Latest / UK Column Radio / China's AI DOMINANCE is Inevitable (Here's Why) | TP Huang Returns

Transcript

China's AI DOMINANCE is Inevitable (Here's Why) | TP Huang Returns · UK Column Radio

0:02
Hello and welcome to another episode of the Silk and Steel
0:05
podcast. I'm your host Carl Zha. 4 month
0:08
ago I had a conversation with AI expert TP Huang on the AI race
0:13
between US and China. This conversation was originally
0:17
released as the audio podcast on Silk and Steel podcast for my
0:22
Patreon subscriber. Now I'm making available on
0:25
YouTube. Take a look.
0:28
Today we have back our returning guest AI expert TP Huang to give
0:34
us AI update and what is going on in China in the tech space.
0:40
Welcome back to the show, TP. Glad to be back Carl, really
0:44
exciting times. Very exciting time.
0:46
You know, last time I have you back, I think it was December
0:50
last year, you said the rare earth, the China's rare earth
0:55
export control, it's going to real cripple US industries.
1:00
It took them six months to notice and, and finally now, now
1:06
everybody's talking about it. Well, I guess it's because
1:09
originally the export control was only Angallian, Germanian
1:13
and fuel metals, and then it got expanded after Trump's
1:18
Liberation Day tariff announcement.
1:21
Yeah. I think also there was a lot of
1:24
a transshipment going on. But I guess the amount of the
1:29
newly added metals, say the amount of they need is so much
1:34
that it's not easily transshipful, so that the auto
1:39
industry noticed. Pretty quickly yes, yes, because
1:42
every car, every car needs those heat resistant magnets you know
1:47
for people who don't know because they need their magnets
1:50
in the car motor because. The the the.
1:55
The motor generated a lot of heat and the heat demagnetized
2:00
magnets, but with these critical minerals magnets, they're heat
2:05
resistant. This is why it goes into every
2:08
cars. So the reason I have you back is
2:12
because we haven't talked for a while.
2:14
Last time I think we talked was February, right, right after the
2:19
DeepSeek a moment and now it's like 4-5 months later and that's
2:27
like a lifetime in AI development cycle.
2:30
And I remember distinctly, distinctively, you mentioned
2:34
when I ask, you know what, what do you think about AGI,
2:39
artificial generative intelligence?
2:42
Are we ever going to get there? You said it's possible.
2:45
It's possible we're possibly just one or two years or what?
2:50
Do you still stand by? That it's, it's hard to say.
2:53
So the way I look at it is like the, the models themselves keep
2:58
getting more useful and they're not like, I don't think they
3:07
have intuition at the moment really, but they're getting
3:11
there. And I think that it is possible,
3:17
you know, one over the next year or two, we're going to get
3:20
something that's not real AGI but appears to be AGI to us.
3:27
Like if hears like if you ask him anything, they can probably
3:30
do most of the stuff, 99% of the stuff that that normal humans
3:34
can't do. But you still need to be there.
3:40
There's still some, there's still value for, you know, for
3:43
for people to be at the other end to know how to use it, how
3:46
to use the models correctly. That makes sense.
3:51
Yeah, I mean, I, I, I heard from people who use AI to do coding
3:55
and they basically say, yeah, you definitely need a human, a
3:59
pair of human eyes to make sure to correct all the mistakes that
4:04
AI is making in the coding. But but what what the AI did it
4:08
just make the process of coding easier because you take out all
4:13
the groundwork and then you know, you can just act as a
4:16
supervisor to to correct the the the the mistakes to prove
4:22
reading. But you still need a human prove
4:24
reader. Yeah, actually that's how I use
4:26
it right now. In fact, I I do use it quite a
4:30
bit. I use DeepSeek, I use ChatGPT
4:33
sometimes to do and I use like cursor to program, so it's
4:37
actually helped my productivity quite a bit.
4:40
For a programmer, you'll notice that if you know how to use it
4:43
correctly, it's pretty helpful. But obviously what we're seeing
4:53
right now is that AI is actually replacing some of the need for
4:57
entry level programmers. So you're seeing a lot of
5:00
layoffs right now in Microsoft, Meta, and possibly Google, but I
5:04
think Microsoft and Meta are the main ones.
5:07
Speaking of Meta, Meta, just recently they, they spent $15
5:13
billion to acquire Scale AI, you know, by led by Alexander Huang
5:20
Huang, the guy who basically run a digital sweatshop like we
5:25
talked about last time. You know, this guy basically
5:27
hired these Filipinos and Kenyans and, and pay underpaying
5:32
them and, and have them label data for the big AI company.
5:37
Now why, why did the Zuckerberg do that?
5:39
Why did he spend $15 million for Scale AI?
5:43
Well, I think mainly because Zuckerberg really desperate and
5:49
he, so you got to understand the Silicon Valley types, they're
5:53
very insulated. They're bubble.
5:56
The only world that they see is in the Silicon Valley bubble.
5:59
So Zuckerberg when he looked at the landscape like two or three
6:03
years ago, he's like, oh, opening eye is going not open
6:06
anymore and Google is not open and Claude is anthropic is not
6:15
open also. So I know what I'm going to do.
6:17
I'm going to make my scene open source.
6:19
So everyone that doesn't trust their data going through Sam
6:25
Altman is going to use my llama models.
6:30
And then what happened in the past year, both Quinn and
6:37
DeepSeek and now you and Kimmy have surpassed llama models and
6:45
he's very desperate to catch up. So and he is just throwing money
6:51
like cuz Mehta has a lot of money.
6:53
So there's he just throwing money out there and he's
6:57
throwing money on investing chips.
6:59
He's throwing money on investing talent hoping that he can catch
7:02
up and be state-of-the-art. Basically cuz he doesn't want to
7:08
lose this AI race either and it's obvious the meta is 1
7:16
trailing everyone else right now.
7:19
So he's OK, So he's just desperate and he's just grasping
7:22
at straws. But how is Scale AI going to
7:25
help him? I mean, my understanding is
7:27
scale AI, it's not really AAI company.
7:31
I mean, it's basically they're, they're doing the kind of the
7:34
groundwork for the other, other AI companies, right by, by data
7:39
labeling. Yeah, the entire seems very
7:43
confusing to me. Like some of his effort in
7:46
paying like $100 million to recruit ethnic Chinese people
7:50
from that made a lot of sense, right?
7:53
Like you need as many talent as possible.
7:56
And Alexander one is definitely not a talent.
7:58
He's just talker. But I think like in America,
8:02
sometimes it's just about talking a big game.
8:05
This is why this country is going to shit because we reward
8:09
people for talking in big game and can't back it up.
8:12
That's, that's sadly true. So I, I wanted to one reason I,
8:17
I wanted to bring you back to the show is to talk about the
8:21
latest development in China. So what?
8:24
What has changed in the last six months after DeepSeek emerged as
8:31
like the big shock into this? I think in general, the Chinese
8:36
AI models were always were pretty good by middle of last
8:40
year already, but the the Silicon Valley crowd hasn't
8:45
caught on to it yet. So there wasn't this hype behind
8:47
it. Behind all that, the Quinn
8:50
models when they came out were really good.
8:52
Like, you know, they, they were very good for open source and
8:56
they were already surpassing Llama models by that point.
9:00
And then the Chinese public also did not appreciate how good the
9:03
Chinese models were. And the reason the Chinese
9:06
models were behind weren't really because they were lacking
9:09
computing power. It was because they just didn't
9:15
go down this stretch. So like opening I release
9:19
ChatGPT. That's it was only then that the
9:22
Chinese AI labs and big tech decided we need, we need to, you
9:29
know, catch up the opening eye. So like the American tech firms
9:33
had like a 2, two or three years advantage in the in the race
9:38
basically. So when Deep Sea came out, it
9:44
was only then the Chinese public caught on to how good the
9:48
domestic AI model was. And it's kind of sad that like
9:52
Chinese people need outsiders to validate their own progress, But
9:56
that's just been like a story of our culture for many years now.
10:00
So what happened is when this came to China, Deep Sea came to
10:06
China, it blew up basically over Chinese New Year's, right?
10:10
And then we start seeing everyone incorporating the
10:17
reasoning models and people are people start finding out the
10:21
reasoning models were really good.
10:22
So then they start putting into all their apps everywhere.
10:25
And it was only at that time did we realize how much compute
10:30
Chinese firms really had because, you know, we had the
10:33
sanctions for the since 2022. So people believed that they
10:38
weren't getting these chips. And it was only at that time all
10:43
these news came out about just how much transshipment was
10:47
happening in the in the AI chips that China was getting them from
10:51
Malaysia, Singapore, and I don't know, Thailand, Taiwan, like all
10:57
these Asian countries, right? And it was shown like a video is
11:03
making as much revenue out of like Asian regions where you
11:07
think most of the chips are going to China from that as they
11:11
were making more money from from that than America itself.
11:15
And it was also at this point I really started tracking the data
11:18
center build up in China. And I realized that the data
11:23
centers were having a whole lot more computation power than I
11:26
was expecting personally. And this is also the, the, the
11:32
really eye opening moment for me was when Tencent opened up their
11:37
AI app Youngbao. And it was from day one, it was
11:46
offering up, you know, service of DeepSeek reasoning models to
11:53
everyone. And you can think of like
11:56
Tencent has, you know, almost entire Chinese population using
11:59
it. So if they open this up in their
12:03
AI app, there's going to be a lot of users.
12:05
And what I, what I heard was actually this, they had enough
12:08
bandwidth to handle 8 million requests at the same time.
12:16
I mean, Can you imagine an American one having that?
12:18
Like it's, it's, it's crazy amount of bandwidth, right?
12:20
So, so this this told me that the Chinese big tech had a lot
12:30
more AI chips than we thought. They were advancing quickly and
12:37
also deep seq was driving all the AI labs in China to move a
12:43
lot faster. So we have the same nature like
12:47
we saw with EVs and and solar panels and everything else.
12:53
And that was also when we saw that renewed commitment by Byte
12:57
Dance, you know, Alibaba making announcements on how much
13:02
investment they're going to put into AI chips and also Tencent.
13:08
So basically we have like 3 or 4 major players in China and a
13:13
couple up and comers making really huge investments.
13:16
AI mean in the end, it's hard to say who the winners will be, but
13:22
I would say that there's quite a few very competitive Chinese
13:26
models and for for me, I personally love using Quint
13:31
because Quinn models, you can run them on your computer, you
13:34
can run them on your edge devices, you can run on your
13:37
little AI robots if you want. And the Quinn 2.5 and all the
13:41
Quinn multimodal models are state-of-the-art, like they're
13:44
the best for what you can use them for, right?
13:48
So if you have a, let's say, let's say you're trying to make
13:55
like a little AI robot in your home to play with your kids, and
14:01
you don't want that information to go to open AI, then you want
14:05
all the stuff to be handled by the AI robot.
14:09
Quinn is like the ideal model for that because it can handle
14:14
voice and the text at the same time and it has different size
14:20
models so that you can fit it according to your the needs that
14:25
you have. So what's happening underneath
14:29
is that a lot of people, if you, if you stay in America, people
14:34
in America still thinks America is ahead, but that's only
14:36
because they only focus on chat bots.
14:39
What they don't realize is that AI development, it's China is
14:43
very focused on the product driven part of it.
14:46
And there is like right now a huge gap between what the
14:53
Chinese industries are making in terms of AI robots versus what
14:58
American companies are making, if that makes sense.
15:04
I know this because I personally worked on a project like this
15:07
recently and I was on call with, so we had a couple of Americans
15:15
on our side and because we needed to do some kind of AI
15:20
robots. So I was on call with this
15:23
factory in Shenzhen and then they had their one sales lady
15:29
talking to us because the sales lady was has we're trying to
15:35
access this Chinese chip to put on our AI robots because the
15:38
Chinese chips are good enough and they're cheap, right?
15:43
You don't wanna buy NVIDIA because they're so expensive.
15:47
So 15 minutes into this call, I realized that America was cooked
15:51
because this lady, this sales lady know more about how to make
15:55
electronic sport and, and the, the, the, the battery and the
16:00
heat dissipation, heat management stuff that all the,
16:04
all the smart engineers that we have in America And I and I,
16:09
that was when I looked at myself, I was like, America's
16:13
really in trouble. If, if you, if you hear all
16:16
these big talking about Silicon Valley about, you know,
16:18
deploying AI to robots, they don't even know how to make a
16:21
board. They don't know that they don't
16:24
have the, the models needed for the board.
16:27
You know, nobody, even I bet the people in Silicon Valley never
16:32
have to deal with what the battery life is for, for
16:34
products that you need to use or how to, how to do it so that the
16:41
your AI robot doesn't burn up from too much heating.
16:46
So what we did was we actually tried different AI inference
16:54
edge AI chips from different companies.
16:58
So what we found was that NVIDIA was the easiest to set up,
17:02
obviously because everything was written for it, but they were
17:04
charging a ridiculous amount of money.
17:07
And then we also tried Qualcomm, Mediatek and two mainland
17:12
Chinese companies, Rock Chip and Acera.
17:15
And we found out that the Rock Chip and Acera, they, they were
17:22
pretty affordable, but more importantly, they had very good
17:25
software support for their, the AI portion of it.
17:29
So when you're running the models, you can use their, the
17:33
NPU, so the AI processing unit to speed up the process of
17:38
generating results. And then we tried Qualcomm and
17:43
Mediatek. So Mediatek is a Taiwanese
17:46
company and Qualcomm, you know, is in San Diego.
17:49
And because we thought like, you know, like if we want to bring
17:53
the product to America, we need to have like a non mainland
17:57
option, right? And then we just found the
18:01
software support from Mediatek and Qualcomm to be so lacking,
18:08
Like their MPU doesn't support the functionality, the
18:11
operations needed for some of the most widely used models like
18:16
Quinn models or stuff like commercial text to speech and
18:21
then speech to text. So you can't actually use these
18:25
chips because their MPU is effectively useless because of
18:28
this. So this is what a lot of people
18:32
don't realize. 1 you know, if you want to have that working AI
18:36
model, you have to use the Chinese.
18:38
You have to use the models themselves have to be Chinese.
18:42
The where you make them, the electronics board and stuff like
18:45
that has to be China because that's where the entire supply
18:49
chain is. That's how the entire place
18:51
where people know how to like put stuff in electronics, put
18:55
the Rams in there, put the audio things in there, hook it up to
18:58
the battery and how to do heat dissipation, maybe put like a
19:03
fan in there and how to deal with the battery life, things
19:05
like that. All the people that know how to
19:07
do that are in Shenzhen right now.
19:09
And then after that, the chip, which people would normally
19:18
think as something that China's behind the West in is actually
19:21
not a problem in this case because the one company that
19:24
came out with this stuff first hash has the which is Rockship.
19:27
They have the best software support out there outside of
19:31
NVIDIA. And then I would get on these
19:37
calls with the Chinese factories and ask them like, you know,
19:42
basically, are there competitors to us in this space?
19:45
And they're like, yeah, yeah, plenty.
19:46
We have so many people interested in doing this stuff
19:48
here, and I'm pretty sure there are many people doing this in
19:53
America. I could talk, talk to in the
19:58
Indonesian factory. They're like, yeah, you should,
20:01
you should have speed it up a little bit because there's
20:02
Chinese companies trying to do the exact same thing.
20:05
Do this product, getting it assembled in Indonesia, get the
20:08
electronic board done in China, do the software in China, then
20:10
get it assembled in Indonesia and ship it to America.
20:14
It's like, OK, yeah. So.
20:18
That that that's kind of the just the lay of the land.
20:21
Basically, if you want to get any kind of electronics done, I
20:23
found out that the hard way. You want to get any kind of
20:26
electronics done with AI, you need to be in Shenzhen for about
20:29
3 or 4 months. Now, now the question I my next
20:32
question is, did you guys have this discussion before or after
20:37
the liberation, the Tariff Liberation Day announcement?
20:42
Oh, definitely. After definitely.
20:44
Really. Really.
20:46
So even after whatever 55% tariff whatever, it will still
20:52
make sense to get the Chinese hardware.
20:55
I mean, you just can't be done in America.
20:58
There's no expertise in doing this right?
21:01
Where am I gonna find a guy that knows how to, to, to design and
21:04
board together and put it together like there's no factory
21:07
in America doing that kind of stuff.
21:09
We had some problem with just just to give people an idea of
21:14
like how much China dominates the electric electrical board
21:18
space. What they called single board
21:20
computers in in the Western country.
21:23
There's this thing called Raspberry Pi.
21:26
I don't know if you used if you used it, but it's like designed
21:29
in UK and made in UK slash Japan slash China, whatever.
21:37
In China they have orange pie, they have banana pie, they have
21:43
Saipede and I'm pretty sure they have another pie.
21:46
Also, I can't remember now, but there's like for every like
21:50
Western option outside of China, there's like a 5 or 6 inside
21:53
China itself. And you don't even know about
21:56
them until you go to like Aliexpress and just look up like
22:01
sports that support this chip. And they all know what they're
22:03
doing. So and no matter where the the
22:07
headquarter of the chip, no matter where the headquarter of
22:10
the chip design company is, even in China, all the factories are
22:14
in Shenzhen. That's amazing.
22:17
We rarely hear about this in the mainstream media because I, I
22:21
just recalled like just not that long ago, like I think a month
22:24
ago there was article coming out either in Wall Street Journal or
22:28
New York Times that says, oh, oh, the, the Chinese are good at
22:32
the assembling electronic stuff because the Chinese women, they
22:35
have small hands. I mean, like I, I can't, I can't
22:41
believe I read that in like the paper of the records and but
22:46
it's Oh my. Anyway, so enough about the
22:49
stupidity of the of the legacy. Media, yeah, like, like America
22:53
is, is undefeated in copium, completely undefeated in that.
22:58
OK. And it's written by a bunch of
23:00
liberal arts majors that don't, that aren't into technology,
23:04
that don't have to deal with this stuff, right.
23:07
Honestly, if unless you try to do this, unless you try to
23:13
produce this, it is actually very hard to know what the lay
23:17
of the land is like. I did not realize Quinn was so
23:20
much better than Llama before I started this.
23:23
I didn't even know rock chip was so so big.
23:26
They're like an industry standard.
23:28
They're like the the de facto single board computer standard
23:31
because there's so much better support for them software wise
23:34
than others. I didn't know ESP 32 was like
23:38
the de facto industry standard before this.
23:41
Right? Like they're just all these
23:43
things that because entire ecosystem is inside China, how
23:48
to make a board. The expertise behind that is
23:52
just not known. People don't know like how much
23:56
the electrical electronics industry is just congregated
24:01
there. I think there was a recently, I
24:05
would say a contact friend of mine, Molson Hart, he was
24:09
returning to Shenzhen because, you know, his toy business does
24:15
a lot of stuff there. And he was saying that
24:19
basically, if you want to do electronics, you have to be in
24:22
Shenzhen. Even if you're designing stuff
24:25
in somewhere in middle of America, if you want to get the
24:27
shit done, if you want to get it produced, you have to be in
24:30
Shenzhen. Because the expertise there is
24:32
just unparalleled elsewhere. Like where in America you're
24:35
going to find like several factories where people know how
24:40
to source different chips, source different batteries,
24:44
source boards, the audio components, the fans, and put
24:52
them, assemble them together at a manageable cost.
24:56
Like, I didn't know how much any of this stuff cost before,
24:58
right? But the people in Shenzhen, they
25:00
all know it. But the argument of the Silicon
25:03
Valley tech bro, is that OK? Yeah, yeah, yeah, yeah, sure.
25:07
China dominating hardware, but that doesn't matter because we
25:10
can just concentrate the design of the software, right?
25:14
And that's as long as we get the AGI out of the door first, you
25:20
know we're going to win. What do you say to that?
25:24
So it depends. I think there is applications
25:26
that doesn't require hardware right?
25:28
Like just right now, the opening AI reasoning models just by
25:36
themselves are pretty useful for if you want to do research
25:40
reports, if you want to have like a summary of things, if you
25:45
want to plan out vacations and, or even if you have like a
25:50
programming question that you have for it, then it can it can
25:53
do that for you. But it's one of those things
25:58
where once America gets to something that looks like AGI,
26:05
China will get there maybe at worst like 6 months later,
26:09
right? And nothing is going to happen
26:11
in the six months that's going to say, oh, China's done.
26:16
So the, there's no in terms of computational resources, it
26:24
seems like China's doing pretty well there in terms of energy
26:28
resources. You know, China has no issue
26:30
with supplying power to the data centers where America is a it's
26:34
a huge issue. And there's not just not just
26:36
data center, not just the power itself, but also the all the
26:39
other part of the supply chain that goes into the data center.
26:41
A lot of electrical stuff that, you know, America relies on rest
26:46
of the world to build its data centers.
26:49
So but you know, aside from stuff like this, there are a lot
26:55
of areas like material science, biotech, auto manufacturing,
27:04
arrow design and things like that where we're already seeing
27:09
Chinese industry really using AI.
27:13
Like for example, EVs companies in China are now are able to
27:18
turn on models in like a year and a half because the process
27:24
got speed up by the usage of AI in designing the cars.
27:29
And whereas America and Western countries, it still has much
27:33
longer development cycles because of that.
27:36
So there is that and I think AI robots and any kind of AI
27:41
robotic stuff is a huge industry in the future.
27:46
And you know, it's very hard for people in America to just do
27:52
software and wait for like the partners in Shenzhen to do the
27:57
hardware part of it because one, you're not in the same time
28:01
zone, you're not next to them. So the just in terms of getting
28:10
help and things like that, you can't iterate product as quickly
28:15
when you're not next to the, the supply chain.
28:21
Fair. And I can, and the reason for
28:23
this is, is pretty obvious, right?
28:25
Like if I'm doing product development and I need someone
28:28
from Shenzhen, I need to be on the phone every night at like
28:32
10:00. And if you have a family, that's
28:35
a real problem. So, so it's very so it's very
28:41
hard to iterate your product at the same speed as as a team in
28:47
China that can speak the same language that can try it out,
28:51
that has more access to a lot of data also and can get the cost
28:58
down because they don't have to go through Indonesia and then
29:01
ship the product over here to America, right.
29:04
So it's a real thing. OK, this is a real thing.
29:09
That's. Thanks to to tariff, right?
29:11
That's thanks to Trump's tariff, right?
29:13
OK, Yeah. And by the way, the Indonesian
29:16
factories, they're running like they're very running very hot
29:18
right now. So, you know, like, it's not
29:21
like they have all infinite capacity that can just get this
29:23
stuff through. Trump is probably gonna have to
29:26
make the tariffs 500% in order for the manufacturing to come of
29:30
that stuff to come back to America.
29:31
Probably. Well, I mean, the, the, the
29:33
problem with the tariff is, you know, tariff could work if they,
29:37
it's come, it's paired with the industrial policy.
29:39
But Trump administration think you, you, you, you can, you
29:43
know, you, you can just do tariff and everything will
29:46
magically resolve itself. But that's not how it works.
29:49
You know, traditionally tariff has been paired with industrial
29:52
policy to actually target a specific segment of
29:55
manufacturing. You know, tariff works if you
29:58
actually have to manufacture. But right now, yeah, this what
30:04
what Trump is actually doing is create creating a whole lot of
30:07
uncertainty in the market and and one thing the business
30:11
really hate is uncertainty. Talk to me about Kimmy because
30:16
this is a new development, right?
30:18
Comments on our our last shows video.
30:21
People are saying, oh, Kimmy is so good is so much better than
30:25
deep sea. I'm just hearing about Kimmy
30:28
right now. This is a This is a new language
30:31
model that just came out of China.
30:33
Yeah, like they've been were there, they've been like, you
30:37
know, kind of under the radar AI lab in China for a while now.
30:41
And I think they, they actually came up with the last model that
30:46
came up was actually pretty good.
30:48
And there was a lot of positive reviews of their model from the,
30:55
the AI community here in America just because they came out with
30:58
good papers. But deepstick took all the,
31:02
sucked all the attention away because, you know, it got hyped
31:06
up by, by Silicon Valley, right? And you know, they kind of
31:10
deserve they're, they're pretty smart guys and stuff like that.
31:14
But it just shows like in China itself.
31:16
And it's there are multiple AI labs who are doing great things.
31:22
And the Kimi model that came out is actually, I haven't
31:26
personally tried it myself, but it's like not a reasoning model.
31:30
So it's like, and it's it, it's better than the the most recent
31:35
deep sea V3 model, which is, which was the, you know, the
31:39
best Chinese open source like non reasoning model up to that
31:46
point. And so I think it'll be
31:50
interesting to see what happens. Like I actually need, I still
31:53
need to test it out. People say it's really good with
31:55
coding and things like that. And possibly it could be really
32:00
good for deep research. But the, the Chinese labs, they
32:04
really have to do something to stand out because there's so
32:07
many good Chinese labs right now.
32:09
Like opening the eye a couple of weeks ago, they were spreading
32:14
alarm about triple. Triple.
32:18
Yeah, yeah. I don't know exactly how the
32:21
Chinese word, I'm not sure exactly what it is, but yeah.
32:25
So there's multiple of them that are really good and they have
32:31
access to the data. They have enough compute power
32:35
right now, from what I can tell, and they're all running on much
32:42
lower cost basis than American competitors because they're not
32:45
like paying $100 million for the top talents.
32:48
So that is so crazy, man. I mean that that's like when I
32:54
find out that the compensation package they're offering, well,
32:58
that's, that's just Zach, right. Oh, you're in the wrong
33:01
industry. Seriously.
33:03
Right. I I mean, but, but the 100
33:09
million compensation package that that's just from
33:12
Zuckerberg, right or is there is more?
33:15
I, I, I think, I, I think that's like a Zuckerberg madness.
33:19
Well, we'll see like how, what, what the top salaries other
33:22
companies start offering. But like I'm, I'm sure like
33:25
quite a few people are getting at least several millions a
33:28
year, but 100,000,000 still. Right, that's that's.
33:31
Not it's still like kind of crazy actually to think about
33:34
that's like, that's like 10 times more than probably what
33:38
Deep Seeker spent during the entire lifetime of the firm.
33:44
Oh yeah, and yeah. I I just wonder when this entire
33:50
stuff is going to blow up because there's the economics
33:55
behind it is just not there at the moment unless you get into
34:01
the, the hard product of it. They were saying all this time
34:03
about like how they're going to cure cancer and things like that
34:06
from a large language models. But you still, you know, that's
34:10
you still need the the the firms to use it to help you to this,
34:15
you know, that only speeds up product development.
34:18
Yeah, it helps you, but like there's still need PPP people
34:23
paying for these things at the end and there's just a lot of
34:27
money going in and NVIDIA is making a whole lot of money.
34:30
But yeah, I I haven't seen a huge pickup in like usage from
34:37
call centers or whatever. I was just going to ask, how is
34:41
NVIDIA now the most valued company in the world when
34:44
they're being cut off on like one of their biggest market,
34:48
China? Is it all due to transshipment
34:51
or because it's AI bubble in United States?
34:55
I think it's a combination of that.
34:57
I think the Chinese transshipment thing is going to
35:01
continue for a while longer. And also I think people don't
35:08
realize that this demand for computation just can't keep
35:11
growing at this pace. People think that a lot of
35:15
investors seem to think you just keep buy into it because this
35:19
this, this tree would just keep growing exponentially.
35:23
But if you look at the actual demand for this from the third
35:31
party cloud providers, whether it's also the hyperscalers, they
35:37
only need so much chips. And there's just I just, I just
35:47
don't, I just don't see the the IT has to be like when when you
35:52
have someone on the on the upstream charging that much
35:55
money, revenue has to be generated on the downstream for
35:59
it to be worth it eventually, right?
36:01
There's just a lot of money going, a lot of hype money going
36:03
in. And when are they going to
36:07
generate the results? And at some point, I do think
36:10
Huawei is going to take almost all the market share inside
36:16
China once their software part of it gets sorted out.
36:21
So the issue with Huawei right now is they the software support
36:27
isn't as good as NVIDIA because a lot of the libraries, a lot of
36:33
the support were designed with CUDA in mind.
36:37
So they're just things you can run CUDA.
36:38
Everyone supports CUDA when they write these Python libraries for
36:43
AI inference, right? So until Huawei gets all the
36:49
software part of it sorted out, there's always like a
36:53
disadvantage in using Ascend chips versus CUDA chips.
36:59
So Huawei has to sort this out. So at some point Chinese firms
37:04
stops using NVIDIA chips as much.
37:08
And at the same time it's expanding all these like data
37:12
centers globally where they were going to use the full like
37:17
Chinese Huawei stack of AI solutions.
37:20
And and I think a lot of the global S countries are going to
37:24
use that and use the Chinese technology instead of the more
37:28
expensive in video technology. Speaking, Speaking of, of that,
37:38
you know, you know, recently there's a Germany just pushed
37:41
for the bang of DeepSeek, right? Yeah, IA lot of these things I
37:50
don't think really matters all that much.
37:55
I think, I think at the end of the day, there's the app
38:05
component of it and there's the API component of it and there's
38:10
application component of it. And it is possible that the,
38:17
that just due to national security fears or whatever, the
38:22
American apps for AI, at least in Western countries, will
38:28
always be dominant because there's an entire TikTok, you
38:31
know, story behind that. Like there's, there's always
38:35
going to be national security fear mongering going on.
38:39
But I think what's going to happen is if, if your industry
38:45
doesn't use AI in these countries, you're going to fall
38:51
behind. And what we're seeing is outside
38:55
of China and USAI adoption isn't at the same level, I would say.
39:04
So the industries that are in these countries are already
39:08
falling behind and they have to make a decision at some point on
39:16
whether they, they use the, the, you know, closed source American
39:25
model versus the open source Chinese model where they have,
39:30
you know, full control over it. And it, and I feel like a lot of
39:36
countries will probably end up picking the close off source
39:40
model just because they have friendly they, you know,
39:44
political reason for more for geopolitical reasons than
39:47
anything else, right? So, but there's still pretty
39:51
large market out there for the Chinese models just because
39:54
inherently developers like to work with open source models.
40:00
Right, fair. And I think all the future AI
40:03
robots, as I said earlier in the show are going to be made in
40:07
China with most likely Chinese AI models.
40:11
That's that's my opinion on this matter.
40:13
What's the temporary Trump ban on CUDA and other AI design
40:19
software to China? Did that have an impact?
40:22
I think probably not because the the EDA's, once you have your,
40:30
once you have your design finalized and sent to the, the,
40:35
the foundries, they can, they can just produce the chips.
40:41
They don't really need additional permissioning to do
40:44
those. But if you're trying to design
40:46
new chips and you don't have the the EDA's, then it becomes a
40:50
problem for you. So I think it was annoyance, but
40:58
not for not for the AI like it doesn't like Huawei would have
41:04
already finished its design of that scent and it's all
41:07
manufactured in China anyways. And they also you do have Edas
41:15
for seven nanometer processes inside China.
41:18
The problem was a lot of the fabs inside China are still
41:22
using American Edas because it's just, it's just there's, there's
41:30
always a switching cost when you go from your current software
41:35
and hardware to a new software hardware, right?
41:38
So people don't necessarily do it until they have to do it.
41:43
So I think there's going to be a greater shift away from American
41:47
EDA's as a result of this, because I think like shifting
41:51
software is still easier than shifting hardware.
41:55
Oh yeah. Oh, yeah.
41:57
But I mean, like, I guess what I meant was did that, did that
42:03
temporary ban light a fire under the Chinese companies in their
42:07
decision to possibly switch over to Huawei software, you know?
42:14
It's, it's hard to say. So you would think it would, but
42:19
I think it's probably going to take another year or so for us
42:22
to really tell how much of A shift that is because we would
42:26
end up seeing what normally happens is this, these
42:31
companies, they have their earnings call and then the
42:34
earnings call, they will ask them questions about different
42:36
regions. And then they will say like, oh,
42:39
we saw this, this, this year that the, this, we suddenly
42:43
start seeing our Chinese revenue dropping because we believe they
42:47
found out like a local supplier for it.
42:51
And that's, that's happening pretty gradually right now.
42:54
Pretty gradually. A lot of the American equipment
42:59
makers and Japanese ones are still making a lot of their
43:02
revenues in China, but for semiconductor.
43:05
But The thing is, it's because China's Chinese market is
43:09
growing so fast that the domestic suppliers just can't.
43:15
They can't expand 4050% a year and there's still a huge market.
43:21
So you're seeing like, I don't know if you've seen in the
43:24
recent months, there was a side carrier.
43:28
They were like facilities. They're now one of the largest
43:34
semiconductor manufacturing equipment makers in China now.
43:39
And they weren't really, they didn't really even advertise
43:42
their product until like to globally until a few months ago.
43:46
But a lot of the Chinese Staffs were already using them.
43:49
And you know, there's, there's a lot of players like that inside
43:52
China, but it still takes time for that to, for the industry,
43:56
entire industry to develop, right.
43:57
We, we want them to like grow overnight, but you can't go from
44:03
1 billion revenue to 100 billion revenue like in a year like
44:08
that's it just takes time to grow.
44:11
That's fair. That's fair.
44:14
Outside of the AI space or what is going on in in the Chinese
44:21
tech space, What, what, what, what are what other new
44:24
development in the Chinese tech space that we should be aware
44:27
of? So I, I want to talk a little
44:31
bit about just the, there's some misconceptions about rare earth
44:36
in general, OK, People think rare earth is just like digging
44:40
stuff out of the ground. Yep.
44:42
So people, there's what people don't realize is China is
44:48
actually a huge net importer of rare earths.
44:52
But what it does is it exports rarest magnets.
44:57
It stopped export and basically stop exporting rare, the raw
45:02
rare earths and it's also stopping exporting refined rare
45:06
earths, which which hasn't been the case until recently.
45:09
Now all these export going out are rare earths, Magnus.
45:12
So China's basically saying to them, you give us your raw rare
45:16
earths and we make the we're moving up the value chain,
45:20
right? And this is a pattern that China
45:24
has executed across wide variety of critical minerals.
45:29
And by doing this, it's moving up the material science supply
45:33
chain right where in the last, I would say 20 years, America has
45:39
shifted away from heart sciences and manufacturing to just
45:43
software and design. China has really spent a lot of
45:47
R&D into this so that it's biotech got really good, it's
45:51
chemistry got really good and it's mechanical engineering got
45:55
really good. And we're seeing the results
45:59
because it is now able to lead globally because I see rare
46:07
earths as material science in certain part of material science
46:11
and the other areas it's also catching up to to globally.
46:17
And the result of that is you're going to see huge improvements
46:23
in Chinese military versus what you see in America because a lot
46:31
of the the development in aerospace and industries like
46:36
that require the most advanced material science.
46:41
You know, some in some cases like rare earths, in other cases
46:45
it's carbon fiber composite and there's other kind of, you know,
46:52
synthetic fibers that China's getting increasingly really good
46:55
at. And the key point is that China
46:59
has all this, all this supply chain.
47:02
So it's had this effect where the, the, the downstream
47:09
industries, because they have this supply chain really close
47:13
to the, to the downstream producers.
47:17
You see these rapid iteration product development across many,
47:23
you know, mechanical engineering or chemical engineering type of
47:26
industries. And those are basically stuff
47:31
like aerospace rockets and things like that.
47:35
So areas that, you know, we used to think Chinese Chinese
47:39
industry is behind in we're going to see them really leap
47:43
ahead in the next 5 or 10 years, right.
47:48
And recently I posted basic a picture of the drones that China
47:56
was making ten years ago versus what they're making now 10 years
48:00
ago. They are their drones currently
48:06
the most their most recent like large drone looks like basically
48:11
a more more stealthy and larger version of B 210 years ago their
48:21
their drones didn't have they're very limited in their range.
48:26
They're very limited in their stealth and they were very
48:30
limited in the the kind of weapons they could carry.
48:34
And all that is just due to the fact that once you have this
48:40
material science supply chain domestically and you got really
48:44
good at the electrical part of things, you get really good at
48:49
the mechanical part of things. And then you can iterate through
48:54
the engineering cycle really quickly.
48:58
And I don't think people outside of China really realize that
49:04
yet. China's export control on real
49:08
earth, is that still on against US defense industries is I mean,
49:14
is there so I, I understand China extended licenses for for
49:20
US car makers that they granted 6 month export license to, to
49:26
this US export. But what about, you know,
49:29
Lockheed Martin? You know they they still need
49:32
Sumerian cobalt for Sumerian cobalt magnets for their F35 and
49:38
missiles right? Yeah, I think that part of
49:42
things we will what were all the articles we read show that they
49:49
have not opened up export licenses for those things.
49:53
Maybe they're looking to trade that for certain reduction US
50:01
export controls of the semi equipments.
50:06
But if I was China, I would just leave that as is and just try to
50:12
make the equipment, you know, develop the equipment myself,
50:15
right? And because it will take many
50:19
more years for America to develop like same level as rare
50:23
Earth's magnets as China has right now.
50:26
Because you know, there are, once you get the rare Earth's
50:33
magnets produced, there's still different degrees on their
50:35
performance level. People would never say a 90
50:41
nanometer chips is as good as a three nanometer chip.
50:46
Yet people think that rare earths magnets are all the same.
50:51
But clearly if you have a higher purity rare earths alloy and
50:56
then convert it to like a really well done magnet, it's going to
51:00
be significantly higher performing in terms of able to
51:05
operate at a higher temperature range and operate under like
51:10
higher vibration than something that was developed with lower
51:14
purity level and designed to like a little, you know, less
51:20
impactful magnets, for example. So in the future you might get
51:24
into a case where China still sells magnets, but you would
51:27
only sell the lower end stuff globally.
51:30
So even if US defense industry can somehow get like
51:35
transshipments of the magnets that's designed for export to
51:40
say car industry, you would not be the same grade magnets that
51:45
we need for military applications, right?
51:47
That's what you're. Saying yeah, yeah, like it would
51:49
be very obvious something like Sumerian, right, That it's
51:52
pretty much only for based on what I read, it's only for
51:57
military industry. There's just no reason for China
52:01
to export it to like some automaker somewhere.
52:04
Makes sense. And and I think it's also very
52:07
short sighted if we have a company that had that actually
52:12
needs rare earth magnets to get it and then give it to the
52:18
defense contractors. Yeah, my, my, I do wonder like
52:22
what's going to happen to a company like Boeing, you know,
52:25
like need need rarest magnets for their civilian products.
52:31
Yeah, but they also have the huge military program, right?
52:33
Like those those those kind of things we don't really know,
52:37
like how tightly China is actually controlling the the
52:41
export licenses. Interesting, interesting.
52:47
So yeah, I mean, United States have been aware of the rare
52:52
earths dependency on China since 2000 O 9.
52:55
There was like congressional report back in 2000, O 9
52:58
screaming that we need to, we need to develop domestic supply
53:04
chain for rare earth because, you know, the over dependency on
53:07
China is dangerous. But now it's what, 20/20/25?
53:12
So it's more than a decade later, you know, 16 years later,
53:16
they still haven't done anything.
53:17
So I don't expect, you know, US will pick up this rare earth
53:22
production and rare earth refining capability anytime
53:26
soon, at least not within 10 years.
53:29
So yeah. Yeah.
53:31
I mean, it's, it's, it's very hard to see how that could
53:34
happen, right. But I, I would say that what's
53:41
more likely to happen is you're end up with a situation where
53:45
America does get some, you know, magnet production that are of
53:49
lower, lower quality, let's just say domestically.
53:56
And China keeps the higher end stuff for itself, which we can
53:59
use for like aerospace hypersonic missiles or or things
54:05
like that. Right, right.
54:09
So that that will have some serious implications for, for
54:12
Pentagon's plans to to contain China.
54:17
Anything else you'd like to cover in terms of, you know, any
54:23
new development in in Chinese tech tech?
54:28
Yeah, I think the, I think just generally AI robot is going to
54:31
be huge going forward. I think that we're already
54:35
seeing the the the humanoid robots really making the Chinese
54:40
factories. We're seeing like wide
54:43
application of different kind of small to large AI robots being
54:48
used for different purposes. And part of the reason is just
54:52
like the Chinese blue collar workforce is declining every
54:57
year, so it needs greater automation.
55:00
And in terms of AII think could be very interesting to see like
55:10
when Deep sea comes out with V4 which is a non reasoning model
55:16
and R2 which is the reasoning model because everyone expects
55:19
R2 to be really great. So we're kind of all waiting for
55:24
it because, you know, the entire open source community actually
55:27
likes open source model regardless of which country it
55:30
comes from. Well, I'm glad we have you to
55:33
give us update because right now, you know, just recently I
55:38
saw a tweet from Scott Adams, you know, the creator of Dilbert
55:42
Khartoum, he tweeted. It's like, oh, the coming of
55:47
robotics and automation will totally destroy China's
55:50
advantage in manufacturing because they will lose their
55:53
edge in cheap labor. I'm like what?
55:58
What? Do you think it's going on?
56:00
China leads the world in the robotic automation.
56:04
You know, China has more manufacturing robots and all the
56:08
rest of the world come by. Where is this guy getting his
56:11
information? And Scott Adams, not some no
56:15
ignoramus from like, I don't know, like he, he's, he's like a
56:19
part of the educated US elite, right.
56:22
Yet this is kind of outdated view they hold about, about tech
56:27
and about China. So yeah, so, so I, I'm glad that
56:32
I have you to set the record straight.
56:34
And we didn't even, we're at a one hour mark and we didn't even
56:37
talk about stuff like Chinese EV development and stuff like that.
56:41
But I hope to have you back to talk about that.
56:45
Next time, Yeah, I think that. I think that's also very
56:47
interesting. But like this entire robotic
56:49
thing, I only found out in the past few months and I think I
56:55
think very few people understand this.
56:58
Yeah, so, well, thank you. Thank you TP again for coming to
57:04
my show. If people want to, you know,
57:07
find you and and where can they find you?
57:11
They can find me on X or Twitter as TP Wang and they can also
57:16
find me on sub stacks as TP Wang.
57:19
OK, I'll have those links. Yeah, I'm pretty.
57:23
Consistent with my name. Everywhere, so I'll have those
57:26
links in the description below and people can check it out and
57:30
please subscribe. TP is I consider one of the
57:35
premier expert on Chinese tech AI and and you know, you name
57:40
it. Thank you again TP for coming
57:43
and thanks to everyone for listening.
57:45
Thanks, Carl. Hello my friends, I'm Carl Zaat
57:47
and you're watching Silk and Steel podcast.
57:51
Leave your comments below on what you think about this
57:54
content and what would you like to see more on my channel.
57:58
Thank you for watching. Bye bye.