1
00:00:02,161 --> 00:00:02,921
How are we doing?

2
00:00:05,901 --> 00:00:11,861
So this is a talk about compute markets, but it's kind of more of a talk about why to make

3
00:00:11,861 --> 00:00:12,661
a compute market.

4
00:00:13,801 --> 00:00:14,021
If

5
00:00:14,201 --> 00:00:15,261
you don't know me, I'm Evan.

6
00:00:15,681 --> 00:00:17,721
We run a company called San Francisco Compute.

7
00:00:18,301 --> 00:00:22,241
We make a physically settled order, physically settled forwards market.

8
00:00:22,921 --> 00:00:30,461
So if you were here for the last talk, the last talk was basically why all the operators of GPU clusters don't really work very well.

9
00:00:31,481 --> 00:00:35,581
When we say physically settled, we mean we're the operator.

10
00:00:35,581 --> 00:00:39,641
We are both the operator of the hardware in the same way that a cloud is the operator

11
00:00:39,641 --> 00:00:43,081
of the hardware, and then we are also the market at the same time.

12
00:00:43,441 --> 00:00:44,441
It's kind of like a grid for compute.

13
00:00:45,301 --> 00:00:52,361
So the general gist of this is that in order to run and operate a supercomputer, we think

14
00:00:52,361 --> 00:00:57,341
problems start right at the design of the supercomputer, right in the design of the

15
00:00:57,341 --> 00:00:57,881
bill of materials.

16
00:00:58,941 --> 00:01:02,001
It's control of the backplane fabric as well.

17
00:01:02,241 --> 00:01:04,541
It's also control of bare metal orchestration.

18
00:01:04,541 --> 00:01:07,281
It's also services on top, VMs and so on.

19
00:01:07,801 --> 00:01:11,061
There's a lot of different parts of the stack that you have to do effectively.

20
00:01:11,581 --> 00:01:18,321
Otherwise, the folks who have a data center and would like to make more money on that

21
00:01:18,321 --> 00:01:24,341
data center by becoming the GPU cloud end up failing for the customer.

22
00:01:24,801 --> 00:01:26,381
So that's what we do.

23
00:01:26,541 --> 00:01:28,801
We offer what's called supercomputing as a service.

24
00:01:29,101 --> 00:01:30,821
We go to folks who want to become a cloud.

25
00:01:30,821 --> 00:01:34,401
that's like folks who want to become what NVIDIA would call an NCP,

26
00:01:35,161 --> 00:01:38,361
or venture firms that want to have clusters for their portfolio

27
00:01:38,361 --> 00:01:41,141
and want to be able to reallocate that capacity

28
00:01:41,141 --> 00:01:43,401
when a company like Instinct is taking off,

29
00:01:44,121 --> 00:01:47,641
or AI companies who would like the residual value on the compute

30
00:01:47,641 --> 00:01:50,241
because they've realized that's really important to their financials

31
00:01:50,241 --> 00:01:52,141
but don't want to operate a supercomputer.

32
00:01:52,521 --> 00:01:53,941
So that's physical settlement.

33
00:01:54,481 --> 00:01:56,781
And then on top of that, we run a live order book

34
00:01:56,781 --> 00:01:58,861
for capacity that you can buy reserved,

35
00:01:58,861 --> 00:02:01,101
served, either right now or in the future.

36
00:02:03,241 --> 00:02:07,381
This together is not at all what we thought we were going to be doing.

37
00:02:08,381 --> 00:02:13,301
Originally, we were this company called JunLark that was trying to make kind of like Suno or

38
00:02:13,301 --> 00:02:15,121
UDO, but prior to Suno or UDO.

39
00:02:15,841 --> 00:02:22,381
At the time, what you needed to do to realistically start a lab was you needed to raise a big

40
00:02:22,381 --> 00:02:26,141
deposit, a huge one, which was $10 million at the time.

41
00:02:26,821 --> 00:02:30,221
We thought that was a crazy amount of money, so we thought, no, that's stupid.

42
00:02:30,401 --> 00:02:33,881
We're going to raise 500k and obviously what we're going to do is rent from somebody else.

43
00:02:35,681 --> 00:02:39,181
Obviously today it's not $10 million, it's a billion dollars.

44
00:02:41,401 --> 00:02:47,021
And nonetheless, we had raised our 500k, we buy a big cluster, and then we're underwater

45
00:02:47,021 --> 00:02:51,581
on our cluster every month, and if we didn't sublease our cluster, we would have gone bankrupt.

46
00:02:52,041 --> 00:02:58,301
If you have heard of GPU brokers, we were the first GPU broker, so the first year the

47
00:02:58,301 --> 00:03:01,221
company was us trying to sell our cluster, otherwise we would go bankrupt.

48
00:03:01,801 --> 00:03:07,801
So I would wander around with these actual pieces of paper at NeurIps, which both say

49
00:03:07,801 --> 00:03:14,121
H100 and A100 at the same time, because that's how I made the posters so quickly.

50
00:03:15,561 --> 00:03:22,641
And so this talk is not really about necessarily how a compute market works, it's how a compute

51
00:03:22,641 --> 00:03:25,261
market gets to exist.

52
00:03:25,821 --> 00:03:29,261
We like to think about it like there's this pamphlet, and on the pamphlet, on the front

53
00:03:29,261 --> 00:03:33,861
side, it's like, man, starting a Neolab sucks, there's all these different financial and

54
00:03:33,861 --> 00:03:34,541
economic problems.

55
00:03:34,761 --> 00:03:37,921
On the back of the pamphlet, it says compute markets.

56
00:03:38,501 --> 00:03:43,061
It's not super clear how one gets to the other until you open the pamphlet and then

57
00:03:43,061 --> 00:03:45,641
And holy shit, there's so much stuff in this pamphlet.

58
00:03:45,781 --> 00:03:46,201
And it keeps going.

59
00:03:46,361 --> 00:03:46,901
There's so many folds.

60
00:03:47,261 --> 00:03:49,241
It's a difficult and challenging problem.

61
00:03:49,581 --> 00:03:53,061
Because you need to solve two miracle problems to get working.

62
00:03:53,301 --> 00:03:57,541
You have to both make a market, which is a uniquely difficult problem because it has

63
00:03:57,541 --> 00:03:58,261
a cold start problem.

64
00:03:58,561 --> 00:03:59,581
You also have to build a cloud.

65
00:03:59,821 --> 00:04:01,101
And you have to not become a cloud.

66
00:04:01,661 --> 00:04:03,481
Which is a whole separate challenge.

67
00:04:04,381 --> 00:04:06,821
But these are the problems that the labs are facing.

68
00:04:07,141 --> 00:04:09,381
And it's about to be the problems that everyone is about to be facing.

69
00:04:09,581 --> 00:04:11,521
So if you work at a start-up, these are your problems now.

70
00:04:12,041 --> 00:04:14,221
Right now, everybody is sold out of compute.

71
00:04:15,141 --> 00:04:19,061
This is probably not going to change for a while.

72
00:04:19,761 --> 00:04:23,241
So that means you're going to have to buy a new cluster, a net new one, and there's

73
00:04:23,241 --> 00:04:23,641
a lead time.

74
00:04:24,061 --> 00:04:27,041
You have no way to pay for that cluster to come online now.

75
00:04:27,221 --> 00:04:28,681
You're going to pay for it to come online later.

76
00:04:28,821 --> 00:04:32,341
There's nothing in the market that can realistically get that cluster online for you now.

77
00:04:33,121 --> 00:04:36,521
Meanwhile, you're going to pay your entire bank balance as a deposit.

78
00:04:37,061 --> 00:04:39,841
The reason you're paying your entire bank balance as a deposit

79
00:04:39,841 --> 00:04:43,101
is because the lender has no idea how to underwrite AGI.

80
00:04:43,861 --> 00:04:45,641
How in the world are you supposed to do that?

81
00:04:45,881 --> 00:04:49,501
You've got two-person startups in San Francisco, or maybe New York,

82
00:04:49,761 --> 00:04:52,001
who say, I'm going to build superintelligence.

83
00:04:53,281 --> 00:04:56,081
No lender knows how to underwrite superintelligence.

84
00:04:56,561 --> 00:04:58,381
So they have no idea what the asset is worth.

85
00:04:59,901 --> 00:05:02,361
Meanwhile, your cluster is very likely to be delayed.

86
00:05:02,561 --> 00:05:05,761
That's because the lender needed to underwrite against the credit risk of the customer,

87
00:05:05,761 --> 00:05:08,021
and so they need to know who is the customer first

88
00:05:08,021 --> 00:05:09,461
before you deploy the cluster.

89
00:05:10,121 --> 00:05:11,561
So right when the customer signs

90
00:05:11,561 --> 00:05:13,921
is the riskiest point in the entire operation.

91
00:05:14,221 --> 00:05:15,721
You wanted it to be the case

92
00:05:15,721 --> 00:05:18,361
that you built the cluster first and then you sell it.

93
00:05:18,581 --> 00:05:21,301
This also helps reduce all sorts of different types of problems

94
00:05:21,301 --> 00:05:22,801
where clusters break.

95
00:05:23,161 --> 00:05:25,361
Meanwhile, most of the people who are setting up clusters

96
00:05:25,361 --> 00:05:27,401
are people who have the actual asset

97
00:05:27,401 --> 00:05:28,661
that makes the thing cheap.

98
00:05:28,981 --> 00:05:31,121
Like you've got the data center space already.

99
00:05:31,621 --> 00:05:34,501
What that doesn't get is supercomputing expertise.

100
00:05:34,501 --> 00:05:37,901
as we saw in the last talk, if you've watched that one.

101
00:05:39,061 --> 00:05:41,361
Meanwhile, you're going to run into the exact same problem

102
00:05:41,361 --> 00:05:42,161
that Junelark did,

103
00:05:42,301 --> 00:05:44,301
which is you're going to either way overbuy

104
00:05:44,301 --> 00:05:45,761
or you're going to way underbuy.

105
00:05:46,261 --> 00:05:48,281
That's because AI doesn't have margins

106
00:05:48,281 --> 00:05:49,441
that look like SaaS companies.

107
00:05:49,681 --> 00:05:51,641
And so you don't have very much margin for error.

108
00:05:51,861 --> 00:05:54,221
And so the chances that you either underbuy

109
00:05:54,221 --> 00:05:56,081
and somebody else wins because they did it correctly

110
00:05:56,081 --> 00:05:59,561
or you overbuy and then you just become stability AI

111
00:05:59,561 --> 00:06:01,881
or you blow up, you end up like character AI,

112
00:06:02,501 --> 00:06:03,541
are really high.

113
00:06:04,501 --> 00:06:06,981
So, these are all the same problem. They're all just different types of risk.

114
00:06:07,421 --> 00:06:09,941
They're technical risk, they're financial risk, they're credit risk.

115
00:06:10,301 --> 00:06:14,321
They're all one big bundled risk problem, which is the problem that we end up solving,

116
00:06:14,541 --> 00:06:17,721
and it's also the problem that we think compute markets broadly, the entire industry solves.

117
00:06:19,121 --> 00:06:22,341
So we'll walk through each one of these and walk how that gets to a compute market,

118
00:06:22,501 --> 00:06:25,821
because if you don't understand these, it's kind of unclear why you would want a compute market,

119
00:06:26,481 --> 00:06:29,041
and they almost seem like this problem that somebody thinks,

120
00:06:29,161 --> 00:06:31,441
oh, wow, compute markets are cool, so I'm going to go start one.

121
00:06:31,441 --> 00:06:34,901
We think they should be derived from the actual problems that people face,

122
00:06:35,081 --> 00:06:37,001
which is, first, everyone is sold out.

123
00:06:37,281 --> 00:06:38,401
Why are they sold out?

124
00:06:38,541 --> 00:06:40,341
Because the lender couldn't finance the cluster

125
00:06:40,341 --> 00:06:42,381
unless you had a long-term off-take agreement.

126
00:06:42,621 --> 00:06:44,941
In CPUs, this wasn't the case,

127
00:06:45,001 --> 00:06:48,001
because CPU buyers are SaaS companies.

128
00:06:48,241 --> 00:06:49,701
SaaS companies have high margins,

129
00:06:49,741 --> 00:06:51,161
and because they have high margins,

130
00:06:51,341 --> 00:07:01,281
you can build a product that rewards you with high margins for your CPUs.

131
00:07:02,801 --> 00:07:03,981
Think about it this way.

132
00:07:04,281 --> 00:07:08,781
If you're a SaaS company and your underlying cloud doubles their prices, you don't go bankrupt.

133
00:07:09,201 --> 00:07:11,541
You hate your cloud, you're mad, but you don't go bankrupt.

134
00:07:12,141 --> 00:07:15,801
If you're an AI company and your underlying cloud doubles their prices, you go bankrupt.

135
00:07:17,741 --> 00:07:21,761
And what that means is that you're actually price sensitive, and because you're actually

136
00:07:21,761 --> 00:07:26,761
price sensitive, there's nothing keeping you at any particular cloud, so you can't sell

137
00:07:26,761 --> 00:07:30,661
on demand because your customers will immediately leave to whoever is cheaper.

138
00:07:30,661 --> 00:07:36,581
And so that means that in order to finance GPU clusters, you have to have long-term take-or-pay

139
00:07:36,581 --> 00:07:39,221
contracts, hopefully with someone who has investment-grade credit.

140
00:07:40,381 --> 00:07:44,341
However, if you don't have liquidity on those, if you're selling reserved and your customer

141
00:07:44,341 --> 00:07:48,121
can't get out, then that also means that some other customer who wants the capacity right

142
00:07:48,121 --> 00:07:52,101
now because they're like instinct and their product is taking off, can't pay for anything.

143
00:07:52,281 --> 00:07:55,581
They have to finance a new cluster which has a lead time because the physical world just

144
00:07:55,581 --> 00:07:56,941
exists and it's going to take a while.

145
00:07:57,321 --> 00:07:58,841
Sometimes you even have to build the damn data center.

146
00:08:00,081 --> 00:08:03,541
And then, because you're financing a new cluster, you're going to pay a deposit.

147
00:08:03,741 --> 00:08:06,661
Because almost nobody in AI has any reasonable credit.

148
00:08:07,241 --> 00:08:10,021
And because this is a brand new technology, all these companies are new.

149
00:08:10,461 --> 00:08:15,301
Meanwhile, even as you start to scale up that ladder, you're still going to run into credit

150
00:08:15,301 --> 00:08:16,401
issues all the way through.

151
00:08:16,521 --> 00:08:18,261
It's the number one bottleneck of AI at the moment.

152
00:08:18,401 --> 00:08:18,981
It's not power.

153
00:08:19,081 --> 00:08:23,801
You hit credit way before you hit power.

154
00:08:25,041 --> 00:08:30,941
So, what that means is every inference company, every neo lab, they are bottlenecked by how

155
00:08:30,941 --> 00:08:32,241
much cash they have in the bank account.

156
00:08:32,521 --> 00:08:36,481
So they go back to the venture capitalist who says, oh, I'm going to fund at the model

157
00:08:36,481 --> 00:08:39,901
layer, the application layer, the inference layer, because I don't want to fund all of

158
00:08:39,901 --> 00:08:40,841
that capex stuff.

159
00:08:41,421 --> 00:08:48,381
And so the venture capitalist then funds all the capex stuff by giving the new neo lab

160
00:08:48,381 --> 00:08:54,661
or new inference company, the check that they then send to the cloud, which is the deposit

161
00:08:54,661 --> 00:08:56,841
for the cluster, and that's how you underwrite the customer.

162
00:08:59,781 --> 00:09:06,481
So the clusters then are delayed because the cloud couldn't actually deploy the cluster

163
00:09:06,481 --> 00:09:10,321
until they had the customer in hand, and so that was the riskiest point.

164
00:09:10,721 --> 00:09:16,441
So therefore, what should have happened was the cloud had bought the cluster ahead of

165
00:09:16,441 --> 00:09:21,081
time, and then waited until it was landed in the data center, and then sold it.

166
00:09:21,361 --> 00:09:23,501
And that means you have way less likelihood of delays.

167
00:09:24,421 --> 00:09:29,041
Meanwhile, the cluster is very likely to break when you get the cluster as the customer because

168
00:09:29,541 --> 00:09:30,321
you don't have liquidity.

169
00:09:30,621 --> 00:09:34,441
And so if the failure rate is higher than what you expect, and you don't have the ability

170
00:09:34,441 --> 00:09:38,461
to transition a node in, then you don't have anywhere you can pull nodes from.

171
00:09:38,621 --> 00:09:42,781
Which means that if you need new nodes to make up for the failed ones, you have to RMA

172
00:09:42,781 --> 00:09:43,041
them.

173
00:09:43,041 --> 00:09:44,401
which means you have to get a new node.

174
00:09:44,901 --> 00:09:47,661
And then that gets us back to several steps before

175
00:09:47,661 --> 00:09:50,401
where you have the lead time problem again.

176
00:09:50,641 --> 00:09:52,981
And so everything breaks for everybody all the time.

177
00:09:54,061 --> 00:09:56,541
Once you have gotten through all of that as a new AI lab,

178
00:09:56,921 --> 00:10:00,801
your problems are you, you know, you, nobody was sold,

179
00:10:00,941 --> 00:10:01,961
like everyone was sold out.

180
00:10:02,361 --> 00:10:03,661
You then had to wait a lead time.

181
00:10:03,781 --> 00:10:04,761
You paid this big deposit.

182
00:10:05,241 --> 00:10:06,841
The thing was delayed and then it broke.

183
00:10:07,001 --> 00:10:08,521
And finally you got through all of that.

184
00:10:08,841 --> 00:10:11,001
And the likelihood that you predict your demand correctly

185
00:10:11,001 --> 00:10:12,061
is very low.

186
00:10:12,061 --> 00:10:14,021
And so you're either going to be way

187
00:10:14,041 --> 00:10:16,701
over-provisioned or way under-provisioned, and you have no way out.

188
00:10:17,001 --> 00:10:20,561
There's no way to sell your capacity to somebody else, and there's no way to buy more capacity if you need it.

189
00:10:20,841 --> 00:10:22,641
So, compute markets.

190
00:10:23,681 --> 00:10:26,961
To solve these problems, you need to solve a bunch of different problems all at once.

191
00:10:27,301 --> 00:10:33,861
You need to underwrite the value of the GPU cluster rather than underwriting the credit risk of the customer.

192
00:10:34,301 --> 00:10:35,721
This is how it could work.

193
00:10:36,161 --> 00:10:38,101
It's not guaranteed to work like this.

194
00:10:38,101 --> 00:10:42,001
In other industries, what happened is everything just consolidated.

195
00:10:42,321 --> 00:10:46,401
You got one big credit-worthy customer that does the whole damn industry.

196
00:10:46,641 --> 00:10:50,001
That's what the path of AI is currently on track to do.

197
00:10:50,221 --> 00:10:54,521
It's instead of going to be a liquid market where anybody can participate,

198
00:10:54,781 --> 00:10:56,241
it's going to look like TSMC.

199
00:10:56,461 --> 00:11:00,401
You're going to get one buyer or a few buyers that are the clouds.

200
00:11:00,681 --> 00:11:03,561
You're going to get a few model labs that are the AI labs.

201
00:11:03,701 --> 00:11:05,381
And then you won't get other startups.

202
00:11:06,501 --> 00:11:08,001
This is what we think is going to happen.

203
00:11:08,101 --> 00:11:12,361
And this is why three years ago, when we were doing our AI lab, we had a chance.

204
00:11:12,621 --> 00:11:20,521
We could have said, no, we're not going to stick it out for several years just trying to keep our company alive.

205
00:11:20,821 --> 00:11:22,801
We're just going to go do a new AI lab.

206
00:11:22,901 --> 00:11:24,761
But we saw this problem.

207
00:11:24,901 --> 00:11:25,841
And we said, oh, shit.

208
00:11:26,581 --> 00:11:30,781
If we don't do anything about this, then we will just hit this three years from now.

209
00:11:31,081 --> 00:11:35,061
And here we are three years from now hitting this problem and where everyone is going to hit this problem.

210
00:11:36,721 --> 00:11:42,821
If you watched the last talk, it's very hard to solve this problem unless you are operating

211
00:11:42,821 --> 00:11:46,581
both the market and the clusters at the same time.

212
00:11:47,141 --> 00:11:47,801
This is Dylan Page.

213
00:11:48,441 --> 00:11:50,721
Dylan is the co-chair of the CNCF infrastructure group.

214
00:11:50,861 --> 00:11:52,061
That's the organization that makes Kubernetes.

215
00:11:52,501 --> 00:11:54,561
He formerly worked at Lambda doing infrastructure there.

216
00:11:55,341 --> 00:12:02,161
And he joined San Francisco Compute because GPU Compute isn't sufficiently reliable for

217
00:12:02,161 --> 00:12:04,321
the market operator to be divorced from the physical infrastructure.

218
00:12:04,761 --> 00:12:07,641
It doesn't work if you aren't the one who is buying the thing.

219
00:12:07,961 --> 00:12:10,561
Sorry, it doesn't work if the person you are buying from isn't operating it.

220
00:12:10,741 --> 00:12:14,381
To give an example of why this doesn't work is I'm sure some of you are flying in from

221
00:12:14,381 --> 00:12:19,801
overseas or out of town and you have a hotel room.

222
00:12:20,801 --> 00:12:24,281
Wouldn't it be great if you could sell that hotel room to somebody else?

223
00:12:24,621 --> 00:12:30,241
I go to you and I want to buy that hotel room and I say, can I get room service?

224
00:12:30,241 --> 00:12:31,921
Can I get my room cleaned?

225
00:12:32,381 --> 00:12:34,481
I've got a problem with my room, who do I call?

226
00:12:35,261 --> 00:12:39,461
If you don't operate the hotel, you can't fix any of those problems.

227
00:12:39,761 --> 00:12:42,081
That's why brokering clusters doesn't work.

228
00:12:42,241 --> 00:12:44,521
You have to operate the cluster like the hotel.

229
00:12:44,721 --> 00:12:45,981
You have to run the whole damn thing.

230
00:12:46,281 --> 00:12:48,261
Otherwise, you can't actually make a market.

231
00:12:48,501 --> 00:12:48,941
It doesn't work.

232
00:12:49,021 --> 00:12:49,801
You can't do physical settlement.

233
00:12:51,841 --> 00:12:53,701
So why does nobody do physical settlement?

234
00:12:54,401 --> 00:12:57,821
The problem is there's a very strong incentive to never do physical settlement.

235
00:12:58,321 --> 00:13:02,601
That is because if you build out physical settlement for compute, what that literally

236
00:13:02,601 --> 00:13:06,281
means is you are building out the entire operating capacity of a GPU cloud.

237
00:13:06,521 --> 00:13:11,121
And if you are building out the entire operating capacity of a GPU cloud, you should just be

238
00:13:11,121 --> 00:13:11,781
a GPU cloud.

239
00:13:12,041 --> 00:13:16,221
Because it looks to be the case that you will make an ungodly amount of money.

240
00:13:16,701 --> 00:13:19,901
You will sign billion dollar deal after billion dollar deal.

241
00:13:20,241 --> 00:13:22,101
And so, fuck it.

242
00:13:22,421 --> 00:13:24,881
Let's not make a GPU cloud, or let's not make a compute market.

243
00:13:24,881 --> 00:13:29,181
Let's instead just become a GPU cloud, which is what basically everybody does, because

244
00:13:29,181 --> 00:13:31,581
the numbers are astronomical and it's a very reasonable thing to do.

245
00:13:32,441 --> 00:13:36,221
The reason why we ended up building physical settlement was because we knew that if we

246
00:13:36,221 --> 00:13:40,041
didn't start this, we would then not be able to start an AI lab, which was the real damn

247
00:13:40,041 --> 00:13:41,301
thing that we wanted to do in the first place.

248
00:13:41,561 --> 00:13:42,501
We wanted to make a lab.

249
00:13:42,861 --> 00:13:44,901
We didn't get to make a lab unless this exists.

250
00:13:46,361 --> 00:13:52,241
So when we started, it was 10 million as the deposit that you had to pay to start a new

251
00:13:52,241 --> 00:13:52,461
lab.

252
00:13:52,461 --> 00:13:54,881
This is how you get over the credit risk problem.

253
00:13:56,461 --> 00:14:00,921
Then the cluster sizes got bigger, the sort of minimum ante into the space got bigger,

254
00:14:01,381 --> 00:14:03,141
then it was 100 million, now it's a billion.

255
00:14:03,481 --> 00:14:05,021
We don't think it can continue.

256
00:14:05,381 --> 00:14:09,161
We don't think there is a world in which you can get a $10 billion round right out the

257
00:14:09,161 --> 00:14:13,361
bat before you have revenue or traction or your product or anything.

258
00:14:14,301 --> 00:14:19,141
And what that means is that if you don't solve this problem, we are currently on track to

259
00:14:19,781 --> 00:14:21,081
not have any start-ups in AI.

260
00:14:21,661 --> 00:14:27,301
Now, one of the things that you could think could happen is maybe there can be new inference

261
00:14:27,301 --> 00:14:28,941
companies and new model labs.

262
00:14:29,081 --> 00:14:33,041
And the way that they solve the credit problem is that instead they just ship the risk to

263
00:14:33,041 --> 00:14:34,041
the application layer.

264
00:14:34,261 --> 00:14:42,041
They say, look, I need to get over my credit problem, so what I'm going to do is I'm going

265
00:14:42,041 --> 00:14:47,921
to only sign customers who are investment grade credit, you know, Fortune 500, I'm going

266
00:14:47,921 --> 00:14:51,061
to only sign customers on five-year contracts for tokens.

267
00:14:51,541 --> 00:14:56,561
I'm going to only sign customers for tokens that require a big deposit up front.

268
00:14:58,521 --> 00:15:00,421
That is the end of everything.

269
00:15:02,021 --> 00:15:08,101
you have inference layer getting bottlenecked by credit, what happens is you can't get tokens

270
00:15:08,101 --> 00:15:10,761
unless you have credit or you have a big deposit.

271
00:15:11,901 --> 00:15:15,701
the problem of that is that tokens are used by the whole damn economy.

272
00:15:15,701 --> 00:15:18,881
There is no rule that says that startups have to exist.

273
00:15:19,981 --> 00:15:24,681
If every startup is suddenly bottlenecked by credit in order to access intelligence,

274
00:15:25,161 --> 00:15:28,721
what happens is existing firms just ship everything.

275
00:15:29,001 --> 00:15:32,861
They themselves get to make the product because they've got credit,

276
00:15:33,081 --> 00:15:36,641
and then they make the revenue from that product, and then it makes their credit better.

277
00:15:36,801 --> 00:15:40,961
And that's a flywheel that makes consolidation happen throughout the entire economy.

278
00:15:41,261 --> 00:15:44,221
If you don't fix this bottleneck, there aren't any startups.

279
00:15:44,221 --> 00:15:46,061
This is why we started doing this.

280
00:15:46,381 --> 00:15:52,661
Because at one point, what we think is going to happen is you're going to end up in three

281
00:15:52,661 --> 00:15:55,861
to five years, you're going to look out and say, I want to start a new startup.

282
00:15:56,861 --> 00:16:03,521
To get there, you need to buy so many tokens in order to compete with the existing players.

283
00:16:04,601 --> 00:16:07,281
You need to buy $100 million of tokens.

284
00:16:08,021 --> 00:16:11,661
We think this is what you're going to have to do in order to start a new company because

285
00:16:11,661 --> 00:16:14,141
other people are spending hundreds of millions of dollars of tokens.

286
00:16:14,321 --> 00:16:16,241
If you just draw the line out, this is what happens.

287
00:16:16,741 --> 00:16:21,561
So you need a big giant deposit because you're two people who have no idea what they're doing

288
00:16:21,561 --> 00:16:22,841
and so nobody trusts you.

289
00:16:23,061 --> 00:16:26,301
So you're going to go and raise $100 million from your venture capitalist.

290
00:16:27,621 --> 00:16:30,521
Meanwhile, nobody wants to work with you because you're a two-person startup.

291
00:16:31,201 --> 00:16:35,101
Then you're going to be asked to pay for that hundred, or you're going to be asked to pay

292
00:16:35,101 --> 00:16:38,701
for a five-year contract on that hundred million if you're not paying it all as a deposit.

293
00:16:39,281 --> 00:16:42,061
And then you just don't get new model labs, you don't get new start-ups.

294
00:16:42,461 --> 00:16:44,021
You get a consolidation in the industry.

295
00:16:45,221 --> 00:16:49,521
So we, at one point, were this little lab.

296
00:16:50,041 --> 00:16:55,681
And what we did is we lived in San Francisco and we knew all the other people who were

297
00:16:55,681 --> 00:17:00,381
also just two little folks who had started a start-up and wanted to be able to participate

298
00:17:00,381 --> 00:17:02,081
in the most important technology of our time.

299
00:17:04,421 --> 00:17:06,381
So I want these people to exist.

300
00:17:06,381 --> 00:17:09,341
I want this person to exist and I want this person to exist.

301
00:17:09,581 --> 00:17:13,001
I want this startup to exist and this startup to exist and this startup to exist.

302
00:17:13,241 --> 00:17:17,301
This one and this one and this one and this one and this one and this one and this one and this one.

303
00:17:17,881 --> 00:17:23,741
If you get these companies going, maybe they'll become big folks that build your cancer cures

304
00:17:23,741 --> 00:17:28,401
or they build something that nobody thought of or something that wasn't just sort of AI slopped out.

305
00:17:29,981 --> 00:17:32,381
This is kind of the problem that compute markets solve.

306
00:17:32,561 --> 00:17:35,541
And if you don't solve it, we don't think there's going to be any startups.

307
00:17:36,841 --> 00:17:37,521
Cool.

308
00:17:38,681 --> 00:17:39,261
That's it.

309
00:17:39,361 --> 00:17:39,601
That's it.

310
00:17:39,961 --> 00:17:40,481
Yeah.