00:00Anthony, talk to us about your assessment of the capabilities of these Chinese AI models.
00:07Yeah, so the technical details for Kimi K3, the Moonshot AI's latest model, is a little bit
00:16kind of dense, but let's go through it step by step. So there is a lot of software innovation
00:21in there in what is objectively a very large model. So the first thing is the ambition and
00:26the scale of the model is a step further, and that's probably why it has caught the
00:31U.S. models on the frontier front. But the second most important aspect is the resource
00:36optimization around the model. It is very memory efficient. In short, there are proprietary
00:42innovations in there to also help with the usage of HBM and GPUs, where China is obviously
00:49constrained. And the fact that this model was trained with those constraints in mind is
00:54the important aspect to markets here. China has found a software workaround to its hardware
01:00constraints. And that is what the market is pricing in here on the AI picks and showers
01:04trade. And that is what will continue to play out. It is also interesting that Alibaba owns
01:10such a sizable chunk of Moonshot AI. So when those two companies come head to head, it's
01:16a bit of a family feud here.
01:19Anthony, how are China's AI leaders able to make such a competitive product without the
01:24access to the cutting edge technology that its U.S. competitors have?
01:30It's software innovation, in a word, right? There is a lot of software that is dealing very
01:35creatively with the constraints around compute and memory. And how they're doing it is by using
01:41extreme probabilistic mathematics, basically, to try to understand how a machine thinks in
01:47a more efficient manner and correlate that with the kind of use cases that the model is
01:53most likely to be used for. So it's not trying to do everything for everyone. It's trying to do the
01:58most probabilistically useful thing with the resources that it has available. And that approach has really
02:04paid off on the benchmarking. As you guys have covered relentlessly, the biggest use case for
02:10AI at the moment is in coding. And in approaching coding, this model is extremely efficient in how it
02:17uses its resources.
02:20How does this affect the demand for memory and perhaps what this also means for
02:25the big chip makers, TSMC and NVIDIA?
02:30So at the margin, the fact that it was trained on a compute-restricted ecosystem has worried
02:36people. Like, how has China done this without access to TSMC's and NVIDIA's top-of-the-line chips?
02:42But overall, the debate has moved on to now the price of these chips are so, price of these models
02:49are
02:49so constrained that it might open up new use cases for AI. And as a result, increase the size of
02:55the
02:55GPU market. So this debate is very live. Does the market grow overall as a result of the Chinese
03:00models? Or does the Chinese model's resource optimization reduce the size of the GPU market?
03:06On memory, the efficiency of memory is a clear negative to the kind of ongoing bull market in
03:14memory. But they are very large models which require a lot of HBM upfront. So for the moment,
03:21the HBM market looks very stable, but there is a concern that the memory market is being pushed
03:26by Chinese innovation towards the DDR5 or the DRAM market where China is more competitive and China
03:32can bring on more capacity to bear. And that capacity is going to become the new touch point for markets
03:38in the memory trade in the days to come.
Comments