Latest / In the Loop AI / Three Billionaires Just Called the AI Top: Elon, Ken Griffin, Chamath and What Each Stands to Gain

Transcript

Three Billionaires Just Called the AI Top: Elon, Ken Griffin, Chamath and What Each Stands to Gain · In the Loop AI

0:00
There's no easy way to say this, but the stock markets are getting crushed because
0:03
three of the richest men in the world might have called the top of the AI bubble.
0:07
One of them is Elon Musk, one of the biggest proponents of AI in the first place.
0:11
He thinks that there's not enough capital and money in the market to support
0:15
the expensive buildout of AI.
0:17
And Ken Griffin, the CEO of Citadel, agrees with them. He says we're almost
0:20
certainly in a bubble. Chamath thinks that 75% of every single dollar spent
0:26
and invested in AI is being subsidized by venture capital.
0:29
And AI stocks are down horrendously, particularly neoclouds,
0:32
who have drawn down on average around 30 to 35% over the last couple of weeks. Absolutely brutal.
0:39
But what doesn't make sense about me and where the alarm bell started ringing
0:42
is each of these three individuals have an incentive to say that thing.
0:46
Elon Musk most famously recently took SpaceX to a pivot where they weren't just
0:51
a rocket company or AI company, they also became a neocloud.
0:54
So it makes sense for him to talk down on his competitors.
0:57
And Ken Griffin's Citadel is rumored to be short billions of dollars worth of semiconductor stocks.
1:02
And Chamath, he says it in his own words, owns a bunch of power and energized
1:07
land. So he might just be talking his book.
1:09
Now, the reason why I'm creating this episode is kind of self-serving in a way.
1:13
I am an investor, I invest in some of these markets, and I'm trying to figure
1:16
out what's actually truthful from what they're saying, and what actually isn't.
1:20
I want to debunk anything and everything and see what actually makes sense.
1:23
And I figured a bunch of you are also investors or investing oriented as well.
1:26
Or maybe just some of you just want to know what the hell is going on,
1:29
because it's affecting pretty much anything and everything from an investment
1:32
perspective out there. It's affecting global macroeconomic indexes.
1:36
So let's get into all of this and more on today's episode brought to you by Qualcomm.
1:43
So it all started off with this one tweet from Chamath, where he basically lists
1:47
out a lot of reasons why neoclouds and investing in compute GPUs,
1:52
the capex intensive part of AI, is currently in a bit of a precarious situation.
1:57
Neoclouds are disorganized and very poorly capitalized. Neoclouds being the
2:01
smaller startups and companies that aim to provide compute for AI labs that
2:05
want to train and inference their models.
2:07
He's saying they're poorly capitalized they can't get access to enough money
2:11
to be able to fund the compute that they're promising these companies and so
2:14
it's ending up becoming a bit of a bubble or bottleneck due to power and energy
2:18
stuff and Elon Musk decided to pile on here and he said as of the past few weeks
2:22
debt markets have almost entirely shut off access to money
2:24
for any company that doesn't have substantial revenue independent of AI data
2:28
centers he then goes on to say that means the 200 plus neoclouds are going to see
2:33
guppy level death and consolidation some will make it through, but not many.
2:38
And obviously, the market's reaction to this was terrible.
2:41
There was an entire near cloud meltdown, as you're seeing on this tweet,
2:44
where stocks just cratered and tanked.
2:47
And it's interesting to see, one, because these guys have such a huge amount
2:50
of influence on social media to influence these markets at such a great capacity.
2:54
But two, this was the second time these guys have done it in a week.
2:58
Last week, I was talking about how the core we've trade kind of became unwound.
3:03
It was down, I think now 35% after Chamath and Elon Musk wrote tweets about
3:09
them being over-levered in the debt market. I made an episode back then basically explaining.
3:14
This was kind of not exactly true. Yes, they're taking out a lot of debt,
3:18
but they had a huge amount of revenue backlog to back it up.
3:22
And Chamath himself even said that it isn't as bad as everyone is making out
3:26
and that the market is actually overreacting.
3:28
Chamath's point is that 90% of AI spending that we see today,
3:32
that Drive, Anthropic's massive ARR that we saw in the S1 yesterday,
3:36
and OpenAI's rumored ARR, comes from 1% of companies.
3:41
And 80% of that revenue, of that spending, comes from VC subsidized dollars.
3:48
So subsidized from venture capital, aka coming from startups that don't actually
3:54
have the revenue or profit to be spending that money. They're just spending
3:57
the funding that they raised.
3:59
And so the point that he's making is if this napkin math is true,
4:03
then we are in a heavily inflated bubble that is bound to either pop or recede
4:08
pretty aggressively over the next year.
4:10
And he makes the case that it's going to happen mainly in 2027.
4:13
We're going to see some form of consolidation.
4:15
Now, Elon Musk piling on, particularly around neoclouds, I find very interesting
4:19
because SpaceX's entire MO,
4:22
the reason why they had blockbuster revenue earnings literally last week.
4:27
Was because they became the number one Neoclod provider. So forgive me if I'm
4:31
just a little bit skeptical as to why Elon might be talking his book.
4:34
This is coming from one of the biggest Elon fans out there. And their words
4:37
have had a pretty gnarly impact. The markets have been wiped of around $1.6
4:41
to $2 trillion over the last 14 days.
4:45
Staggering and it's part of a broader drawdown that we haven't seen for a significant time.
4:50
The last couple of years has been up only and you could consider this nearing
4:54
a recession type indicator.
4:55
But the sector that's suffering the most are neoclouds and the three main ones
4:59
CoreWeave, Nebius and IREN are down around 30 to 35 percent over the last couple
5:03
of weeks absolutely shattering the market caps now the reason why this is a
5:07
problem in their situation if you were in their boots is
5:10
they need as much capital as they can get they need as much favorable credit
5:14
financing aka money that they can borrow that they can get to be able to buy the gpus to
5:21
set up the data centers to then do their entire business.
5:25
So you can see how this could end up in a house of cards or domino style collapse.
5:29
But at the same time, I see some form of justification behind what Ken Griffin,
5:34
Elon Musk and Chamath are saying.
5:36
The number one indicated being that there are currently millions of dollars
5:40
or rather billions of dollars worth of GPUs sitting collecting dust in data centers.
5:46
And if you don't believe me, take Amazon's word for it, where they have openly
5:49
said that they have millions of dollars worth of GPUs collecting dust.
5:53
Because they can't quite find the power and electricity and access to the grid
5:59
to be able to switch them on.
6:00
Now, this is okay, typically, but eventually it becomes cumbersome,
6:04
where you end up actually losing money to the tune of $1 to $5 million a day.
6:09
And I saw something recently, which basically indicated that now these companies
6:13
who have been promised data centers stood up by a certain time,
6:15
if it gets delayed, they are owed $1 to $5 million per day.
6:20
And insurance markets are coming out of nowhere saying, okay,
6:23
we'll back this kind of debt.
6:24
And these are the same insurance providers that are lending the money in the first place.
6:28
Now, I'm not trying to get too complex here, but what I'm trying to describe
6:32
is this is a very circular loop that is forming.
6:36
And this is not something that we have been secretive about in previous episodes.
6:41
I've spoken about it quite blatantly and honestly here, but we're starting to
6:44
see the actual evidence of these deals, what's being kind of signed on behind
6:49
the scenes actually forming and what might give this a bit of substance.
6:53
And even banks and chief economists, I think the chief economist of Apollo Global
6:57
actually went online over the weekend and posted his thoughts.
6:59
And he basically said that, listen, aside from the top seven companies in the
7:03
world who are actually using AI profitably, so per employee,
7:07
they're making actually more money than they did, let's say a year ago.
7:10
The other 493 companies in the Fortune 500 are not doing this at all.
7:16
In fact, they're probably losing money. So I'm a bit at odds here.
7:20
Now let's take Chamath's first claim, which is 75% of every dollar spent in
7:24
AI is heavily subsidized by venture capital.
7:26
Now, I'm from the start going to say I don't necessarily disagree with this
7:31
take. I definitely do agree that there's a bunch of subsidies being made here,
7:35
but I don't think it's as extensive as he's making out.
7:39
And I thought about that immediately yesterday when Anthropic released their
7:43
S1 file, which is their pre-IPO financials. And I looked into it.
7:47
And one astounding thing for me was, one, how much cash they had on the balance
7:51
sheet. It's around $132 billion, by the way.
7:54
And two, how much they've burned on compute over the last year.
7:58
We're talking about one of the leading AI labs here who have spent the most
8:01
to train and inference their models to create the number one leading model.
8:05
Guess how much they spent of that?
8:06
Only $7 to $9 billion.
8:09
So they're very capital rich at this point. But there was also another important
8:12
detail in the S1, which said that of all of Anthropic's main enterprise customers
8:17
who were spending anywhere between $100,000 a year to $10 million a year,
8:22
this only accounted for 0.1%,
8:27
of their entire annual company budget.
8:30
0.1% of the entire company's money that they spend in a year is being spent on clawed models.
8:37
I don't know about you, but I thought that number would be way, way bigger.
8:41
And it just highlights that not only is companies or labs
8:45
generating the revenue that we hypothesize and speak about, but they're also
8:49
making it in a way where it's not really too damaging to the books,
8:52
which implies that there's probably more to grow here and more money to be spent.
8:56
And maybe I just can't wrap my head around what that number might look like.
8:59
Now, the second claim, which is showed beautifully by this Apollo chart over
9:03
here, basically says that of the 90% of money that is being spent in AI.
9:08
It's all focused on 1% of companies. And I guess this doesn't surprise me too much.
9:13
If you are AI native or trying to pivot your business to become AI native,
9:17
it makes sense that you are spending as much time experimenting with these different
9:21
AI models and tools to try and figure out how to accelerate and automate parts of your business.
9:25
It kind of makes sense that the concentration of spending is happening in these few companies.
9:30
And that gray box might actually signal future potential growth for the AI sector as a whole.
9:35
But equally speaking, if they are right, and it is all coming from VC subsidized
9:39
dollars and only from 1% of companies in the world, the question is,
9:44
why aren't the other 99% of companies actually using this thing?
9:47
And my answer is, I don't think they're not trying to use the thing.
9:50
I just don't think they've realized whether these models are useful to them in the first place.
9:55
And so there's this weird gap that's widening behind the people who are AI obsessive
10:01
and are spending all of their money, capex and time trying to figure out how
10:04
to pivot to become AI native.
10:06
And then the rest of the world that just don't really know what's happening
10:10
that interact with a version of chat GPT, or maybe even an older model of Claude.
10:14
And they're like, yeah, this thing hasn't really changed over the last year.
10:17
Now Elon's point around neoclouds not being able to access financial capital,
10:21
I think is starting to materialize now. It maybe wasn't the case six months
10:26
ago, but it is now because only investment grade lenders. So these are.
10:32
Providers that are able to gather the amount of capital, which is a lot,
10:35
like we're talking about tens of billions of dollars to back up all the promises
10:39
that these neoclouds collectively have made.
10:42
Is hard to come by because capital used to be abundant, but is no longer abundant.
10:48
You've got OpenAI that is raising rumored to be around $30 billion that I think
10:52
they've recently closed.
10:53
You've got SpaceX, which announced that they officially closed $40 billion on
10:57
some higher interest rates than they did a few months ago.
11:00
And you've got Anthropic that closed $63 billion worth of compute,
11:04
but actually required Broadcom CEO of Hocktan to walk into the banks in the
11:09
last hour to try and convince them to actually close the deal.
11:12
So the question I ask myself is, well, if they're tentative on the biggest AI
11:16
lab in the world to give them that compute,
11:19
then why the hell would they give that to the neocloud smaller providers who
11:23
don't have the AI research expertise and revenue backlog that a bigger company like Anthropic has.
11:29
And when I look at the last year, the explosion of neoclouds,
11:32
the pivots of all the Bitcoin miners and crypto miners to become neoclouds, I do end up thinking...
11:39
Maybe this place has become a little too crowded. I see a lot of small cap stocks
11:43
and I think, yeah, maybe we are due some consolidation. There are way too many
11:46
players in the game. So I do think that we will see some of these neoclouds go bust.
11:51
I do think that some of their contracts will need to be given up.
11:55
But if I had to guess, they'll just get sucked up by bigger hyperscalers like
11:59
Amazon, like Google, like Microsoft, who have the credit score,
12:02
the capital cash accounts, and the financing and reputation to be able to do so.
12:07
And the reason why I'm confident of that is because, as the Anthropic S1 and
12:11
many other indicators show, the demand for AI is still at an all-time high.
12:15
You look at the average price of a GPU that was debuted 10 years ago,
12:20
and it's selling for 2.5x the price than it did back then, even though there
12:23
are much better versions of GPUs out there.
12:25
You look at NVIDIA selling their brand new GPUs, the Vero Rubin,
12:28
and it's completely sold out to one or two particular people,
12:32
Mark Zuckerberg and Elon Musk.
12:34
And there are a ton of other companies out there that want the same GPU.
12:36
So we are very limited and constrained, but the capital itself is not abundant.
12:41
Now, on the point of hyperscalers, so these are companies like Amazon,
12:45
Google, Microsoft, who have very huge cloud service provider units,
12:48
cash-rich balance sheets, and are buying up as many GPUs as they can to stand up data centers.
12:53
These hyperscalers have a different type of problem.
12:57
They're quite bureaucratic and they're slow moving, which is the point that
13:01
Chamath makes. He says that they can't make decisions quick enough in order
13:05
to help build out these GPUs, except for one provider. Guess who that provider is?
13:10
SpaceX, of course, which Chamath is invested in. So you can start to see the
13:14
kind of cohesion that these two are starting to have and the impact that they're
13:17
having on the market and where they might be actually talking up their book.
13:20
Because of course, SpaceX has been one of the quickest companies to scale up GPUs and data centers.
13:26
Elon Musk has famously spun up Colossus 1, 2, and 3 and are bringing online
13:30
200,000 to 300,000 GPUs, $13.1 billion worth of compute every single month.
13:36
This is currently unmatched by any other company in the world,
13:39
so much so that Anthropic is
13:41
buying compute directly from SpaceX AI to the tune of around $3.1 billion,
13:47
per month, committed over the next couple of years until, I think, 2032.
13:51
And Elon Musk is charging a premium for this. He's the only company and only
13:54
man that is able to pull this off. And you might be asking, well,
13:57
why is Elon the only person to be able to pull this off?
14:00
Well, it's because the constraint right now is in the physical...
14:05
Challenges that are required to build a data center. Let me try and put this into context for you.
14:10
To build a normal data center, so forget about the GPUs, just normal data centers
14:14
as they existed pre-chat GPT.
14:16
That took a long time, a lot of regulatory burden, a lot of legal red tape.
14:21
Now you add in GPUs, incredibly expensive, but also so, so complex to set up.
14:26
You need to train technicians from scratch to now try and teach them a completely
14:30
other different architecture, a different cooling system, a different type of
14:34
transformer, power lines that are able to scale at an industrial level.
14:39
You need to retrain all of these people and then get them to work.
14:42
But then you're faced by all the legal burden, the red tape.
14:45
Data centers want to be switched on tomorrow. They want to be set up in a blink of an eye.
14:49
But obviously, that isn't how the world currently works. And even miraculously,
14:53
if we get to that point, there is not enough electricity, power or gridline
14:59
infrastructure to support this thing.
15:00
So what ends up happening is it slows the entire movement of setting up data
15:05
centers down, which means that the chips are ready before the electricity and
15:09
power is available to power them on. So what do you end up with?
15:13
GPUs collecting dust, burning lots of money, and the whole slowdown of the AI market right now.
15:18
And so you could argue that the drawdown in the market was because too many
15:23
investors were excited or anticipated a much quicker build-out.
15:27
And to be honest, if you listen to the news, you'd be led to believe that exact
15:30
hypothesis because OpenAI would Stargate, Anthropic with all their data center
15:35
efforts, and SpaceX, Google, and all the likes implied that their data centers
15:38
would be set up like that.
15:39
And in fact, that's what they were promised. But what we're actually seeing
15:42
materializing is a slowdown of effort, and Chamath and Elon are simply calling that out.
15:47
All right, so is Elon right or wrong, or where is he right, and where is he
15:50
wrong? Let's get into it on today's episode brought to you by Qualcomm.
15:57
So let's start off with where I think he's right, the debt window,
15:59
which basically states that there's not enough money out there to
16:03
give attractive enough rates of lending to these neoclads to be able to pay
16:07
off all their GPU and CapEx expenses so that they can fulfill the contracts
16:10
that they sign with Microsoft Meta and all the other likes.
16:13
I think there's some truth here simply because the interest rates for lending
16:18
or borrowing this money has skyrocketed over the last two months.
16:22
CallWeave is a prime example where it was sub 10% two months ago,
16:26
and now it's at 16%. And the next two months, it's probably going to increase even more.
16:30
If you are a neocloud, which had less bargaining power because you're a smaller
16:36
company, you don't have a reputation versus some of the bigger hyperscalers,
16:40
you're probably borrowing this money at very poor or bad, risky rates.
16:44
Now, of course, you can supplement this by charging more money on the back end,
16:49
which is what a lot of these neoclouds do.
16:51
But at the end of the day, if the rates become higher than what you're actually
16:54
charging your client, you end up losing money and you end up being in a worse deal.
16:59
Which leads me to the second point that I think Elon's right about,
17:01
which is there will be stranded compute from a lot of these different neoclouds.
17:06
I don't think it'll be all of them, but it'll certainly be some of them.
17:09
And I think two things are going to happen here.
17:12
Companies like SpaceX or other hyperscalers are going to suck up the compute
17:15
from the smaller fishes in this ocean, or some of the bigger neoclouts are going to merge.
17:20
It wouldn't surprise me if I saw a Nebius and CoreWeave merger, for example.
17:24
Now, I have to add on that Elon Musk, when he was quizzed about this,
17:27
said SpaceX has no interest in acquiring compute.
17:30
Why would that make sense? And I found that pretty hilarious because his long-term
17:34
vision, I guess, is to launch these GPUs out into space and harness the energy
17:39
of the sun at the lowest cost possible, which begs the question
17:42
for the 40 billion dollars that you just raised Elon why on earth aren't you
17:45
spending that just on focusing on that mission and getting the GPUs out into
17:49
space now I'm sure the answer that he would give you is,
17:51
we need to make enough money in the short term by selling compute on earth at
17:55
a higher premium to the likes of Anthropic and OpenAI so that we can fund the
17:59
expensive CapEx build out
18:01
of AI in space so that's the balance there now where I think he's wrong and
18:04
maybe talking his own book is obviously because he is officially a NeoCloud
18:09
competitor anyone who says SpaceX is primarily a space company
18:13
hasn't read the news recently because most of their money, I think it's something
18:17
along the lines of 40 to 70% of revenue that they generate.
18:21
Is now coming from their AI business, their last quarterly earnings.
18:23
I think it was about $13.7 billion or $14 billion around there.
18:28
And the point being is this is now more than Starlink, which was their major
18:32
revenue earner. So it's obvious that Elon wants to double down on this to keep the company alive.
18:37
The second thing is he just raised $40 billion.
18:40
Now, of course, you don't want your competitors to then go out and do the same
18:43
thing or to get the idea to go out and do the same thing. So he's trying to
18:46
shut that down immediately.
18:48
And the third thing is his accomplice, Chamath, is an investor in SpaceX.
18:52
And listen, I love both of these guys. I love their takes. But Chamath is definitely
18:55
talking his book when it comes to power, energy, and land.
18:59
That being said, I think land power shell is the biggest bottleneck that AI
19:03
faces. And to be honest, maybe I should just make an episode explaining my thoughts
19:06
here because I think it's important.
19:08
But the idea here is this whole AI CapEx buildout is going to slow down because
19:14
we don't have enough physical atoms to move to set up these data centers.
19:17
We are constrained by regulations, power, and just not having enough energy
19:22
to deliver to the GPUs at the right time. The chips are coming first,
19:26
the energy is coming later. We are at an undeniable delay.
19:30
And that's not necessarily a bad thing because it stops the bubble from inflating.
19:34
But people just need to be careful about how much money they borrow and what
19:38
interest rates they do so.
19:39
And finally, to maybe end on a positive note, I put out an episode yesterday
19:44
talking about the Anthropic S1. And one of the main things that I called out
19:47
is I was surprised at how lean the revenue is that they're making.
19:51
One, that they're making so much. But two, the...
19:54
Average that they're making per dollar spend on compute has actually turned
19:59
positive. For every dollar of compute that they spend or invest,
20:02
they make around $2.08. And that's the first flip that we've seen a step towards
20:07
them becoming profitable.
20:08
If this is the case for the likes of OpenAI and other AI companies out there
20:12
that are spending so much money on this compute that Elon Musk and all these
20:16
neoclads are selling, that might actually end up being a bull case, optimistically.
20:20
That might reverse this entire trend and actually justify a lot of the spend that we'll see.
20:25
At the end of the day, I don't know, but I hope this is a balanced enough take
20:28
to help you guys understand what's good and what's bad happening in the markets right now.
20:33
And it's important to point out that it's not just the Western market.
20:36
The Chinese market, as of this week, is down 40% from its June highs. Catastrophic.
20:43
Other parts of Asia, the Indonesian stock index down terribly.
20:46
Now, you might be thinking, well, our markets aren't relevant to them.
20:49
We are, because a lot of their markets are propped up by the AI bubble,
20:52
by the valuations that we're seeing with the Anthropic IPO and Open Air's rumored IPO.
20:57
So it's all part of the same cohesive thing.
21:00
And AI is the biggest driving factor right now, which is why watching three
21:05
individuals tweet out and tank the market is an interesting thing to see and
21:09
I guess talk about. So I hope you enjoyed this episode.
21:12
If you're watching this on YouTube and you aren't subscribed,
21:15
what are you doing? Please subscribe, turn on notifications,
21:17
leave me a comment. It helps us out massively.
21:20
Also, share it with a friend. If you're listening to Spotify or Apple Music,
21:23
if you haven't given us a rating, please do so. Leave us a comment.
21:25
I love reading your comments. I love responding to them. And also,
21:28
if you have any ideas, thoughts, or perspectives, even if you disagree with me, let me know.
21:32
I'm always trying to find the holes in my thinking. Thank you for listening
21:35
to today's episode brought to you by Qualcomm.