Skip to playerSkip to main content
  • 2 days ago
Transcript
00:00A social media account affiliated, we're talking DeepSeek now by the way, affiliated with Chinese
00:03state media, talking about DeepSeek and the delayed release of its V4 model pointing to,
00:10the report says, to a shift toward deeper integration with the chip ecosystem in China.
00:17Now the startup, in case you missed it by the way, on Friday released this preview version
00:21of its long-awaited new model, but Bloomberg Intelligence thinks it actually fails to narrow
00:25the gap with leading U.S. products, thanks to access to advanced Invita chips.
00:30Robert Lee, author of that report, joins us now to talk us through, you know, I think
00:35we've been spoiled because of the initial DeepSeek moment.
00:38This is 1.5 or 2.
00:40What do you make of this?
00:41Okay, well, compare and contrast to last January when DeepSeek made the big splash.
00:46It was off most people's radar.
00:48That isn't the case this time.
00:50So the new model has been well-previewed, well-trailed.
00:54It's performance and the characteristics largely met with expectations, so hence the lack of
01:01major surprise.
01:02But where it does lead is in terms of it's, again, forging ahead and pushing the barriers
01:09in terms of cost down and software optimization, which is the key strategy that China as a whole
01:15is pursuing at the moment.
01:16And I think the most interesting thing about it, as was detailed in their accompanying white
01:21paper, is that the model was trained and developed almost exclusively on Huawei chips.
01:27So again, highlighting China's ongoing push towards semiconductor self-sufficiency and
01:32reducing any reliance on NVIDIA, et cetera.
01:36So therefore, the chance of China taking whole-scale orders from NVIDIA going forward, as we've
01:43been writing over the recent months, is probably pretty limited.
01:47Well, I think you highlight a good point.
01:48In terms of how we need to be looking at how this one pushes the boundaries, is it more
01:53on, to your point, cost, or is there something here on the tech side that catches your eye?
01:57Well, as you said, I mean, very kindly in the opening quote from our report, China lacks
02:02access to leading-edge NVIDIA chips.
02:04Therefore, its ability to compete in terms of frontier model development is, you know,
02:09they're on the back foot.
02:10And they're not trying to do that.
02:12So they're tying in with the government's national strategy to proliferate AI across the
02:18entire economy, across every segment and every province, which is defined by what's referred
02:23to, well, as one of the government's major white papers.
02:28So the focus for Chinese companies, including DeepSeek, is really on efficiency.
02:33It's by using clever software optimization techniques to squeeze more performance out
02:38of what they have at the moment.
02:40And I think if you look at the characteristics and the technical competency of that model,
02:44that clearly illustrates that China is developing, I think, a real lead in software-optimized models
02:50and in cost down.
02:51And if you look at the pricing in terms of token pricing that they're charging external clients,
02:56they're still, compared to the leading anthropic models, they're an approximate 80% to 90% discount.
03:03Now, I would suggest that the pricing is a loss-leading price.
03:07There's obviously some degree of subsidization going in there.
03:12But they're also benefiting from this cost optimization, which allows them to push ahead in terms of pricing.
03:19So, yeah, China is clearly emerging as a global leader in software-optimized models.
03:28It lacks access to NVIDIA and, therefore, its ability to compete on frontier model developments
03:34is curtailed.
03:35So the strategy, given the constraints they work within, is, I think, the right one.
03:39So, yeah, I think the right one.
Comments

Recommended