They don't even do anything about companies open sourcing their SOTA models /s
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
If I'm going to burn $10k in tokens trying to one-shot a 3 hour task (for which human?) it better either give much better than 80% success rate or give me a refund if it doesn't lol
In practice the ability to use as much Deepseek V4 Pro Max as I want like it's free (it is practically) is far, far more useful than having to budget how much Opus 4.6 to use or if I want to gamble some money on 4.7. And if Deepseek struggles (it almost never), there's always Kimi.
What I've come to realize now is the number of tokens you get per $1 is a FAR more useful metric for AI utility than probability of the model one-shotting a problem.
Multiple recent datapoints showing that domestic cards are now being extensively used in training. The previous training bottleneck is gradually being solved by domestic cards.
Baidu just announced that their Kunlunxin P800 chip has succcesfully deployed in multiple 10K+ clusters and was used training their latest EARNIE 5.1 model (800B parameters)
Recently, there was another report than Meituan's LongCat-2.0-Preview model (1.6T parameters) was fully trained on a 50-60k cluster of domestic cards.
There was also a recent interview with an AI industry expert close to Chinese top labs, who stated that Ascend chips are now used by these labs for pre-training
Baidu just announced that their Kunlunxin P800 chip has succcesfully deployed in multiple 10K+ clusters and was used training their latest EARNIE 5.1 model (800B parameters)
Recently, there was another report than Meituan's LongCat-2.0-Preview model (1.6T parameters) was fully trained on a 50-60k cluster of domestic cards.
There was also a recent interview with an AI industry expert close to Chinese top labs, who stated that Ascend chips are now used by these labs for pre-training
Attachments
-
Screenshot 2026-05-12 120150.png
Last edited:
I think we should calm down on this a little. Chinese companies with global ambitions need to be careful with data of their overseas operations not being sent to China. This is an area many countries frown upon (including China), plus with the US/China confrontation ,this complicates things even more. So it’s normal for private companies who want to have significant presence in western markets to have a global headquarters outside China just like bytedance does in Singapore, to differentiate their China operations from Foreign operations . I see nothing wrong here. Since Chinas market is often closed and different from global market specially in the I.T/internet sphere .
moreover , I think Chinese leaders understand the stakes are high and China is in a delicate phase of her growth. So I think they are right in adopting a cautious approach and restrain. Afterall, despite Chinas growth and economy, the US/West still dominate the world global financial landscape , so it won’t be in Chinas interests to completely rock the boat at this stage. Not until China is far stronger and the dominant power globally . Strategic patience has served the country well so far. So why change a winning solution ?
So I don’t think it’s that serious to necessitate such drastic actions . China is doing well so far. No point disrupting my the system .
Tencent's latest 2026 Q1 earning call had interesting remarks on Chinese AI chip capacity
Tencent expects Chinese AI chip supply to increase second half this year, allowing them to buy more chips. As far as I'm aware, the only Chinese AI chip being fabbed domestically is Ascend, which indicates Ascend production bottlenecks are solved and is scaling up production, matching DeepSeek's comment on 950DT availability in 2026 second half.
Tencent expects Chinese AI chip supply to increase second half this year, allowing them to buy more chips. As far as I'm aware, the only Chinese AI chip being fabbed domestically is Ascend, which indicates Ascend production bottlenecks are solved and is scaling up production, matching DeepSeek's comment on 950DT availability in 2026 second half.
Cambricon as well.
Cambricon posted record revenue of 2.89 billion yuan in first quarter of 2026.. 160% increase over the same period last year. they are second only to Huawei in terms of Ai chip sales/units..
-----------------------------------------------------------------------------------------------------------------------
this chart shows how local companies taking over market share from foreign OEMs not all fabbed domestically but top 2 firms fabbed locally like Huawei and Camribcon and other small firms which i mentioned down below.
Moore Thread, MetaX and Biren too fabbed on domestically but on small scale.. all these companies are blacklisted by the US commerce department but now gradually returning back as more advanced capacity coming online this year.
Last edited:
well, it's been well known, but there are far more domestic AI chip supplier than just Ascend and Cambricon. There are also quite a few that fab their AI chips with Samsung (Tencent is not one of them).
sea bed AI data center. Save 1/3 energy
Directly connected to nearby wind turbines, achieving a 95% direct green electricity supply rate, that's cool