A couple of years ago Bloomberg reported about spy chips/hw backdoors in SuperMicro mainboards but to my knowledge without a smoking gun proof. Maybe they had to settle outside of court and also had to sign papers to help protect the company from further damage in the future. Using (other) Bloomberg material may have triggered this. Of course this is a wild speculation. I have no evidence or insider knowledge.
GN used Bloomberg clips of US Gov officials speaking on AI chip matters, fully under fair use.
And Bloomberg did a DMCA takedown through youtube, copystrike in parlance which pulled the video down for a week. GN had no recourse other than to wait and counterclaim.
Week timed out, Bloomberg did nothing but be the bully.
They did have the video uploaded to archive.org (or at least link to someone else who did) and gave permission to anyone else to repost it. Which is how I saw it, some rando burner account on YouTube :)
It's sad to see what's happened to SuperMicro. They were one of the few vendors of server-grade hardware fitting standard ATX, mATX, and ITX form factors. In my experience their hardware was always better than the others who attempted to do the same (Gigabyte, Asus, ASRock). These days, motherboards with the features I want are going to be on AliExpress. Ironic considering this latest news is about putting trade barriers between the US and mainland China.
Remember when Singapore buyers were an abnormally high percentage of nvidia's revenue? You have to wonder if these companies are this brazen because they know the DoJ will have political pressure not to nuke the bubble which is more important than being China hawks.
Yep, same how the sales of German industrial CNC, machines, tools and lathes exploded in Russian neighbouring countries after 2022. Kazakhstan must be an industrial powerhouse by now with all that German machinery.
Can someone shed light on why China still couldn't copy the Nvidia GPUs in some form?
I understand its complex and there many parts to it, but which is the most complex part making it difficult for China to copy it?
Let's say they don't have access to 3nm process, what if they just use 12nm and create GPUs with much bigger size but comparable performance with CUDA compatibility? Or other option could be less tensor units, training will take longer, but they might be able to produce it cheaply
Copying CPUs isn't really a thing: they are too complex.
If you could steal all the designs at TSMC, and you had exactly the process that TSMC uses, you could definitely make counterfeits. If you didn't have TSMC's specific process, you could adapt the designs (to Intel or Samsung) with serious but not epic effort. If you couldn't make the processes similar (ie, want to fab on SMIC), you are basically back to RTL, and can look forward to the most expensive and time-consuming part of chip design.
This is nothing like copying a trivial, non-complex item like a car. Copying a modern jet engine is starting to get close (for instance, single-crystal blades), but even they are much simpler. I mention the latter because the largest, most resourced countries in the world have tried and are still trying.
If engines are hard to build, why not build a car 3x the size of a normal one, well you can but due to things like aerodynamics, etc etc you'll never match the speed or fuel economy of cars.
Same with chips, efficiency, speed, etc all depend on good design, and cutting edge factors, if the main reason your chip isn't faster is because of the distance between your L1 cache and your core is far, then having a bigger node process but bigger chip won't make it quicker.
But there's a distinct time/value of investment equation with the current AI boom. The jury is at best still out on what that equation is for the goals of capital (it's increasingly looking like there's no moat), but if you're a national government trying to encourage local bleeding edge expertise in new fields like this it's quite a bit more clear.
They can copy it. And no, the software moat is not there if someone choose the blatant copy route. They just can't build it in the scale they want yet.
> what if they just use 12nm and create GPUs with much bigger size but comparable performance
well, physics does work that way, depending on what you mean by performance.
(in the sense that power is normally part of performance when we're talking about chips).
you could certainly use a larger process and clone chips at an area and power penalty. but area is the main factor in yield, and talking about power is really talking about "what's the highest clockrate can you can still cool".
so: a clone would work in physics, but it would be slow and hot and expensive (low yield). I think issues like propagation delay would be second- or third-order (the whole point of GPUs is to be latency-tolerant, after all).
The timing is brutal - SMCI already had the accounting restatement scandal in 2024, spent months fighting delisting, finally got somewhat rehabilitated in the AI infrastructure boom... and now this. 25% single-day drop on a company that was already trading at a discount to peers tells you the market was still pricing in tail risk. For anyone tracking institutional holdings - the 13F filings from Q4 showed several funds adding back SMCI after the accounting mess cleared up. Those bets just got very painful.
I'd been assuming that the Chinese AI labs producing excellent LLMs despite the NVIDIA export restrictions was due to them finding new optimizations for training against the hardware they had access to.
I wonder if any of those $2.5B of smuggled chips ended up being used for those training runs.
combination of both, they published papers so we can clearly see they are not just duplicating old methods but coming up with new optimizations. ... yet we can't rule out that they used Nvidia. I don't even see how the export restrictions work, it's stupid. A Chinese company can go to another country, say France or Canada, setup a business buy a bunch of GPUs then make it available to their subsidiary in China. The export restrictions doesn't restrict usage/sharing/renting as far as I know...
They don't work. Chinese are skilled enough to desolder and smuggle just the ships themselves. They make the rest of GPU in-house. With more VRAM than the nvidia offers, comically, in case of 4090.
If they were using banned chips they wouldn't declare them in public papers. There have been multiple documented/alleged cases of chips being routed through Singaporean shell companies.
Did you think the hesitancy of westerners engaging and relying on Chinese labs was due to vibes? There are fundamental cultural differences at play, wether we are comfortable admitting that or not.
Simon, love your work. Hope this is sarcasm. If not, imagine the opposite: Sam Altman and co suddenly started producing tons of content about how smart they are in Mandarin. Why do they even need a story to begin with, let alone one they push halfway around the world?
The $2.5B number is just these guys. It could be 10x in total.
Well, also had other pen testers come forward saying that they had found implants on supermicro servers and had talked to federal authorities who had said it was a known relatively large issue they were trying to get a handle on whole keeping under wraps.
And if it were posted to move the market, that would have been about the most cut and dry SEC violation possible, posted at a time when the federal government still enforced such things.
Whenever some soylent-drinking, impossible foods-eating dilettante says "debunked" I find myself not fully believing them. And Supermicro has always been sus. I can't believe people are only just now noticing.
If you do, you could protect yourself with a sell stop below $17.25... because if it breaks that on weekly candles, next are $14 and $10. Or you could buy some calls instead when the volatility calms down. If you do it now, the volcrush could happen even if you're correct.
Not investment advice, do you own research. I'm just someone on the Internet.
I've had my own dealings with this awful company. Including Wally.
Let's just say that none of this comes as any surprise.
Now, what people should be asking is how much Jensen knew. In May he said there was nothing going on. But the videos of the Chinese guy holding H1/200's ... never got to him?
Also interesting how they waited until just after GTC...
Those claims were never confirmed, no? Some of it might be true or trueish but I'm not talking Bloomberg's anonymous sources word for it, and with so much supermicro gear out there you would think some other evidence would show up.
It depends on what you consider confirmed. It was kind of corroborated, at least. There was a CEO of a hardware security firm that came forward after the original article. He claimed that his firm had actually found a hardware implant on a board during a security audit. It wasn't exactly as Bloomberg described, though.
His take was that it was very unlikely that it impacted exclusively Supermicro, though.
I don't think it was a confirmed story. That is, the tiny "grain of rice" size Ethernet module that CEO of a security audit company allegedly found, was not present in other SuperMicro servers. SuperMicro itself, as well as it's buggest customers did not confirm the findings.
From what i recall, the story was very vague, there were no pictures of the specific chip, no pictures of the motherboard of the motherboard that would include serial, i.e. no details that would accompany a serious security research.
A supply chain attack similar to Supermicro's would be much more targeted and recalls with national security implications do get flagged via a separate chain.
Bloomberg's tech coverage is not great from what I've seen. Last year they published a video which was intended to investigate GPUs being smuggled into China, but they couldn't get access to a data center so they basically said we don't know if it's true or not. Meanwhile an independent Youtuber with a fraction of the resources actually met and filmed the smugglers and the middlemen brokering the sales between them and the data centers. Bloomberg responded by filing a DMCA takedown of that video.
What Bloomberg proposed - sniffing the TTL signal between BMC and boot ROM and flipping a few bits in transit - is far from science fiction. It would be easy to implement in the smallest of microcontrollers using just a few lines of code: a ring buffer to store the last N bits observed, and a trigger for output upon observing the desired bits. 256 bytes of ROM would probably be plenty. Appropriately tiny microcontrollers can also power themselves parasitically from the signal voltage as https://en.wikipedia.org/wiki/1-Wire chips do. SMBus is clocked from 10khz to 1mhz, assuming that's what the ROM was hanging off of, which is comfortably within the nyquist limit on an 8 - 20mhz micro.
Something similar has been done in many video game console mod chips. IIRC, some of the mod chips manage it on an encrypted bus (which Bloomberg's claims do not require).
"On PsNee, there are two separate mechanisms. One is the classic PS1 trick of watching the subchannel/Q data stream and injecting the SCEx symbols only when the drive is at the right place; the firmware literally tracks the read pattern with a hysteresis counter and then injects the authentication symbols on the fly. You can see the logic that watches the sector/subchannel pattern and then fires inject_SCEX(...) when the trigger condition is met.
PsNee also includes an optional PSone PAL BIOS patch mode which tells the installer to connect to the BIOS chip’s A18 and D2 pins, then waits for a specific A18 activity pattern and briefly drives D2 low for a few microseconds before releasing it back to high-impedance. That is not replacing the BIOS; it is timing a very short intervention onto the ROM data bus during fetch."
I don't believe that there was ever extra chips being added to the boards, but what I could believe is that they shipped with firmware on specific chips that enabled data exfiltration for specific customers and due to a game of telephone with non technical people it turned into "they're adding chips inside the pcb layers!"
Schneier was simply taking at face value the contents of the Bloomberg article, especially the statement by Mike Quinn who claimed he was told by the Air Force not to include any Supermicro gear in a bid.
There also was a CEO of a hardware security company that came out and said that his firm had found an implanted chip during an audit. IIRC, he was convinced that it was very unlikely to be limited to Supermicro hardware.
> he was convinced that it was very unlikely to be limited to Supermicro hardware
Yep. This was why there was a significant movement around mandating Hardware BOMs in both US and EU procurement in the early 2020s.
Also, the time period that the Bloomberg story took place was the late 2000s and early 2010s, when hardware supply chain security was much less mature.
There was a security auditing firm that came out a few days later claiming they'd found a chip, similar to the one Bloomberg described, during a security audit.
It's still nothing concrete, though. Their CEO basically said that they'd found one and that they couldn't say much more about it due to an NDA.
I'd like to think that modern centers are water cooled so it'd be more quiet these days unless you are implying that this application of theirs is running on legacy hardware? :P
And the entire Bloomberg takedown drama added fire to the flames.
And Bloomberg did a DMCA takedown through youtube, copystrike in parlance which pulled the video down for a week. GN had no recourse other than to wait and counterclaim.
Week timed out, Bloomberg did nothing but be the bully.
Louis Rossmann's excellent explainer video here on the Bloomberg bit: https://www.youtube.com/watch?v=6RJvrTC6oTI
You either take a gamble on something and hope it's good, or try to buy the same thing that someone else bought and reviewed.
If some market has large margins, it means it has some inefficiencies.
Remember the 2018 accusations of spy chips implanted in supermicro motherboards that everyone denied so strongly?
Can someone shed light on why China still couldn't copy the Nvidia GPUs in some form?
I understand its complex and there many parts to it, but which is the most complex part making it difficult for China to copy it?
Let's say they don't have access to 3nm process, what if they just use 12nm and create GPUs with much bigger size but comparable performance with CUDA compatibility? Or other option could be less tensor units, training will take longer, but they might be able to produce it cheaply
If you could steal all the designs at TSMC, and you had exactly the process that TSMC uses, you could definitely make counterfeits. If you didn't have TSMC's specific process, you could adapt the designs (to Intel or Samsung) with serious but not epic effort. If you couldn't make the processes similar (ie, want to fab on SMIC), you are basically back to RTL, and can look forward to the most expensive and time-consuming part of chip design.
This is nothing like copying a trivial, non-complex item like a car. Copying a modern jet engine is starting to get close (for instance, single-crystal blades), but even they are much simpler. I mention the latter because the largest, most resourced countries in the world have tried and are still trying.
Same with chips, efficiency, speed, etc all depend on good design, and cutting edge factors, if the main reason your chip isn't faster is because of the distance between your L1 cache and your core is far, then having a bigger node process but bigger chip won't make it quicker.
But there's a distinct time/value of investment equation with the current AI boom. The jury is at best still out on what that equation is for the goals of capital (it's increasingly looking like there's no moat), but if you're a national government trying to encourage local bleeding edge expertise in new fields like this it's quite a bit more clear.
They can copy it. And no, the software moat is not there if someone choose the blatant copy route. They just can't build it in the scale they want yet.
> what if they just use 12nm and create GPUs with much bigger size but comparable performance
Physics do not work this way :/
you could certainly use a larger process and clone chips at an area and power penalty. but area is the main factor in yield, and talking about power is really talking about "what's the highest clockrate can you can still cool".
so: a clone would work in physics, but it would be slow and hot and expensive (low yield). I think issues like propagation delay would be second- or third-order (the whole point of GPUs is to be latency-tolerant, after all).
"sorry guys, I did something token-bad a while ago that got you more money."
that's the sort of meaculpa I'd expect to get rewarded these days...
I wonder if any of those $2.5B of smuggled chips ended up being used for those training runs.
Gamers Nexus did a whole deep dive which basically proved that Chinese researchers had access to whatever they wanted.
https://youtu.be/1H3xQaf7BFI?si=ojlxOC7uiPqZxv0N
edit: not sure if this was sarcasm
DeepSeek v3 was trained on 2,048 NVIDIA H800s. https://arxiv.org/abs/2412.19437
MiniMax M1 used 512 H800s. https://arxiv.org/abs/2506.13585
The H800 wasn't banned in the first round of export controls - but was after October 2023: https://www.cnbc.com/2023/10/17/us-bans-export-of-more-ai-ch...
Z.ai say they used Huawei hardware: https://www.theregister.com/2026/01/15/zhipu_glm_image_huawe...
Qwen and Kimi haven't disclosed their hardware as far as I can tell.
For example: https://www.tomshardware.com/tech-industry/artificial-intell...
Did you think the hesitancy of westerners engaging and relying on Chinese labs was due to vibes? There are fundamental cultural differences at play, wether we are comfortable admitting that or not.
The $2.5B number is just these guys. It could be 10x in total.
Edit: Officially-debunked, I should note
(Allegedly) just some Bloomberg (alleged) bullshittery, (allegedly) posted to move the market.
And if it were posted to move the market, that would have been about the most cut and dry SEC violation possible, posted at a time when the federal government still enforced such things.
Not investment advice, do you own research. I'm just someone on the Internet.
Let's just say that none of this comes as any surprise.
Now, what people should be asking is how much Jensen knew. In May he said there was nothing going on. But the videos of the Chinese guy holding H1/200's ... never got to him?
Also interesting how they waited until just after GTC...
[0] - https://www.bloomberg.com/news/features/2018-10-04/the-big-h...
[1] - https://www.bloomberg.com/features/2021-supermicro/
[2] - https://www.schneier.com/blog/archives/2021/02/chinese-suppl...
[3] - https://www.theinformation.com/articles/apple-severed-ties-w...
His take was that it was very unlikely that it impacted exclusively Supermicro, though.
It was covered various places, including The Register https://www.theregister.com/2018/10/09/bloomberg_super_micro...
From what i recall, the story was very vague, there were no pictures of the specific chip, no pictures of the motherboard of the motherboard that would include serial, i.e. no details that would accompany a serious security research.
Bloomberg's tech coverage is not great from what I've seen. Last year they published a video which was intended to investigate GPUs being smuggled into China, but they couldn't get access to a data center so they basically said we don't know if it's true or not. Meanwhile an independent Youtuber with a fraction of the resources actually met and filmed the smugglers and the middlemen brokering the sales between them and the data centers. Bloomberg responded by filing a DMCA takedown of that video.
Something similar has been done in many video game console mod chips. IIRC, some of the mod chips manage it on an encrypted bus (which Bloomberg's claims do not require).
Here's one example of a mod chip for the PS1 which sniffs and modifies BIOS code in transit: https://github.com/kalymos/PsNee
"On PsNee, there are two separate mechanisms. One is the classic PS1 trick of watching the subchannel/Q data stream and injecting the SCEx symbols only when the drive is at the right place; the firmware literally tracks the read pattern with a hysteresis counter and then injects the authentication symbols on the fly. You can see the logic that watches the sector/subchannel pattern and then fires inject_SCEX(...) when the trigger condition is met.
PsNee also includes an optional PSone PAL BIOS patch mode which tells the installer to connect to the BIOS chip’s A18 and D2 pins, then waits for a specific A18 activity pattern and briefly drives D2 low for a few microseconds before releasing it back to high-impedance. That is not replacing the BIOS; it is timing a very short intervention onto the ROM data bus during fetch."
Multiple security companies looked into this and found nothing malicious.
Yep. This was why there was a significant movement around mandating Hardware BOMs in both US and EU procurement in the early 2020s.
Also, the time period that the Bloomberg story took place was the late 2000s and early 2010s, when hardware supply chain security was much less mature.
It's still nothing concrete, though. Their CEO basically said that they'd found one and that they couldn't say much more about it due to an NDA.