Black Market A100s and Twitter’s China Tech Connection
1,000 words or less: two tech gems mined from the deep underworld of the Chinese internet
Irene—How to Buy Banned Nvidia Chips … on China’s Instagram
The Wall Street Journal reported three days ago that the Commerce Department is considering new export restrictions to make shipping Nvidia and other high-end chips to China even harder. If put into effect, the regulations would require customers to obtain a license before a Nvidia shipment was authorized.
But Chinese buyers want still want their chips — and as Reuters noted, despite US export controls, Chinese buyers are still sourcing advanced chips via a flourishing underground market.
You can find your way to banned semiconductors at, for example, various sprawling electronics markets in Shenzhen. Or you can go online — specifically, to Douyin 抖音 (TikTok’s Chinese version) or Xiaohongshu 小红书, the Instagram of China.
So, I tried to buy an Nvidia A100 on Xiaohongshu.
The hashtag #A100, which is mostly about the Nvidia chip, had 176,000 views on the app, and the #英伟达 (Nvidia) hashtag had 2.8 million — a notable niche.
While scrolling, I kept coming across a user we’ll call Chen. Chen is located in Beijing and has only forty-six followers — but since February 2023, Chen has posted dozens of times about selling various GPUs and associated gadgets. A significant portion of Chen’s stock appears to be covered under US export controls.
On Goofish/Xianyu, an Alibaba-owned eBay-like app popular for electronics, A100s go for as high as 230,000 yuan (US$31,710) apiece:
I was thus pleasantly surprised to find out that Chen was selling them for 56,000 yuan ($7,730) each — roughly comparable with the price on Amazon! Chen had only nine left, though.
I told Chen I was helping someone who’s working on AI video generation, and Chen proceeded to recommend the Nvidia GeForce RTX 4090 to me. With plenty of 4090 in stock, Chen was willing to sell the chip for 16,000 yuan ($2,210) each — 15,000 if I bought ten in one go.
It all sounded great, but things got tricky when I asked for proof that the chips are authentic. Chen didn’t want to discuss where the chips came from, and said that the GPUs were in “industry packaging” — that is, they won’t have labels or retail wrapping, and I won’t be able to get any Nvidia customer service.
Chen promised to get back to me with photos to prove the chips aren’t fake. I’m still waiting.
But who are the people buying banned GPUs on Chinese Instagram? According to my forays into various comment sections, they are a mix of individual hobbyists, AI developers training models, computer science students, and crypto miners. The #Nvidia hashtag is a whisper network of simmering anxiety where hoarding GPUs insures oneself against an uncertain future. As one user put it, “Buying chips now is like buying a house in Beijing circa 2000.”
ChinaTalk is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.
In March, Twitter open sourced its recommendation algorithm — that is, the code which takes the 500 million tweets posted every day and turns them into the personalized list you see when you open the app or site. I was surprised to find at the algorithm’s core an implementation of a 2021 paper by a team at Sina Weibo Corp. Weibo (as it is ubiquitously known) is basically the Twitter of China — and it surpasses Twitter in monthly active users: 586 million, to Twitter’s 368 million.
The largest single neural network model in Twitter’s algorithm is called the “Heavy Ranker,” which takes 1,500 candidate tweets proposed by other parts of the algorithm and determines the order in which you see them. In its code, the Heavy Ranker uses a neural network architecture named “MaskNet,” proposed by Zhiqiang Wang, Qingyun She, and Junlin Zhang of Sina Weibo Corp in their paper titled, “MaskNet: Introducing Feature-Wise Multiplication to CTR (Click Through Rate) Ranking Models by Instance-Guided Mask.” That’s a mouthful, so let me break that down a bit.
“Features” are the properties of individual tweets (e.g. is the tweet about kittens? or does the tweet contain a photo?) that the neural net uses to rank a tweet. Discussing kittens or containing a photo could raise a tweet’s rank — but you might want a tweet that combines the two features (discussing kittens with photos) to rank much higher than what the two features would contribute individually.
MaskNet is part of a long line of neural-net architectures that model interactions between different types of features, which gets difficult when you work with thousands of features. The most basic architecture is something you might learn in undergrad (a linear regression model with interaction terms). The MaskNet paper builds on more advanced architectures from teams at Huawei and Google.
And the “click-through rate” measures how many users who see a tweet actually click on it. Traditionally, CTR is what social media companies try to maximize — but boosting CTR alone just spreads clickbait. So the Twitter implementation of MaskNet instead optimizes for ten types of social network behaviors, including the probability you share a tweet, report a tweet, or reply to a tweet and get engagement from the original author (all of which hopefully result in showing you tweets you’re genuinely interested in).
The MaskNet paper didn’t have a big academic splash — only thirteen citations — and I couldn’t find much on the authors online. Even so, given that Twitter chose MaskNet as its architecture, MaskNet must have been rigorously tested against alternatives — including the “Deep & Cross Net” proposed by Google and also mentioned in Twitter’s codebase — and come out on top.