflate" about their revenue decline or they get in serious violation with SEC.

They can't lie to investors about real financial data, but they can certainly make a biased case when talking to a government researcher who's writing a report about the impact of sanctions. Especially when the report is just talking about aggregates.

We see this all the time, like when it comes to the job market. The industry will claim vacancies are one thing, but they're using a methodology that's lazy and biased towards overcounting. If you had an infinite pool of qualified candidates, you'd quickly find their quoted number are overestimates, because it's in their interest to inflate the scale of the problem and not really worth any effort, on their part, to provide more accurate figures.

Tell that to the EUV lithography machine that China is building.

Sure, they can build it. But, let's see how long it takes to start using in production, and then how well it actually works, once they do.

If China achieves an EUV, it's really game over.

In the time it takes them to catch up to where the rest of the world was, the rest of the world will have advanced as well. Even when they can do EUV, they'll still be behind. The estimate was it'd take them 10 years to catch up.

 

[phead128]

In the time it takes them to catch up to where the rest of the world was, the rest of the world will have advanced as well. Even when they can do EUV, they'll still be behind. The estimate was it'd take them 10 years to catch up.

Do you really think 2026 is 10 years behind?

That's when the article says they expect full production. They will trial EUV in Q3 of 2025, and full production in 2026.

They actually said it will take 10 years for China to reach 7nm. They said it was impossible to achieve EUV... we'll see very soon.

 

[George³]

I don't know whether to trust the reasoning of someone who is biased towards a certain propaganda. Of course, expected revenues are not the same as realized. This does not mean that the companies are lying. It is correct to compare the revenues by year and to take into account all the factors for their growth or decrease. It cannot be denied that there are at least a few more major factors during the described period that reduce the percentage of sanctions.

 

[bit_user]

Do you really think 2026 is 10 years behind?

That's when the article says they expect full production. They will trial EUV in Q3 of 2025, and full production in 2026.

I'm not the expert, but let's say they hit 7 nm in 2026. They can't proceed beyond that at a pace much faster than everyone else did. By next year, everyone else will be ramped up on 2 nm. It took TSMC about 6 years to get from 7 nm to 2 nm. Let's say China does it in 4, so by 2030. Where will TSMC be, by 2030?

 

[Pierce2623]

Tell that to the $545 billion in lost revenue by US semiconductor companies due to export controls.


The problem is Chinese semiconductor companies never gave domestic options as chance because foreign semiconductor companies were much more advanced, cheaper, reliable. Now the Chinese semiconductor equipment and suppliers is thriving thanks to the US sanctions, in ways not previously possible despite many decades of state-led investment and top-down orders.

As Gina Raimondo has said, it's merely a "speed bump", not enduring slow down, at the cost of $545 billion lost revenue for US semiconductor companies, and China still has been able to achieve 7nm nodes and D1z HBM3 which is more than sufficient for advanced GPU inference compute for LLM models.


SMIC's CEO has said they have designs solidified for 5 and 3nm node has been completed in 2020 (source), it's a matter of implementing them using EUV lithography, which is the chokepoint, not designing new nodes. You don't need GPU compute to design new process nodes. Even sanctioned Huawei can design 3nm or 2nm nodes a long time ago, it's EUV that's the limiting step.

A “completed design” for a lithography machine you can’t build is a minimum of 5 years away from useful yields once you’ve actually successfully built a working prototype. They will also require tons of compute to optimize production level yields once they do have a successful production machine. Everything else is political stuff I have no interest in arguing. I couldn’t care less if massive companies lose revenue from import controls unless it negatively affects my retirement. Nothing you say can convince me that making compute harder to get and more expensive for China is hurting the US.