WEBVTT - Monologue: Concentration Risk

0:00:03.080 --> 0:00:07.840
<v Speaker 1>Today in sentences that are not in the Bible, YouTuber

0:00:07.880 --> 0:00:10.900
<v Speaker 1>Mark Plyer bought a major stake in camera maker GoPro

0:00:11.160 --> 0:00:14.560
<v Speaker 1>and then GoPro turned into an AI data center neocloud.

0:00:14.800 --> 0:00:17.230
<v Speaker 2>I'm sick and tired of the goddamn AI bubble, I

0:00:17.270 --> 0:00:27.610
<v Speaker 2>swear to Christ. Bird up, this is Better Offline and

0:00:27.650 --> 0:00:30.900
<v Speaker 2>I'm your host Ed Zitron. So today we're going to

0:00:30.960 --> 0:00:33.010
<v Speaker 2>talk through a term you may or may not have

0:00:33.050 --> 0:00:36.670
<v Speaker 2>heard before, concentration risk. It's a term that refers to

0:00:36.690 --> 0:00:38.949
<v Speaker 2>having all your eggs in one or a few baskets,

0:00:39.250 --> 0:00:42.690
<v Speaker 2>becoming overly reliant on a few investments, customers, or particular

0:00:42.710 --> 0:00:45.410
<v Speaker 2>business lines to the point that without them, your business

0:00:45.470 --> 0:00:49.690
<v Speaker 2>or portfolio would suffer massive harms or just explode. In

0:00:49.770 --> 0:00:53.690
<v Speaker 2>banking specifically, to quote the National Credit Union administration, It

0:00:53.750 --> 0:00:57.210
<v Speaker 2>refers to any single exposure or group of exposures with

0:00:57.230 --> 0:01:00.080
<v Speaker 2>the potential to produce losses large enough relative to capital,

0:01:00.220 --> 0:01:03.620
<v Speaker 2>total assets, or overall risk level to threaten a financial

0:01:03.680 --> 0:01:08.080
<v Speaker 2>institution's health or ability to maintain its core operations. I

0:01:08.120 --> 0:01:09.660
<v Speaker 2>bring this all up because you're going to hear this

0:01:09.700 --> 0:01:11.840
<v Speaker 2>term or variations of this term a lot in the

0:01:11.900 --> 0:01:14.600
<v Speaker 2>next few months and years as the AI bubble unravels,

0:01:14.959 --> 0:01:17.759
<v Speaker 2>because just about every part of the industry involves its

0:01:17.860 --> 0:01:21.630
<v Speaker 2>own flavor of concentration risk. Let's start at the top.

0:01:22.450 --> 0:01:26.130
<v Speaker 2>Per data from fintech firm Ramp, 80% of OpenAI and

0:01:26.170 --> 0:01:30.020
<v Speaker 2>Anthropix Enterprise revenues come from 1% of their customers, a

0:01:30.060 --> 0:01:33.500
<v Speaker 2>number that hasn't improved over the last three years. Ramp's

0:01:33.540 --> 0:01:37.160
<v Speaker 2>lead economist, Ara Karazian, notes that the top 1% skews

0:01:37.459 --> 0:01:40.520
<v Speaker 2>heavily towards the tech sector and AI products and services,

0:01:40.780 --> 0:01:43.720
<v Speaker 2>and that this was a level of concentration risk unseen

0:01:43.780 --> 0:01:47.760
<v Speaker 2>in any other software category they tracked. The dataset, which

0:01:47.820 --> 0:01:50.600
<v Speaker 2>includes big companies like Visa and Cursor, as well as

0:01:50.640 --> 0:01:53.740
<v Speaker 2>a great deal of startups and regular-sized companies, is very

0:01:53.800 --> 0:01:56.380
<v Speaker 2>indicative of the overall spend of the AI industry, with

0:01:56.400 --> 0:01:59.100
<v Speaker 2>the caveat that it doesn't include massive players like Microsoft

0:01:59.140 --> 0:02:01.760
<v Speaker 2>or major banks. To be clear, I'm guessing about Visa

0:02:01.800 --> 0:02:05.050
<v Speaker 2>and Cursor. Any customer on RAMP can opt out of research.

0:02:05.190 --> 0:02:06.850
<v Speaker 2>I have no idea, but I'm going to assume that

0:02:06.890 --> 0:02:09.519
<v Speaker 2>there are big companies in there. I also want to

0:02:09.540 --> 0:02:12.660
<v Speaker 2>be specific that when RAMP says AI products and services,

0:02:12.960 --> 0:02:17.160
<v Speaker 2>that includes AI startups that sell subscriptions with subsidized token spend,

0:02:17.500 --> 0:02:20.120
<v Speaker 2>meaning that users can burn far more than their subscription

0:02:20.160 --> 0:02:22.690
<v Speaker 2>price in tokens. So on a $ 20 a month subscription,

0:02:22.730 --> 0:02:25.730
<v Speaker 2>you can burn $ 30, $ 40, $ 100. This means that the

0:02:25.770 --> 0:02:28.470
<v Speaker 2>money made by Anthropic or OpenAI from an AI startup

0:02:28.510 --> 0:02:32.120
<v Speaker 2>in that 1% spend is contingent on their continued ability

0:02:32.160 --> 0:02:35.520
<v Speaker 2>to raise venture capital dollars. To simmer all this down,

0:02:35.970 --> 0:02:38.510
<v Speaker 2>It means that the vast majority of enterprises, which is

0:02:38.710 --> 0:02:40.590
<v Speaker 2>where the real money is in software and the real

0:02:40.630 --> 0:02:43.570
<v Speaker 2>growth is, just don't spend that much money on AI.

0:02:44.230 --> 0:02:46.130
<v Speaker 2>Those that do spend the most on it are heavily

0:02:46.190 --> 0:02:48.850
<v Speaker 2>concentrated in either AI companies that either use a lot

0:02:48.880 --> 0:02:52.080
<v Speaker 2>of tokens internally because they're bankrolled by venture capital, AI

0:02:52.139 --> 0:02:54.800
<v Speaker 2>companies that allow their users to blow unsustainable amounts of

0:02:54.860 --> 0:02:58.139
<v Speaker 2>money on tokens, bankrolled by venture capital, tech companies that

0:02:58.180 --> 0:03:00.860
<v Speaker 2>are currently under heavy pressure to spend money on AI tokens,

0:03:01.280 --> 0:03:03.520
<v Speaker 2>and I assume a few whale customers of some sort.

0:03:04.440 --> 0:03:08.230
<v Speaker 2>This means that 80% of OpenAI and Anthropix Enterprise revenues,

0:03:08.310 --> 0:03:10.710
<v Speaker 2>which make up the vast majority of their total revenues,

0:03:11.190 --> 0:03:14.070
<v Speaker 2>are dependent on what are likely hundreds of customers spending

0:03:14.210 --> 0:03:17.710
<v Speaker 2>outsized amounts of money on AI tokens, with an indeterminately

0:03:17.750 --> 0:03:20.180
<v Speaker 2>large chunk of them being AI startups that can only

0:03:20.240 --> 0:03:22.320
<v Speaker 2>do so as long as venture capital allows them to.

0:03:22.340 --> 0:03:25.280
<v Speaker 2>I also, and this is a gut feeling, wouldn't be

0:03:25.320 --> 0:03:28.080
<v Speaker 2>surprised if the AI startups spend way more on tokens

0:03:28.100 --> 0:03:31.440
<v Speaker 2>for writing LLM code internally, considering how every time I

0:03:31.480 --> 0:03:35.250
<v Speaker 2>see somebody going nuts on AI Twitter, It's usually a

0:03:35.290 --> 0:03:39.450
<v Speaker 2>VC-backed startup. It also means, as I've hinted, that outside

0:03:39.470 --> 0:03:41.950
<v Speaker 2>of the tech and AI world, very few companies are

0:03:41.970 --> 0:03:44.950
<v Speaker 2>willing to pay very much for AI, which is catastrophic

0:03:45.010 --> 0:03:47.410
<v Speaker 2>on just about every level, with no clear sign as

0:03:47.440 --> 0:03:50.240
<v Speaker 2>to how you reverse that trend. AI has been in

0:03:50.320 --> 0:03:53.060
<v Speaker 2>every media outlet and discussed in every boardroom and company

0:03:53.080 --> 0:03:55.500
<v Speaker 2>for the last three years. Every single company has on

0:03:55.580 --> 0:03:58.500
<v Speaker 2>some level dabbled in using AI. Most businesses have been

0:03:58.520 --> 0:04:00.200
<v Speaker 2>given the green light to spend a bunch of money

0:04:00.480 --> 0:04:02.220
<v Speaker 2>on AI. And in the end, it seems that the

0:04:02.280 --> 0:04:04.260
<v Speaker 2>only people the tech industry can get to spend money

0:04:04.310 --> 0:04:08.470
<v Speaker 2>on AI is the tech industry itself. This is OpenAI

0:04:08.510 --> 0:04:11.910
<v Speaker 2>and Anthropic's underlying exposure because these customers are also prime

0:04:11.970 --> 0:04:15.490
<v Speaker 2>targets to move to either cheaper models that they train

0:04:15.530 --> 0:04:18.589
<v Speaker 2>themselves because they're open source or eventually on device models.

0:04:19.370 --> 0:04:22.290
<v Speaker 2>Even if these customers choose to stay with Anthropic and OpenAI,

0:04:22.529 --> 0:04:25.250
<v Speaker 2>a chunk of this spend is contingent on venture capital funding,

0:04:25.290 --> 0:04:27.350
<v Speaker 2>like I've said, and the rest is contingent on whether

0:04:27.390 --> 0:04:31.170
<v Speaker 2>tech firms continue to be willing to spend money at scale. 80%

0:04:31.170 --> 0:04:34.050
<v Speaker 2>of the revenue concentration depends on spending and capital that

0:04:34.089 --> 0:04:38.929
<v Speaker 2>varies from unreliable to actively unstable. Meanwhile, these two AI

0:04:38.990 --> 0:04:43.800
<v Speaker 2>labs represent a massive concentration risk for Microsoft, Google, Amazon, Oracle, CoreWeave,

0:04:44.120 --> 0:04:46.460
<v Speaker 2>and anyone else that sells compute to them, with Anthropic

0:04:46.500 --> 0:04:51.130
<v Speaker 2>and OpenAI standing over 1.1% trillion worth of compute commitments

0:04:51.480 --> 0:04:53.820
<v Speaker 2>based on demand that's mostly coming from a very small

0:04:53.880 --> 0:04:56.440
<v Speaker 2>subset of customers. These are, from what I can tell,

0:04:56.760 --> 0:05:00.000
<v Speaker 2>take-or-pay agreements where they agree to buy that compute capacity

0:05:00.279 --> 0:05:03.080
<v Speaker 2>regardless of how much capacity they actually end up using

0:05:03.120 --> 0:05:06.040
<v Speaker 2>and how much revenue they actually bring in. As a reminder,

0:05:06.200 --> 0:05:09.340
<v Speaker 2>both Anthropic and OpenAI are woefully unprofitable and lose tens

0:05:09.400 --> 0:05:11.850
<v Speaker 2>of billions of dollars a year. To give you an

0:05:11.930 --> 0:05:15.380
<v Speaker 2>idea of the concentration risk, OpenAI's compute spend and revenue

0:05:15.420 --> 0:05:19.060
<v Speaker 2>share represent about 70% of Microsoft's AI revenues in fiscal

0:05:19.080 --> 0:05:22.200
<v Speaker 2>year 26, which just ended in June, or a little

0:05:22.260 --> 0:05:26.380
<v Speaker 2>over 70% of Microsoft's entire fiscal year revenue that year.

0:05:26.900 --> 0:05:30.710
<v Speaker 2>And UBS estimates that OpenAI and Anthropic's compute spend will

0:05:30.730 --> 0:05:34.950
<v Speaker 2>account for 48% of Google Cloud's entire revenues next year,

0:05:35.350 --> 0:05:36.910
<v Speaker 2>or somewhere between $ 84 billion and $ 100 billion in 2027.

0:05:36.890 --> 0:05:43.820
<v Speaker 2>That's on top of, per Barclay's, OpenAI and Anthropix estimated

0:05:43.820 --> 0:05:46.690
<v Speaker 2>$ 40 billion spent on Amazon Web Services and at least

0:05:46.690 --> 0:05:50.300
<v Speaker 2>$ 50 billion that both of them will spend on Microsoft

0:05:50.339 --> 0:05:53.640
<v Speaker 2>Azure in calendar year 2027, which I note because of

0:05:53.839 --> 0:05:56.849
<v Speaker 2>Microsoft's old fiscal year system. On the low end, that

0:05:56.910 --> 0:06:00.510
<v Speaker 2>means that Anthropic and OpenAI account for over $ 174 billion

0:06:00.550 --> 0:06:05.010
<v Speaker 2>worth of expected revenues from Microsoft, Google, and Amazon in 2027,

0:06:05.010 --> 0:06:07.630
<v Speaker 2>which is contingent on their ability to raise venture capital

0:06:07.670 --> 0:06:10.089
<v Speaker 2>or debt, which is contingent on the continued growth of

0:06:10.110 --> 0:06:13.130
<v Speaker 2>their businesses, which is contingent on growing AI spend from

0:06:13.170 --> 0:06:15.890
<v Speaker 2>a small subset of customers, many of whom are funded

0:06:15.920 --> 0:06:19.380
<v Speaker 2>by venture capital. The reason this hasn't been a problem

0:06:19.420 --> 0:06:21.620
<v Speaker 2>yet is that when you sign these contracts, you tend

0:06:21.640 --> 0:06:23.560
<v Speaker 2>to pay a little upfront fee and the capacity in

0:06:23.600 --> 0:06:26.160
<v Speaker 2>question is yet to come online. That's going to start

0:06:26.200 --> 0:06:28.940
<v Speaker 2>happening next year and get dramatically worse month after month

0:06:29.000 --> 0:06:32.600
<v Speaker 2>as capacity starts powering up and they start actually having

0:06:32.620 --> 0:06:35.700
<v Speaker 2>to pay for it. A really shittily written piece from

0:06:35.720 --> 0:06:38.500
<v Speaker 2>an outlet called Groundbreaker that people keep emailing me did

0:06:38.540 --> 0:06:41.120
<v Speaker 2>make a good point about this, comparing it to when

0:06:41.160 --> 0:06:43.840
<v Speaker 2>the rates on millions of mortgages exploded as they hit

0:06:43.860 --> 0:06:47.409
<v Speaker 2>a reset wall in 2027. where the low teaser interest

0:06:47.470 --> 0:06:49.969
<v Speaker 2>rates ended, so when you signed a mortgage, you would

0:06:50.370 --> 0:06:53.390
<v Speaker 2>get like 1%, 2%, 3%, very low, exploding the monthly

0:06:53.430 --> 0:06:56.510
<v Speaker 2>mortgage payments to unsustainable highs, with customers assuming when they

0:06:56.550 --> 0:07:00.030
<v Speaker 2>signed it, incorrectly, that their houses would keep appreciating, they'd

0:07:00.070 --> 0:07:02.159
<v Speaker 2>be able to refinance, or they could simply sell the

0:07:02.180 --> 0:07:04.860
<v Speaker 2>bloody thing, which they obviously could not do when everyone

0:07:04.900 --> 0:07:07.580
<v Speaker 2>was trying to do the same thing. In other words,

0:07:07.720 --> 0:07:11.080
<v Speaker 2>OpenAI and Anthropic's massive compute commitments are the subprime mortgages

0:07:11.120 --> 0:07:14.240
<v Speaker 2>of the AI bubble. They signed big, beautiful deals that

0:07:14.280 --> 0:07:18.200
<v Speaker 2>helped hyperscalers and neoclouds post massive revenue backlogs under the

0:07:18.250 --> 0:07:21.470
<v Speaker 2>belief that nothing bad would ever happen. That growth would

0:07:21.570 --> 0:07:24.490
<v Speaker 2>happen unabated, and of course the money would always be

0:07:24.530 --> 0:07:27.950
<v Speaker 2>available for everyone involved. Finally, at the top of the

0:07:27.990 --> 0:07:31.390
<v Speaker 2>pile sits NVIDIA, whose concentration risk lies with the hyperscalers

0:07:31.430 --> 0:07:35.690
<v Speaker 2>and neoclouds themselves, who justify buying further GPUs based on demand,

0:07:35.710 --> 0:07:38.310
<v Speaker 2>and I put that in air quotes here. from OpenAI

0:07:38.330 --> 0:07:41.350
<v Speaker 2>and Anthropic, with said demand for services contingent on whether

0:07:41.470 --> 0:07:45.250
<v Speaker 2>they can continue to raise money. Even those buying GPUs

0:07:45.270 --> 0:07:47.750
<v Speaker 2>to build AI data centers for other customers are doing

0:07:47.830 --> 0:07:50.620
<v Speaker 2>so because they believe there's some sort of crazy demand

0:07:50.660 --> 0:07:53.380
<v Speaker 2>for AI compute, with their reference point being the massive

0:07:53.420 --> 0:07:56.460
<v Speaker 2>revenue backlogs for Core, Weave, Iron, Nebius, and other neoclouds,

0:07:56.940 --> 0:08:00.380
<v Speaker 2>who primarily sell compute to either OpenAI, Anthropic, or one

0:08:00.400 --> 0:08:04.730
<v Speaker 2>of the hyperscalers backing them. Oh, and NVIDIA's customers are

0:08:04.890 --> 0:08:07.550
<v Speaker 2>no longer able to buy its GPUs through cash flow alone,

0:08:07.590 --> 0:08:11.410
<v Speaker 2>so all of those purchases are contingent on their constantly

0:08:11.670 --> 0:08:15.730
<v Speaker 2>availability of debt. None of this is very good at all.

0:08:16.110 --> 0:08:18.470
<v Speaker 2>I should also add that Broadcom added on their latest

0:08:18.550 --> 0:08:21.530
<v Speaker 2>earnings that Anthropic and OpenAI are going to be their

0:08:21.770 --> 0:08:25.590
<v Speaker 2>top two customers. It's all very good. It's all very normal,

0:08:25.650 --> 0:08:31.140
<v Speaker 2>very good. Everything's fine here, okay? Nobody freak out. Even

0:08:31.180 --> 0:08:32.500
<v Speaker 2>when you put it all in a line, it all

0:08:32.520 --> 0:08:34.579
<v Speaker 2>sounds really fucking bad. I'm sorry, I'm not trying to

0:08:34.600 --> 0:08:37.100
<v Speaker 2>be alarmist, but even at the end of my own monologue,

0:08:37.140 --> 0:08:37.620
<v Speaker 2>I'm kind of like.

0:08:38.500 --> 0:08:42.120
<v Speaker 1>Anyone else fucking think about this? Anyone else worried? No.

0:08:42.320 --> 0:08:45.110
<v Speaker 2>The answer's no. If you ask most sell-side analysts or

0:08:45.160 --> 0:08:47.030
<v Speaker 2>financial journalists, they'll tell you that all of this is

0:08:47.090 --> 0:08:49.290
<v Speaker 2>totally fine and it's nothing to worry about. They will

0:08:49.330 --> 0:08:51.890
<v Speaker 2>assure you that these are the smartest people in the world,

0:08:52.170 --> 0:08:54.590
<v Speaker 2>the most powerful companies, that they wouldn't spend all this

0:08:54.650 --> 0:08:57.260
<v Speaker 2>money for no reason. that the demand for both AI

0:08:57.280 --> 0:09:00.120
<v Speaker 2>compute and AI itself is real, and that the AI

0:09:00.160 --> 0:09:04.480
<v Speaker 2>skeptics are cherry-picking data. Well, we're going to fucking find out,

0:09:04.700 --> 0:09:07.360
<v Speaker 2>aren't we? And when we find out, I think it's

0:09:07.400 --> 0:09:09.240
<v Speaker 2>going to be the thing I've been warning about. And

0:09:09.280 --> 0:09:12.740
<v Speaker 2>when that happens, I've been keeping really detailed notes about

0:09:12.820 --> 0:09:16.630
<v Speaker 2>all the people that tried to hand-wave this away. Because

0:09:16.650 --> 0:09:20.550
<v Speaker 2>I think this is a catastrophic misallocation of capital, but

0:09:20.630 --> 0:09:27.020
<v Speaker 2>also just the largest miss in journalism history. Just unbelievable

0:09:27.240 --> 0:09:31.120
<v Speaker 2>to me that when this eventually falls apart, and I

0:09:31.200 --> 0:09:34.400
<v Speaker 2>am literally looking at my fucking Bloomberg terminal, and what

0:09:34.440 --> 0:09:38.829
<v Speaker 2>just popped up says, Crusoe signs roughly $ 13 billion Jane

0:09:38.890 --> 0:09:41.770
<v Speaker 2>Street deal for cloud computing. Now, you may think, wow,

0:09:41.790 --> 0:09:44.470
<v Speaker 2>that's a different customer, Jane Street, a hedge fund. How

0:09:44.510 --> 0:09:47.070
<v Speaker 2>could they possibly be involved in this? Well, you never

0:09:47.130 --> 0:09:50.200
<v Speaker 2>guess what. Jane Street's a major customer of CoreWeave and

0:09:50.240 --> 0:09:53.200
<v Speaker 2>an investor in CoreWeave. I bet they fucking invest in

0:09:53.220 --> 0:09:56.660
<v Speaker 2>Anthropic at some point. Jesus fucking, did they invest in

0:09:56.740 --> 0:09:59.030
<v Speaker 2>anthropic All right, no, I got to end this. I

0:09:59.050 --> 0:10:01.880
<v Speaker 2>got to end this goddamn monologue. Look, I'll be back

0:10:02.360 --> 0:10:04.720
<v Speaker 2>next week. I still have yet to come up with

0:10:04.760 --> 0:10:06.760
<v Speaker 2>what I'm going to do, but it's going to be great.

0:10:07.179 --> 0:10:09.480
<v Speaker 2>My cat just knocked over an empty Diet Coke can

0:10:09.540 --> 0:10:13.059
<v Speaker 2>and that's very annoying. But nevertheless, I will be back.

0:10:13.440 --> 0:10:16.200
<v Speaker 2>I love you all. I appreciate you listening. I'm Ed

0:10:16.240 --> 0:10:18.100
<v Speaker 2>Zetron and this has been Better Offline.