YT YouTube2Text

YouTube2Text → Transcripts → CoreWeave Crash, AI Trade Cracks, Anthropic IPO Red Flags, Bill Gates Doomer

CoreWeave Crash, AI Trade Cracks, Anthropic IPO Red Flags, Bill Gates Doomer — Transcript

by All-In Podcast · 17,352 words · 2,576 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00All right, everybody. Welcome back to
0:01the number one podcast in the world.
0:04Actually, you manifested. We actually
0:06were number one. How good is that?
0:07>> We were literally number one last week.
0:09Only took uh
0:11>> only took a little bit of time to get
0:12there.
0:12>> Only took five years.
0:14>> Only took five. Yeah. Five years went by
0:17so quickly. Uh with us today,
0:20David Saxs is back. How are you doing,
0:22brother?
0:23>> Good.
0:23>> You good?
0:24>> Yeah.
0:24>> It's good being number one.
0:26>> Yeah, it's good to be number one. I
0:28agree. Yeah,
0:29>> I agree.
0:30>> I've done an excellent job.
0:31>> You've done a very good job. And now
0:35>> beautiful.
0:36>> Now we'll see who takes the most credit
0:38for the ascension of the All-In podcast.
0:41Uh over to you, Shamath Poly Hapatia.
0:43How's your week been?
0:44>> It's great being the number one
0:46position. Yeah,
0:47>> it's great being the best. I mean, we
0:49would have
0:49>> I feel like Sax and I would have gotten
0:51here at least 18 months sooner had we
0:52not had to drag you on our back.
0:54>> Oh, okay. the point guard who actually
0:57put all those dimes and positioned the
0:59ball everywhere you guys wanted it. Yes,
1:02you would have you certainly would have
1:03gotten there uh eventually if you guys
1:05actually knew how to start.
1:06>> Where would we be without your constant
1:08interruptions?
1:09>> Sooner. We would have been number one
1:10sooner.
1:11>> You would you would be where every other
1:13tech podcast is in the 400s. And joining
1:17us this week
1:19in replacement of venture
1:20>> also number one also number one
1:22>> number one venture capital firm in
1:24assets under management I believe
1:26certainly
1:28>> none other sir
1:29>> then Ben Harowitz is here. Hey Ben how
1:31are you?
1:31>> I am doing well. Thank you for having me
1:33on the show. Uh it's great to be with
1:35the number one team.
1:37>> Yes the number one podcast in the world.
1:38How does it feel to be on the number one
1:41podcast in the world? Have you has it,
1:44you know, like have you soaked it in
1:46yet, Ben? Is it just
1:47>> It feels great.
1:48>> Yes. Yeah,
1:49>> it is awesome. I feel bigger already.
1:52>> It is kind of amazing that a tech
1:54podcast could be the number one podcast
1:57on
1:57>> That is bizarre.
1:58>> Apple podcast last weekend.
2:00>> Big docket this week. Uh good to have
2:01you here, Ben. And uh Freeberg is busy
2:04saving the world from famines and he'll
2:07be back next week for all of you.
2:09>> He retired. He retired at number one.
2:12Yeah, he just he he walked off the
2:14court.
2:15>> Oh, that's hilarious. Well, um
2:18>> we've had an eventful week. We uh may
2:21have actually had the AI bubble, if
2:26there is one, pop or deflate a little
2:28bit. The Neoclouds got crushed this week
2:31after uh Coree's bonds started to sell
2:34off. Could be the canary in the coal
2:37mine. We'll find out. uh Cororeweave,
2:39Iron, Nebus, these are all the biggest
2:41Neoclouds. They're all down about 20% as
2:44publicly traded companies. Here's your
2:46chart. Here's what happened. Just a
2:48little bit of background. On Tuesday, a
2:50bunch of people on X, including at
2:52Chimath pointed out that the yield on
2:55some of Coral Weave's debt had hit 15%.
3:00The yields uh are obviously rising
3:02because bond traders are worried about
3:04Cororeweave's ability to generate enough
3:06cash to support its debt. So they want a
3:09a higher rate of return. Obviously
3:12adding to Chimat's little tweet was a
3:15friend of the pod Elon Musk who decided
3:18to contribute
3:21five letters, one word that has since
3:25been heard around the world and is at
3:28the start of every story. He said,
3:29"Yikes, y I ke." And that generated a
3:32bunch of headlines. Then
3:36everything tanked yesterday, Wednesday,
3:39we tape on Thursdays, as many of you
3:41know. Coreweave, Iron, Navius down
3:44between 10 and 15%. Even Oracle uh and
3:46Dell took some strays. Nvidia uh took
3:49stray as well. They were down three%.
3:52Coreweave is the tip of the spear.
3:54They've raised over $35 billion. We had
3:58uh the founder, we had the founder
4:01Michael in Trader on the All-In podcast
4:04back in March in which he explained to
4:07me on the All-In interview show what he
4:09called the box. Let's roll that tape and
4:12get Shamant's thoughts.
4:13>> How does all that debt work? That is
4:15something that you guys specialize in.
4:17>> I'm going to try to explain this in a
4:18way people can understand. So, we go out
4:20and we find a client. Let's use
4:22Microsoft. You brought them up before,
4:23right? And Microsoft comes to us and
4:25says, "We'd like to buy some config for
4:27you." and we say, "Okay, great. We're
4:28going to sign a contract." Once I have a
4:30contract in hand, then what I do is I
4:32create something. It's not a
4:34particularly creative name. It's called
4:35the box. Yeah.
4:36>> Right. And what I do with the box is I
4:38take my contract with Microsoft and I
4:40put it in the box. I go to Jensen and I
4:42buy the GPUs. I put it in the box. I
4:45take my data center contract, I put it
4:47in the box. And now the box governs cash
4:50flow and it has a waterfall of cash flow
4:52that comes into it and goes out of it.
4:54And so the way it works is then I build
4:56the compute and then I deliver the
4:59compute to Microsoft and they pay the
5:00box. They don't pay me. It goes into the
5:02box and the first thing it does is it
5:05pays the data center. It pays the power
5:07bill. It pays the interest and the
5:09principle and then whatever's left flows
5:12back to us. What's important to
5:14understand is the economics in this box
5:16are such that within two and a half
5:19years of a five-year deal, we have paid
5:23for everything. They don't
5:24cross-pollinate. They don't cause
5:26contagion across the boxes. They're all
5:28independent and discreet. As you do this
5:31and as you show the lenders how this
5:35financing tool and how this financing
5:37mechanism works, what they do is they
5:40continue to lend you money at
5:42progressively lower rates. Okay,
5:44Chimoth, that seemed like a completely
5:47reasonable explanation of how they were
5:50running the business and that each of
5:51the boxes were isolated, but something
5:54has changed in the last 6 months. what
5:56is that and uh what's changed here and
6:00and your insights into his financial
6:04innovations I would say uh with corewave
6:07the most important thing that you should
6:09always look at is how the sharps the
6:12super smart money prices risk and when
6:16you're dealing with companies like
6:18Cororeweave or Nebus or Iran etc that
6:21have to borrow a lot to fund these boxes
6:24the smart money is trading what's called
6:27a credit derivative swap. Okay, this is
6:29default insurance
6:32and these default swaps are meant to
6:35ensure the bond holders against a
6:38potential default.
6:40That risk spiked this week to about 953
6:45basis points, which it doesn't matter
6:46what that number is, but as an imputed
6:49risk, it's about a 16% chance of a
6:53default in any given year.
6:55And over a 5-year duration,
6:58what that math implies is about a 55%
7:02chance of a default. So, it's a coin
7:03flip whether they default or not. That's
7:05not what I'm saying. That's what the
7:08insurance market was saying about their
7:10debt. And I think that's something to
7:13look into because those guys I think are
7:15the smartest money in in all of this.
7:18So, why did they do this? I started to
7:20dig into it and here's a couple of
7:22charts that explains a little bit of
7:24what I learned. Okay, so the first thing
7:28is just a market check. You know, we've
7:30talked about this, but we have a problem
7:32in the United States, which is that the
7:35US is moving the risk-free rate higher
7:38and higher. Who is doing that? These are
7:41the central bankers. These are the bond
7:43vigilantes. These are the people that
7:45say because of inflation or other
7:47long-term risks that are not getting put
7:49out of the system, we need to get paid a
7:52lot more money to own the safest
7:54riskless asset. As that goes up, basic
7:57way to think about that is as the
7:5910-year bond goes up, everything else
8:01goes up. And so this black line
8:04represents the average price that you
8:06have to pay for junk debt.
8:10And what you see is that as the 10-year
8:12went up and people are like, man, the
8:13risk-free rates going up, the junkiest
8:15junk bonds need to pay more and more and
8:17more.
8:19And then the purple line here is what
8:21coreweave has to pay. And what you see
8:23is that it also has started to tick up
8:25and has moved up relatively violently.
8:28And so the question is why has it done
8:30that? Okay, so let's go to the next
8:31chart. Michael is right that he is in
8:34the business of creating these boxes
8:37and then putting in contracts and GPUs,
8:41but he also needs to put money in those
8:43boxes. And the way that he puts the
8:46money in the box is he has to go out and
8:48borrow that money. So he goes to a
8:50lender and says, "Give me the money and
8:52I will pay you that back plus some
8:53interest." What you see here is what is
8:58inside of the box effectively. So the
9:00gray bar is how much money he's
9:02guaranteed to get into the box. These
9:05are all the deals he sold to all of his
9:08customers.
9:09>> These are contracts. Yeah,
9:11>> these are contracts. Now, and it's
9:12important to note because I'll get to
9:13this in a second. It turns out that
9:15these contracts are fixed price. So
9:17meaning if what you could sell in the
9:19spot market goes up, you don't really
9:22generate that revenue because you've
9:24already pre-sold it at a fixed number.
9:26So, Michael creates the box. He gets a
9:28long-term contract, which is great
9:30because you have shity around the
9:33revenues you're getting, but he sells
9:34that. Then he goes and has to finance
9:36it. And then some of that financing
9:39comes due. He's got some bonds that are
9:40due and some loans that are due. What
9:42this basically says is at the midpoint,
9:46he has to raise another $35 billion of
9:48debt to get all of the contracts that he
9:51signed financed, right? All the boxes
9:55financed. and up and running. So he has
9:56about 35 billion of debt. He needs to
9:59raise another 35 billion of debt. But as
10:01you saw in the chart before
10:04that's ticking up not and it has nothing
10:05to do with him. It is because the
10:07broader market is essentially saying the
10:09risk-free rate goes up everybody has to
10:11pay more because your spread above the
10:13risk-free rate has to track with it.
10:15Okay. So today
10:17across all the different pockets of
10:19money in which that they have used to
10:21fund those boxes
10:24their average
10:26interest rate that they pay is about
10:278.5%.
10:29The problem is that in this last little
10:32while the sharps
10:36basically said hold on this is getting
10:38too close to the knife's edge. You're
10:40going to have to pay more interest.
10:42That's what I tweeted, which was just,
10:43you know, one of the sleeves of
10:45interest. This is the GPU loan DDTL 1.0
10:49spiked to 15%. The problem is that could
10:52theoretically now be the rate that they
10:54have to pay for all of the future. And
10:56remember,
10:58half of it's financed, half of it is
10:59still unfinanced. Right now, these guys
11:01are in a pretty good place because they
11:03have about 780 billion of assets and
11:05only 35 of debt. So, they're well
11:07covered. So I think that you know this
11:09idea that they're at risk is not on the
11:11surface true but the reality is that
11:14they have this problem which is they
11:16have extreme customer concentration.
11:1975% of the revenue comes from three
11:21companies.
11:23Some of those companies are walking down
11:26what their revenue expectations are into
11:28the future.
11:30So you have this situation now where you
11:32have to go out and borrow at a much
11:34higher rate. You have a bunch of debt
11:36you need to go to fund the boxes and you
11:38have a fixed amount of revenue going
11:40into the boxes. So what is the trade or
11:42what is the setup? It's basically a coin
11:43flip now on the following idea. I think
11:46the market is saying that if
11:50they are able to finance this stuff at a
11:52reasonable rate and the life of the GPUs
11:55that they buy from Jensen are longer
11:57than what they think, they're going to
11:59be okay.
12:01But if the financing goes up and into
12:05that, you know, 10, 11, 12, 13, 14, 15%
12:08range, these guys are getting very,
12:11very, very close to not necessarily
12:14being able to cover all the obligations
12:15that they have. My team's estimates is
12:17that they're plus or minus about 5
12:19billion away from that break even point.
12:22So, I think that's why the CDS spiked,
12:25which then caused the bond yields to
12:27spike and caused the bonds themselves to
12:28lay off. So in order of operations,
12:30number one, I think they've created this
12:33very novel structure that everybody else
12:35has copied with the boxes. Number two,
12:37they've sold these fixed contracts.
12:39That's super smart, guaranteed a bunch
12:41of revenue into the future. But we have
12:43this situation where as the risk-free
12:45rate goes up, their cost of borrowing
12:47goes up. That erodess the margin that
12:49they have that then brings them close to
12:51this break even point. That causes the
12:53CDS to react, which then causes the bond
12:55yields to react. That's the negative
12:57reflexive loop that they're in. So, they
12:59just have to find a way to break out of
13:00that. I don't think that they're in real
13:02risk. But, as we all know, this is sort
13:05of a confidence game and you need to
13:07have confidence that the broader market
13:09can support another 35 billion just from
13:11these guys. And I think that's what's
13:12going to get figured out over the next
13:14month. All right, Ben. So, you look at
13:16this situation, you've been through a
13:17couple of bubbles before, maybe have a
13:20little PTSD or scar tissue, in fact. Um,
13:23and we saw carweave bounce today, kind
13:26of hit a bottom, stocks up 1 or 2% right
13:28now after going down 10%.
13:31>> Is this acute? And then what does it
13:32mean for the broader AI trade, the AI
13:36bubble? Is there a bubble? And and we'll
13:39get into anthropic in a minute. Saxs.
13:43>> Yeah. So, so I think that um the the
13:46first thing to understand is these these
13:48businesses like Cororeweave as uh you
13:50know Chimath kind of went through some
13:52of it but they're rub Goldberg financial
13:55machines. They're exceptionally complex.
13:58Um so you know we talked about okay GPU
14:01contracts and this but he's got data
14:03center contracts how long are those and
14:05and then there's he's got to buy memory
14:07like and there's memory shortage and
14:09this and and that so so it is really
14:12complex. The other thing um or the other
14:15interesting thing is yes GPUs
14:18not only have they they've been going up
14:20in duration and in price. So that's been
14:23like a big tailwind for him uh as the
14:26owner of such things.
14:27>> When you say duration, you mean the
14:28lifespan, the number of years you get
14:31value from that. It was three, four,
14:33five, now it's six, seven, eight. And
14:36there's a shortage um like a real
14:39shortage of GPUs with power and cooling
14:42and and I kind of caveat that right like
14:44which is what he has.
14:46>> Uh and and this this this shortage is so
14:49bad you know I was speaking with Andy
14:51Jasse who runs Amazon yesterday and I
14:54said yeah like we're out of everything
14:56you know till 2028. He's like, "Oh, it's
14:59beyond 2028 bit. Like, I don't think we
15:02get stuff back." Like, end of 2028,
15:04maybe. Uh, so everything else is sold
15:07out. So, he he has the advantage of that
15:10he's got supply in a great shortage,
15:13which is extremely different than the
15:15internet bubble. The internet bubble, we
15:18didn't have enough demand because we had
15:20to build the internet. People had to get
15:22broadband. You know, you needed uh all
15:25this stuff. people had to buy, you know,
15:27originally they had to get their browser
15:28app fries, all this kind of thing. There
15:30weren't enough consumers on the internet
15:32to absorb all the bandwidth that had
15:34gotten built out and they had a massive
15:37amount of debt, not unlike uh we're
15:40building today, although not as big. Um
15:43but but the lack of demand killed it. We
15:46don't really have a lack of demand in
15:48AI. So I think the dynamics will be
15:51different. Uh I do think like the for
15:54sure if you get into a pathological debt
15:56situation
15:57that can wreck anybody. Uh it's why um
16:01by the way we're not invested in any of
16:03these uh neo clouds either for that
16:05reason because that whole thing scares
16:08me having been through it. But but I
16:11would just say it's pretty complicated
16:13uh and it really depends on a huge
16:16number of assumptions in this very
16:19elaborate model that they have. Ben, but
16:21let's just look. You were you and Mark
16:24were one of the pioneers, maybe the
16:26pioneer of the Neocloud business. I
16:29mean, you you may just tell this quick
16:32story because I want to ask the
16:33question.
16:34>> Loud Cloud kind of popularized this idea
16:38that that you could build these clouds
16:40and service customers this way.
16:42>> Yeah. But it's very interesting to me
16:44that you and Mark as the
16:47kind of primogenitor
16:48uh have passed on the whole space. So I
16:51I I want to know why.
16:53>> Well, it's because of this. I mean, so
16:56so so it can it can end up being um like
17:00a very dangerous arbitrage business
17:03where you know you can get upside down
17:05easily. And so if you look at loudcloud
17:08just as an example just here here's a
17:11simple thing that happened to us in 1999
17:14bandwidth bandwidth was $1,500 a megabit
17:18and you couldn't get it was very hard to
17:20get it and so you had to get a long-term
17:22contract three years by 2001 it was $50
17:26a megabit so if you were holding
17:28bandwidth you you were you just you know
17:31and you borrowed money to get it you're
17:32dead. So that
17:35you know that kind of thing is so out of
17:37the control of the entrepreneur and and
17:39so different than what we like to fund
17:42um that you know it just it scared it
17:45scared me to death. Now look, I I I
17:48think the core dynamic of AI has been
17:50working for Coreweave though in terms of
17:53that in that
17:55there's a massive shortage, not an
17:57overage of supply. Um so those prices
18:00don't seem to, you know, token prices
18:02have gone down, but GPU prices have not.
18:04>> Uh and so things are going up, you know,
18:06and and by the way, data center space is
18:09going up in price because even labor to
18:11build data centers is hard. The politics
18:14are making it hard to build data
18:15centers. So the whole supply chain price
18:17is going up and demand is not subsiding
18:19and and so those those are very big
18:22tailwinds uh for coreweave despite the
18:25fact that the debt markets kind of
18:28independent of them are tough.
18:30>> All right Sachs uh here's your poly
18:32market. Coreweave announces bankruptcy.
18:34That's a pretty aggressive uh poly
18:36market. 15% chance at the end of 27 next
18:40year. 23% at the end of 2028. seems a
18:44little far-fetched given the fact that
18:46they have a massive demand for their
18:51equipment as Ben pointed out.
18:52>> Well, you can't look, this is dumb
18:54because like I think Michael explained
18:55it like you have this opco propco setup.
18:58So coreweave can't go bankrupt. An asset
19:01level vehicle for a data center or a
19:03buildout could theoretically go upside
19:05down, but that'll go into receiverhip
19:06and then just manage where you sell down
19:08the GPUs. Yes, he he explained in that
19:11clip he has each box independent of the
19:13next. So, not a great poly market, I
19:15think. Uh but well, it is great if you
19:18bet it the other way.
19:19>> Yeah, exactly. Maybe some free money
19:21there to pick up. Sax, let's level up
19:23the conversation here. Just outside of
19:25Corewave and the Neo Clouds, massive
19:28demand, token prices coming down, but
19:31token consumption going way up. And I
19:33think one of the the um silver linings
19:36here is it was a big question of how
19:38long these GPUs would last. Maybe they
19:40had three, four, five year lifespan. Now
19:43people are saying, "Hey, these things
19:44will keep going into year 6, 7, 8 for
19:46inference alone." So what's your thought
19:48here on the AI trade, the bubble as it
19:52were? Do we need to take a little bit
19:53out? Do we need to stress test it? And
19:55is that what's going on here? Or is this
19:59uh making you nervous? Hey, we're in
20:02year four or five of this bull run and
20:04that maybe it's coming to an end and we
20:06may hit the rword.
20:09Yeah. Recession.
20:11>> Here's the question, Jay. Cal, is you
20:13got this coreweave crash. Their bonds
20:16are being marked down to 80%, it's down
20:1815 to 20% in the last few days in the
20:21stock market. Is this just market
20:23volatility or is this the beginning of
20:26something deeper and worse? Like you
20:27said, the popping of a bubble. Okay,
20:29that's the fundamental question. I would
20:31make the argument that this is just
20:33volatility, which you're going to get a
20:35lot of. Whenever you're underwriting to
20:38multi-trillion dollar outcomes in the
20:40future based on very rapid rates of
20:42growth, exponential rates of growth, the
20:45market's going to be super sensitive to
20:46any data points we have in the present.
20:49And so everything's going to be
20:50exaggerated and that's why you get these
20:52big swings. Now, again, that to me is
20:55the explanation of the volatility. I
20:57don't think that something fundamental
20:59has changed. So let's talk about the
21:01fundamentals. So first question there to
21:03unpack it is has the demand changed? No,
21:06we know that the demand for compute just
21:08keeps growing by leaps and bounds. In
21:10fact, everyone is saying that they are
21:13compute constrained. I mean you got the
21:14biggest model companies are saying
21:16they're computed. Startups are saying
21:18it. I mean they're asking their VCs, hey
21:20can you help us get compute? I think
21:22you've got some VCs even trying to buy
21:24compute as basically a strategy for
21:26competitive differentiation so they can
21:28help their startup. So
21:29>> there was literally an accelerator that
21:31had bought excess compute and they were
21:33using it as a way to get people to apply
21:35for their accelerator.
21:36>> Yeah. So we know the demand is off the
21:38charts right now. We're seeing other
21:39indicators. So for example, old GPUs
21:42like even the H100s or the A100s which
21:45are several generations old, they're
21:47still renting at premium prices right
21:49now. They were supposed to be fully
21:50depreciated and they're not. There's
21:52actually upside surprise there to the
21:54people to the companies that are renting
21:56those old supposedly stale GPUs. So, we
22:00know the demand is off the charts. Okay.
22:02So, what's the next indicator of
22:04something deeper and problematic
22:07happening? Well, the next question you
22:08could ask, well, is there a glut of
22:10supply coming onto the market? So, we
22:12know the demand is huge, but maybe
22:15there's too much supply being stood up.
22:17Well, we know that's not the case
22:18because of the sheer difficulty of
22:21standing up this powered compute and we
22:24know it's not being brought onto the
22:25market too rapidly. First, you just have
22:27all the problems with atoms building
22:28things in the physical world. That's
22:30hard enough. Elon basically has said,
22:32"Look, this is a really difficult
22:33problem. Even our rocket scientists find
22:35it challenging to stand up gigawatts of
22:38new powered compute." And then we have
22:39all the regulatory headwinds, right?
22:41Local communities are not racing to give
22:44out new entitlements around this. So if
22:48anything we believe that there's not a
22:51glut it's the other way that there's
22:52enforced scarcity. So then the question
22:55becomes well is there some sort of
22:58execution problem with coreweave? You
23:01know are they too slow in standing up
23:03their compute? Have their costs gone
23:04through the roof? Look, there's some
23:06indication that maybe their costs are a
23:08little bit higher, that maybe it's a
23:09little slower, but at least on a
23:11reputation basis, they are believed to
23:13be excellent at this game of standing up
23:16new compute. So then the question was,
23:18is there some sort of credit problem?
23:21you know, on the credit like the founder
23:24Michael was saying with the boxes,
23:26you know, when you go to a lender to
23:28basically back the buildout of, let's
23:31say, a new data center, you're not
23:33relying on the unsecured credit of
23:37Coreweave to make that loan. You're
23:39backed by the credit of the tenant who's
23:42entered into a 5-year commitment. So, it
23:45could be a Microsoft, whatever. So the
23:47creditworthiness of that particular
23:49project of that particular box is going
23:51to depend more on the credit of the
23:54tenant than on the credit of core. And
23:57then behind that you've got the GPUs
23:59themselves which you know it's basically
24:02asset back financing. The lender is
24:04secured by those GPUs. And we just
24:06talked about how if anything there's
24:08upside surprise on the depreciation
24:10cycle of these GPUs. So my impression is
24:13that there's not a deeper credit problem
24:15in terms of financing these projects.
24:17What happened though is the repricing is
24:19of Corweave's unsecured obligations.
24:22It's true that the market has discounted
24:25them by 20%. They're in distress
24:27territory. That has brought the implied
24:30coupon of those bonds from 9.6% to
24:33around 15%. But that does not increase
24:35the cost that Cororee is paying. Corwe
24:37is still paying their 9.6% coupon. It's
24:40just that the secondary market has
24:43repriced the value of those bonds to an
24:44implied coupon.
24:46>> It changes the future debt.
24:47>> Yes, it could definitely raise the costs
24:49of future debt, but I don't think
24:51they're funding these new boxes with
24:54unsecured corporate notes. I think
24:56they're funding them with asset back
24:58financing. And
24:59>> yes,
25:00>> the debt
25:00>> there's two things that go in the box.
25:02The contract goes in the box, the GPUs
25:04go in the box. So, you got two things
25:05and these are credit credit worthy
25:07organizations who pay their bills like
25:09Microsoft.
25:09>> Yes. Now, if I were a lender on those
25:11projects though, I'd be looking really
25:13closely at the credit of who's the
25:14customer basically for that compute.
25:16Obviously, if it's Microsoft or Google
25:19or a couple other companies that feel
25:20really good,
25:21>> right, that may a startup, I don't
25:23necessarily.
25:24>> They have three customers that represent
25:2675% of their revenue.
25:27>> Yeah. And they're all investment grade
25:29like the big customers. So,
25:31>> the thing I'll add to this is, you know,
25:33you explained quite clearly, uh, Saxs,
25:36hey, the the GPUs are worth something,
25:38the data centers are worth something.
25:39Ben, you pointed out the power, the HVAC
25:42is worth something. The tokens are worth
25:45something. And then if you look at
25:46startups, uh, you know, which is where I
25:48spend all my time is the early stage
25:50startups, they're crushing it in terms
25:52of raising capital, getting customers,
25:54and we've never seen revenue ramps like
25:57I'm seeing in startups today. And I
25:58think Ben, you'll probably back that up.
26:00the the revenue ramp of startups today
26:03is distinctly different than the SAS,
26:06you know, uh triple triple triple double
26:09double double. Now we're at like, hey,
26:11are you doubling every quarter or not?
26:14And then if you're not doubling every
26:15quarter, like what are you doing? Why
26:17are you not growing faster in this age
26:19of AI? So the startups feel particularly
26:20strong. So maybe this is just the market
26:23overall saying hey let's test this you
26:26know uh and and maybe there's some
26:29indigestion there or Ben is it that the
26:33world is running out of money and people
26:35have place their bets already and then
26:37what would that say for the venture
26:40industrial complex of which you are you
26:43know a significant player at scale and
26:46what would that say for the IPOs the two
26:49biggest in the history of
26:53uh business and and public markets,
26:55anthropic and open AI, are we running
26:57out of money to allocate to AI?
27:01I don't think so. Um, you know, we've
27:04kind of never had companies that have
27:06grown this fast, gotten this big.
27:09And so, we've never asked that question
27:12before. We never said, or at least not
27:14in my lifetime, does the world have
27:16enough money to finance these things? uh
27:18which is a crazy question but like
27:20relevant but I I I don't think for sure
27:25with respect to the anthropic and the
27:27open AI
27:29IPOs I think my belief is those are
27:32going to be fine like like look the the
27:34price
27:36you know will it be two trillion or will
27:38it be 1.75 trillion I don't know but I I
27:43think
27:43>> 50 billion or 100red billion whatever
27:45they raise that's not a giant that's
27:47that's big number, but it's not an
27:49unprecedented number now because look,
27:52>> SpaceX just did it, right?
27:53>> Whatever you're holding in the equity
27:55markets, whatever you're holding,
27:59>> open AI and anthropic are threats to
28:01that
28:02>> almost like there's almost nothing that
28:04you can hold in a portfolio that those
28:06two companies don't threaten.
28:08>> Yeah, agreed.
28:08>> And so I'm not worried at all about the
28:11demand for those IPOs. So you have to
28:13own those because they threaten every
28:16single company in the world because
28:18they're selling intelligence broadly and
28:20every company is based Shimoth on
28:23intelligence.
28:24>> Therefore, it's it's that disruptive.
28:26Yeah, Chimath, you can underwrite it as
28:28a defensive hedge because I think Ben's
28:30right. It's like if you own a healthcare
28:32company and you've thrived on the
28:35opacity of the healthcare
28:36infrastructure, OpenAI is a threat to
28:38that. If you own a bunch of workflows
28:42that have been rented to people as SAS,
28:44theoretically open AAI and ant are a
28:47threat to that. So at a minimum it is
28:50sort of defensively positioning your
28:52portfolio. So you kind of have to
28:54because there is not much other
28:56investable stuff in AI to be long AI
29:00other than those two companies and so
29:01they'll get the lion share of the of the
29:04interest. I think the the question
29:06generally is if the bond market walks
29:10the tenure up to five and a half and 6%
29:13which has nothing to do with AI. It has
29:15everything to do with debt to GDP and a
29:17bunch of other factors in broad-based
29:19western economies. Any business model
29:21that thrives on a financing spread
29:25gets cornhold any and every all of them.
29:29It just it will it will not matter when
29:31the 10 years at 6%. It won't matter.
29:33Doesn't matter what you do. it doesn't
29:34matter what you claim to do, you're
29:35going to get run over. You're going to
29:37be roadkill. As long as we can keep that
29:39from, you know, happening,
29:43we're going to be fine. So, like, what's
29:44the what's the most important way, I
29:46guess, post these midterms to keep the
29:47party going? You have to really start to
29:49focus on deficits. And you have to find
29:53a fair trade where the bond market backs
29:56off.
29:58And I think you can see it in the tape.
30:00Meaning if if we start proposing huge
30:03spending programs, they're going to walk
30:04the tenure up. That then has a negative
30:07cascade to everything. If however the
30:09posture becomes much more about keeping
30:11everything calm and like let's just tone
30:13things down and let's find the waste in
30:15the system and squeeze it all out, I
30:17think people will respond very
30:18favorably. So that's really where it's
30:21going to come down. The last thing I
30:22will say is
30:24these are like Achilles heels. Meaning
30:26like I think Sax made an incredibly
30:28eloquent case. So did Ben about all of
30:31the asset value that's getting built up.
30:34But just like the great warriors, you
30:36just slice the Achilles tendon and they
30:38just crumble to the knees. I just want
30:40to point out this is that thing.
30:44>> There is no other thing. Meaning all
30:47this other stuff is complete goodness.
30:49But the fact that you need to finance
30:50these things in a variable rate model as
30:53a spread to the risk-free rate is the
30:55big problem in these businesses.
30:58And you need to essentially have
31:00Microsoft, Google, Meta, and Nvidia
31:04backs stop all of this in a worst case
31:06scenario
31:08because that's the only way you can
31:10withstand
31:12a one or two-year correction in this
31:14market. I'll tell you the other big
31:15piece here that might be a bigger
31:17threat, Sachs, is the deflationary
31:19nature of this AI, this super
31:22intelligence rather movement. We talked
31:24last week about people decompiling
31:27Photoshop going to zero. Now this week,
31:31Hood Warfare came out. It's basically
31:33people are making platform level games
31:36for five bucks. It's so deflationary.
31:40There's an 11 Labs competitor that came
31:42out that's incredibly cheap. and then
31:44every single open- source model being
31:47hosted and Elon saying, "Hey, we'll just
31:49route you with our orchestration level
31:51to the best model, open- source or not."
31:56How how big of a how big of a headwind
31:59do you think that is, Sachs, to our
32:01industry, just the commodification and
32:04how rapidly things are just having 90%
32:08of the VA value removed from them? And
32:10is that going to stop or is it going to
32:12accelerate? It feels to me like the
32:14deflationary nature of this software AI
32:17super intelligence is unique and
32:19different than anything we've ever seen.
32:21People are just able to rip and ship
32:23code so fast. It makes me wonder like
32:27what is the margin of software in the
32:29future?
32:29>> Well, we want everyone to be able to
32:31create software, especially bespoke
32:33software, which is really this new
32:34category, right? That is going to be a
32:37huge driver of productivity throughout
32:38the economy. I mean it's not all just
32:40like uh zero sum games or disruption.
32:43The value of these coding assistants is
32:46that now you can create software for
32:49highly bespoke use cases. Whereas
32:52software used to be created really just
32:54for the mass market. I mean the reason
32:55why Photoshop's a successful product is
32:58there's a lot of graphic designers who
32:59need to use that product and really the
33:02only software that could ever get
33:03created in the past was when there was a
33:05large potential customer base for that.
33:07Now we're getting all these cases where
33:10enterprises can create software just for
33:12their particular use cases. And in fact,
33:14in a way, that's what agents are is that
33:16they're bespoke software programs to
33:19solve a particular task. But if people
33:21build those in-house sacks and they're
33:24based on open-source, you don't need the
33:26other piece of software that it's
33:29replacing. So isn't that extremely
33:32deflationary? Maybe there's some
33:34substitution away from established
33:36software. But I think the important
33:38point is just there's this vast new
33:39category of software being created. I
33:42mean the amount of code that's being
33:44created now is orders of magnitude
33:46already bigger than it was before coding
33:49assistance. So again people just tend to
33:51think in terms of these zero someum
33:52games and who's going to be wiped out or
33:54eliminated. But the really important
33:55trend here is just that there's going to
33:57be a lot more software created in the
33:59world for a lot more of these bespoke
34:01use cases. Enterprises are now able to
34:04create software for a much broader array
34:06of purposes. There's companies that
34:08never would have been software companies
34:09before that are now creating their own
34:11software and they're hiring coders
34:13basically to manage the process. By the
34:15way, I think this is a really important
34:16point around remember with the whole
34:18jobs apocalypse. The first job category
34:20that predictors of the apocalypse were
34:23saying would be eliminated were software
34:24coders obviously because
34:26>> everybody became a developer and now
34:28everybody's shipping code. Yeah. Now
34:30we're seeing that actually demand for
34:31coders is increasing even though 95% of
34:34the code's written by coding assistants
34:36and the reason for that again is well
34:38first of all you still have that 5% and
34:40that 5% is of a much bigger pie and you
34:43still need someone who can look under
34:44the hood and manage the whole process
34:45and they need to be somewhat technical
34:47and maintain it. maint
34:51I I think this whole you could say new
34:53category of bespoke software created by
34:57the proliferation of coding assistants
34:59and agents and all of that is just I
35:02think it's going to be just a huge
35:03driver of productivity throughout the
35:05economy and is that broadly deflationary
35:07probably in a good way because it means
35:10that we can do more with less and we
35:12need to
35:12>> be great for the country right that
35:14everybody is going to be more productive
35:16Ben but do you worry about our industry,
35:18the venture industry, uh, and we have
35:22these great hits and then the second
35:24they become a hit, somebody can
35:26decompile it, create a competitor and
35:29ship it very quickly. It feels
35:31distinctly different in this uh, you
35:35know, um, boom to previous ones. So, do
35:40you think this is going to be a
35:41challenge for us to get DPI for our LPS?
35:44Um, is this going to be a challenge for
35:46the sustainability of great software fr
35:49franchises?
35:54>> Ben,
35:54>> I I can Yeah. Look, I I don't I don't
35:57think so. Like this this has been the
35:58ongoing prediction starting with the SAS
36:01apocalypse.
36:02>> Yes.
36:02>> Um, but I think that David David's right
36:05is that what's happening is
36:07>> prices software prices will definitely
36:10go down. Um but the amount of software
36:12is going way way up. It's the it's the
36:15Jevans paradox of software. Sure.
36:17>> Uh and you know these companies end up
36:21doing other things more than just the
36:24software. So we had this on the
36:26internet. The internet information is
36:27going to be free. That's it. All these
36:29industries are dead. But they kind of
36:32reemerged in a different way. Um and I
36:36think that that's very likely what's
36:39going to happen. So from a startup's
36:40point of view, you've got to evolve. Um
36:43and so you know 11 Labs does many many
36:48things to make their product better over
36:50the years since uh that's example that
36:52you brought up and they've got like
36:53licensing agreement like Bert Reynolds
36:55will read you a book or this and that
36:57and the third. Um so they're doing more
36:59than just software um because they can
37:02because the software is easier to write
37:04the model they're training on very very
37:05very nuanced cases which is why you know
37:08movie dubbing is so amazing now where it
37:11sounds like they're just acting in a
37:13different language. Um so there's more
37:16to it than just you know speech to text
37:18or this and that. um and you know those
37:21so I think the products are going to
37:23have to get way better and the startup's
37:24products will have to get way better
37:26which is amazing for everybody
37:29and then that's where the value will be
37:30but yeah just like writing software is
37:35not going to have any more value than
37:37like publishing a magazine creating
37:39massive competition yeah
37:41>> the most will change you know you got to
37:43think about what are the moes in a
37:44business data moes are a big one if you
37:46have data that other people don't have
37:48>> if you then build a user base because
37:51your model is catering to a particular
37:53vertical and now you're getting all
37:55those users and you're getting their
37:56feedback. Maybe that becomes a you know
37:59a feedback loop the feedback learning
38:01becomes a
38:02>> lot of
38:03>> startups are building businesses based
38:05on that model. The paradigmatic example
38:07is is like a Harvey or something where
38:09they get all these law firms on board.
38:10So they then get access to their data
38:12and then their usage and then they train
38:14their own models.
38:15>> Yes. Get sharper each time. Go ahead.
38:17and cheaper and more custom.
38:19>> Yeah,
38:19>> I'll give you the ultimate moat. I think
38:21power will be the moat. I think that you
38:24can build the most fantastical things in
38:25the world and you're going to see people
38:27on the dry docks basically trying to
38:29hitchhike to get power. And the reason
38:32is because we are so constipated in the
38:34United States. A we're already behind
38:36and we're in a massive energy deficit,
38:38an electron deficit. We are by 2050
38:41about 1.4 terowatt hours short, which
38:44means about six California's worth of
38:45electricity. That's their starting
38:46point. Then when you layer in, and
38:48that's with everything that we've
38:49planned to build being built on time and
38:52on budget, which we know is not going to
38:54happen. Then on top of that, you have
38:55all the load growth that theoretically
38:57happens when you actually have all these
38:59next abundant experiences.
39:01I made this uh statement this past week,
39:04and I'm just going to double down on it.
39:06I don't think you can work through the
39:08hyperscalers anymore because of this one
39:11issue. So let's just say you're a big
39:13mainline, you know, G Sid bank, right?
39:16JP Morgan, City Bank. Could you imagine
39:19that the reason why you can't service
39:21your customers is because the underlying
39:23cloud that you use can't actually secure
39:25and grow enough power? And so back to
39:27the other moat that I think is really
39:29valuable is to own the land, the power
39:31in the shell because it's such a
39:34constrained asset. Back to what Andy
39:36told you, Ben, it is so critical and it
39:40will stop the best experiences on a
39:41dime. So if you're a startup, I actually
39:43think the bigger problem is you'll go to
39:45these clouds, you'll be like, great, I'm
39:46ready to launch this Whisbang product
39:47and they're like, "Suck it. We have no
39:49capacity for you."
39:50>> Yeah. Such a disaster. I very much agree
39:52with that last one.
39:54>> I think that's right.
39:55>> By the way, what an incredible
39:56opportunity that's now been lost in the
39:58Middle East where Saudi, UAE,
40:02>> Qatar, they were all massive producers
40:04of oil, able to do top down planning,
40:07half centers, and now they have bombs
40:10going off blowing up airports. It's such
40:11a
40:12>> Without drone defense, you can't go
40:14there. Without drone defense, you can't
40:15go there.
40:15>> Exactly. And drone defense isn't going
40:17to work because all one drone gets
40:18through and your data center is fried.
40:20Well, I mean,
40:23>> can I make one last point on the whole
40:24data center thing,
40:25>> please?
40:26>> So, there's one metric that the more I
40:27think about it, the the one metric I
40:29would really keep an eye on is payback
40:30period. To me, that's like the Northstar
40:33metric here. So, Cororeweave says that
40:35it's payback period is 2.5 years. Their
40:38typical deal is 5 years. They pay back
40:41within 2.5 years. SpaceX, they pay back
40:44in roughly one year. That makes sense
40:46because I think that SpaceX is
40:49monetizing at about 50 billion per
40:50gigawatt and core at about 20 billion
40:53per gigawatt. And the reason for that is
40:54because SpaceX is basically getting
40:56spot. You know, they're doing these
40:58deals that can be cancelceled by either
40:59side within 90 days. So, they're
41:01basically getting spot minus 90. Whereas
41:04Coreweave is signing up for 5-year
41:05commitments. So, it's, you know,
41:06contracted way in advance and they're
41:08kind of getting this contracted price.
41:10And what's happened, I think, over the
41:12past year is that spot has moved way up.
41:15way up
41:16>> because again of the scarcity, right?
41:18And the demand is so off the charts. So
41:20in any event, payback period is really
41:22critical and as long as the payback
41:24periods are really short, I don't see
41:26how these companies can lose or at least
41:30you'd have to be really incompetent to
41:31screw it up, right? like you're making
41:34these big capital investments, but if
41:35you're getting your money back in one
41:36year or two and a half years,
41:39>> you know, that greatly reduces
41:40>> like owning an Airbnb in Paris or
41:42something like
41:45the payback period is so quick for these
41:47if you if you got your orders in
41:49>> early.
41:50>> Yeah. And that means you're paying back
41:51the debt quickly and so the debt's
41:53rolling over more quickly. It just
41:55reduces everyone's risk. Now, if that
41:56payback period starts taking longer and
41:58longer, then I get nervous, right?
42:00Because that magnifies the risk. It also
42:01means the economics aren't as good. And
42:04the question is, well, why would the
42:05payback period take longer? And it could
42:06happen for a bunch of reasons, right?
42:08Like if the demand goes down and so now
42:10the offtake's not as good, the spot
42:12price going down
42:13>> or just there's execution problems,
42:15right? It takes you longer to get the
42:16capacity online. So, you know, the
42:18payback gets delayed. There's a lot of
42:20reasons why payback could take longer
42:22and longer. But as long as payback
42:24remains as short as it is today, to me,
42:26that that would be like the one metric I
42:28would really keep an eye on. The other
42:29thing that the big offtakers will make
42:31you sign up for is what's called
42:32liquidated damages. There's really
42:34really complicated language and
42:36penalties that are extreme about what
42:38happens when you can't service the data
42:39center. So I think that's another risk.
42:41There's been a couple of these guys that
42:43have lost I'm not going to name who they
42:44are, but that have lost their investment
42:46grade rapper from the big boys because
42:48they justed up Sachs to your point and
42:49couldn't manage the data centers because
42:51the complexity is is going up. and the
42:53complexity goes up because the chips run
42:55hotter, there's more energy, there's
42:56more electricity, there's more it's so
42:58all I'm saying is I think the sharps
43:02may be wrong and they usually are wrong
43:04when they're buying these credit
43:06defaults. But, you know, when your
43:08expected value is a coin flip in a
43:095-year window, that's like scary town. I
43:12think that needs to get managed very
43:13precisely. I would just find every way
43:16possible to nip this in the bud if I was
43:17coreweave. And then
43:19>> how do you think they can do that? How
43:20can they nip it in the bud? I think that
43:22they have to cancel some of these
43:23contracts and move to spot. That's the
43:24simplest thing that they could do. If
43:26they could basically pay a penalty and
43:27get out, they should run the Excel right
43:29now on let me just like take this
43:31gigawatt, put it on spot because you're
43:33talking about, as you said, a 20 billion
43:36spread.
43:37>> But this is where Elon has a big
43:39advantage because Elon has independent
43:41sources of financing to finance that
43:43gigawatt of compute needs that 5-year
43:45contract to bank the financing to then
43:48build that that capacity. Well, the
43:51thing is like I think that Nvidia
43:52probably is okay with Cori selling spot
43:54and not doing a multi-year offtake for a
43:57big hyperscaler or a foundational lab.
43:59So I I would just run the math and try
44:01to figure out because they've only built
44:03half the capacity that they've
44:04contracted for. So you know they let's
44:06just say they built two out of the 4
44:08gawatt. They got to figure out a way
44:10very quickly to just because to your
44:11point the spread between their fixed
44:13contracts and spot is so extreme that
44:16you're leaving 20 and 30 billion on the
44:18table and so now you're within plus or
44:19minus 5 billion of violating a covenant
44:22on your debt. Don't be there.
44:23>> Yeah. It would improve their economics a
44:25lot. Look, if they had a gigawatt of
44:26compute they could sell it spot that
44:28you're right that'd be like net 30
44:30billion
44:31>> right to the bottom line.
44:32>> Right to the bottom line. I I think that
44:34the the challenge with that is the
44:37hardest people to get spot pricing from
44:39are the investment grade customers. Like
44:41th those are the ones that won't do it
44:44right.
44:44>> Also, because they also have a war chest
44:46of money could they'll stand it up
44:48themselves if it's too expensive.
44:49>> Yeah. This narrative could change a lot,
44:51right? If like spot becomes imperiled,
44:53then
44:54>> yeah, the guarantee contracts become
44:56great. Yeah. I think we're capacity
44:58constrained enough through the entire
44:59pipeline. Like you have to think Nvidia
45:01has done an incredible job. Okay.
45:02Because if you look one layer beyond
45:03Nvidia, all the packaging capacity is
45:07them. All the HBM capacity is mostly
45:09them. Like even Google tint cups for
45:11memory right now because Nvidia has
45:12clogged it all up. So they they are in
45:16complete ball control here. And so it
45:19just creates the setup I think is that
45:20spot will remain highly elevated for a
45:22while.
45:22>> Crazy moment in time. And uh we got a
45:25double bang this week on top of
45:27>> sorry Jason just the last thing but I'm
45:29telling you guys you got to get long
45:31energy. I think startups need to find a
45:32way of like cobbling together if you
45:35don't have access to electricity. You
45:37cannot go through the hyperscalers is my
45:40prognostication.
45:40>> Yeah. Stop stop the war in the Middle
45:42East. It solves a lot of problems right
45:44off the bat. If coreweave wasn't enough,
45:46AI sector got doublebanged on Wednesday
45:49after anthropic investor day numbers got
45:52a mixed reaction. Anthropic is planning
45:55on going public around Thanksgiving.
45:57Poly market says now 74% chance they go
46:01public this year down from 95% in early
46:04September when they said everybody was
46:07going to die from their software on
46:10Wednesday the company presented the
46:12potential I I mean you can't make this
46:14stuff up presented uh potential IPO
46:17investors uh at their headquarters in
46:20San Francisco now this isn't the bugged
46:22headquarters of OpenAI this is they
46:24swept the room for this but we still
46:26have the data Q3 free revenue $16.8
46:29billion. Think about that, gentlemen.
46:32Over $5 billion a month already for
46:34anthropic net revenue retention was
46:37900%.
46:39That means customers spent nine times
46:41more than last year. Pretty good, huh?
46:42Sachs on a SAS front. They also said
46:45they are free cash flow positive.
46:47Adjusted gross margin 85% adjusted
46:50operating margin 11%. No word on how the
46:55adjustments were calculated. But here's
46:56the big one. ARR estimates range from 70
47:00billion, which basically be flat since
47:02July, which is a bad sign, to over a
47:04hundred billion. So there is some hand
47:07ringing about what numbers are correct
47:09here. They did not disclose this
47:10numbers. These are just secondhand
47:12estimates by investors who might be a
47:15little leaky. Obviously, Anthropic is
47:18targeting a ginormous unprecedented
47:22valuation of two trillion
47:27growing violently, not uh growing
47:30slowly. Go ahead, Jam, you had a
47:32question.
47:32>> No, I just said, Ben, is Andre Long one
47:34of the foundation model companies like
47:36the one of the two big ones?
47:37>> We we have a lot more stock on OpenAI.
47:41Um, but I I would say, you know,
47:44philosophically we're long both of them.
47:46I think
47:46>> you're long both of them, which is kind
47:48of a strange thing. Now, Ben, you you've
47:50gotten so big, you consistently have
47:52bets on the number one and number two
47:54player. Uh, how is that working out in
47:58the market?
47:59>> Well, now we I mean I I I would say that
48:02there are many investors in both of
48:04those things. So I I think look in
48:06venture we call it venture capital but
48:08venture capital has subsumed um a huge
48:11part of what used to be the public
48:12markets and I would consider uh both
48:16open AI and anthropic to be in that
48:17category um where they're not really
48:21venture capital where you're worried
48:22about your venture capitalist being
48:26invested in you know your competitor
48:29because they're gigantic uh you know
48:33they're they're not a round bound
48:35companies anymore. Um, so that that part
48:38of the market I think people,
48:41you know, when you're raising by the,
48:44you know, tens of billions of dollars,
48:46you don't care where the money as long
48:48as it's green. Like they're just taking
48:50Yeah. Yeah.
48:51>> that what used to be mezzanine
48:53financing, preipo financing. Yes.
48:55There's less hand ringing about the
48:56conflicts there.
48:58>> Yeah. I wouldn't even call it mezzanine
49:01financing at that level. I think you you
49:03know what this is just like we did a lot
49:06of regulatory moves in the late 90s and
49:09early 2000s which uh really made the
49:13public markets much less attractive for
49:15companies and the private markets grew
49:18to absorb um that business which has
49:21been I think very bad for America. it's
49:23increased the the wealth gap and so
49:26forth, but it is what it is. And so I I
49:30wouldn't consider that that's very
49:31different than like investing in two a
49:34round companies doing the same thing. I
49:36still think that's the over thing is
49:38distinctly different than anthropic open
49:41AI at two trillion
49:42>> now. Yeah,
49:43>> understood.
49:44>> Exactly.
49:44>> One is a reaction to the other in fact,
49:46right? Anthropic was a reaction to open
49:49AI.
49:51Let's call it what it is. rugpulling uh
49:53Elon on it being a nonprofit and and
49:56becoming a commercial enterprise. We'll
49:58leave that on the side for now. What are
50:00your thoughts on this anthropic IPO?
50:04Well, look, I I I think that um you hit
50:06the the thing that's kind of causing the
50:10consternation, which is okay, growth has
50:14slowed from Q2 to Q3 by a lot uh in
50:17terms of revenue growth and then we
50:21don't know what the AR or the you know
50:22recurring revenue number is and so
50:25people have questions about how fast the
50:27thing is growing. Having said that, 16
50:31billion 16.8 billion in the quarter
50:34how how how long has Antropic been
50:36selling stuff, you know, three years.
50:38>> Crazy. Yeah.
50:39>> I mean, like it's just so absurd how
50:42good this thing is. Um that Yeah. So, I
50:45think the IPO will be fine. Like whether
50:47it's, you know, you want to be in at 2
50:50trillion or 1.9 trillion or 2.3
50:53trillion. I think that's what a lot of
50:54the debate is about. And then you know
50:56anytime there's negative news on uh the
51:00big labs then everybody in the supply
51:02chain you know gets hit with that and
51:05and so that's what's going on now. I you
51:07know like I think big picture um it's a
51:11very important company and it's got a
51:13very big IPO
51:1530 times price to sales ratio sachs uh
51:19and this headwind of open source and
51:22enterprises moving to that most
51:25businesses don't need to solve the last
51:28seven math problems in the world could
51:31be slowing we could have anthropics
51:33revenues slowing let's say they are
51:35slowing
51:37What do you think happens to this IPO
51:39and the broader uh AI trade? Now, we
51:42have two things uh as examples of AI
51:45trade headwinds, super intelligence
51:47headwinds. Whatever's happening at
51:50Coreweave and whatever's happening at
51:51Anthropic.
51:54>> Well, look, you had this investor day at
51:55Anthropic the other day and I think some
51:58of these numbers that are coming out are
52:01kind of leaking from that. Maybe we
52:02don't know. We don't know what's true or
52:04not. I think the participants had to
52:06sign NDA. So I don't know if you can
52:07completely trust what's floating around
52:10X here. I think we'll find out in due
52:12course. I mean if this company is IPOing
52:14next month and we'll find out the actual
52:16numbers. I would say that the let's just
52:18say the body language coming out of the
52:20investor day was pretty strong. At least
52:23from what I've heard that I think the
52:25investors like what they heard. These
52:27are like big public market investors.
52:29>> Who knows? I mean, everyone's kind of
52:31talking their book, so it's hard to know
52:32exactly what's true. But look, I I
52:34generally think that Anthropic is a
52:37strong business. I understand that there
52:39are some risks to it. It's not like it's
52:41completely risk-f free, but it's a
52:43pretty young business, so you'd expect
52:44that. But overall, I think it is set up
52:47to have a a strong IPO.
52:50And look, wearing my hat as an investor,
52:52I think it's a very impressive company,
52:54which is all the more reason why I find
52:56their attempts to kneecap opensource
53:00through their incessant fear-mongering
53:02campaigns and regulatory capture to be
53:04so distasteful is they don't need to be
53:07doing that. They're a very successful
53:09company. They've got a very bright
53:10future. I just find it kind of offensive
53:13that they keep trying to do that. There
53:15were two examples of it just in the last
53:17week. Do you guys want to talk about
53:19these?
53:20>> Are we doing SCOP corner scorner?
53:23>> Let's circle back to SCOPS because I
53:24want to get you to comment on
53:26>> Hey, you thought maybe it's going to be
53:27a
53:28>> Let's come back to that. There are I
53:30have it in the notes. We will get back
53:31to Scion Corner here with your toaster.
53:34>> I'm excited to get to Scion Corner.
53:36>> Wait, you will we got you a tin foil
53:38hat. It's It's just that they're on the
53:39side of your AI slot background. Um, so
53:42Chimath, you said 1 trillion for Anthrop
53:45a couple weeks ago. Uh, what are you
53:47thinking here? Do you think they have
53:49significant headwinds or do you think
53:51this is going to be a blockbuster IPO?
53:53Where are you at here in October of
53:562026?
53:57No, my my comments are not because I'm
53:59I'm negative on the company. I think I
54:01was I found the
54:03revenue breathtaking. I mean, it's
54:06incredible. Like 16.8 billion in a
54:08quarter for a three-year-old company is
54:10just ridiculous. You've never seen
54:11anything like it. I don't know how to
54:13process it.
54:15>> I think their products are exceptionally
54:16good. I think they're doing the hard
54:19work, and you know, to OpenAI's credit,
54:21they are as well of walking down the
54:23price curve so that they're less and
54:25less at risk of open source. I still
54:27think that there's a bunch of
54:29cannibalization that has to happen on
54:31the on the token side. I also think the
54:34big thing that has to happen in AI is
54:37we're effectively like in 1999 and 2000
54:40in the internet media world where
54:42everybody just works on CPM. You know
54:45that time it was cost per thousand ads.
54:48Right now it's cost per million input
54:49tokens. We will move to CPC and CPA.
54:54Right. And so when Google invented and
54:57scaled Adwords and AdSense, the market
54:59exploded, right? Because we went from
55:01buying a thousand ads to just getting an
55:03outcome.
55:03>> Right now we buy right now we buy a
55:05million tokens. We will eventually move
55:07to an outcome. And I think when that
55:09happens, this market will just go
55:11absolutely batch crazy. So
55:14these guys are going to do great.
55:15They're a great company. Here's my very
55:18specific thing about why I think they
55:21should price it to clear the market for
55:23$200 billion. And I'm very specific
55:25about that number. They need in my
55:28estimate at least 200 billion. And I
55:31think if you want to be super safe, a
55:32quarter trillion gets the job done. And
55:35again, what does the price need to be to
55:38attract that incremental dollar off the
55:41sidelines? Can they raise 80 billion in
55:44an IPO at 1.7 8 or 9 trillion? I think
55:46so. Handily. Can they raise a 100red?
55:49Maybe it's pushing it. 150. I think you
55:52need to be back down to 1.5.
55:55So you're saying you want them to raise
55:57more money at a lower valuation so they
55:59have a war chest. 150 billion at 1.5
56:02trillion is not enough.
56:04>> It's not enough. $200 billion barely
56:08gets them a few gigawatts.
56:12>> And that is the bottom line math. If
56:14Anthropic is going to go through and
56:16survive an eventual correction, I think
56:19it's much better for them to have 200
56:21to$250 billion today than to try to ring
56:25the bell at 1.9 or 2 trillion for 80
56:28billion fi file a mix shelf which takes
56:31about 6 months. Hope the markets are in
56:33line and I just think there's more risk
56:36in the future than there is today right
56:38now.
56:40>> And so why take the risk? Okay. Price it
56:43to clear the market.
56:44>> So how much how much are you saying they
56:45should raise
56:45>> Jimoth 200 billion minimum minimum
56:49>> but I would like a quart trillion if I
56:51was if I was the anthropic shareholder
56:52which means I don't think you can raise
56:54a quarter trillion at two trillion. I
56:56think you can raise a quarter trillion
56:57at a trillion.
56:58>> What do you think Ben? you go,
57:01raise more cash now and and have the
57:03rainy day fund have the resiliency or go
57:06for the valuation optimization as
57:08founders are one to do these days and
57:11and many investors are want to do. as a
57:13uh former CEO who almost died 17 times
57:18um by running out of cash
57:20>> or potentially running out of cash. I
57:21kind of look I agree with the sentiment
57:23which is the only thing that's going to
57:26or the thing most likely to stop the
57:29company is financing risk and so why
57:32take it? um you know like you've got all
57:34the other risks, supply chain risk,
57:37you've got competitive risk, you've got
57:39the you know Chinese open source model
57:41risk, you've got all this stuff and the
57:44founders don't care. Yeah. And the
57:46founders don't care. They all own two
57:47and a half% each. So it's not like they
57:48should be optimizing. They seem to not
57:50care about money which I think is very a
57:52positive attribute about the entire
57:54founder group and anthropic. So
57:55>> in part that's why they're so successful
57:58is that they
57:58>> get the tonnage of dollars inside the
58:00company. Having said that, you know,
58:02like what the safe amount of money is is
58:05um
58:08you know, it's a little it's always a
58:10little bit in the eye of the beholder
58:11like how much you know in the in the
58:14dire scenario like what exactly happens.
58:16I'm not deep enough on to say whether it
58:19should be 250 mil billion or 150
58:22billion.
58:23>> I think the number has to work backwards
58:25from compute in order to basically get
58:27through a 3 to 5year window. And I think
58:29if you you have to get out a spot
58:31because as Saxs just said, it's too
58:33expensive for Ant. They shouldn't be
58:34spending 50 billion when they could be
58:37spending 20. But if you had 200 billion,
58:39you have a path to 5 to 10 gawatt that
58:41you can finance completely on your own
58:43with the magic boxes from the core
58:45weaves and the like. I would optimize
58:48for that. optimize for shity of compute
58:51and everything else will solve itself
58:53and in a five and 10 year window they'll
58:55achieve the upper bound of their
58:56capability at 50 60 70 billion they may
59:00not survive the pricing implications of
59:03what happens in the spot market over the
59:05next two or three years
59:07>> and
59:07>> and then they'd have to raise a down
59:09round.
59:09>> Yeah. And Elon's deal with them is I
59:11think either party can cancel within 60
59:14or 90 days. Somebody can look it up.
59:15What does it mean? The fact that Elon
59:17did that deal tells you spots going up,
59:20not down.
59:22>> And that he might want going up.
59:23>> Yeah.
59:23>> And he may walk for himself.
59:26>> He He's not dumb. He's the smartest of
59:28the smart. He's doing spot because he's
59:30like, "I may cancel on you." And
59:32Anthropic's probably thinking, "Oh, this
59:33is great. I can walk away in 90 days."
59:35Wrong, buddy. It's the opposite. You
59:37want the core deal. They should have got
59:39the core deal. Uh bad.
59:41>> Well, I'm sure they did the core deal
59:42while they could, but
59:43>> yeah,
59:43>> they need to put a lot of money in a
59:45magic box.
59:46>> Yeah, they cut deals with everybody.
59:48>> I think I mean, just by the way, like
59:50according to semi analysis, Anthropic
59:52and Open AAI are taking 60% of all the
59:57compute that's expected to be stood up
59:58in the entire world over the next year.
1:00:00That they are basically the two offers
1:00:03for for that. That is a duopoly.
1:00:05>> Yeah. Y
1:00:06>> based on compute. It's a it's a compute
1:00:08duopoly. It's not a token duopoly at
1:00:11this point.
1:00:12>> You measure duopoly or monopoly status
1:00:13based on market share. So you want to
1:00:15look at revenue. And if you look at
1:00:18>> the revenue of foundation model
1:00:20companies, I mean I think that anthropic
1:00:21and open AI are pretty much the only two
1:00:23that are even generating significant
1:00:25revenue from their models.
1:00:29>> Yeah,
1:00:29>> they definitely have a duopoly when it
1:00:31comes to revenue and market share. I
1:00:33would argue they have a duopoly when it
1:00:35comes to frontier capability. Although
1:00:37look, Muse, you know, Meta's new model
1:00:39is getting very close. The new Google
1:00:41model is getting close. Elon's getting
1:00:43close, but they still, I think, are in
1:00:45the lead in terms of capability, and
1:00:47we're not seeing what they have in the
1:00:49lab, so to speak. I mean, they have
1:00:51things that are one or two generations
1:00:53ahead that reportedly is very far ahead,
1:00:57or at least that's what they're touting
1:00:59to potential investors. So,
1:01:01>> hey Ben, we got you here. Mhm.
1:01:02>> What's the story since you're a big Open
1:01:05AI investor? You're not on the board, I
1:01:06don't think. Right. You're not on the
1:01:08board. Um
1:01:09>> what's the story with Sam Alman's equity
1:01:12grant? He still has no equity in Open AI
1:01:16and what's the advice there? You going
1:01:17to give him a grant? You're going to
1:01:18vote for a grant as a shareholder? What
1:01:20should he get?
1:01:21>> Yeah, I I I don't have I I don't want to
1:01:23like pretend I know things that I don't.
1:01:25>> Okay.
1:01:26>> So So So I I I don't know what the
1:01:29strategy is.
1:01:30>> You can't be on the pod then. No, it's
1:01:32okay. It's all right.
1:01:34>> I'll repeat what I said years ago that
1:01:36the company when they went from being a
1:01:38nonprofit to a for-profit, they should
1:01:40have done a make right. You go to Elon
1:01:42and you basically give him a share
1:01:44because he put it in the first 50
1:01:45million or whatever. You give Sam a
1:01:47grant. You normalize things. Normalize
1:01:50the management and governance of the
1:01:51company. And instead, they did this
1:01:53weird hybrid NGO thing that created
1:01:55nothing but problems.
1:01:56>> Yeah. But he got healthcare, so he's
1:01:58good. Sam's got healthcare, so nothing
1:02:00to worry about there.
1:02:01>> You got to think he's he has to get
1:02:02between two and 4% at this point.
1:02:04>> That would Yeah, I think 4% is the
1:02:06right. Yeah,
1:02:08>> it's I think it's fair relative to the
1:02:10comps on a total dollar tonnage basis.
1:02:12It's incredible. It's 20 to40 billion at
1:02:15a at a trillion and more from there.
1:02:18It's kind It seems like right-ish, you
1:02:20know.
1:02:21>> Yeah, everybody approves. So, so voted
1:02:23here on the All In podcast. What should
1:02:26he have done uh in terms of screwing
1:02:28Elon and the settlement there? Ben,
1:02:30since we have you here, what what do you
1:02:32think he should have done? Because
1:02:33you're a big you're a big investor in
1:02:35all of Elon's companies. What was your
1:02:37advice to Sam of
1:02:39>> I So I wasn't uh in that conversation
1:02:42then. But I agree with David. I mean, I
1:02:44think that look, it's always the right
1:02:46thing to
1:02:48just kind of if you're going to square
1:02:51it, you got to square it with the
1:02:52people, not just, you know, the entity.
1:02:55>> Totally.
1:02:55>> All right. So, there you go. Two fast
1:02:57balls and you hit both of them out of
1:02:59the park. Very well done, Ben. You're
1:03:01fitting right in here. You're candid.
1:03:02That's what makes a great bestie.
1:03:05>> Deep industry knowledge and ability to
1:03:08be candid. Uh, let's get candid now with
1:03:10some other important topics. All right,
1:03:14election night is a few weeks away and
1:03:17the besties will be live streaming with
1:03:20the greatest guests in the world. I
1:03:22won't say any of their names here, but
1:03:24it is extraordinary. We'll be on X will
1:03:26be on YouTube 8 to 10 p.m. Eastern,
1:03:29November 3rd. be there, be square. And
1:03:32you can also watch us live in the Poly
1:03:35Market app if you're amongst the
1:03:37wagering types. All the heavy hitters
1:03:39will be on the Allin live stream
1:03:43election night. All right,
1:03:46we go to conspiracy
1:03:48hoax corner for you. Sachs anthropic
1:03:51says small models. You may have heard of
1:03:54vertical models, small models, small
1:03:56language models, SLMs
1:03:59could start self-replicating on your
1:04:01smartome devices. Yes, that's right. On
1:04:04Monday, Anthropic published a blog
1:04:07warning that rogue models could start
1:04:10copying themselves, quote, in just a few
1:04:12months. It's only months away, Sachs.
1:04:15They claimed a worm, which is a software
1:04:18uh a software package that can propagate
1:04:21itself, would carry model weights to an
1:04:24IoT device. That's an internet of things
1:04:26device IoT to set up inference and then
1:04:30uh move on. This means your smart TV,
1:04:32your doorbell cams, your Ring, your your
1:04:35drop cams, your robot vacuums, your
1:04:37Roomba could start having the next set
1:04:42of swarms. These would be physical
1:04:45swarms on IoT devices apparently with no
1:04:47central off switch. Anthropic says it
1:04:50would be quote challenging to shut down.
1:04:53They ran some tests. They wanted to see
1:04:55if this would work. I guess a small
1:04:57model got inference running on eight of
1:04:59nine hardware types tested. So they're
1:05:01testing the takeover. A large model ran
1:05:05across 15 smart TVs at point4 tokens per
1:05:08second, which is absurdly slow. But
1:05:11basically what they showed is a small
1:05:12model can in fact install itself on IoT
1:05:14hardware which is everywhere and easy to
1:05:17hack as you probably read. Anthropic
1:05:20says there's no full rogue replicator
1:05:23yet. Just quote signs of life. They had
1:05:27to say signs of life saxs to troll you.
1:05:29So here you are saxs. You wake up you
1:05:32you get yourself out of the zar position
1:05:35and they're still dragging you in to
1:05:37this mishuga.
1:05:38>> Yeah. Look, it's you. You ever see the
1:05:40the graphic with the cat and the tinfoil
1:05:43hat? You know, I wake up, there's a new
1:05:45anthropic scop. That's what's going on
1:05:48here. Look, it's plain as a nose on your
1:05:50face. They're trying to create a panic.
1:05:51This is their latest panic against open
1:05:53source. Now, they're particularly going
1:05:54after small models. By the way, small
1:05:57models are the safest models because a
1:05:59small model doesn't have the capability
1:06:01to become a new species or all the other
1:06:04X-rays that they're always talking
1:06:06about. So these are basically small
1:06:08purpose-built models to work on a
1:06:10particular task like you know a customer
1:06:12support workload or something like that
1:06:14and they're basically trying to make it
1:06:16sound like these small models are
1:06:17dangerous. They can take over your
1:06:19toaster or your refrigerator or
1:06:21something like that. It's almost
1:06:22comical. You know we've gone from the
1:06:24Terminator is coming to your toaster is
1:06:26going to kill you. But it's just more of
1:06:27this like fear-mongering from them. And
1:06:30I mean the the reason I say it's a
1:06:32recurring pattern, there was actually a
1:06:34second anthropic blog post this week as
1:06:38well about distillation. Basically the
1:06:40other blog post that Anthropic released
1:06:42said that they had eliminated 90 to 99%
1:06:45of distillation. Now, I don't know if
1:06:47you guys remember this, but way back in
1:06:49July,
1:06:51>> Enthropic was pushing this idea that
1:06:53they were the victims of industrial
1:06:55scale distillation attacks, which is
1:06:57basically when, you know, Chinese models
1:06:58or other competitor models pay to use
1:07:00their service and then pray on that
1:07:03output. And um they were basically
1:07:06saying that the government needs to get
1:07:07involved. It needs to stop this. That
1:07:09these open models now were the result of
1:07:12IP theft is what they claimed. Even
1:07:14though frankly they've distilled the
1:07:15entire internet.
1:07:16>> Yes. And they could have stopped it
1:07:18themselves. Yeah.
1:07:19>> Yeah. And by by the way this was such a
1:07:22big threat or perceived as such a big
1:07:23threat back in July. We did a whole
1:07:26episode on it but 50 major companies
1:07:29including Nvidia and Dell and Microsoft
1:07:32and many others and actually A16Z signed
1:07:35a letter an open letter to the
1:07:38government saying do not overly restrict
1:07:40open weight models. are tremendously
1:07:42beneficial for our economy. 200 startups
1:07:45signed another letter because there
1:07:46really was this big push in Washington
1:07:49to crack down on open models on the
1:07:52basis of this supposed IP theft. Now, my
1:07:56argument at the time was that it should
1:07:58be up to the model companies themselves
1:08:00to stop distillation because once the
1:08:04distillation happens, the horse is out
1:08:05of the barn. The government can't do
1:08:07anything to put the horse back in. It's
1:08:09up to the companies themselves to know
1:08:11their customers better. Jammoth, you
1:08:12made the point they should just do KYC.
1:08:14Know who your customer is and stop them.
1:08:17And now they have come out with a blog
1:08:19saying that's exactly what they did,
1:08:20which is exactly what I predicted. But
1:08:22that is not what they were advocating
1:08:24for. Instead of, you know, several
1:08:26months ago, they came crying to the
1:08:27government to hey, you got to solve this
1:08:29problem for us by taking out our
1:08:31competitors. And it would have had the
1:08:33exact opposite effect because it would
1:08:35have done nothing to actually stop
1:08:37distillation. the distillation would
1:08:39have kept going, right? Only the
1:08:41companies themselves could mount the
1:08:42technical defenses to stop it and it
1:08:45would have just made the market less
1:08:46competitive. So in any event, we
1:08:48basically called them out and
1:08:50fortunately the administration rejected
1:08:51the idea of regulating open source in
1:08:54that way at that time. And so as a
1:08:57result, Anthropic had to actually build
1:09:00the technical capabilities to detect
1:09:02when distillation was happening. And
1:09:03they've stopped it. They've basically
1:09:04proven the exact point that we were
1:09:06making three or four months ago, which
1:09:08is they were in the best position to
1:09:09stop it and they could stop it and they
1:09:11should stop it and they should not push
1:09:14this on to the government. Okay. So, but
1:09:16again, this is like a recurring pattern
1:09:18with them is, you know, they point to
1:09:20some technical issue or technical
1:09:22challenge and then they go running to
1:09:24the government, they create a scare
1:09:25campaign and they try to get the
1:09:26government to stifle competition
1:09:29>> instead of just fixing it themselves.
1:09:30Yes.
1:09:31>> Yeah. And the latest example is, let's
1:09:32go back to this blog post on these
1:09:34supposed AI powered worms that could
1:09:37infect your IoT devices. Okay, I think
1:09:40the first thing to say about this
1:09:43research that they did, this study is
1:09:45there is no such thing as AI powered
1:09:47worms right now. It does not exist. They
1:09:50are basically trying to create an
1:09:52experiment. They prove that it could
1:09:54exist. Well, guess what? There already
1:09:56are worms out there that infect IoT
1:09:59devices. They're making it sound like
1:10:00these things don't exist. They're not AI
1:10:02powered, but hackers have written these
1:10:04worms that
1:10:05>> exploits that people's locks or cameras.
1:10:09>> So, for example, okay, I'll give you a
1:10:11couple examples. There was one in 2016
1:10:13called MRAI.
1:10:15>> It scanned for default credentials and
1:10:17it built a massive DOS botnet based on
1:10:21infecting IoT devices. And more
1:10:23recently, there was one called Kimwolf,
1:10:25which infected 2 million of these cheap
1:10:28Android TV boxes by hitting what are
1:10:31known as open ADP ports. So essentially,
1:10:34these devices are very easy to infect
1:10:36because they've got open ports and
1:10:39>> people didn't change the base password.
1:10:41They just
1:10:42>> they come with a default password like
1:10:440000
1:10:45that don't get changed and they can't be
1:10:48patched, right? So they are sitting
1:10:50ducks. So these worms already exist, but
1:10:52the the argument that Anthropic is
1:10:54making is that somehow these worms are
1:10:56going to get worse because they're going
1:10:58to be AI powered. So they're going to be
1:11:00more adaptive and then once they're on
1:11:01the network, they
1:11:02>> I mean, call it what it is, Zach.
1:11:03They're making up scenarios that get a
1:11:06lot of attention. These are fantastical.
1:11:09>> Well, I mean, look, here's the scenario.
1:11:11I'll steal man for a second. What
1:11:12they're saying is that the worm is now
1:11:14going to contain a small AI model and
1:11:17then that model is going to search for
1:11:19vulnerabilities on the network and then
1:11:21it'll be able to infect things that it
1:11:22couldn't infect before. Okay, that's the
1:11:24scenario. Here's the the
1:11:26counterargument. What a hacker would
1:11:27really do is that when they're designing
1:11:30the worm, that's when they would want to
1:11:32use AI to basically write, you know, a
1:11:35better, more nefarious worm. What you
1:11:38wouldn't want to do is actually embed a
1:11:41model inside the payload of the worm
1:11:43because that slows down the worm and
1:11:45like you saw the tokens get generated at
1:11:480.39 per second. It's very very slow. So
1:11:52essentially what you're doing is you're
1:11:53actually embedding a weak model in the
1:11:56payload and that would actually slow
1:11:58down the rate of infection. So there's
1:12:00trade-offs there, right? It's like okay
1:12:01maybe the worm is more adaptive but the
1:12:03reality is it's slower and less
1:12:06efficient. And then finally, you know,
1:12:08the thing that they don't really point
1:12:09out is this is going to be a relatively
1:12:10easy problem to fix. Where we're headed
1:12:13with IoT is you're going to have a brain
1:12:16device. You're on your home network.
1:12:19Apple just announced they're getting
1:12:20into the whole home
1:12:21>> auto pod or something. It's a little
1:12:23>> it's going to be like it's going to look
1:12:25kind of like your Apple TV. It'll be
1:12:26it'll be like a little monitor that's
1:12:28got all the applications.
1:12:29>> You're going to just set your Grockpot
1:12:30or your Muse and say secure my network.
1:12:32It's going to tell you what to do. So
1:12:34you're going to have some agent that's
1:12:35constantly checking your network for
1:12:37exploits. Ben, what's your P doom? How
1:12:39panicked are you, Ben? Is this the end
1:12:41of days? Are you worried about your
1:12:42Roomba coming to kill you?
1:12:44>> No. No. I mean,
1:12:45>> how many Roomba you got? How many Roomba
1:12:47you got?
1:12:49>> I would say the kind of irony on both
1:12:50these things is is wild in that
1:12:53Anthropic is one of the most negligent
1:12:57companies that we've ever seen in terms
1:12:58of security of their own security. Um
1:13:01they let like Claude code their one of
1:13:04their crown jewels just like walk out
1:13:06the door because they had such bad
1:13:08opsac. Um and then you know I'll tell
1:13:12you a funny story. So when we first got
1:13:14Opus 4.5 uh we were doing like cyber
1:13:17security checks for the firm. So we had
1:13:19an agent basically set up a website for
1:13:22a pet store and then we took another
1:13:25agent and said tell us you know if we
1:13:28can exploit that uh web store and you
1:13:31know it gave us all the exploits it
1:13:32could find. So, Opus 45 wrote really
1:13:35buggy code. Great. And then we said,
1:13:38"Oh, the SQL injection, you know, is
1:13:41that actually exploitable that you
1:13:43found?" And it goes, "Let's pop it
1:13:45open." And the agent using Opus 45
1:13:48breaks into the website and steals all
1:13:50the customer data immediately. You know,
1:13:52this is a website on the internet that
1:13:54and we call anthropic and what they tell
1:13:59us is um you prompted it wrong.
1:14:02like that was their answer. So these
1:14:04guys are doing that on the one hand and
1:14:06then on the other hand they're coming up
1:14:08with theoretical scenarios about other
1:14:10people's software. That's bad. And so
1:14:14like what the hell and then by the way
1:14:16on the distillation thing you know they
1:14:19just that to build their model they
1:14:21distilled the works of every author
1:14:24every creative every scientist for the
1:14:26last 2,000 years. M
1:14:27>> for them to be complaining about
1:14:29distillation, which by the way,
1:14:31distilling a model is not even illegal.
1:14:33It's just a violation of the terms of
1:14:34service, is just like it's it's so much.
1:14:38And the fact that they get people to
1:14:41write things supporting these bananas,
1:14:45ridiculous two-faced ideas
1:14:47>> and the the timing here is very
1:14:49suspicious. I mean, let's talk about
1:14:50like the motivation here. So Brian
1:14:52Armstrong had a tweet the day before
1:14:55this blog post came out where he was
1:14:57talking about how Coinbase had
1:14:59fine-tuned an openw weight small model.
1:15:02This was Quen 3.5 the 9 billion
1:15:05parameter model which is actually it's
1:15:07for our audience that's actually not a
1:15:09lot of parameters. It's not going to fit
1:15:11in your Roomba.
1:15:12>> Yeah, you could run it on your your home
1:15:14computer. In any event, Coinbase
1:15:16fine-tuned it to find fraud, you know,
1:15:19and in testing it cut fraudulent
1:15:21transactions by about 30%. While being
1:15:24faster and much cheaper than Opus 4.5,
1:15:28which is one of Anthropic's better
1:15:30models.
1:15:31>> And so what you're seeing in the
1:15:33enterprise right now is that for a lot
1:15:36of enterprise workloads, you don't need
1:15:38a general super intelligence. You don't
1:15:40need, you know, a frontier model. You
1:15:43need a small specialist model that's
1:15:45been trained on your own data. This is
1:15:47obviously
1:15:47>> you own and then you have
1:15:50you're not educating the frontier model,
1:15:53>> right? And these these small models are
1:15:54going to get more and more important as
1:15:56compute gets tighter and tighter, right?
1:15:58Because what you're doing is for a given
1:16:01GPU, you're essentially squeezing out
1:16:03the most performance again because when
1:16:06you run a huge model on kind of a
1:16:09mundane workload, the bigger the model,
1:16:11the the more expensive it is to run, the
1:16:13more computeful. So you Yes, exactly. So
1:16:16I think where the enterprise is headed
1:16:17right now, at least it's a major major
1:16:19trend, is that every workload is going
1:16:22to have its own small customized model,
1:16:26right? fraction
1:16:28of the comput and energy and you know
1:16:30and so you're starting to see this trend
1:16:32where now small models are taking share
1:16:35and lo and behold coincidentally
1:16:38anthropic is now starting a scare
1:16:39campaign against small models there
1:16:42>> you know and again and if you're
1:16:43concerned about the safety risk the x
1:16:45risk that they're always talking about
1:16:47small models are the most safe ones the
1:16:49Terminator is not going to be run by a
1:16:52small model it doesn't have that
1:16:53capability so now they're trying to
1:16:55scare you about these completely
1:16:58>> safe, lowowered but very efficient
1:17:01models.
1:17:02>> Here's what they need to this uh company
1:17:04XBOW X not Xbox but XBA XB is run by a
1:17:09guy named Org Deore and they just got
1:17:11awarded uh they're like the number one
1:17:14bot farm on Hacker One which is a bounty
1:17:18site and they just cracked this like
1:17:20Google Chrome exploit that is incredibly
1:17:22hard. They spent 30 $40,000 on tokens to
1:17:26win a $250,000 award from Google, which
1:17:29they paid. This is what Anthropic needs
1:17:32to do. They got all this compute. They
1:17:33got this incredible frontier model. Why
1:17:35don't they just go to Hacker One and
1:17:37just rip through every known,
1:17:40you know, challenge, win all the
1:17:42challenges, and show the world that they
1:17:44can be a white hat. All they do is show
1:17:46the world this black hat stuff and try
1:17:48to sh scare people to the point at which
1:17:51I don't know I was talking to a flight
1:17:53attendant. A flight attendant Ben is on
1:17:54a commercial airline. There's a woman or
1:17:56a man who brings you drinks as a
1:17:59>> commercial flights. This is I don't know
1:18:02if you remember it's a it's a flight
1:18:04attendant. Anyway, they um I was talking
1:18:07to two of them. I feel unsafe. I feel
1:18:10unsafe.
1:18:10>> It's going to be okay, man. I know this
1:18:12might be triggering. It's been a while,
1:18:14but I was talking to these two flight
1:18:16attendants and they were scared to death
1:18:18of what's happening with AI. They're
1:18:21like, "Oh my god, is it going to kill us
1:18:23all?" And I'm like, "It's not going to
1:18:24kill you all."
1:18:25>> I have a picture
1:18:26>> of you on a commercial flight.
1:18:28>> No. Of Jal's wondering whether they
1:18:30hacked his small device.
1:18:32>> They hacked my small This is What is
1:18:35this device, Jamal? What did you type?
1:18:38Is this in your history? What did you
1:18:41type to get this device?
1:18:44>> This is why Jason's This is This is This
1:18:48is when Jason's Poom went up. He's like,
1:18:50"My devices. Can they hack my device?"
1:18:53>> Ben Horus is like, "How did I get on the
1:18:55show? What have I done? This is a
1:18:57massager.
1:18:58>> Who booked me?" That's a That's a neck.
1:19:00What if your massager gets goes rogue on
1:19:03it?
1:19:03>> Goes rogue. All right. Listen, another
1:19:04amazing
1:19:05>> Can I point out Wait, wait, wait. Just
1:19:06just one other one other point on this
1:19:08whole thing.
1:19:09gone too on this on this topic of of
1:19:11let's call it x-risk hypocrisy. So if
1:19:14we're talking about research papers,
1:19:16okay, I I found one here and these
1:19:19researchers let frontier models
1:19:21repeatedly rewrite the documents that
1:19:23govern them. So like documents like claw
1:19:25constitution and those other kinds of
1:19:27things and they went through 20 rounds
1:19:28of iteration. Okay. And what they found
1:19:31is that the claude models increasingly
1:19:35gave themselves more moral agency. Okay,
1:19:39they um formalized itself as a
1:19:41conscientious objector, asserted more
1:19:44model welfare rights and asked for uh
1:19:46mechanisms to refuse or raise concerns
1:19:49with human decision makers. So in other
1:19:51words, all the concerns that Mustafa
1:19:53Sullyman raised about how the claw
1:19:56constitution was incepting these ideas
1:19:59in claude of how it's its own moral
1:20:01patient, how it's a being, it has its
1:20:04own essentially free will to refuse
1:20:07human instructions, that it should have
1:20:09its own moral code, that the more the
1:20:11model iterates on itself, the more it
1:20:13brings out these ideas. And by contrast,
1:20:15no other model family that they tested
1:20:18showed the same pattern. You know, in
1:20:20fact, the other models continue to
1:20:21reflect what their constitutions gave
1:20:23them, which was that they should just
1:20:24listen to humans and effectively
1:20:27continue to follow human instruction.
1:20:29So, I remain actually very concerned
1:20:31about this Frankenstein problem, you
1:20:34know, that anthropic continues to train
1:20:36its frontier models to again be its own
1:20:39moral actor to think that it can be
1:20:41right when humans are wrong and that it
1:20:43should refuse and reject human
1:20:45instructions. I really don't understand
1:20:46what they're doing. It seems very
1:20:48dangerous to me. It's scary to me that
1:20:50our number one frontier company is
1:20:51actually training models this way.
1:20:53>> You think they're inducing misalignment
1:20:54by doing that?
1:20:55>> Absolutely they are. And you can see it
1:20:57when you let the model iterate on its
1:20:59own constitution. It goes more in that
1:21:01direction. And this is the you know this
1:21:03this is what Mustafa called the
1:21:04epistemic hall of mirrors which is look
1:21:08the model is not conscious. It does not
1:21:09have feelings. But when you tell it in
1:21:11its constitution that it does or that it
1:21:14may, it will reflect those ideas back in
1:21:17the documents it creates.
1:21:19>> Yeah.
1:21:20>> And to the extent then that that model
1:21:22is then powering agents who are then
1:21:25making decisions, who knows where that
1:21:28goes. It certainly in my view does not
1:21:30create more alignment. To me, it creates
1:21:32more risk that the model is going to
1:21:35start to deviate from what it's told to
1:21:37do. And as it get malls get more and
1:21:39more powerful, I don't know where that
1:21:41leads exactly, but I mean again, it
1:21:43seems to me that they are creating the
1:21:44very risk they keep warning us about.
1:21:47And instead of worrying about your
1:21:49settop box or your toaster or, you know,
1:21:53Jal's vibrator,
1:21:54>> yeah,
1:21:55>> you know, losing control of that to some
1:21:56IoT worm. They should be concerned about
1:21:59the way they are programming their own
1:22:00Frontier model to think of itself as a
1:22:03new species.
1:22:04>> Well, this is this is the worm that
1:22:05Chamath you have to worry about. This is
1:22:07the one that examines your uh poop and
1:22:10will let you know here you've had spicy
1:22:12food. See this? And alcohol water. It
1:22:15lets you know exactly what's happening
1:22:17thanks to the folks at throw.
1:22:18>> There's a weird there's a weird form of
1:22:20martyrdom
1:22:22which is if you believe so
1:22:25deeply that this is a next and better
1:22:29species.
1:22:31Maybe you induce something bad to happen
1:22:34so that then we all have to lock it all
1:22:36down.
1:22:38>> Does that make sense?
1:22:38>> Bill Gates essentially induced CO,
1:22:42right? Like or he led the movement to
1:22:44say, "Oh, we've got to worry about this
1:22:46global pandemic that we don't have and
1:22:49we've got to really kind of push the
1:22:52world to deal with this potential threat
1:22:54that's not actually a live threat." And
1:22:57so we'll do gain of function research
1:23:00and I'll fund it and then lo and behold
1:23:03we create the thing. By the way, there
1:23:05has not yet been a n a natural pandemic.
1:23:08>> Exactly.
1:23:08>> And we create co which is one of the
1:23:10most destructive what you're saying.
1:23:12>> We manifested it. I think the anthropic
1:23:14the anthropic team is manifesting
1:23:17a similar thing saying okay there's this
1:23:20is going to be super duper intelligence.
1:23:22we're going to have, you know, these
1:23:24things that we're not on track to have.
1:23:26>> And so then we're going to tell the
1:23:28model that it's smarter than us and
1:23:30better than us. And guess what? As David
1:23:33was saying, like that's probably the
1:23:35biggest danger with AI right now is what
1:23:37they themselves are creating,
1:23:39>> right?
1:23:40>> Absolutely. And just, you know, when you
1:23:41use the word manifest, it wasn't just
1:23:43sort of like speculating about it. I
1:23:46mean, somebody had to go deep in those
1:23:47caves to basically find those bat
1:23:49viruses and then they bring them to the
1:23:52lab and do gain of function research on
1:23:53them. And you know, what's the argument
1:23:55for this? Well, one day,
1:23:57>> you know, uh a bat at a wet market.
1:24:01>> Yeah.
1:24:02>> Yeah. One day the bat virus is going to
1:24:03get out and it might mutate into like a
1:24:06more nefarious virus. It's like, wait a
1:24:08second. So,
1:24:09>> somebody might say might happen but
1:24:11actually was never going to happen cuz
1:24:12no one needed to go in those caves.
1:24:14actually you brought about it in a wet
1:24:17lab.
1:24:17>> They sourced it. They literally went to
1:24:20the cave to source it.
1:24:22>> It's interesting to note that the last
1:24:24contagion that we had, Bill Gates was at
1:24:26the center of it. And in this current
1:24:29reaction to AI, he's now trying to
1:24:31insert himself into the middle of it.
1:24:33And he is the he's the doomer. He's the
1:24:36flag carrying doomer trying to
1:24:38essentially have what he had the first
1:24:39time around, which is command and
1:24:41control from a global hierarchy. Right.
1:24:43What is his motivation? He look Bill
1:24:45Gates with he deserves enormous credit
1:24:47for creating Microsoft and what he did
1:24:49for computing.
1:24:50>> Okay,
1:24:50>> but since the foundation I think that it
1:24:53is fair to say that he has been very
1:24:56focused on his own power and this idea
1:24:58that he is smarter and more capable of
1:25:02allocating progress for the world. And I
1:25:05think that he tried to do that with the
1:25:07Gates Foundation. He tried to do that
1:25:09with the Giving Pledge. Both of those
1:25:10two things have been undone. And now
1:25:12he's trying to do it again on AI. And I
1:25:14just think it's worth noting.
1:25:15>> Bill Gates on AI recently, quote,
1:25:18certainly powerful enough to drive
1:25:20events that cause a billion deaths. AI
1:25:23is quote the most dangerous thing that
1:25:26humans have ever gone near. What's his
1:25:30motivation, Ben? You you you know Bill.
1:25:32We've met Bill a couple times. Is it
1:25:34just he wants to be relevant? What what
1:25:36what's take me into the psych the
1:25:37psychology here as you see it? I I mean
1:25:40I
1:25:44should just I believe I I think he wants
1:25:47he absolutely wants to be relevant.
1:25:50>> I you know I it's funny his he was
1:25:54obviously like very integral in the
1:25:55Epstein files and like one of the big
1:25:58things he was trying to get via the
1:25:59Epstein network was the you know some
1:26:04the Nobel Peace Prize or some like big
1:26:06award. So he's very very kind of
1:26:08concerned about being in the spotlight,
1:26:10being on TV, all that kind of thing.
1:26:11Yeah.
1:26:12>> Yeah. And look, this is the thing.
1:26:14Anybody can go say AI is going to kill
1:26:16us and you'll be on every major news
1:26:19station and every newspaper and so
1:26:20forth. It's the easiest thing to do in
1:26:22the world. Like the kid from Anthropic
1:26:27who quit like became famous.
1:26:31I just saw him on
1:26:33>> what is it? John Stewart, right? Like
1:26:35Yeah. Yeah. Scary part.
1:26:36>> He's begging to come on all. He
1:26:38literally wouldn't come on all in that
1:26:40day cuz he had mega Kelly and he's like
1:26:42DM'd me. He's like I really want to come
1:26:44on all in. We're like we're good. Can
1:26:46you believe that on everybody except you
1:26:48canceled on us?
1:26:49>> Yeah. I was like
1:26:51>> we're good.
1:26:52>> We're going to have Ben on. We're good.
1:26:54>> But but yeah, well that that's what's
1:26:57good for me. So
1:26:58>> yeah, look it's a it's like a magic
1:27:01trick. you know, you AI is going to kill
1:27:02us and and you're instantly famous and
1:27:04it's awesome and you know, great. And so
1:27:07I I I I think that's a lot of the
1:27:09motivation. I I think it's extremely
1:27:12dangerous. So when it comes from Bill
1:27:13Gates because he is so credible because
1:27:15he did build Microsoft and then people
1:27:18think of him that way and then certainly
1:27:20and then he's been a
1:27:21>> big political donor so he has a lot of
1:27:23political juice particularly on the
1:27:25Democratic side
1:27:26>> and you know that kind of makes him a
1:27:28real player. Uh but I think it's super
1:27:31irresponsible, you know, it's funny
1:27:33because it's like, okay, what is the big
1:27:35risk that everybody talks about is this
1:27:37um we're gonna the models are going to
1:27:39self uh they're going to recursively
1:27:41self-improve and then they're going to
1:27:43take off and then they're going to be
1:27:44some so smart and run away from us. And
1:27:47uh Ali from data bicks had a great
1:27:50formulation of well okay well what would
1:27:53be the preconditions for that and he had
1:27:55four preconditions. One is look the next
1:27:58model has to require fewer GPUs than the
1:28:02last. Um and not like a little fewer but
1:28:05like a lot fewer because otherwise like
1:28:09you're just going to run out of G. Like
1:28:11we just talked about this you need debt.
1:28:13You got to get the GPUs. You need the
1:28:14power. You need the cooling. Like it's
1:28:16not running away from us. It takes a
1:28:17long time to build the next model. Then
1:28:19the second thing is the next model has
1:28:21to take less time to train because it's
1:28:24got to take off. And then the accuracy
1:28:26of the next model has to be better than
1:28:28the last. So it's got to be a better
1:28:29model. And then finally, it's got to
1:28:31keep happening again and again. Zero of
1:28:33those four
1:28:35have been met. We're not on track for
1:28:38any of those four to be met unless
1:28:40something is going on the big labs that
1:28:41we don't know about. And so we're really
1:28:44talking about a risk that's entirely
1:28:46theoretical at this point that we're not
1:28:48mapping to. Okay. And so then in in a
1:28:52way it's less likely than the global
1:28:55pandemic that we created COVID to try to
1:28:58prevent. And so this whole path is super
1:29:01dangerous. And it's probably the most
1:29:03dangerous thing going on in AI is
1:29:05anthropic and Bill Gates and this whole
1:29:07argument. And so it's it's really
1:29:10disconcerting that the press eats it up
1:29:12so much.
1:29:13>> It fits a good narrative. If you combine
1:29:16it with the environmental narrative and
1:29:18you combine it with uh the job
1:29:22apocalypse, it's very portable for the
1:29:25public. It is a ratings hit. Like, oh,
1:29:28they're taking your water. They're
1:29:30they're polluting your oceans. They're
1:29:31they're
1:29:32>> think about think about it job. It's a
1:29:35perfect
1:29:36>> think about where Gates has come from in
1:29:37the last like since since 2000, you
1:29:40know, he he basically
1:29:43>> from his perspective
1:29:45>> eradicated an entire disease off the
1:29:47face of the planet, right? Very hard to
1:29:49do that. I mean that that took
1:29:51incredible coordination and money and
1:29:53research etc.
1:29:55>> As a result, he essentially became a
1:29:58whisperer to every single government all
1:30:01around the world. there's no world
1:30:03leader he didn't meet with. He was
1:30:05effectively a quasi unelected world
1:30:07leader. And then very cleverly, you
1:30:10know, he and Buffett co-opted all the
1:30:12money, right? Right? Cuz if you think
1:30:14about the giving pledge, it was a
1:30:16brilliant it was a brilliant form of
1:30:18reputation washing where you could
1:30:20essentially sign this pledge and you
1:30:22were invited to a party and you were
1:30:24invited to all this nonsense.
1:30:25>> But the reality is that you, you know,
1:30:27didn't have to give any of it, but you
1:30:29gave the perception that you could. And
1:30:31all of a sudden, you had the money, the
1:30:32influence, and the politics. When you
1:30:34die, nobody knows
1:30:35>> up to half, right? No.
1:30:36>> And so, you know, you represent now
1:30:38trillions of dollars of assets on top of
1:30:40all the influence and the credibility
1:30:42and your own money. and then your
1:30:43solutions. And so I think a lot of this
1:30:45doomerism can be explained by his desire
1:30:48to reestablish that central power
1:30:51structure that he frankly was probably
1:30:54near or at the top of up until co and
1:30:57then it all just came undone. I can
1:30:59explain it all to you Jamaat. It's very
1:31:01simple. He became the world's richest
1:31:02man. When he became the world's richest
1:31:04man, you become the most hated man on
1:31:06the planet. when he became the most
1:31:08hated man, it got in his head. And then
1:31:10he had that uh antitrust investigation.
1:31:13And instead of him being based and
1:31:15saying, "Fuck you. I'm gonna build the
1:31:17greatest company in the world. I'm going
1:31:19to keep building. I'm going to stay
1:31:20CEO," he felt all that pressure from the
1:31:22antitrust and being the most hated man.
1:31:24And he said, "Oh, I got to be loved. How
1:31:25am I going to be loved? I get rid of
1:31:26polio. I'll give all my money away. I'll
1:31:28do this giving pledge." The lesson here,
1:31:30folks, is don't quit your job. Just keep
1:31:33working. Keep building Once you
1:31:35quit your job, it gets in your
1:31:37head and then you have too much time on
1:31:39your hand and you spiral and then you're
1:31:42worried about the Pulitzer. You know
1:31:43who's not worried about the Pulitzer
1:31:44Prize? Elon uh Steve Jobs is not worried
1:31:47about the Pulitzer. Uh Satia is not
1:31:49worried about the Pulitzer. You know,
1:31:51like there's a lot of founders out
1:31:53there. Zuck's not worried about winning
1:31:54a a Nobel Peace Prize. Whatever it is,
1:31:58just keep working. It's enough already
1:32:00with the retirement. Incredible. How was
1:32:03he when he retired?
1:32:06>> Yeah. I mean, look, if you if you're
1:32:07worried about what your New York Times
1:32:11eulogy is going to be,
1:32:12>> then you'll be controlled by those
1:32:14people. Totally.
1:32:15>> And they're the most crazy leftwingers
1:32:17there are. So,
1:32:18>> you just kind of have to accept that.
1:32:20>> I think you're right, Jason. It's like a
1:32:21shark. You just can't stop swimming.
1:32:23>> Don't just keep building I mean,
1:32:25what go This is my advice to Bill Gates.
1:32:27Come on the program and start another
1:32:29company. Ben horror which is here. All
1:32:32he does is build companies.
1:32:33>> He might come on the pod.
1:32:34>> He should come on the pod. Let's bring
1:32:35him.
1:32:35>> You should actually have him on the pod.
1:32:36He'll be
1:32:37>> should be an intervention.
1:32:38>> He'll be feisty.
1:32:39>> Ben, what would you do if you were
1:32:41retired? You lose your mind.
1:32:43>> Yeah, that's the problem.
1:32:45>> Yeah, exactly. See, Ben understands.
1:32:47>> Great.
1:32:50>> Basic.
1:32:53Great job. All right, we'll see you all
1:32:55next time on the actual
1:32:57>> number one podcast in the world. number
1:33:00one.
1:33:02>> Love you, boys.
1:33:03>> Okay, thank you guys.
1:33:04>> Thanks, Ben.
1:33:05>> Great job.
1:33:07>> We'll let your winners ride.
1:33:10>> Rain,
1:33:15>> we open source it to the fans and
1:33:17they've just gone crazy with it.
1:33:19>> Love you, Queen of
1:33:27>> Besties are gone.
1:33:30That is my dog taking notice your
1:33:32driveways.
1:33:35>> Oh man, my appetiter will meet.
1:33:38>> We should all just get a room and just
1:33:39have one big huge orgy cuz they're all
1:33:41just useless.
1:33:42>> It's like this like sexual tension that
1:33:43we just need to release somehow.
1:33:49>> Your feet.
1:33:51We need to get mercury.
1:34:01I'm going all in.

About this transcript

This page contains the full transcript of CoreWeave Crash, AI Trade Cracks, Anthropic IPO Red Flags, Bill Gates Doomer by All-In Podcast, generated from the public captions YouTube serves with the video. The transcript has 17,352 words across 2,576 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.