Latest / UK Column Radio / US AI Chips to China? Huawei's Rise, DeepSeek, and China's Tech Future | Silk and Steel

Transcript

US AI Chips to China? Huawei's Rise, DeepSeek, and China's Tech Future | Silk and Steel · UK Column Radio

0:01
Hello and welcome to another episode of the Silk and Steel
0:04
podcast. I'm your host, Carl Zah.
0:08
Today we have back on the show returning guest TP Huang.
0:12
TP is a expert on Chinese AI and general tech development.
0:18
First, first of all, welcome back to the show TP.
0:21
Oh, glad to be back, Carl. So a lot of has happened since
0:25
we last time talked. I think we like talk every four
0:28
or five months and five month in like AI development is like it's
0:34
like years, decades. So I first I want to talk about
0:40
the latest news, which is Trump tweeted out that he's going to
0:45
lift the ban on NVIDIA SGAI chip H200 to China.
0:52
I think that we are putting ourselves at risk by basically
0:57
assisting the CCP with their military modernization and the
1:01
perpetration of human rights abuses by selling these highest
1:05
end chips which kind of power their AI models, whether it's in
1:11
hypersonic missiles or facial recognition algorithms used to
1:17
go after the Uyghurs, Tibetans, Hong Kongers, or.
1:20
I think this is probably the result of the intense lobby by
1:24
NVIDIA and also the consideration for the US
1:28
government that they don't want Huawei to take over the entire
1:33
Chinese AI market. At the moment, we are 100% out
1:37
of China and so China is 0% of we went from 95% market share to
1:43
0%. And so I can't imagine any
1:49
policy maker thinking that that's a good idea that whatever
1:52
policy we implemented caused one of caused America to lose 1 of
1:58
the largest markets in the world to 0%.
2:01
But anyhow, and all of our forecasts, if there are any
2:03
shareholders out there, all of our forecast worse, we're
2:06
assuming 00 for China. If anything happens in China,
2:10
which I hope it will, it'll be a bonus.
2:12
But it's a large market. China's the second largest
2:16
computer market in the world. It is a vibrant ecosystem.
2:20
I think it's a mistake for United States to not
2:22
participate. I want your take, you know,
2:26
what's your assessment on this move and how it will impact the
2:31
AI, both the AI development in China and as well as the semi,
2:36
the semiconductor industry? Yeah.
2:38
So I think in general, this is a sign that one, the people that
2:46
are in the US and like Trump administration, mostly David
2:49
Sacks have won the battle about, you know, making sure that
2:55
Huawei doesn't get a much larger market share because there's,
3:00
it's been pretty obvious that H20 itself is just not a
3:04
competitive product in the Chinese market.
3:07
So they have to provide something better in order to be
3:15
accepted, be bought by the by the Chinese producers.
3:18
I think with this news, you know, China's still doing this
3:22
very taking this news cautiously.
3:25
It will. My opinion is it will do some
3:29
kind of assessment to see whether or not the on banding of
3:35
H200 will actually make it easier for the domestic AI labs
3:42
to actually do their training of their models.
3:45
And they will only allow it if they think there's actually a
3:50
tangible benefit in this transaction.
3:55
And my guess is based on talking to some of the people on this is
3:59
they probably will allow some some importance.
4:03
It's not going to be like some crazy amount because they
4:06
already have, you know, smuggled in so much Hoppers and
4:10
Blackwells. And salvage valuable silicon
4:13
components from dead PCBS of banned GPU, innovatively hand
4:18
modifying them to be better than and have more VRAM than Nvidia's
4:22
own official products use I. Make one right now.
4:41
NVIDIA We even spoke with AUS based Chinese national buying
4:49
video cards in the US to strip them and ship the GPU's to
4:53
companies in China, violating the US export control law.
4:57
I haven't seen someone have a computer in a car for a long
4:59
time. Black market is normally a
5:02
phrase that's associated with drugs or guns, but there's a new
5:05
kind of black market and it's high end AIGPUS.
5:09
And this particular black market is worth billions of dollars a
5:12
year, and it's hiding in plain sight.
5:16
It's it's more like some of the some of the bigger institutions
5:19
like, you know, Alibaba and Tencent, people like that, they
5:24
don't dare to buy the the smuggle stuff directly.
5:30
So they've been buying stuff and you know, and putting them
5:34
oversea database data centers and stuff like that, right.
5:37
So they probably will buy some when this becomes available.
5:43
So I think that's like Byte dance Alibaba are are some of
5:47
the bigger customers that that could see be interested in this
5:52
and maybe maybe deep sequel would want some, I don't know.
5:56
But I don't really think this will affect the the long term
6:01
trajectory because if you think about it, right, the only reason
6:04
that you're able to get the more advanced chips is because the
6:10
domestic option became so much better, right.
6:14
So if your domestic option doesn't get supported, then
6:19
you're not going to get access to the to the better stuff.
6:21
So, but hopefully in the long term, your domestic stuff gets
6:24
so well supported that you just don't need the the Nvidia's and
6:28
AM DS of the world. And I think it's very important
6:32
to just say that if you look at the Chinese compute clusters,
6:36
the data centers and stuff like that, they've really grown a lot
6:40
recently. So, and that the, I don't think
6:45
that their expansion plans rest on what U.S. government might
6:50
decide to do. Because, you know, in the end,
6:54
you have to depend on yourself for everything if you're China.
6:57
So you can't depend depend on your think, wait for your rival
7:01
to give you something when they when they when they might not.
7:04
So H200 probably in my opinion, will be bought in some quantity,
7:09
but it's not going to be a game changer.
7:11
And some of the hysteria you'll see online from the US think
7:16
tankers are just just way out there.
7:20
And yeah, like that. I think it's, I think it's, I
7:27
think it's probably on that a good thing gives, you know,
7:30
Chinese AI labs a, an additional option, but it's not like a
7:34
super game changer that some people are concerned about.
7:39
OK. Well, then what about is the
7:43
potential impact on Chinese domestic domestic semiconductor
7:47
industry? You know, like how would this
7:49
affect Huawei's Ascent chip and its future market share?
7:55
So I think it's very interesting to see what Huawei is doing
7:58
because it is working on the software component of it.
8:02
So one of the things that you know, like that's as a software
8:05
developer, one of the things that you really hate is
8:08
switching text stacks, right? Like that is really annoying if
8:11
you roll everyone started off by writing their software's text
8:16
stack around NVIDIA and CUDA. Like if you ever tried to do use
8:22
some of the Pytorch libraries online, you'll find out very
8:26
quickly that the NVIDIA stuff is really well supported by all the
8:29
libraries, but the other stuff is not.
8:32
So I think a lot of the DeepSeek work in the past year, and they
8:38
might have even been slowed down by this is actually devote
8:42
resources so that their stuff is compatible with Ascend for
8:47
training and inference. So, and, you know, and it's,
8:51
it's kind of good for them too, because they get to work
8:53
intimately with Huawei to so that the so that they know all
8:58
the optimizations that they can do to take advantage of all the
9:01
little compute. What people don't realize about
9:04
NVIDIA is like, it's, it's architecture itself is like not
9:09
the most fantastic for, for AI computation.
9:12
It just happened to be the most the best fit available at the
9:15
time. So that's why everyone decided
9:19
to use them. But whether or not like, you
9:23
know, and then they just kept going on and and they're the
9:27
most well supported. And now, you know, you've seen
9:30
the news recently about how Google trained their Gemini
9:34
models on the TP use. And that's because Google
9:38
itself, the ASIC might be actually a better architecture
9:41
for AI computation than than NVIDIA CUDA is.
9:45
So, you know, like there, there is a very good chance that with
9:52
a lot of like, you know, human resource software help on the
9:55
Huawei and working together with deep sea, they can get a sense
10:00
the entire software stack to be, you know, as good in China as as
10:07
CUDA NVIDIA is. And that, and that's probably
10:09
the the most important part because you know, they long term
10:16
you everything is going to be running on Huawei long term, I
10:20
would say. And they're building a lot of
10:24
fabs in China right now. And it's probably going to be at
10:29
some point it's going to be all running on the Chinese
10:31
semiconductor manufacturing equipments also, right.
10:35
So like eventually it's all going to be domestic because
10:38
they just don't want foreign companies to wreck their plans
10:44
and and whatever. So it is very important that
10:50
Deep Sea is working with Huawei on this.
10:52
And and I don't believe that the other Chinese labs are working
10:56
as intimately with Huawei. So, so it's, so it's, so it's
11:00
very important that Deep Sea was doing this and that at some
11:05
point next year, I think you'll see like a really much better
11:09
acceleration in the deep Sea models just because the fruit of
11:14
these results will happen. And then they will be getting a
11:16
lot more compute power behind it.
11:19
And, and you know, and then, and then we'll see what happens.
11:24
Because in essence, like DeepSeek itself is still like,
11:28
it's very much like a pure AI lab kind of things.
11:31
It's not thinking too much about the commercialization of the AI
11:35
technology. It's, it's more just thinking
11:37
about, we just want to push the frontier, you know, like they,
11:40
they're very much about that. They're able to keep the cost
11:43
down. Like, you know, and the, the
11:47
Chinese government probably told like Huawei that you should work
11:50
with, with DeepSeek as much as you can and give them like cheap
11:53
compute resources so they can do get all the stuff they need to
11:56
be done finished. So I think that I think that's,
12:01
I think that's very bullish for the, for the Chinese market long
12:04
term. And you know, if they need to
12:09
have some H200, it's not end of the world.
12:12
It doesn't. It doesn't change the long term
12:14
trajectory of things. How have Huawei been making
12:18
progress getting people and, and companies to adopt their AI, to
12:24
switch over to their architecture and AI chips?
12:27
I, I remember there's some Western financial press
12:30
reporting, you know, some big Chinese companies are kind of
12:33
resisting the government pressure to switch over to
12:37
Huawei because they're so used to, you know, use NVIDIA and,
12:42
and CUDA. Have you heard anything you know
12:46
through the Grapevine on the level of Huawei adoption among
12:51
Chinese companies? Yeah.
12:55
So it depends on the lab we're talking about.
12:59
Alibaba uses its own chip. It's called like it has its own
13:04
chip design from tea head or in Chinese, it's called Pintolgua
13:11
and it designs its own AI chips that uses and it's, it's kind of
13:16
like not as good as Huawei's chip, but it's, it's, it's
13:19
internal. So they can design their own,
13:22
their chips to fit their own team.
13:23
Like, you know, when you design your own chip, you can work with
13:26
your internal AI lab to figure out like what they need to to
13:31
fine tune it, right. And I think that they probably
13:38
don't use ascend. My guess is that bike dance uses
13:44
a combination of NVIDIA Cambrikong and also ascend
13:51
chips. They're probably like 5050 ten
13:56
cent. They probably use some NA Huawei
13:59
chips also because there's there's there's there is in
14:02
China. There is some kind of like deep
14:04
de trust between the between the major hyperscalers because they
14:10
don't want to work too, too closely with Huawei.
14:15
That's part of the issue. Baidu also designs their own
14:19
chips. It's called, they have their own
14:21
team called Quinlanching and that they're, you know, has been
14:27
designing AI chips for like many years now.
14:30
So they got their own option also and most of the.
14:35
Baidu and Baidu and and Alibaba, they're they're not on the
14:42
entity list, so they can have their AI chips be manufactured
14:47
in TSMC by TSMC. I'm not sure if they're
14:52
necessarily manufacturing TSMC or Samsung.
14:55
They might be. I think they might be
14:57
manufactured locally, but if they are able to manufacture
15:03
theirs in TSMC, then there is a ceiling of how much compute
15:07
there can be. So you can design your own chip,
15:11
but like it can't exceed whatever NVIDIA is is able to
15:17
sell to to China. So, so there is something like
15:20
that. I don't know the exact details,
15:22
but you know, it's something like that.
15:27
The other thing is DeepSeek is probably the one that's working
15:33
that's most frontier and also use Huawei, but the Huawei chips
15:37
themselves are pretty widely deployed in like a lot of the
15:40
big public data centers and also like a lot of the big Chinese
15:46
cloud providers like China Mobile, China Telecom, China
15:51
Unicom and people like that and they're perfectly fine for
15:57
inference. So you know like you can run
16:00
Quinn models on them. You can run DeepSeek models or
16:05
Zai models. Zai, I'm pretty sure, and I just
16:09
heard about this today. They they basically train their
16:12
stuff on 10,000 H 100 cluster that they have and they might be
16:18
one of the people that would be interested in buying some H200
16:21
from from NVIDIA. So, so like, yeah, so I was
16:27
just, I was just going to say that the training a lot of
16:29
people still do their so bad point is a lot of people still
16:32
do their training on the each on the Hopper series and maybe
16:36
Blackwell, when that's ready. But you know, the in terms of
16:42
training on SSN, it's basically they're still trying to make
16:48
sure that it functions as well as the NVIDIA stack, text stack.
16:52
So that's that's where things are at.
16:54
Speaking of different models, there's so many Chinese AM model
16:59
now. There's DeepSeek, there's Q win,
17:02
there's Kimi. Which one is the best nowadays?
17:07
I mean like they they all have pretty good reviews what I
17:12
heard. Yeah, so they all have pretty
17:13
good reviews. The most interesting ones to
17:16
watch for are probably the Quinn models and also the the deep Sea
17:21
models. The Quinn models are kind of
17:23
interesting. They are probably the most
17:27
downloaded open source models now and they're also the most
17:32
the models that people use for doing fine tuning and doing
17:36
derivative models. So like if you look at the
17:40
download chart, what's cool about Quinn 3 is that it has
17:44
different size models from like the smallest model size to the
17:49
to the largest size to like several 100 billion parameters.
17:53
And the small models are really useful to run AI on X devices
17:59
and things like that. So what you're going to see is
18:03
that for a lot of people, even in the West, they will use Quinn
18:09
models because they can fine tune it really easily and then
18:13
they can adapt it for their own use.
18:15
So there was a story that from a 16 ZA while ago about how like
18:20
80% of the the open source models that Silicon Valley uses
18:25
are like Quinn, Quinn models, right?
18:29
So, and the reason for that is because everyone knows when you
18:33
say use open source model in in the valley, you're using clan
18:36
models because it's so easy to find you for your robotic
18:41
applications and things like that.
18:44
Is Chinese AI development AI models in your assessment?
18:49
Are they still behind the, the, the USAI models or, or if they
18:55
are, how much are they still behind?
18:57
Because I hear different assessment.
18:59
You know, you, you hear people saying they're only nanoseconds
19:02
behind. Have they?
19:04
Have they almost? Have they caught?
19:06
Up. So it depends on the area,
19:07
right. Like I think in terms of the
19:09
latest LAR Frontier models, they're they're still probably a
19:14
couple months behind. I would say when the deep sea,
19:18
deep sea 3.2 came out, that one was really good.
19:22
That one was actually, I think something close to
19:25
state-of-the-art. Like it was as good as the IT
19:29
was almost as good as the Gemini models, but you know, lack some
19:32
tool calling stuff, but it's really good.
19:35
But it was better than the Gemini and GBT models on the on
19:40
some of the math solving and like deep reasoning science
19:43
research and things like that. But it was worse than some of
19:46
the tool calling that AP is that people care about these days.
19:50
So depending on what you're looking at, you know, the, the
19:53
DeepSeek models are already at that level and the, the Kimi K2
19:57
was also really good when it came out.
19:59
And there's, you know, there's some rankings on that.
20:01
I personally done testing myself of the, the Gemini models I have
20:05
access for free online versus DeepSeek.
20:08
And maybe I'm sure the, the Gemini 3 pro might be a little
20:13
better than what I was testing against, but I found the, the,
20:15
the DeepSeek models to work better for me.
20:18
So you know I use it for programming and crash.
20:21
So what's the argument for people to to adopt the Chinese
20:27
models? Is it just because it's open
20:29
source? Yeah, so I think the, the, the,
20:32
the cost is part is a big part of it.
20:34
But the open source is a big deal, right?
20:36
Like if you give developers a chance to to use open source
20:41
models, developers will always use open source models when they
20:44
can, because then they can do their own fine tuning.
20:48
They have more control over it. If you have a model that you can
20:51
run off your computer and you don't want to like send your
20:54
data across, then it's always better that way.
20:57
Like if you're a big corporation, right, and you
21:00
don't want your data to go to to open AI and they can see
21:05
everything you're sending through or Google, then you will
21:09
prefer to have it in house somewhere that you can set up
21:13
backups for, right? So hello, one thing, one thing
21:20
that, yeah, sorry, one thing that I've noticed is that open
21:25
AI constantly have these reliability issues also.
21:29
So you will use them and then because Cloudflare goes down or,
21:33
or something like their servers can't handle the computer or
21:36
whatever, they just can't handle requests.
21:40
So there's reliability issues if you use like some of these AI
21:45
labs and you know, if you use it and if you do it internally and
21:51
you have your open source models, then you can have your
21:53
own primary computers that run them and then you can have the
21:56
backup to run them. So for a lot of corporations,
22:02
that's actually more preferable to actually sending your data
22:05
across to open AI. And you can also see that people
22:11
are don't trust open AI at all because there's more, there's,
22:17
they're losing market share even in America to like Anthropic and
22:21
Google. So that's, so that's a big part
22:26
of it. And there's a lot of
22:28
applications out there you can only use small open source
22:31
models for like if you have a AI robot, the only thing you can,
22:36
you can put in there is is a small model you that can run
22:39
locally. So it pretty much because Llama
22:45
gave up on, you know, whatever they were doing.
22:48
So right now Quint is the only thing out there that people can
22:51
use. So like every small model out
22:54
there, any derivative model are based off like Quint 3 Series.
23:01
Like there's a chart recently of of of the AI open source model
23:05
download between China and US and China has surpassed US few
23:10
months ago. And the gap is.
23:12
Growing based on our conversation or now our past
23:15
conversation, you are saying currently the the AI labs in
23:21
China have enough compute to do what they need.
23:24
But going forward they will still need to breakthrough this
23:30
semiconductor bottleneck, which is, you know the the whole
23:36
purpose of the sanction is to keep, to make sure China do not
23:43
have the ability to manufacture advanced ship.
23:46
Have China made any progress in last year or so on that front?
23:51
Yeah. So there's there's two elements
23:54
to that, right. There's one is whether they will
23:56
have enough compute and 2 is how they're doing, how well they're
24:00
doing in the semiconductor part of things.
24:03
So I think that they will have enough compute even if they
24:08
don't buy, you know, don't get official access to NVIDIA chips
24:13
just because of the, you know, the build out and the growth of
24:17
the, the China's advanced semi capacity.
24:22
And there's going to be a lot more 7, you know, 5 to 7
24:26
nanometer capacity that come online in the next couple years.
24:30
And they will definitely be making the AI chips.
24:35
If you look at Huawei's like recent Ascend, like from a few
24:39
months ago, they did like an Ascend presentation where they
24:43
showcase their plans for the next three to four years, right?
24:47
And it's very interesting because they're already
24:52
showcasing these systems that are basically it's, you know,
24:59
they take advantage of like Huawei's advanced networking and
25:08
optical connection technologies. So what that means is that per
25:16
individual chips and due to the fact that there are limitations
25:20
to the Chinese semi fabbing stuff, you might not have the
25:24
best chips, but by connecting them all together in a system,
25:30
it functions better. So they so they'll do these
25:33
things called super nodes, which is basically you link everything
25:37
together. You have these really quick,
25:40
really fast interconnect that can so the so the chips can talk
25:43
to each other. So as a whole, they function,
25:46
they get utilized better and they can they can process stuff
25:49
better. And then you connect these super
25:52
nodes together in a super cluster.
25:55
So you can have like these clusters that require like one
26:00
GW of electricity. And then not only would it, but
26:06
it's like really powerful. You can do like 2000 E flops of
26:11
a computation or whatever, something like that.
26:13
It's, it's some of the numbers are pretty, pretty crazy
26:16
actually. And I think that in order for it
26:19
to be even talking about these plans, they must have capacity
26:22
to produce the chips, right? Because why would you, why would
26:26
you present these plans saying that you can, that you can build
26:29
data centers with 2000 E flops of compute if you can't actually
26:35
make the chips And, and a lot of these things would require like
26:39
millions and millions of chips. So like, you know, that's,
26:42
that's the scale we're talking about here.
26:45
So that's so I think in terms of compute, you know, they will
26:49
have it. The maybe some of the labs still
26:53
prefer to use NVIDIA as a training.
26:55
And that's where the, the continued smuggling of, you
26:58
know, Blackwell will come in to play.
27:01
But in terms of inference, which is where most of the demand is,
27:05
they will be probably using ascent clusters on these public
27:10
data centers long term. Well, I saw on Twitter, you
27:15
know, of course, the most reliable information on the
27:18
Internet, someone made the case that US will stay far behind,
27:24
far ahead, because US had built thousands of data centers who
27:29
were as it completely dwarfs in numbers that's currently in
27:34
China. Now, is that comparison fair?
27:37
No, no, no, no. They, they, they have no idea.
27:39
Like how much computer China has the the the, the, the China has
27:44
this EDWC projects, which is East data W compute project.
27:50
What basically the idea is like the demand for compute is all in
27:55
the East Coast. So they need to build like data
27:58
centers in the western part of China where all these free and
28:03
cheap renewable energy is. And then they can build these
28:06
huge clusters in Qinghai, Gansu places, Inner Mongolia and
28:13
places like that, right? So that they can get these cheap
28:19
compute resources. And there was my, my friend Dave
28:24
Fishman actually posted like went to a recent conference
28:27
where the Chinese, you know, state grid showcase what they
28:33
expect, where they expect the, the electricity growth to come
28:36
from for the next 5, for the 15th 5 year plan.
28:41
And the interesting part that the plan says is that about 20,
28:45
they expect about 21% of the, the gross and electricity demand
28:48
to come from AI and computation servers and things like that.
28:54
And the number that they gave was like something like 642
28:59
terawatt hours. Just to put just put 642
29:03
terawatt hours into perspective. That is a, that is basically
29:08
having about 75 GW of data centers running 24/7 with no
29:13
interruption. I think in reality, that's
29:15
probably, you know, the data centers are never going to run a
29:19
hundred 100% capacity, you know, full year round, right?
29:23
There's always going to be clusters that need to be taken
29:26
down for maintenance or whatever.
29:27
So we're probably talking about like they, they, they believe
29:33
that they can build 80 to 90 GW of data centers, which is
29:36
basically what the Silicon Valley people think that America
29:40
will build, except America doesn't have the the grid to
29:44
support it. So I'll give you a story, right,
29:49
because my, my friend works on works on Meta and they're,
29:52
they're building the data center out in Louisiana.
29:55
I think they're already saying that some basically their their
30:02
their data center plants. They're only going to get to
30:06
like maybe 50% of the data center plants they were planning
30:10
to for 2026 because they were told that the the grid wasn't
30:14
ready. Again, on Twitter, I see all
30:16
these people saying, oh, not a problem, we can build gas
30:20
generators. You know, we can build these AI
30:25
company can build their own gas power generators, which
30:28
generates the electricity. Do you think that's realistic?
30:31
Well, that's there's a reason why the gas turbine companies
30:34
are G Renova is doing really well right now, right?
30:37
Because the demand for gas turbines is off the hook.
30:40
But but you have to produce a gas turbines and there's only so
30:44
much capacity out there for the gas turbines, right?
30:46
Like they're trying to raise it. But it's not like something
30:50
where you can, you can quickly increase in a short term and
30:54
then you also need the backup diesel generators.
30:58
And there's also a capacity factor for that also.
31:01
And there's also some people would use, there's a lot of
31:07
constraint out there that people are not appreciating.
31:10
When you increase capacity in short term there is going to be
31:20
production quality issues. And I, I was told specifically
31:25
from my friend a matter that they're seeing production
31:28
quality issues with the Corning tables and some of the other
31:34
items. Everyone wants to use the
31:39
optical network now, which use requires Indium phosphide and
31:46
that the production in China of this thing is, is just not going
31:49
up as much as people think that they are.
31:52
So there's, there's, there's a, there's many choke points in the
31:58
supply network for building like an large data center network.
32:03
And large part of that supply chain is in China.
32:08
And you know, the Chinese providers are going to, you
32:11
know, favor themselves first before the Western providers,
32:16
right? There's only so much cooling
32:18
capacity that are available. You need copper for the
32:24
electricity for the data center. You need aluminum, right?
32:28
America is seeing like these huge price jumps in copper
32:32
because the demand for it is so high.
32:34
And then they put a 50% tariff on the on copper.
32:38
So then, so then you can see like I was posting a chart
32:41
recently, the import of copper from China into America went
32:46
down 25% year on year. And then, and then you wonder
32:50
why the, that the copper prices are all time high and that's
32:53
terrible for the AI data centers, right?
32:55
Like they're, they're, it's almost comical.
33:00
Some of the, some of the, the factors involved here.
33:03
And people just don't understand that when you increase the
33:08
capacities, not only do you need more AI chips and and HBMS, but
33:14
you also need these power chips, heat sinks, the the cross
33:19
connect the the fiber networks, you need the what they call the
33:26
optical transceivers. And the optical transceivers
33:29
need the raw material that the wafers and rafers need the raw
33:33
materials and lot of that stuff only get sourced in China
33:38
basically. So this is a question I asked
33:42
you last time. You know, I remember I asked you
33:45
maybe 6-7 months ago that how close are we to AGI with all
33:55
these investment going in to AI right now and you said it could
34:00
be as short as one year. Are we, are we still on that
34:04
schedule to AGI you think or? It depends on what you define
34:09
AGI I think. If it's something that consumes
34:11
off the Internet and tells you an answer right?
34:15
And pretty smart doing that. Like, you know, it might appear
34:20
to people that we have AGI when you're dealing with it in, in a
34:24
couple of years because it's really smart and answering
34:27
things that scour the Internet for.
34:29
But a lot of these like models, all these models still don't
34:33
know what day of the year it is. They still don't know what, what
34:36
the date is, right? Like that's, that's, that's the
34:39
same ball AI models, like whatever they're trained at that
34:43
they think the date is. So there's some still some basic
34:46
stuff they just don't understand.
34:47
So there's so there's been some questions recently about whether
34:50
or not HGI itself is actually quite far away and that the
34:55
scaling laws that people use right now just aren't the
34:57
solution. So in my opinion, a lot of the
35:05
Silicon Valley people are too obsessed by HGI itself.
35:10
It's, it's more about how you can put it into application that
35:13
really matters. So stuff like drug discovery,
35:19
stuff like advanced manufacturing, robotics, and,
35:25
and that's where the money is to be made.
35:27
We really, you know, it's, it's kind of like most of the actual
35:35
use cases for AI right now are not all that useful.
35:39
Like it's basically just keeping people hooked on to their
35:42
phones. It's like AI slops.
35:45
That's where, you know, a lot of the demand is.
35:49
And yeah, if you talk to a lot of the finance and Silicon
35:55
Valley people, they use AIA more productive case and they think
35:59
it's the greatest thing out there, right?
36:02
Because if you use it, you can see you're doing smart things.
36:05
You think there's just so much value in it, but you know, the
36:09
vast amount of the consumer demand right now is still just
36:15
AI slots basically. Yeah, well, I do.
36:21
I do use AI for writing up my YouTube, YouTube show notes.
36:27
It's pretty good at that. You know, I use DeepSeek.
36:30
But but now, now recently I start to explore other options
36:33
because, you know, deep sea default client has too much
36:38
censorship building. And I talk and a lot of my talk
36:43
topic, podcast topic has to do with China.
36:46
So it triggers like the censorship very easily.
36:50
And I'm just too lazy to install DeepSeek locally.
36:53
I don't want to take half a day to do that.
36:56
And so I'm anyway, OK, So what do you think right now?
37:03
So, so you so in your assessment, China right now is
37:09
pretty close in AI development with on par with US?
37:14
Yeah, like I think it's pretty close all like across the board,
37:19
right? It's closer than it was a year
37:22
ago. There's, there's, there's this,
37:24
there's this belief by like all these finance people, you just
37:27
need to limit China's compute. And then, oh, the, the, the
37:32
compute advantage will work on the gap will expand.
37:35
Again. There's been no evidence of
37:37
that. Because if you think about
37:39
where, where the gap was a year ago versus now, it's, it's
37:42
clearly a lot smaller. And in terms of like the
37:47
diffusion of AI into robots and things like that.
37:50
So small models, China is way ahead on that now just because
37:56
the US labs are not working on it.
37:59
They're very singly focused on people using their, their,
38:05
their, their API services, They're super smart research
38:09
platforms and things like that. They're not thinking about the
38:11
robotics part of it at all. Seems like, I mean, Trump talks
38:15
about it, but it's like, yeah, it's, it's it's not the focus.
38:21
So there's been some kind of panic, I guess in Silicon Valley
38:26
when they realized that the that the China is way ahead in the
38:30
robotics part of it. You hear a lot a lot about that
38:33
recently and. Mark and Jason talk about it all
38:36
the time, yeah. Not just him, Eric Schmidt, our
38:39
favorite, our buddy there. But The thing is like these
38:43
people don't understand manufacturing.
38:46
They they look at the humanoid robot and they think that's what
38:48
they need. And, and I'm pretty sure that
38:52
the that the Elon Musk, there's a reason that people think
38:57
humanoid robot is what we need. And I think he only gets does
39:02
humanoid robot because he eventually want a sex bot.
39:05
I really think that's what it is, but they're actually not but
39:09
but. China is actually making a lot
39:11
of progress on the humanoid robots right now.
39:14
I mean, they're, they're breaking the Internet with all
39:17
these robots doing Kung Fu, you know, You saw that T800 video?
39:21
Yeah, came out. Yeah, engine AI is actually the
39:24
best at the the maneuvering part of it.
39:27
They're actually the leader in the in my opinion at least.
39:29
But I don't know if you remember like when we first talked about
39:33
EVs, I was already talking about the humanoid robots.
39:36
But I'm actually I cool down my opinion on the humanoid robots a
39:39
lot by now. I think the the actual robots we
39:43
need is especially in factories are not 2 legged robots.
39:47
We need probably stuff that can move around much more easily,
39:52
you know, like whether it's 4 legged or like, you know,
39:55
wheeled bottoms. And then you might need a robot
39:58
with like 2 that can go up and down that has two arms.
40:02
That's kind of what you need. And at home, you know,
40:05
personally I'm scared by the idea of having a joy walking
40:08
around. Especially like something as
40:13
intimidating as AT-800 that can do Kung Fu kicks.
40:17
It's like I don't. I don't want that thing to
40:19
malfunction and kick my ass man. Yeah, I I remember like this is
40:24
maybe a few years ago and I had AI robot at home.
40:30
And then I realized because I got it because I thought it
40:35
would be good to cleaning my house.
40:36
And in the middle, middle of the night, this thing just
40:38
malfunctioned and start making noises and loud noises and woke
40:42
my wife and I up and start moving around cleaning my home.
40:48
So after that I got rid of that fucking robot.
40:52
I've never bought one since. That's pretty scary actually.
40:57
So you can. Think about how how likely I am
40:59
to get a Droid at home after after that.
41:01
Absolutely. But but I actually think there
41:05
there are robots that people would love to have at home.
41:08
So I think it was UB tech recently showcased these cute
41:13
emotional robot robot robot toys.
41:17
They're, they come in like like these little fluffy little
41:19
packages and then they can talk to you maybe with your kids and
41:23
stuff like that. And those are kind of the robot
41:26
that are not threatening that people would want in their
41:28
homes. And the other kind of robot that
41:30
people don't think about very much is actually the, the, what
41:35
I was talking about the cleaning iRobot types.
41:37
And this is 1 area that China has really just destroyed U.S.
41:43
companies in. There was a recent report about
41:46
how iRobot is basically close to bankrupting and, and they're
41:55
basically they've outsourced all the work to a Chinese company
42:00
called Pichia and they haven't been able to pay them.
42:04
So, so basically, I don't know if you, I don't know if you use
42:09
Dyson products at home. It's Dyson's, this really famous
42:14
cleaning vacuuming company a lot of people use.
42:18
That little little disk robot that, that that do the cleaning,
42:22
self cleaning, the vacuum stuff. My, my parents, my my sister
42:26
have them. My I think my my relatives in
42:29
China also have them. That entire industry right now,
42:32
even the famous companies like Dyson, they're they are not only
42:37
manufacturing their their robots in China, but their their their
42:40
stuff is being designed, designed in China because they
42:43
aren't able to design they don't have the knowledge and they
42:46
don't have the experience with with AI to designing themselves.
42:50
So my impression is there's going to be a lot of things that
42:54
are just getting smarter and there's going to be cleaning
42:57
robots that can move around your home, climb stairs that people
43:01
are going to prefer over just having a Droid around.
43:04
So at the end of the day, this, the people that are talking
43:09
about the embodied AIS in the Silicon Valley, they most likely
43:12
don't have an idea. They've never spent any time
43:15
dealing with heat sinks. They've never spent any time
43:19
dealing with battery life and PCBS and things like that.
43:26
They just think everything can be solved with software.
43:32
So that that's kind of where I am with the with that part of
43:39
it. And I think, I think the, the,
43:42
the, the only major area left for China now is just getting
43:46
the false domestic supply chain for like Uvs and entire maybe 3
43:53
to 4 nanometer production lines. And then?
43:56
Do you do, do you think you know they they will be able to do
44:02
that in five years? Yeah, yeah, really, of none of
44:06
that, yeah. Wow, OK, so look, 'cause I, I, I
44:11
remember I, I said maybe two years ago, I said, you know,
44:15
give them five years. But I, I, I don't know, you
44:18
know, if they're say any 'cause they, they haven't really
44:21
announced anything. I, I, so I, I'm everybody's just
44:25
guessing at this point, you know, when they're going to be
44:28
come out with their indigenized supply, semiconductor supply
44:32
chain. But, but it's good to hear that
44:34
that you you think they can do it in five years.
44:37
Yeah. So like I think one thing is to
44:39
realize is that the in terms of three to four nanometer
44:43
production, the the limiting factor is not just UV's, right?
44:47
There's the entire supply chain that they needed to build out.
44:51
And I'm not even sure if UV the one that they're that the that's
44:56
going to be the the biggest folding factor.
44:59
It probably is, but I'm not sure.
45:01
So recently there was the news that YMTC, so the the memory
45:06
chip producer that does all the solid disk drive, 3D NAND for
45:11
solid disk drives. And the news is that they've
45:14
been, you know, fine tuning, they've been test trialling the
45:20
on basic and all domestic production line and that this is
45:25
expected to really start production in early 2026.
45:29
And they're like, they're pretty close to it already.
45:32
And so that went when you know, when you say that, that means
45:35
uses domestic lithography machines also, right?
45:38
So that, that gives you an idea of like how good, good it is.
45:41
So there's also the FT report about the Smick and Huawei
45:48
building fabs that are using like all Chinese supply, all
45:52
Chinese production lines for like the same 7 nanometer.
45:56
And that's probably going to come in the next year or two.
45:58
Also, my sense is that the EUV project is at most two years
46:07
behind this at most. Wow, wow.
46:12
So I, I, I've never been too concerned about the, the, the
46:17
entire topic because there's, there's, there's a lot of stuff
46:22
that they have to fix up. And then if you've seen where
46:25
they are now in the memory chips, they're basically in the
46:28
YMTC is basically probably the leader in 3D NAND technology.
46:34
They just need to ramp up the production.
46:37
CSMT made huge progress recently.
46:40
Now they have like the fastest DDR5 product and they actually
46:49
improved like hugely in the past year.
46:51
And and it's it's expected that they're going to get to the DDR6
46:55
next year. So they're going to be right
46:58
there. They're basically one year
46:59
behind Samsung at the moment, which was quite amazing.
47:04
So I don't know if you read a recent news about the memory
47:07
chip prices going up a lot. No.
47:13
Why is that? So the story is that all the big
47:19
three, so Microns, SK and Samsung, there's so much demand
47:24
for the HBMS and the server chips that they basically don't
47:29
have enough capacity for for the consumer drams.
47:36
Still, a lot of the consumer tech products is going to see
47:39
price jumps like Raspberry Pi came out with with a PCB with
47:43
just one GB of RAM. I don't even know what you can
47:45
do with a with a board 1 GB of RAM, but you can install the OS
47:52
in there and it starts up and it dies.
47:54
So that's my opinion of 1 GB of RAM does right.
47:58
But so, so there so so that's the thing.
48:05
Micron actually exited out of the consumer market.
48:08
They're there. They just said we're only going
48:10
to, you know, focus on the, the server and the business market
48:15
because that's where all the money is to be made.
48:18
And that's giving the Chinese firms like a great chance,
48:21
YMTCCXMT and whoever Huawei supports like a huge chance to
48:27
catch up in the space. And so, so the memory chips are
48:32
where China is going to catch up.
48:33
The first the and then it's going to be logic chips because
48:39
logic chips is, you know, it requires more advanced
48:44
lithography machines when the memory chips patch up.
48:47
That just tells you that, yeah, in terms of like the etching
48:51
machines, the deposition machines and a lot of the
48:54
testing and cleaning product there that they're up to par.
49:00
And one of the things that I've really focused on in the past
49:05
quarter is just looking at the the R&D numbers and also the
49:09
revenue numbers coming into some of these major Chinese equipment
49:15
makers. And they're still growing by
49:16
like 50, around 50%. Some of them grow by like close
49:20
to 100% a year depending on what they're doing, right.
49:24
And the reason for that is because once every time you move
49:28
up from like a lower node to A to a higher node, the revenue
49:33
drops significantly. So if you can, for example, get
49:39
to the, if you can get to the three nanometer, your, your tool
49:43
can be used for a three nanometer product production
49:45
line, then you're going to get paid a lot more than if it was
49:49
used for a 28 nanometer, for example.
49:51
So, so not only in terms of more unit they're selling, but better
49:55
product that they're selling that they're able to just see a
49:58
huge jump in the revenues and over maybe space of five years,
50:05
you will see a lot of these Chinese equipment makers catch
50:09
up to the Applied Materials, Lamb and KLA of this world.
50:17
Wow, that's good to hear. Yeah.
50:19
So I think what China is doing right now is not just in terms
50:23
of like the the production level, they they are building a
50:26
really solid foundation and a lot of the foundational
50:30
technology that that money went into can be cross used across
50:34
many industries. Like there was this one thing
50:36
recently. It's it's what's called a laser
50:40
interferometer and it's some, it's a technology used to be
50:45
able to measure very precisely where things are at.
50:54
So this is a technology that's needed to to do the most
51:02
advanced lithography machines. So that because it's precise to
51:07
like picometer kind of precision level.
51:10
But what's interesting about it is it's also very useful for
51:13
like, you know, the in the machine tool.
51:17
So CNC machine tools. So because they're making
51:21
process progress on the laser technology required for the
51:26
lithography machines and a lot of the testing tool for
51:28
semiconductor, the same technology can be applied to
51:33
improve like something that's very important like the CNC
51:37
machines. And, and in terms of like a lot
51:41
of the mechanical stuff that that really require a very
51:46
highly precise sensors. So you're seeing a lot of, so
51:51
you see like the cross breed of innovation in one sector helping
51:55
another sector. Because prior to, I would say
52:01
the last couple years, these technologies just weren't
52:04
available to China because, you know, American, American
52:08
companies are just not allowed to sell those kind of technology
52:11
to China, right, because of export restrictions.
52:15
So we are finally getting to the point where a lot of the the
52:20
fundamental technologies needed to make these machines are now
52:26
able to be controlled by the Chinese tech sector.
52:30
And then and that's when we're going to see the real result of
52:33
the real fruit of this work. And if you're not focused on
52:37
these things, like a lot of the Skull Silicon Valley people,
52:40
they think it's just all magic. And a lot of finance guys, they
52:44
just think it's all magic. Then you don't really understand
52:49
how much progress they made on this.
52:51
So what I'm hearing from you is that there will be some real big
52:56
breakthroughs in the coming five years in terms of in the
53:00
semiconductor industry in China. Yeah, like I wrote a series
53:05
about this recently about why China was able to do the rare
53:08
earth, rare earth export controls.
53:12
And the reason my fundamental argument for that is because
53:16
China has shored up so many weak points in the past, just past 18
53:22
months. And I think it's because of it
53:25
just takes several years from for, you know, like when when
53:30
the Huawei sanctions started back in 2018-2019, that's when
53:34
China really woke up, right. But it's taken this many years
53:37
for the for the science to catch up, for all the R&D to catch up.
53:42
And a lot of the real fruit of the R&D is only showing up in
53:52
the past year and half or so. But the you can see the
53:56
acceleration in the military sector, you'll also see the
53:59
similar acceleration the civilian sector side of things,
54:03
I think. Wow, thanks for that.
54:08
That's, you know, this make me a lot more optimistic about the
54:13
future. You should have nothing about
54:16
optimism because the problem with the West is that, say,
54:21
America has focused all its energy on AI and it has an
54:27
energy grid problem. Europe is spending all its money
54:33
on Ukraine, so it has no money left for for R&D.
54:38
The the smartest people in America are not going to
54:41
chemical engineering or mechanical engineering and the
54:44
basics, hard sciences, the traditional stuff.
54:48
China has so much resource that is going into the chemical and
54:52
mechanical engineering side of things that once it catches up
54:56
to the West in these areas, it's going to speed run the
55:00
technology because this kind of stuff takes so much longer R&D
55:04
cycle to develop. So you know, when, when Western
55:09
firms or Japanese firms spent decades to develop these things
55:15
that they still have an Rd. like just because they, you know,
55:19
they did develop the stuff already.
55:21
But but the cadence at which China is moving in this in these
55:25
fields is just so fast that once it gets to that point, it's
55:31
going to accelerate away. But TP, but TP the US is going
55:38
to develop AGI 1st and then it's, they're going to have
55:42
equivalents of 10s of thousands of pH DS working 24/7.
55:48
You know that, that they will, they will, they will, they will
55:51
destroy the Chinese. You know, the, the, the Chinese
55:53
are still made from flesh and blood.
55:55
They still have to rest. The the, the, the, the AGI that
55:59
the AI will will be, you know, they'll be running circles.
56:04
In the Chinese, do you know, that's the argument someone used
56:07
recently to explain to me why the American biotech sectors
56:10
will be Chinese biotech sectors long term.
56:12
I said the Chinese biotech sector was nothing like five
56:15
years ago. Now it's like now like American
56:19
drug companies depend on the Chinese biotech sector.
56:22
What's your explanation for that?
56:23
He's like, oh, we have alpha fold now.
56:26
Alpha fold will fix everything and that's how these people
56:29
think. And OK.
56:36
So, so you don't you don't think that's realistic?
56:38
Just just to make it clear, you don't think that's realistic
56:41
that the AI, you know, the coming out of Silicon Valley is
56:45
going to, you know, run circles around all these Chinese STEM
56:50
grads? No.
56:50
No, no. I think the majority of the
56:53
usage cases for AI where they're going to make all their money on
56:56
is AI slop. That's that's where the demand
56:59
for everyone is. It's the same thing in China
57:02
too. People invent AI to improve
57:05
productivity, but everyone just spends more time looking at
57:07
their phone like we're all the same.
57:11
Well, I mean, Sam Altman already pointed out the direction he's
57:16
gonna use AI to generate adult content.
57:20
I mean, isn't that like that? Like I don't want to go too far
57:23
into this, this loop, but like, isn't that where the initial
57:26
Internet was made for like a third of the traffic was like
57:30
porn And and that's like where, where all the money, like
57:35
seriously, I, I, I really look at a lot of the humanoid robot.
57:38
I think that's where the the end purpose end purpose is.
57:41
Oh yeah, definitely, definitely. I mean, like I I don't think
57:44
they even disguise the fact. All right, great.
57:48
This is has been amazingly educational session.
57:53
Is there any other topic that you wish to cover?
57:58
Yeah, yeah, I, I know. I think that's it.
58:00
I think they're, they're doing really well.
58:01
And I think there's like, I, I feel like there's a bit of like
58:06
a panic right now going on. Like when the Blackwell stuff
58:11
would, when, when they were talking about, you know, giving
58:14
export licenses to the blackboard ships to China, There
58:18
was this, this hysteria going online about how this is a
58:21
civilization type of mistake by America.
58:27
I was just like, how desperate do you have to get for this to,
58:31
for people to say stuff like this?
58:34
It's really it's AI is AI is just a tool to help you improve
58:38
your life, you know, and these people I, I, I, I, I, I really
58:44
think that this got way out of control so.
58:47
Well, I think it has to do with all that cash that's going
58:51
through the AI bubble. I, I think that's, that's what's
58:54
really at stake here. And well, I, I mean, I, I, I
58:59
think it's just a matter of time.
59:02
That bubble will burst. It's it's just a matter of when.
59:08
You know, like when that irrational exuberance will run
59:11
out on AI. Well NVIDIA already had its own
59:14
moment right recently because people realize you don't need
59:17
NVIDIA chips to train stuff. When Google, when Google came
59:21
out with the TPU news and how the Gemini model was as good as
59:26
was better than the GPT 5.1. So I, I think the, yeah, I think
59:34
the NVIDIA bubble has to burst at some point because it's just
59:38
like they're the margins are too ridiculous.
59:41
And also their chips itself is not even that special.
59:47
And they're, they're playing with funny money, right?
59:49
I mean, NVIDIA is investing in open AI, they're all buying each
59:53
other's stocks. So so to keep the bubble going.
59:59
Yeah, then they buy all these chips.
1:00:00
I just collect dust like. Yeah, and so far the only, only
1:00:05
company that's making money so far is actually NVIDIA, because
1:00:09
they're the ones selling the GPU chips, but.
1:00:12
Oh, I'm pretty sure DeepSeek is making money too, because just
1:00:15
because they're just because their cost is so low.
1:00:20
I mean how do DeepSeek make money?
1:00:22
I thought it's open source. Yeah, but people use their API
1:00:26
services, so. Oh, yeah.
1:00:30
OK. So even though they don't charge
1:00:32
a lot of money, but because they're so efficient with
1:00:34
everything that they're, they're not wasting money around that
1:00:37
they're probably still making they're they're still probably
1:00:40
making money at this point. Yeah.
1:00:42
That that that's like the Chinese model like like high
1:00:47
volume, low margin. Yeah, yeah.
1:00:51
And then the crazy thing about the Chinese market, a lot of
1:00:53
these Chinese people say is that even at the current prices, the
1:00:57
Chinese this is still think it's too expensive.
1:01:00
I need to get cheaper. Wow.
1:01:04
Wow. Yeah, OK, that's great.
1:01:07
Thank you very much, TP, for making the time to speak to us.
1:01:12
If people want to find you and follow you on social media,
1:01:16
where would they go? Yeah, they can reach me on X or
1:01:20
Twitter, formerly known as Twitter at TP Wong or they can
1:01:24
find me on Sub Stack, also under the same name, TP Wong.
1:01:28
I'm, I post stuff on there and I and I really enjoy talking about
1:01:31
all this semiconductor and AI stuff.
1:01:35
Yeah, I, I would love to have you back.
1:01:38
Hopefully, you know, less in less than four or five month
1:01:42
time. Yeah.
1:01:43
And because right now the, the whole tech sector is evolving so
1:01:49
fast, you know, like we, we we need we need, we need more
1:01:54
regular updates. So I look forward to forward to
1:01:57
have you back on my show again TP and and thank you again for
1:02:03
for for making the time to talk to us and for the rest of the
1:02:06
audience tuning in. Thank you for listening.
1:02:09
I'll to you next time. Bye bye.
1:02:13
OK, that was.