Research papers and staff interviews detail the workarounds Chinese companies like Huawei, Baidu, and Alibaba use to develop AI systems without the latest chips

Wall Street Journal 2023-05-08

Context & Ripple Effects

This WSJ report is the baseline document in an arc the outlet has kept extending: at the time, tighter US rules were squeezing Alibaba and Baidu onto stockpiled 2017-era Nvidia V100s, and the paper-and-interview method here was how the workaround playbook first got documented company by company.

Since then the same coverage thread has shown those workarounds industrializing — broker networks sourcing top chips anonymously, training data driven to foreign data centers, and Beijing stepping in to allocate scarce domestic output — which is why this early accounting of Huawei, Baidu, and Alibaba's constraints matters: it names the techniques the entire ecosystem now runs on.

First-order effects

Alibaba and Baidu are building AI on aging, hoarded silicon, so software efficiency rather than raw compute becomes the binding constraint on their model roadmaps.
The reporting hands US export-control enforcers a documented map of the exact loopholes — research papers and staff interviews spelling out what Huawei, Baidu, and Alibaba each do to compensate for missing hardware.

Second-order effects

A services layer grows around the scarcity: brokers using cryptocurrency smart contracts to source top-tier chips anonymously, and hard drives of training data physically carried to Malaysian data centers housing advanced Nvidia hardware.
Constrained access redirects demand toward substitutes — Huawei's domestic silicon and open-source releases like Alibaba's Qwen models become the compensating levers, with researchers crediting open-source technologies for narrowing the China-US gap.

Third-order effects

If the pattern holds, US chip controls do not halt Chinese AI development but restructure it around heterogeneous compute, domestic substitution, and an offshore arbitrage industry — with the state later intervening in SMIC's output allocation to prioritize Huawei.
The longer-term effect is a bifurcated AI compute market: frontier-scale training migrates to wherever unrestricted hardware sits, while domestic stacks absorb everything else, entrenching two parallel toolchains.

The trend: US chip export controls are producing not a stop but a rerouting of Chinese AI development — through brokers, offshore data centers, domestic substitution, and open-source compensation.

Discussion

@prestonjbyrne Preston Byrne on x
Sanctions seem self-defeating as they give adversary states a heads up that they need to start developing their own stuff https://twitter.com/...
@_karenhao Karen Hao on x
US sanctions are driving Chinese firms, like Huawei, Alibaba, Baidu, to seek ways to advance AI without cutting-edge chips. We dug through open-source research papers and spoke with employees & analysts to understand what they're doing. w/ @raffaelehuang https://www.wsj.com/...
@jchengwsj Jonathan Cheng on x
@_KarenHao @raffaelehuang Using workarounds to catch up on AI remains a challenge, researchers says. Some experiments have shown promise, however, and if successful could allow Chinese firms to weather U.S. sanctions and make them more resilient to future restrictions, they said.…
@minhasnajma Najma Minhas on x
Necessity is the mother of invention- was expecting this as USA takes futile & ultimately self defeating action against itself by stopping various IT exports to China- Chinese tech companies seek ways to get state-of-the-art performance with existing chips https://www.wsj.com/...
@alecolarizi @alecolarizi on x
A WSJ review of research papers and interviews with employees found that Chinese companies are studying techniques that could allow them to achieve state-of-the-art AI performance with fewer or less powerful semiconductors. https://twitter.com/...
@wsj @wsj on x
Chinese companies are also researching how to combine different types of chips to avoid relying on any one type of hardware https://www.wsj.com/...
@jchengwsj Jonathan Cheng on x
@_KarenHao @raffaelehuang A WSJ review of research papers and interviews with employees found that Chinese companies are studying techniques that could allow them to achieve state-of-the-art AI performance with fewer or less powerful semiconductors. https://www.wsj.com/...
@rwang0 @rwang0 on x
MyPOV: can't do it without compute power even if you have everyone's data and unlimited energy. Hit by U.S. sanctions, Chinese tech companies seek ways to achieve state-of-the-art AI performance with existing chips https://www.wsj.com/...
@jchengwsj Jonathan Cheng on x
@_KarenHao @raffaelehuang Huawei Technologies, Baidu and Alibaba Group are among those seeking ways to milk more utility out of existing computer chips. They are also researching how to combine different types of chips to avoid relying on any one type of hardware. https://www.wsj…

Chronicles