1
00:00:00,020 --> 00:00:06,140
Speaker1:
At the end of last year, Meta announced a $2 billion acquisition of an AI startup called Manus.

2
00:00:06,260 --> 00:00:10,580
Speaker1:
They make AI agents very similar to OpenCore, but they were founded in China.

3
00:00:10,840 --> 00:00:15,560
Speaker1:
Then in March recently, the two founders visited China on business and were

4
00:00:15,560 --> 00:00:17,500
Speaker1:
told that they couldn't leave the country.

5
00:00:17,640 --> 00:00:21,280
Speaker1:
Their passports were taken away from them and they were blacklisted from travel.

6
00:00:21,660 --> 00:00:25,460
Speaker1:
Then today, China revealed that they have completely banned the acquisition

7
00:00:25,460 --> 00:00:30,280
Speaker1:
deal by Meta and have requested that Meta reverse the entire deal.

8
00:00:30,480 --> 00:00:33,900
Speaker1:
Now, this is unprecedented from China. For the last 20 years,

9
00:00:34,020 --> 00:00:38,120
Speaker1:
they've happily accepted US capital, but this marks the turning point after

10
00:00:38,120 --> 00:00:42,480
Speaker1:
three years where China has built their own independent AI stack and they don't

11
00:00:42,480 --> 00:00:44,120
Speaker1:
need to rely on America anymore.

12
00:00:44,380 --> 00:00:47,980
Speaker0:
Yeah, those two people from that meeting last month, that was the CEO and the

13
00:00:47,980 --> 00:00:49,960
Speaker0:
chief scientist of Manus.

14
00:00:50,120 --> 00:00:52,460
Speaker0:
They got a summons from the Chinese

15
00:00:52,460 --> 00:00:55,940
Speaker0:
government body called the National Development and Reform Commission.

16
00:00:56,080 --> 00:00:58,860
Speaker0:
Basically, it's the agency that runs the country's economy. They go to the meeting,

17
00:00:59,000 --> 00:01:02,040
Speaker0:
they get questioned about the meta deal, and then they say, you are not allowed

18
00:01:02,040 --> 00:01:03,880
Speaker0:
to leave. Now, fast forward to today.

19
00:01:04,080 --> 00:01:06,960
Speaker0:
Four weeks, I will say, before Donald Trump was supposed to land in Beijing

20
00:01:06,960 --> 00:01:10,920
Speaker0:
to meet with Xi Jinping, China killed the deal. And they did so in one sentence.

21
00:01:11,060 --> 00:01:13,760
Speaker0:
They said, just undo it. The two men still can't leave.

22
00:01:13,960 --> 00:01:18,200
Speaker0:
So now, not only are the two people stuck in China, but the deal...

23
00:01:18,760 --> 00:01:21,860
Speaker0:
Meta has signed with manis is now

24
00:01:21,860 --> 00:01:24,520
Speaker0:
kind of in a weird gray area now for those

25
00:01:24,520 --> 00:01:27,940
Speaker0:
who aren't familiar with manis manis is meta's response

26
00:01:27,940 --> 00:01:30,840
Speaker0:
to open claw open claw is the claw it is the agentic system

27
00:01:30,840 --> 00:01:35,220
Speaker0:
that kind of runs your life for you it performs agentic tasks long form it has

28
00:01:35,220 --> 00:01:39,080
Speaker0:
context does memory it was a big deal for meta and now meta has this problem

29
00:01:39,080 --> 00:01:42,320
Speaker0:
on their hands where china is saying no when the reality is is there's already

30
00:01:42,320 --> 00:01:48,620
Speaker0:
an entire singaporean office set up and the ceo of manis is on the C-suite of Meta already.

31
00:01:48,800 --> 00:01:51,660
Speaker0:
So there's a lot to unpack here. There's a lot to unwind. The tension between

32
00:01:51,660 --> 00:01:53,960
Speaker0:
China and the United States has never been higher.

33
00:01:54,080 --> 00:01:58,280
Speaker0:
I mean, this is pretty unprecedented as it relates to the dynamics between these two countries.

34
00:01:58,900 --> 00:02:03,240
Speaker1:
Yeah, this all started around the time last year that Zuck started acquiring

35
00:02:03,240 --> 00:02:04,820
Speaker1:
or started doing his spending.

36
00:02:05,000 --> 00:02:10,640
Speaker1:
I think he spent in total around $25 billion, $15 billion, which was spent on just one single man.

37
00:02:11,260 --> 00:02:15,280
Speaker1:
But Manus was part of the spending spree. And the timeline on this is pretty

38
00:02:15,280 --> 00:02:16,840
Speaker1:
insane. I've got it pulled up here.

39
00:02:17,140 --> 00:02:21,800
Speaker1:
So in March in 2025, Manus launched their AI agent, right? And it's what you

40
00:02:21,800 --> 00:02:24,860
Speaker1:
just mentioned. It's like its own AI operating system.

41
00:02:25,400 --> 00:02:31,440
Speaker1:
And within like 24 hours or within seven days, rather, they had 2 million people on the waiting list.

42
00:02:31,880 --> 00:02:38,080
Speaker1:
And people were like, I remember this, reselling invite codes for like $1,500.

43
00:02:38,740 --> 00:02:39,420
Speaker0:
$1,000 is crazy.

44
00:02:39,420 --> 00:02:43,180
Speaker1:
I actually tried to acquire one because this is before AI agents existed.

45
00:02:43,320 --> 00:02:49,260
Speaker1:
This is before GPT and Claude had built their own agentic systems. Manus was the first.

46
00:02:49,560 --> 00:02:53,560
Speaker1:
So the founders, with all the success, was like, oh, we've got to kind of kill

47
00:02:53,560 --> 00:02:56,720
Speaker1:
the Chinese image so that Americans will use our product.

48
00:02:56,880 --> 00:02:59,560
Speaker1:
So they ended up moving from Beijing, where they were based,

49
00:02:59,720 --> 00:03:03,780
Speaker1:
to Singapore, and they raised a massive round from Benchmark Capital,

50
00:03:03,900 --> 00:03:07,000
Speaker1:
which is one of the biggest funds in America, I think it was like $70 million.

51
00:03:07,460 --> 00:03:12,020
Speaker1:
And within six months from then, they hit $90 million of revenue run rate.

52
00:03:12,120 --> 00:03:16,280
Speaker1:
So things are looking amazing. And they thought that they'd subverted Chinese

53
00:03:16,280 --> 00:03:20,880
Speaker1:
authority simply by moving their company to Singapore, which is obviously not in China.

54
00:03:21,120 --> 00:03:25,840
Speaker1:
Then in March 2026, China stops the Manus co-founders from leaving the country.

55
00:03:26,060 --> 00:03:29,860
Speaker1:
And today, they've blocked the entire acquisition. Now, the main question that

56
00:03:29,860 --> 00:03:33,040
Speaker1:
pops into my head, Josh, and I don't know if you're able to like answer this for me is,

57
00:03:33,630 --> 00:03:37,370
Speaker1:
are they able to actually do this? I thought the deal was already signed.

58
00:03:37,470 --> 00:03:38,530
Speaker1:
I thought the ink had already dried.

59
00:03:38,690 --> 00:03:41,930
Speaker1:
Like I'm kind of confused how they unwind a deal that's already been dealt.

60
00:03:42,210 --> 00:03:44,830
Speaker0:
I think this is where there's going to be a problem, right?

61
00:03:44,930 --> 00:03:48,070
Speaker0:
It's because like, like we mentioned, there is an entire Singaporean office

62
00:03:48,070 --> 00:03:52,350
Speaker0:
where there's over 100 workers who are from Manus working with Meta.

63
00:03:52,590 --> 00:03:57,410
Speaker0:
The CEO of Manus is on the C-suite in Meta.

64
00:03:57,570 --> 00:04:02,670
Speaker0:
There's a lot of already kind of like baked in progress that has been made after this deal.

65
00:04:02,890 --> 00:04:06,750
Speaker0:
And China pulling the cord creates a lot of tension. And I'm not sure where

66
00:04:06,750 --> 00:04:11,210
Speaker0:
this is going to land because does Meta want to actually go to war with China

67
00:04:11,210 --> 00:04:13,810
Speaker0:
to prevent this blockage from happening?

68
00:04:14,310 --> 00:04:18,650
Speaker0:
Maybe. This is a huge bet for the company and a big bet on the future of the

69
00:04:18,650 --> 00:04:19,630
Speaker0:
company, what this looks like.

70
00:04:19,750 --> 00:04:22,950
Speaker0:
They were starting to roll this out. It was very deeply integrated. It fell apart.

71
00:04:23,170 --> 00:04:25,670
Speaker0:
I don't know what's going to happen. This is actually a pretty insane story.

72
00:04:25,850 --> 00:04:30,370
Speaker1:
Okay. So there are like some crazy parts of this on both sides,

73
00:04:30,510 --> 00:04:32,670
Speaker1:
right? So for Meta, you've spent $2 billion.

74
00:04:33,110 --> 00:04:37,970
Speaker1:
You have hundreds of employees from Manus that now are under Zuck's employment.

75
00:04:38,110 --> 00:04:40,670
Speaker1:
Just to be very clear, it's like under the American company.

76
00:04:40,850 --> 00:04:42,650
Speaker1:
Inc has been sealed. The deal has been done.

77
00:04:43,030 --> 00:04:46,770
Speaker1:
Now, to your point, Manus is being used a lot in Meta.

78
00:04:46,950 --> 00:04:50,410
Speaker1:
But specifically, if you haven't used it as a user, it's being used by advertisers.

79
00:04:50,410 --> 00:04:53,430
Speaker1:
We spoke about this on a previous episode, I think like two weeks ago,

80
00:04:53,750 --> 00:04:57,830
Speaker1:
where Manus is being used by advertisers to spin up their advertising projects.

81
00:04:57,970 --> 00:05:02,090
Speaker1:
And it's actually resulted in a 30% quarterly jump in revenue alone,

82
00:05:02,090 --> 00:05:03,630
Speaker1:
just by people using Matters.

83
00:05:03,790 --> 00:05:08,310
Speaker1:
It's cut out a lot of repetitious manual tasks. So it's a very good product.

84
00:05:08,470 --> 00:05:11,190
Speaker1:
And just because you haven't used it doesn't mean it isn't valuable to Meta.

85
00:05:11,270 --> 00:05:13,990
Speaker1:
They're building it, it's implemented, it's core to their system.

86
00:05:14,070 --> 00:05:16,210
Speaker1:
Advertising is like a lot of Meta's revenue.

87
00:05:16,390 --> 00:05:20,610
Speaker1:
So now they theoretically have to unwind this. But I think the alarm bell that's

88
00:05:20,610 --> 00:05:23,510
Speaker1:
ringing for me, Josh, is China's never done this before.

89
00:05:23,900 --> 00:05:27,040
Speaker1:
They have never unwound a closed deal.

90
00:05:27,200 --> 00:05:30,340
Speaker1:
Let me repeat that again. They've never unwound a closed deal that they have

91
00:05:30,340 --> 00:05:34,620
Speaker1:
like signed off on that is all completely legal, which tells me that they are

92
00:05:34,620 --> 00:05:38,660
Speaker1:
now treating these AI assets, whether you want to call them workers, researchers,

93
00:05:39,060 --> 00:05:42,980
Speaker1:
The actual product itself as a kind of national asset.

94
00:05:43,140 --> 00:05:47,400
Speaker1:
And they're treating it as kind of like a geopolitical asset against the US.

95
00:05:47,400 --> 00:05:50,740
Speaker1:
They feel like if they don't own this thing, then the US is going to wipe them out.

96
00:05:50,840 --> 00:05:55,400
Speaker1:
So I started looking at like other parts or trends And I realized China's kind

97
00:05:55,400 --> 00:05:58,700
Speaker1:
of been like moving in this direction over the last couple of months.

98
00:05:59,080 --> 00:06:05,220
Speaker1:
Specifically, they've like refused to use NVIDIA's GPUs to train their own AI

99
00:06:05,220 --> 00:06:07,640
Speaker1:
models amongst the top Chinese AI labs.

100
00:06:07,740 --> 00:06:12,520
Speaker1:
And they've started mandating that all these Chinese labs start building and

101
00:06:12,520 --> 00:06:16,700
Speaker1:
using the AI models within the China itself and not really open sourcing or

102
00:06:16,700 --> 00:06:20,220
Speaker1:
open sourcing only parts of it, but not the research itself to the US.

103
00:06:20,220 --> 00:06:24,660
Speaker1:
So there's been this cultural shift where China now doesn't really depend on the U.S.

104
00:06:24,700 --> 00:06:27,600
Speaker1:
As they had before in the past or over the last six months.

105
00:06:27,720 --> 00:06:31,120
Speaker1:
They now kind of have their own independence stack. And I think that's probably

106
00:06:31,120 --> 00:06:33,400
Speaker1:
part of what is playing into this story today.

107
00:06:34,080 --> 00:06:37,440
Speaker0:
Yeah, man, I kind of feel so bad for Zuck because he's tried this over and over

108
00:06:37,440 --> 00:06:38,700
Speaker0:
to get this agentic system.

109
00:06:38,840 --> 00:06:43,540
Speaker0:
He first tried to buy Ilya's safe superintelligence, if you remember, and Ilya said no.

110
00:06:43,760 --> 00:06:47,240
Speaker0:
And then he went to another OpenAI co-founder. He went to Mira Mirati's Thinking machines.

111
00:06:47,540 --> 00:06:50,920
Speaker0:
She gave him a no. He reached out to perplexity. They gave him a no.

112
00:06:51,060 --> 00:06:56,040
Speaker0:
And then they wound up paying $15 billion for half of scale AI just to get Alexander Wang.

113
00:06:56,200 --> 00:07:01,220
Speaker0:
So they really were struggling to get some help in this agentic world.

114
00:07:01,240 --> 00:07:04,340
Speaker0:
They finally acquired it through Manus, and now they're having this huge problem.

115
00:07:04,500 --> 00:07:07,460
Speaker0:
And an important caveat that's probably worth mentioning is...

116
00:07:07,980 --> 00:07:11,640
Speaker0:
That it's a little disingenuous and I could see why China's upset.

117
00:07:11,900 --> 00:07:15,280
Speaker0:
Because, I mean, Manus started off in Beijing, then they transferred over to

118
00:07:15,280 --> 00:07:17,280
Speaker0:
Hong Kong, and then they transferred over to Singapore.

119
00:07:17,600 --> 00:07:21,380
Speaker0:
And this is because, I mean, American venture capitalists are not going to invest

120
00:07:21,380 --> 00:07:26,180
Speaker0:
in Chinese companies and China does not want to engage with American acquisitions.

121
00:07:26,220 --> 00:07:27,720
Speaker0:
So they moved to a neutral place.

122
00:07:27,820 --> 00:07:31,160
Speaker0:
When they did this, I think 80 Chinese employees were laid off.

123
00:07:31,360 --> 00:07:32,880
Speaker0:
All of the Weibo posts were deleted.

124
00:07:33,100 --> 00:07:37,000
Speaker0:
All of the new offices were opened up. Everything was a clean slate.

125
00:07:37,000 --> 00:07:39,900
Speaker0:
So the people who helped them get there, they were cut off.

126
00:07:39,980 --> 00:07:42,780
Speaker0:
They started fresh in Singapore. But now China's like, wait a second. No, no, no, no, no.

127
00:07:43,410 --> 00:07:46,770
Speaker0:
You started here. You are Chinese. You're coming back to us.

128
00:07:46,890 --> 00:07:50,270
Speaker0:
And what a disaster for Zuck. This poor company, they can't figure it out.

129
00:07:50,350 --> 00:07:51,230
Speaker0:
They're spending so much money.

130
00:07:51,410 --> 00:07:54,110
Speaker0:
They haven't been able to launch a good product. And they finally got Manus,

131
00:07:54,190 --> 00:07:57,170
Speaker0:
like you mentioned, who was selling these invites for thousands of dollars because it was that good.

132
00:07:57,310 --> 00:08:00,530
Speaker0:
And now it's ripped out from their hands. So really brutal turn of events.

133
00:08:00,610 --> 00:08:03,510
Speaker0:
And it's going to be fascinating to follow Meta and see how they react to this.

134
00:08:04,010 --> 00:08:07,730
Speaker1:
Yeah, I think that the playbook that Manus founders tried to pull off,

135
00:08:07,730 --> 00:08:13,590
Speaker1:
but failed at was, oh, yeah, if we just wipe ourselves or rather the premise

136
00:08:13,590 --> 00:08:17,330
Speaker1:
of the actual company HQ to another country, we can evade Chinese law.

137
00:08:17,470 --> 00:08:21,090
Speaker1:
But remember, China is more of a dictatorship in this sense.

138
00:08:21,230 --> 00:08:24,450
Speaker1:
And they were like, hey, you can't actually do this. So we can pull your passports

139
00:08:24,450 --> 00:08:29,130
Speaker1:
and unwind this entire deal because it is still effectively Chinese. It was founded in China.

140
00:08:29,590 --> 00:08:33,910
Speaker1:
Now, I just want to comment on China's AI strategy in general,

141
00:08:33,950 --> 00:08:37,090
Speaker1:
because I think it plays very well into this particular story today.

142
00:08:38,230 --> 00:08:41,710
Speaker1:
Manus doesn't actually own any AI models.

143
00:08:42,090 --> 00:08:45,110
Speaker1:
I don't know if you guys knew that or the people listening to this knew that,

144
00:08:45,310 --> 00:08:47,650
Speaker1:
but they're kind of like a wrapper over a bunch of different AI models.

145
00:08:47,790 --> 00:08:52,070
Speaker1:
And typically, the AI models that they would use are American AI models, right?

146
00:08:52,390 --> 00:08:58,090
Speaker1:
But since, well, actually, over the last seven days, China released three AI models.

147
00:08:58,250 --> 00:09:05,570
Speaker1:
I think it was DeepSeek V4, Kimi K2.6. I'm just retesting my a model knowledge here now, a Qen 3.6.

148
00:09:06,090 --> 00:09:08,590
Speaker1:
And people have heard about Chinese models all along, right?

149
00:09:08,650 --> 00:09:10,510
Speaker1:
They're like, oh, it's open source, but it's never as good.

150
00:09:10,710 --> 00:09:18,050
Speaker1:
These three models are at parity, if not as good, as Claude Opus 4.7 and GPT 5.4.

151
00:09:18,250 --> 00:09:23,570
Speaker1:
So pretty much 90% of Frontier model capability or 95% of Frontier model capability

152
00:09:23,570 --> 00:09:25,070
Speaker1:
you could now access for free.

153
00:09:25,370 --> 00:09:28,670
Speaker1:
100% open source, download, run it on your own devices or GPUs if you happen

154
00:09:28,670 --> 00:09:31,190
Speaker1:
to have a couple laying at home for whatever reason.

155
00:09:31,580 --> 00:09:35,700
Speaker1:
And they're all Chinese models. Chinese models are actually also being used

156
00:09:35,700 --> 00:09:40,400
Speaker1:
in a bunch of Silicon Valley startups who can't afford to kind of run and fine

157
00:09:40,400 --> 00:09:44,500
Speaker1:
tune their own models, which of course, Claude and GPT are close source.

158
00:09:44,620 --> 00:09:45,920
Speaker1:
So they can't actually do that.

159
00:09:46,040 --> 00:09:48,100
Speaker1:
And it takes hundreds of millions of dollars to train your own model.

160
00:09:48,200 --> 00:09:52,440
Speaker1:
So the point is, Chinese models are more accessible and more easily fine tuned.

161
00:09:52,660 --> 00:09:56,220
Speaker1:
So why wouldn't you use it if you could just access maybe 95% of frontier capability?

162
00:09:56,460 --> 00:10:01,640
Speaker1:
And a large reason for this is because China has started forcing their AI labs

163
00:10:01,640 --> 00:10:04,760
Speaker1:
to train on Chinese-made chips.

164
00:10:04,920 --> 00:10:09,580
Speaker1:
Now, I just want to spend a very quick moment to explain this. DeepSeek, Moonshot,

165
00:10:10,220 --> 00:10:16,580
Speaker1:
And Alibaba have been mandated by the Chinese authorities to not train their

166
00:10:16,580 --> 00:10:18,820
Speaker1:
AI models on NVIDIA chips.

167
00:10:19,220 --> 00:10:22,820
Speaker1:
Now, typically, America had banned China from buying the chips.

168
00:10:23,000 --> 00:10:26,280
Speaker1:
That got reverted, and America was like, cool, you can have these chips.

169
00:10:26,540 --> 00:10:29,840
Speaker1:
China said, nah, we're good. We're going to train our own chips to be better

170
00:10:29,840 --> 00:10:32,020
Speaker1:
so that we aren't reliant on America at all.

171
00:10:32,120 --> 00:10:35,020
Speaker1:
And it seems that Huawei specifically has caught up.

172
00:10:35,180 --> 00:10:38,500
Speaker1:
That's what all these latest models that I just mentioned are trained on and

173
00:10:38,500 --> 00:10:41,900
Speaker1:
are inferenced on. And now it seems like that gap has closed.

174
00:10:42,080 --> 00:10:42,820
Speaker1:
So it doesn't really matter.

175
00:10:43,040 --> 00:10:48,400
Speaker1:
This is worrying, from my perspective, at least, because America could keep tabs on China.

176
00:10:48,540 --> 00:10:51,220
Speaker1:
And now they're kind of given the opportunity to leapfrog us.

177
00:10:51,340 --> 00:10:53,020
Speaker1:
They have a heck ton more energy.

178
00:10:53,380 --> 00:10:57,660
Speaker1:
So this is like the alarm bells ringing from today's story. Like Manus isn't

179
00:10:57,660 --> 00:11:00,860
Speaker1:
just about like one startup. It's about China's entire strategy.

180
00:11:01,420 --> 00:11:04,860
Speaker0:
Yeah. And I mean, everyone knows the name DeepSeek now. DeepSeek was the ones

181
00:11:04,860 --> 00:11:08,440
Speaker0:
that released that paper early on, they kind of changed the game as it comes

182
00:11:08,440 --> 00:11:12,640
Speaker0:
to, I guess, like reasoning and AI models, where OpenAI kind of figured it out,

183
00:11:12,780 --> 00:11:16,320
Speaker0:
but DeepSeek realized it and then published it publicly and then everyone followed on.

184
00:11:16,420 --> 00:11:19,660
Speaker0:
DeepSeek has always been the frontier of this open source race.

185
00:11:19,880 --> 00:11:25,220
Speaker0:
And just on Friday, DeepSeek released V4, which is kind of a scary benchmark

186
00:11:25,220 --> 00:11:26,960
Speaker0:
relative to the other open source models.

187
00:11:27,080 --> 00:11:30,440
Speaker0:
I mean, I'm looking at the chart here and it actually performs better on a few

188
00:11:30,440 --> 00:11:32,780
Speaker0:
benchmarks than GPT 5.5,

189
00:11:32,960 --> 00:11:35,800
Speaker0:
which is OpenAI's brand new frontier model that just came

190
00:11:35,800 --> 00:11:38,960
Speaker0:
out last week and i mean just last week alone china

191
00:11:38,960 --> 00:11:41,600
Speaker0:
dropped three huge models we got deep cv4 which is

192
00:11:41,600 --> 00:11:44,500
Speaker0:
probably the flagship but then there was that new quen model the new kimmy model

193
00:11:44,500 --> 00:11:48,760
Speaker0:
they're publishing at a much faster cadence and you have to imagine that part

194
00:11:48,760 --> 00:11:53,520
Speaker0:
of this comes from just their their boldness and willing to distill the larger

195
00:11:53,520 --> 00:11:57,260
Speaker0:
models from the united states but also their capabilities to train these at

196
00:11:57,260 --> 00:12:01,320
Speaker0:
an increasing rate i mean they're clearly innovating on the software side but something's working on,

197
00:12:03,270 --> 00:12:04,890
Speaker0:
at this cadence with this type of power.

198
00:12:05,090 --> 00:12:08,650
Speaker1:
On that note, Gavin Baker put out an amazing analysis.

199
00:12:09,010 --> 00:12:13,170
Speaker1:
Did you watch the Duakesh episode where he's like grilling Jensen and Jensen

200
00:12:13,170 --> 00:12:17,110
Speaker1:
just says like, I'm not a loser, like NVIDIA is not this, right?

201
00:12:17,330 --> 00:12:22,490
Speaker1:
So Gavin Baker commented on this and he said, actually, Jensen makes a really

202
00:12:22,490 --> 00:12:29,170
Speaker1:
good point where the chips that NVIDIA builds are specifically built with the

203
00:12:29,170 --> 00:12:31,210
Speaker1:
understanding that the U.S.

204
00:12:31,350 --> 00:12:33,550
Speaker1:
Has a very limited amount of energy.

205
00:12:33,890 --> 00:12:37,950
Speaker1:
That's number one. Number two, that a lot of these chips are gonna be based

206
00:12:37,950 --> 00:12:41,750
Speaker1:
on pre-training, which is like the part that comes, I guess,

207
00:12:41,850 --> 00:12:43,310
Speaker1:
before post-training, obviously.

208
00:12:43,510 --> 00:12:46,890
Speaker1:
But the point is, it's like a large training run, right? It's the bulk of the expense.

209
00:12:47,290 --> 00:12:51,270
Speaker1:
Now, China took a different approach. They said, well, we're not energy-constrained.

210
00:12:51,350 --> 00:12:53,330
Speaker1:
We have 3X more energy than the U.S.

211
00:12:53,450 --> 00:12:59,710
Speaker1:
And also, we think that building the best AI model happens after pre-training,

212
00:12:59,830 --> 00:13:00,930
Speaker1:
so in the post-training thing.

213
00:13:01,250 --> 00:13:07,010
Speaker1:
Now, one thing that Dario Amode of Anthropic and Sam Altman of OpenAI have announced

214
00:13:07,010 --> 00:13:10,810
Speaker1:
with their latest models, including Claude Mythos, is inference has been the

215
00:13:10,810 --> 00:13:12,090
Speaker1:
key to unlock a smarter model.

216
00:13:12,190 --> 00:13:14,370
Speaker1:
So they're confirming what the Chinese has also confirmed.

217
00:13:14,730 --> 00:13:18,410
Speaker1:
And so the reason why Chinese chips specifically are doing so well is because

218
00:13:18,410 --> 00:13:19,650
Speaker1:
they're designed around inference.

219
00:13:19,830 --> 00:13:23,810
Speaker1:
So the architecture for a Chinese chip, if you were to try and train like a

220
00:13:23,810 --> 00:13:25,530
Speaker1:
US model, looks very different.

221
00:13:25,610 --> 00:13:29,070
Speaker1:
And it actually wouldn't work. The hardware actually kind of delineates at this point.

222
00:13:29,210 --> 00:13:32,530
Speaker1:
So China, although it seems like we're competing, is building their own kind

223
00:13:32,530 --> 00:13:36,070
Speaker1:
of like stack themselves that wouldn't necessarily operate or work in the US.

224
00:13:36,190 --> 00:13:39,150
Speaker1:
And I think that's a story that most people like aren't addressing at this point.

225
00:13:39,170 --> 00:13:42,830
Speaker1:
And we're giving China the freedom to go do that right now.

226
00:13:43,170 --> 00:13:45,830
Speaker0:
Yeah, that perhaps talks to like the increased velocity, right,

227
00:13:45,910 --> 00:13:48,870
Speaker0:
that they've been having is because they're not spending that much time on pre-training.

228
00:13:48,970 --> 00:13:51,530
Speaker0:
They're all in on inference, all in on distillation.

229
00:13:51,730 --> 00:13:53,670
Speaker0:
And we've kind of seen this, I guess,

230
00:13:54,220 --> 00:13:57,820
Speaker0:
A way that you could track this back to something that we've seen is how something

231
00:13:57,820 --> 00:14:05,660
Speaker0:
like GPT 5.7 Nano will actually be far more superior than the following model,

232
00:14:05,740 --> 00:14:06,960
Speaker0:
even though it's cheaper and smaller.

233
00:14:07,100 --> 00:14:09,580
Speaker0:
And it's because it takes advantage of this distillation. And we see this all

234
00:14:09,580 --> 00:14:13,540
Speaker0:
the time where with Nano Banana, Google's image generation model,

235
00:14:13,980 --> 00:14:17,540
Speaker0:
Nano Banana Nano, like the small one, was actually better than the pro one,

236
00:14:17,720 --> 00:14:19,380
Speaker0:
even though it had much less model weights.

237
00:14:19,380 --> 00:14:22,180
Speaker0:
And I think what we're seeing here is a lot of innovation on that inference

238
00:14:22,180 --> 00:14:25,060
Speaker0:
side, on just like a faster compute.

239
00:14:25,240 --> 00:14:28,200
Speaker0:
And Google's been addressing this recently. They split their most recent TPU

240
00:14:28,200 --> 00:14:30,440
Speaker0:
into two, one just for pre-training, one for inference.

241
00:14:30,620 --> 00:14:33,660
Speaker0:
It's clear that China is very much over-indexing on inference.

242
00:14:33,760 --> 00:14:35,500
Speaker0:
And right now, that's where a lot of the gains are happening.

243
00:14:35,900 --> 00:14:39,440
Speaker1:
I just wanted to pull up this post because you just reminded me of something.

244
00:14:39,640 --> 00:14:44,980
Speaker1:
For those of you who say like no one uses Chinese models, the answer is just you're wrong.

245
00:14:45,680 --> 00:14:49,400
Speaker1:
So Kimi K 2.6, which was released last week, and it was actually the first model

246
00:14:49,400 --> 00:14:52,360
Speaker1:
of the three models that we mentioned, DeepSeek v4 was the last one,

247
00:14:53,500 --> 00:14:56,700
Speaker1:
Is now the number one most used model on OpenWriter.

248
00:14:56,820 --> 00:14:59,640
Speaker1:
OpenWriter is this website where you can basically get access to all the models

249
00:14:59,640 --> 00:15:03,440
Speaker1:
sometimes before they actually officially release, and it can track token usage.

250
00:15:03,840 --> 00:15:07,680
Speaker1:
Now, token usage is the metric that you can kind of track to figure out

251
00:15:08,260 --> 00:15:12,900
Speaker1:
which model actually is worth its weight in gold and what are people using it for?

252
00:15:13,240 --> 00:15:17,880
Speaker1:
OpenRapta reveals that Chinese models take the number one spot almost the entire time.

253
00:15:17,980 --> 00:15:21,320
Speaker1:
And part of that is probably because a lot of the closed-source models don't

254
00:15:21,320 --> 00:15:22,540
Speaker1:
want to reveal their full metrics.

255
00:15:22,700 --> 00:15:27,120
Speaker1:
But the point being is the token usage of Chinese models have now surpassed U.S.

256
00:15:27,340 --> 00:15:30,560
Speaker1:
Token consumption, at least from publicly available data.

257
00:15:30,740 --> 00:15:34,020
Speaker1:
So the data would then indicate that these Chinese models are obviously getting

258
00:15:34,020 --> 00:15:37,820
Speaker1:
not only better, but more accessible to any and all developers that want to

259
00:15:37,820 --> 00:15:39,480
Speaker1:
use it. It's not just enterprises that use it.

260
00:15:39,620 --> 00:15:44,240
Speaker1:
It is just open source developers or just developers that can't afford to train

261
00:15:44,240 --> 00:15:45,600
Speaker1:
or pre-train their own models.

262
00:15:45,760 --> 00:15:49,680
Speaker1:
So the Chinese models are way more powerful at this point. They've almost caught

263
00:15:49,680 --> 00:15:50,900
Speaker1:
up to parity with the US models.

264
00:15:51,100 --> 00:15:56,580
Speaker1:
I think this probably suggests, if I had to guess, that the next couple of iterations

265
00:15:56,580 --> 00:16:00,940
Speaker1:
of Chinese models will probably end up leapfrogging some of the US ones.

266
00:16:01,100 --> 00:16:05,040
Speaker1:
This is just a guess from my end. And so it'll probably switch to becoming closed

267
00:16:05,040 --> 00:16:09,660
Speaker1:
source. I can't see China wanting to open source these things or continue to

268
00:16:09,660 --> 00:16:12,660
Speaker1:
open source these models if they're doing things like unwinding Manus deals

269
00:16:12,660 --> 00:16:14,720
Speaker1:
and trying to keep the assets on mainland China.

270
00:16:15,100 --> 00:16:18,940
Speaker0:
Well, you know who's probably happy about these open source models doing well?

271
00:16:19,240 --> 00:16:21,740
Speaker0:
And I'm like, this is a sneaky one, but probably Jensen.

272
00:16:22,080 --> 00:16:25,300
Speaker0:
And after I watched that interview that he had with Dorkesh,

273
00:16:26,100 --> 00:16:30,460
Speaker0:
I think one thing that became clear to me is that open source models benefit

274
00:16:30,460 --> 00:16:32,480
Speaker0:
NVIDIA more than anybody else.

275
00:16:32,480 --> 00:16:37,780
Speaker0:
And I think that's why NVIDIA has really been leaning in heavily to be the leader

276
00:16:37,780 --> 00:16:40,280
Speaker0:
in the United States, at least in terms of open source AI,

277
00:16:40,500 --> 00:16:46,620
Speaker0:
because it requires the world to train on an open hardware stack,

278
00:16:46,840 --> 00:16:48,060
Speaker0:
right? If everyone is...

279
00:16:48,460 --> 00:16:52,600
Speaker0:
Creating these models and they're publishing the weights, you can train them on NVIDIA GPUs.

280
00:16:52,980 --> 00:16:56,700
Speaker0:
Closed source models, like someone like Anthropic or OpenAI,

281
00:16:56,880 --> 00:17:01,080
Speaker0:
who we're seeing really heavily sway towards moving to TPUs or accelerators

282
00:17:01,080 --> 00:17:05,320
Speaker0:
that are not made by NVIDIA, they have the ability to build this whole custom

283
00:17:05,320 --> 00:17:08,100
Speaker0:
stack built on hardware that isn't NVIDIA's.

284
00:17:08,200 --> 00:17:11,120
Speaker0:
And that seems to be like a real threat to the NVIDIA platform.

285
00:17:11,220 --> 00:17:14,560
Speaker0:
So when Jensen got a little bit heated on the podcast, I have a,

286
00:17:18,230 --> 00:17:21,850
Speaker0:
Open source really benefits them. It makes everyone want to use NVIDIA chips.

287
00:17:21,990 --> 00:17:24,910
Speaker0:
The second these closed source models win, and if everything becomes closed

288
00:17:24,910 --> 00:17:30,570
Speaker0:
source, there's a strong bias towards these closed loop kind of hardware stacks

289
00:17:30,570 --> 00:17:32,170
Speaker0:
that NVIDIA is no longer in the loop on.

290
00:17:32,230 --> 00:17:35,170
Speaker0:
And that could create a little bit of an issue for the company.

291
00:17:35,550 --> 00:17:38,010
Speaker1:
Yeah, I think you hit the nail on the head. In this interview,

292
00:17:38,170 --> 00:17:40,550
Speaker1:
actually, for those of you who haven't watched it, definitely go watch it.

293
00:17:40,970 --> 00:17:46,770
Speaker1:
Jensen says, or was asked, hey, you dominate on GPUs. Why don't you move up

294
00:17:46,770 --> 00:17:48,290
Speaker1:
the stackers and dominate there as well.

295
00:17:48,390 --> 00:17:50,610
Speaker1:
You obviously know so much about how this is going to play out.

296
00:17:50,690 --> 00:17:56,350
Speaker1:
And he answers very simply, I am so hyper-focused on GPUs and the hardware architecture,

297
00:17:56,350 --> 00:18:01,270
Speaker1:
and this is the game I want to play, dominate, and own, right? That was his response.

298
00:18:01,490 --> 00:18:06,650
Speaker1:
So he's hyper-focused on winning GPU architectures and selling the picks and

299
00:18:06,650 --> 00:18:08,130
Speaker1:
shovels for this entire race.

300
00:18:08,430 --> 00:18:14,430
Speaker1:
NVIDIA also announced Nemo Claw, which is their enterprise version of Open Claw.

301
00:18:14,670 --> 00:18:17,230
Speaker1:
Now, you might be asking, okay, why is NVIDIA launching that?

302
00:18:17,330 --> 00:18:20,910
Speaker1:
Why are they launching a bunch of open source models, which seemingly like no one is using?

303
00:18:21,170 --> 00:18:24,370
Speaker1:
It's to your point, he wants to push adoption of AI in general,

304
00:18:24,690 --> 00:18:30,130
Speaker1:
whatever is built on top of his NVIDIA GPUs will demand more NVIDIA GPUs in

305
00:18:30,130 --> 00:18:31,430
Speaker1:
the future. And that's how he's going to make this money.

306
00:18:31,570 --> 00:18:35,590
Speaker1:
And guess what happened at the end of last week, NVIDIA once again surpassed

307
00:18:35,590 --> 00:18:38,610
Speaker1:
a $5 trillion market cap, they went above, then they went below.

308
00:18:38,770 --> 00:18:42,810
Speaker1:
And now they're sitting pretty above it. So his plan is paying off.

309
00:18:42,990 --> 00:18:46,690
Speaker1:
He couldn't explicitly say clearly on the podcast that I'm going to make a lot

310
00:18:46,690 --> 00:18:47,850
Speaker1:
of money if I sell GPUs to China.

311
00:18:48,010 --> 00:18:51,190
Speaker1:
Obviously, he can't say that. It's in poor taste. It would go against the philosophical

312
00:18:51,190 --> 00:18:54,090
Speaker1:
argument that he was making. But that is the truth, ultimately.

313
00:18:54,680 --> 00:18:58,720
Speaker0:
Yeah. So the Game of Thrones is getting a little weird now because now we're

314
00:18:58,720 --> 00:19:02,740
Speaker0:
seeing direct combat in a way between China and the United States.

315
00:19:02,740 --> 00:19:06,980
Speaker0:
We're seeing a lot of that happen just domestically across these companies.

316
00:19:06,980 --> 00:19:09,100
Speaker0:
It is getting higher and higher stakes.

317
00:19:09,280 --> 00:19:12,860
Speaker0:
And you could see as the stakes get higher, people get a little more emotionally

318
00:19:12,860 --> 00:19:14,440
Speaker0:
charged. People make larger decisions.

319
00:19:14,700 --> 00:19:18,020
Speaker0:
And China pulling the rug on this deal sets a pretty crazy precedent.

320
00:19:18,240 --> 00:19:21,200
Speaker0:
I mean, this has never happened before where they've actually interfered with

321
00:19:21,200 --> 00:19:22,400
Speaker0:
something on a material scale.

322
00:19:22,540 --> 00:19:25,480
Speaker0:
$2 billion is a lot of money. and it's been months that

323
00:19:25,480 --> 00:19:29,160
Speaker0:
they have been integrating madness into the platform now meta

324
00:19:29,160 --> 00:19:32,140
Speaker0:
if they do decide to actually unwind this deal is kind

325
00:19:32,140 --> 00:19:35,340
Speaker0:
of sitting in a little bit of an uncomfortable position because

326
00:19:35,340 --> 00:19:38,640
Speaker0:
they still haven't released their product and they now

327
00:19:38,640 --> 00:19:42,040
Speaker0:
don't have a clear path to getting a claw-like operating

328
00:19:42,040 --> 00:19:44,840
Speaker0:
system for their software that has all this

329
00:19:44,840 --> 00:19:47,760
Speaker0:
rich data about all its users so we'll be following the story closely

330
00:19:47,760 --> 00:19:50,480
Speaker0:
we'll be following the china story closely we will be testing out

331
00:19:50,480 --> 00:19:53,940
Speaker0:
all these models including deep seek v4 which i haven't had a chance to play

332
00:19:53,940 --> 00:19:56,880
Speaker0:
around with but seems like it's pretty powerful and i think that's probably

333
00:19:56,880 --> 00:20:01,100
Speaker0:
the update for for china right now it's just kind of wait and see like we have

334
00:20:01,100 --> 00:20:04,440
Speaker0:
this news that came out we're going to see how meta responds we'll see how china

335
00:20:04,440 --> 00:20:07,860
Speaker0:
responds and if they're actually willing to unwind this deal for the sake of

336
00:20:07,860 --> 00:20:09,500
Speaker0:
doing what the chinese government wants.

337
00:20:10,310 --> 00:20:15,710
Speaker1:
Yeah, can we get a response from Meta, please? It's been like 12 hours since this announcement.

338
00:20:16,110 --> 00:20:19,750
Speaker1:
And I'm on the Manus website right now. It says Manus is now a part of Meta.

339
00:20:19,870 --> 00:20:21,390
Speaker1:
So they haven't updated it yet.

340
00:20:21,830 --> 00:20:25,110
Speaker1:
They've probably got a ton of lawyers putting together a response.

341
00:20:25,190 --> 00:20:27,590
Speaker1:
But if you're listening to this from Meta and you want to give us Undisclosed

342
00:20:27,590 --> 00:20:32,070
Speaker1:
information via an anonymous tip, our DMs are open. Feel free to reach out.

343
00:20:32,310 --> 00:20:36,910
Speaker1:
But it'll be sad to see something like this go. I think OpenCore started a movement.

344
00:20:37,530 --> 00:20:42,310
Speaker1:
Manus is probably number two. Maybe Anthropics called Cowork is number two as well.

345
00:20:42,410 --> 00:20:45,410
Speaker1:
But it is a good enough product and it gave Meta the upper hand.

346
00:20:45,490 --> 00:20:48,870
Speaker1:
They were actually making money from this thing. So it's sad to see it happen.

347
00:20:48,870 --> 00:20:53,850
Speaker0:
I have a question for you, EJS. So Trump is going to China in like two or three

348
00:20:53,850 --> 00:20:54,930
Speaker0:
weeks to meet with Xi Jinping.

349
00:20:55,530 --> 00:20:59,490
Speaker0:
So how important of a topic do you think this is? Like, is this something that

350
00:20:59,490 --> 00:21:01,370
Speaker0:
would be raised to that level?

351
00:21:01,650 --> 00:21:04,710
Speaker0:
Is Trump going to discuss this with Xi Jinping to try to work out a deal?

352
00:21:04,710 --> 00:21:08,330
Speaker0:
Like, is AI and meta that important that this acquisition go through?

353
00:21:08,470 --> 00:21:10,150
Speaker0:
I think that's going to be an interesting thing to follow too.

354
00:21:10,250 --> 00:21:13,530
Speaker0:
Like, does this become a topic on the debate table between the two of them?

355
00:21:13,990 --> 00:21:17,250
Speaker1:
Okay. So I don't have a tinfoil hat, but I'm putting it on, right?

356
00:21:17,410 --> 00:21:19,590
Speaker1:
To answer this question. Okay, cool.

357
00:21:20,220 --> 00:21:24,640
Speaker1:
I think China is going to use this as a bargaining chip, pun intended,

358
00:21:24,920 --> 00:21:28,540
Speaker1:
to get access to bleeding edge NVIDIA GPUs.

359
00:21:28,680 --> 00:21:30,580
Speaker1:
I think this is a chess move, right?

360
00:21:30,880 --> 00:21:35,080
Speaker1:
Because right now, NVIDIA is selling them GPUs, but they're like the old ones.

361
00:21:35,320 --> 00:21:38,500
Speaker1:
They're the ones like in the warehouse share, the discarded items, right?

362
00:21:38,840 --> 00:21:42,460
Speaker1:
They want to like keep China on a leash. I think Xi Jinping is going to go to

363
00:21:42,460 --> 00:21:47,060
Speaker1:
Trump and say, hey, listen, we can do this ourselves, but we're willing to pay

364
00:21:47,060 --> 00:21:50,940
Speaker1:
for your GPUs as long as you give us access to the best one. no frills.

365
00:21:51,180 --> 00:21:53,620
Speaker1:
And I think he's going to use this as a bargaining chip. Otherwise,

366
00:21:53,660 --> 00:21:56,320
Speaker1:
we're not going to take any more investment from America. Yeah.

367
00:21:56,980 --> 00:21:59,260
Speaker0:
That could be. It's going to be an interesting dynamic. We'll see.

368
00:21:59,380 --> 00:22:01,320
Speaker0:
I love that this has elevated to the global stage.

369
00:22:01,460 --> 00:22:05,620
Speaker0:
This is now the single hottest topic in the world. So we will be here to follow

370
00:22:05,620 --> 00:22:08,480
Speaker0:
it as always. Thank you guys so much for watching this episode.

371
00:22:08,840 --> 00:22:12,260
Speaker0:
I hope you enjoyed. We had an episode that just released yesterday talking to

372
00:22:12,260 --> 00:22:17,480
Speaker0:
the founder of USVC, an interview, a rare interview with the general partner

373
00:22:17,480 --> 00:22:20,400
Speaker0:
from this fund that gives you access to Anthropix, SpaceX,

374
00:22:21,160 --> 00:22:25,480
Speaker0:
OpenAI, a lot of the large stocks that you would want to be investing in pre-IPO.

375
00:22:25,560 --> 00:22:26,140
Speaker0:
That was pretty interesting.

376
00:22:26,360 --> 00:22:28,460
Speaker0:
And then we'll have a whole bunch of new episodes coming out this week.

377
00:22:28,540 --> 00:22:30,180
Speaker0:
So if you enjoyed, please don't forget to share with your friends,

378
00:22:30,880 --> 00:22:35,500
Speaker0:
like, comment, leave a message, give us a five-star review, whatever you want to do.

379
00:22:35,660 --> 00:22:39,360
Speaker0:
But as always, thank you guys so much for watching and we will see you guys in the next one.

380
00:22:39,540 --> 00:22:39,960
Speaker1:
See you guys.