Tuesday, eleven forty-five in the morning.
The flight from Beijing to Haicheng had just landed at Pudong Airport and was taxiing.
Sitting in business class, Zhao Wenyuan took out his phone, switched off airplane mode, and sent a message to the administrative staff member picking him up.
Then he opened Zhihu and found a question that interested him.
"Are any domestic companies developing large models using Chinese-made chips for training? Do they actually work?"
There was still some time before the cabin door opened, so Zhao Wenyuan began typing.
Questions like this attracted a moderate amount of attention.
They certainly couldn't compete with gossip questions like "What do you think of xxx?" or "How do you feel about xxx's scathing critique of xxx?" But they drew considerably more attention than ordinary technical discussions.
After all, AI was a hot topic, and this question was about Chinese-made chips.
With Glutinous Rice Ball dominating the rankings for a month, people had gradually accepted that Chinese companies had very capable models, even though plenty of critics still accused them of gaming the benchmarks.
After all, if you claimed they were gaming the scores, you could do it too. Why hadn't you knocked them off the top?
But the training ecosystem was another matter. Anyone with even a little knowledge knew that the Americans and their allies had Chinese chip production locked down tight, while Nvidia had used its first-mover advantage to build a vast ecosystem empire. Apart from Google, there was currently no competing ecosystem on the market, and Google only used its own internally rather than selling it.
So the popular answers below were almost uniformly pessimistic.
"Training large models on Chinese-made chips? Completely unfeasible right now. You can't close an ecosystem gap in a year or two. Nvidia's CUDA ecosystem has