ByteDance plans to spend $7 billion on the chips in 2025, according to reporting from The Information, citing inside sources. If ByteDance follows through, it will become one of the world’s top owners of Nvidia chips, despite U.S. efforts to restrict Chinese companies from buying U.S. AI chips like these.
In 2022, the U.S. announced export restrictions on certain kinds of AI chips to countries, including China, where ByteDance is headquartered. These restrictions have gotten tighter multiple times since.
ByteDance is technically adhering to these restrictions by using a loophole: The company isn’t bringing the chips directly to China and is instead storing them in data centers located in other regions like Southeast Asia, according to The Information’s reporting. This doesn’t technically violate U.S. restrictions.
Snigger. Just stick the chips in a data centre not in China.
Then, bugger about with routing back through the Great Firewall…
Thing is, it really doesn’t matter much anyway.
Even if you can’t set this up, Nvidia did some tweaked products that slipped through the net that are almost as good. And ultimately, this isn’t some Indiana Jones magic artefact. You can train an AI model on anything. The latest Nvidia stuff just does it a bit more efficiently.
The ban has led to the Chinese getting nervous about relying on US chips and subsequently there’s been a boom in Chinese companies investing in the open RISC-V architecture to ensure continuity of operation. Alibaba or JD don’t want to find themselves not being able to function because they can’t buy chips. And what’s now starting to happen is RISC-V chips being exported to the USA. It’s opened up a whole new area of competition to Intel, ARM, AMD and Nvidia. Meta are already using some custom RISC-V for AI model training.
As I’ve been saying for some time. AI will move from GPUs to dedicated chips. Eventually, to ASICs. Which will be, err, cheap as chips.
TPUs from Google or Hailo, M.2 connections, about 60 quid from Raspberry Pi resellers including adapters.
Ducky,
You’re ahead of me on this but I assume that Google aren’t going to sell their TPU but just rent it out as a service on their cloud.
A good point, but it doesn’t appear to be the case given the very small amount of work I spent looking at it.
That is; it’s all local. Hmm
“storing” them in remote data centers? That wouldn’t be of any use, they need to be plugged into something and operating.
That GPU was just resting in my data centre….
I think it depends on the work they want the GPUs to do.
Anything actually sensitive they’ll want on Chinese soil if they’re at all careful about others listening in, and the cables as well.
This looks more like the Chinese doing a “nyah-nyah” at the US.
M,
This is Tiktok. It’s mostly videos of stupid pets or kids doing dances, not playing games of Global Thermonuclear War.
And as I said, not magic artifacts. The Chinese could just build a supercomputer to do codebreaking or war simulations by building a cluster out of the sort of boards they make and sell to Google. Would cost a little more than Nvidia stuff, need more energy, but it would be piss all compared to the total cost of the Chinese military.