YT YouTube2Text

YouTube2Text → Transcripts → Anthropic à 2 000 milliards : opportunité du siècle ou bulle IA ? - Martin Pavanello

Anthropic à 2 000 milliards : opportunité du siècle ou bulle IA ? - Martin Pavanello — Transcript

by La Martingale · 15,155 words · 2,278 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Two trillion dollars, that would be
0:02Anthropic's valuation at its IPO, which
0:04is expected for mid-November 2026. It
0:07is the biggest anticipated IPO of all
0:09time for a company that is barely five
0:10years old. Are we talking about a
0:12bubble, or is it the opportunity of the
0:14century? That is the question Martin
0:15Pavanello and I asked ourselves. In
0:17this episode, we break it all down.
0:19Where does this value come from? Who is
0:21actually making money, and what should
0:22we expect in the coming years between
0:24mass adoption and a fierce war between
0:26the different models? Martin is the
0:28co-founder of Mister IA, which supports
0:29more than 1,400 companies in their AI
0:31deployment. We offer you an episode to
0:33understand what this changes for your
0:34investments and to look at the
0:36opportunities that no one is watching.
0:40>> Welcome to the Martin Gal.
0:42>> Thank you.
0:42>> So, Martin Pavanello, you are the
0:44co-founder and CEO of Mister IA. Uh,
0:47well, you aren't exactly in personal
0:48finance, but you know the world of
0:50artificial intelligence well. Remind me
0:52, explain to me today, excuse me, to
0:54explain what I would like to talk about
0:56with you and everyone else. Uh, well,
0:58it's not about talking about AI in its
1:00practice, but you do a lot of that at
1:02Mister IA, but rather its consequences,
1:04particularly its consequences for
1:06companies, stock market consequences,
1:08how there are so many things happening
1:09within it. In fact, we talk about
1:13SpaceX, we think of rockets, but inside
1:15, in fact, there's Grok, there's, I
1:17believe, Starlink that I was mentioning
1:19earlier. There is obviously SpaceX, but
1:22so all of that is very linked. Starlink
1:25and AI will be very closely linked in
1:27many ways. Uh, we also see, finally,
1:30the likely upcoming IPO of Anthropic,
1:33possible and probable from OpenAI, and
1:35how the costs are distributed within
1:38these companies; what is happening
1:40inside? What is valued, overvalued,
1:43undervalued? Anyway, we're going to try
1:45to break all that down. Does that sound
1:46good to you?
1:46>> Very good.
1:47>> OK. Why do you know these sectors well?
1:48Explain it to me.
1:49>> I know them well because I founded,
1:50yeah. Mister IA three years ago now,
1:52with a very simple goal: to say how do
1:54we deploy this super ChatGPT that
1:55everyone is starting to use a bit but
1:57which is still a little nebulous; how
1:59do we help companies use it. So, we
2:01created a consulting and training firm,
2:03and for the past three years, we’ve
2:04been deploying AI for our clients every
2:06single day. In a very pragmatic way, I
2:08choose the right tools, train them, and
2:10manage their ChatGPT, Claude, and other
2:12licenses. So today, with 1,500 clients
2:14—SMEs and mid-sized companies across
2:16France, plus some major accounts—we
2:18see on the ground every day what will
2:20be reflected in the stock market
2:22tomorrow regarding Anthropic and OpenAI
2:24: which one is actually being used? Are
2:26people really paying? If so, are they
2:28all paying € 5 a month or € 200?
2:30What is actually happening in these
2:31companies? We see all of this on a
2:32daily basis, and that’s why we dig
2:34deep into these topics and have such
2:35good insight. Very clear. Listen, one
2:39of the very first questions—and I
2:41think it's interesting that you have
2:43your hands in the grease like I do—is
2:46to understand what is ultimately the
2:48value of a token, the value of
2:50intelligence. Because right now, I’m
2:54using Claude and ChatGPT a lot, and
2:57I’m using them more and more; I’ve
2:59done my "harnessing," I won't go into
3:01the details on that. I’m starting to
3:05use Grok quite a bit. I tested Mistral
3:08a bit, and it’s definitely getting
3:09better. I also tested DeepSeek. I’m
3:12talking about all of this over, let’s
3:14say, the last six weeks. Uh And the
3:18further I go, the less difference I see
3:20. So, if they all offer more or less
3:22the same thing, some even in open
3:24source, where all I need is a bit of
3:26electricity and a Nvidia chip for a Mac
3:28Mini,
3:30>> well, a token isn't actually worth much
3:33.
3:33>> It’s worth less and less. It’s
3:34worth less and less. That much is clear
3:36. If you look back three years, there
3:37was certainly a huge gap between an
3:39Anthropic or an OpenAI, who had
3:40super-models, and the rest of the world
3:41, if I may caricature, who couldn't
3:43manage to build an intelligent model.
3:44Today, well, we’ve seen Chinese open
3:46source catching up at full speed, and
3:47so what we’re seeing in companies, at
3:48any rate, is that the token is becoming
3:50less and less valuable, while the
3:51product is becoming more and more
3:52valuable. That’s clear, and it’s
3:54also why these companies will have an
3:55IPO—we’ll talk about that later—
3:57an Initial Public Offering, so, going
3:58public on the stock market. Yeah,
4:00>> that's it as well. Exactly. Whereas
4:02with Chinese models, we'll talk about
4:04that a bit too, that's not the case yet
4:05today; these companies, all these
4:06Chinese labs, earn much less money. So
4:08indeed, we're starting to see that
4:09we've reached a level, in my opinion
4:11anyway, we've reached a level with
4:12these models, in their performance,
4:14that makes it increasingly hard to
4:15distinguish them. In any case, if we
4:17really take the highest-performing
4:18models, I wouldn't even talk about a
4:20GPT-4 today, but rather an Opus 3.5.
4:23Which in its most max or extra version,
4:25I think it's max at Anthropic,
4:26>> compares more to GPT-4o on the OpenAI
4:29side. The two, eyes closed, even using
4:31them all day long, I can't tell you
4:32which is which. It's just impossible.
4:34However, in terms of interface, tools,
4:35and which one is connected to more
4:36tools in my workflow, then I can tell
4:38you which one performs better between
4:39GPT and Claude.
4:40>> And so, all this is important for
4:41several reasons, right? Uh, you tell me
4:42if I'm right or wrong in your opinion,
4:44but the first is that, one, there is
4:45competition. I mean, when you take
4:47Google or Instagram—well, Instagram
4:50now has a bit of TikTok—but for
4:51Google, there's no competition. I mean,
4:54you want to search, it's Google. Those
4:56who use Bing, you know, you laugh,
4:58right? You go, "Oh, you're on Bing,
4:59that's funny." But Ecosia, oh yeah,
5:01well there you just think, "You're an
5:03idiot." But yeah. And so, the idea is
5:06that by having competition at roughly
5:07the same level, it can only drive
5:09prices down, because at some point, the
5:10best can't say, "Well, I'm better than
5:12the others, so it's normal that I'm
5:14twice as expensive, come on." And
5:15besides, there's a double cool effect,
5:17which is that some are open source.
5:20Once again, this is a bit geeky, but
5:21once you've done it, you understand
5:23quickly. Uh, if you have a Mac, or even
5:26certain PCs with an Nvidia chip inside,
5:28or there are AMD ones that allow for
5:31this, you can install a model on it and
5:33you don't pay anymore; you pay zero.
5:36It’s running locally. So it's often a
5:38bit slower, but hey, if you want to do
5:40tasks that are easy and you're not in a
5:43hurry, it's free.
5:44>> It works really well, too. So, this
5:45intelligence is surely destined to be
5:47worth zero. And that is the big
5:49question when we talk about companies
5:51like Anthropic, with their 75 billion
5:53valuation or 70 billion, and those that
5:55reached a 100 billion valuation in
5:57their latest round. We don't know what
5:59the IPO will be. Latest round at 100
6:00billion valuation, what's hiding behind
6:01that? Because if the model costs you
6:03zero or if it’s being competed with
6:04by a free Chinese model, then really,
6:06why would we pay for Anthropic tomorrow
6:08? So, there’s the tool side, but
6:09there’s still an issue to look at
6:11with them, because what we’re saying
6:12is very true regarding the gap between
6:13open source and closed models up until
6:153 or 4 months ago. Today, the
6:16performance gap between the best
6:18open-source models and their
6:20closed-model equivalents has never been
6:22wider. If you look at today
6:24>> between the Chinese and the Americans,
6:25>> Exactly. Because if you look at the top
6:2710, top 15 best models, there’s the
6:29Artificial Intelligence Index, which
6:31has a fairly serious benchmark for
6:33comparing models. And in the top 15,
6:36you have 5 or 6 Anthropic models, you
6:37have three or four Meta models now that
6:39are very, very good. You have three or
6:41four OpenAI models, and you have to
6:43wait until the 13th, 14th, 15th, or
6:4416th place to find an open-source
6:46equivalent. So we're already starting
6:47to see a bit of a lag there. So,
6:49they’re going to address that.
6:50Secondly, one of the major issues, in
6:52my opinion, that keeps prices from
6:53dropping to zero even with competition,
6:55is that we're starting to see
6:56open-source models, especially from
6:58Chinese labs, that potentially refuse
7:00to release their best models to the
7:02market. It has been stated by the
7:03government. I mean, we are also at a
7:04point where, when I see American
7:06competitors generating 50 or 100
7:08billion in cumulative revenue, or
7:09rather 100 billion in cumulative
7:11revenue for the labs in a year. Well,
7:12me, as a Chinese player, I'd like to be
7:14nice. I like using soft power and I
7:15like trying to challenge them with open
7:17source if I can generate revenue from
7:18the outside, because it's huge, I mean
7:20not just, well there’s American money
7:22too, but you still have a lot of
7:23Europeans paying for all of this. If
7:25you can take 10, 20, 30, 50 billion a
7:26year from them, well, that's money
7:27coming into our economy. And I think
7:29they’re going to close things off a
7:29lot. However, it’s going to raise a
7:31lot of questions, like if we no longer
7:33have access to open source models,
7:35won’t we just go straight back to the
7:37Americans and give our money to them?
7:39>> And I think yes. And I think yes. And
7:42in any case, I am convinced that this
7:43revolution, at least for the mass
7:44market, for the general public, we are
7:46stuck with the Americans and that’s
7:47just too bad for us. We haven't
7:49invested anything in data centers for
7:5020 years. For 20 years, we haven't been
7:51training as many engineers as they have
7:53. For 20 years, we’ve under-invested
7:54in this tech while they’ve been
7:55investing heavily. So today, we’ve
7:57reached a point where competing with an
7:59Anthropic or an OpenAI is almost
8:00impossible. Now, is that a risk or not?
8:02I don’t think so. And notably,
8:04there’s a stock market angle I like
8:05regarding this, which is the idea that
8:07the big risk of putting everything in
8:08American hands is the Cloud Act. It
8:09means everything we use with OpenAI or
8:11Anthropic is subject to the Cloud Act.
8:13Which means the Americans can get their
8:14hands on our data at any time. It’s
8:16legal, it’s contractual, and it’s
8:17the reality. But the real reason behind
8:20all that and the real question is, what
8:22would be the U.S. government's interest
8:24in seizing Microsoft data, or
8:25specifically ChatGPT data found on a
8:27Microsoft server? If they did that,
8:29Microsoft would lose half its market
8:30valuation in a single day. If they did
8:32that, Amazon via AWS would lose half
8:33its stock market value in a day.
8:35That’s trillions of dollars that
8:37could go up in smoke if they really
8:40invoked the Cloud Act. Anyway, we all
8:43have the same reflex. When an
8:44investment works, we want to be part of
8:46the journey. It’s logical, it’s
8:47human. The only problem is that by the
8:49time everyone realizes something is
8:50working, well, generally the bulk of
8:52the progress has already been made. A
8:54crowd confirms that something is
8:56popular, but it doesn't guarantee it's
8:58a good deal. And that's why I'm
9:00thrilled to introduce Carmignac, who is
9:01supporting La Martingale for this new
9:03season. Carmignac is an asset
9:05management firm founded in Paris in
9:071989, so more than 35 years ago. And
9:09there are two things that characterize
9:10them. The first is active management.
9:13In concrete terms, there are people who
9:15make decisions, analyze, invest, and
9:17take responsibility. We aren't talking
9:18about an index that's automatically
9:20replicated. And the second, which is
9:21the one that interests me the most, is
9:23that they are independent. They don't
9:25belong to any banking group. What does
9:27that mean? It means there is no one
9:29above them telling them they must
9:30follow this or that strategy that
9:31everyone else is following, or to sell
9:33this or that in-house product. And when
9:34they have a conviction, they can see it
9:37through, even when it is unpopular.
9:38They manage equity funds. They manage
9:41bond funds, diversified funds, and they
9:42are distributed through most networks.
9:45So, there's a good chance it's already
9:46accessible to you. So, if you want to
9:48take a look, mention me, it's at
9:50go.com/lamartingale. I want to remind
9:54you that all investment involves risks
9:55and that the information provided does
9:57not constitute investment advice. So,
9:59remember that name well; I want to say
10:00thank you to Carmignac. And now, let's
10:03go back to the episode. I don't know if
10:05I really share that with you, Martin.
10:06You see, I tell myself that if power
10:09exists at some point, we aren't
10:11necessarily obligated to know about it.
10:13Meaning that there are still strong
10:15suspicions regarding certain large
10:16tenders, certain things where today,
10:18right now, some large companies tell
10:20you, "Well, here it is, there are ways
10:22to make it sovereign." Once again, even
10:24some of these private companies don't
10:26sell it to you as open source, but
10:28privately, their model that you can run
10:30on your own servers. So, are there
10:32backdoors or not? I don't know, it's
10:33always a bit the same. Are there risks
10:35of data leaks on one hand? I don't know
10:37. Now, with a Donald Trump at the helm,
10:39all that, you have no idea at all. You
10:41know, he's too unpredictable. You see,
10:44the day he tells you—and this is what
10:47some leaders or political analysts
10:49point out at times—he says, "Hey you,
10:52Denmark, give me Greenland because I
10:54want it." No. "Well, I'm going to cut
10:57off all your data, then your hospitals
10:58won't work, your cars won't work, and
11:00you won't be able to do anything." I
11:02mean, he is capable of getting into
11:03that kind of standoff, and that might
11:05hurt the stock market valuations of
11:06those people. But what choice do we
11:08have? Once again, we're not going to go
11:09to the Chinese, you see.
11:10>> That's the problem. I personally think
11:11we don't have many other choices,
11:12unfortunately, and that's why I was
11:14talking about, quote-unquote, we should
11:15have acted sooner, because we could
11:16have invested long ago to be better at
11:17this. And I personally think he's ready
11:18for anything, but he's never done that.
11:20I think he won't do it. And once again,
11:21I think behind all that, there's still
11:23the savings of the American people.
11:24Microsoft wouldn't be worth 5 trillion
11:26or 4 trillion on the stock market like
11:27it is today if it weren't
11:29oversubscribed by Americans. So the
11:30risk is very, very, very high for them
11:32to mess with our data, all the same.
11:35>> I hope you're right, but somehow I
11:38mostly hope that Europe gets moving and
11:40that we finally push companies like
11:43Mistral or others. We're going to try
11:46to analyze a bit, precisely, since
11:49we've done a preamble on the value of
11:51intelligence. Which is interesting for
11:53those listening who don't know these
11:55topics well. This intelligence is
11:58dependent on quite a few things. So
12:01what's interesting is also to see the
12:02underlying factors. So we're going to
12:04talk today about the value of Anthropic
12:05as an example, if that works for you.
12:06Yeah. And potentially also about what's
12:09underneath, because for Anthropic, if
12:11tomorrow in France, for example, I cut
12:13off your electricity, that's its fuel.
12:15If I cut off your data center somewhere
12:19else, that's its engine. So somehow, if
12:22you don't have the engine—that would
12:24be more like Nvidia, AMD, or some
12:26others—or the fuel, which is the
12:28fluid that allows all of this to run.
12:31There you go, if the price of fuel goes
12:33up, well, the price of...there's a huge
12:35amount of interdependence, and in the
12:36end, why, and what does it get me?
12:39That's the whole question. that we need
12:42to ask ourselves. So, basically, we're
12:44going to talk about data centers, we're
12:45going to talk more about Nvidia, we're
12:47going to talk about electricity and
12:48energy, and so on. Does that work for
12:49you?
12:50>> Very well.
12:50>> OK. Uh, where do we start?
12:52>> Valuation, I think it's going to happen
12:54, it’s going to hit us in the coming
12:55weeks. Maybe in the next few hours when
12:57we...But we have the Anthropic IPO
12:59that’s coming up. So
13:01>> let's go, I'm with you. So, Anthropic
13:03IPO, the question is as follows. You
13:06mentioned a trillion to me; those were
13:08valuations from funding rounds.
13:10>> For the IPO, are we looking at closer
13:11to double that, is that right?
13:13>> We're looking at more like two trillion
13:15for Anthropic, and for OpenAI, we were
13:16also talking about two trillion, and it
13:18turns out OpenAI is currently raising
13:20funds to be announced in the coming
13:21days privately at a 1.4 trillion
13:23valuation. OK,
13:24>> that's quite consistent with the fact
13:25that within 6 months, because for them,
13:26it should happen in 6 months. We will
13:27reach two trillion.
13:28>> Just for the younger ones listening or
13:30who aren't familiar with the stock
13:32market, from memory—and I’m
13:33estimating here—the first company to
13:35pass the 1 trillion mark in valuation,
13:37I believe it was Apple, and I think it
13:39was more or less during Covid, so we're
13:41talking about something 5 or 6 years
13:42old. You see, so here we are at 2
13:45trillion for companies that were 5 or 6
13:47years old...Anthropic, in any case,
13:49didn't exist. The whole thing is just
13:51enough to bang your head against the
13:53wall. It's just, it's nonsense.
13:55>> Yeah, completely. Though it is
13:57impressive, but if you look at the
13:59numbers in a bit more detail, it's not
14:01that crazy. I mean, specifically the
14:03ARR; we're talking about ARR, which is
14:05Annual Recurring Revenue.
14:06>> It's not necessarily a bubble.
14:08>> No, I don't think it's a bubble at all
14:09today. For one, I think the underlying
14:10foundation isn't a bubble at all. The
14:11underlying point is: are we
14:13overestimating the use of AI today and
14:14so on? No, not at all.
14:16>> Is AI going to work? So it's not going
14:17to work.
14:18>> But you see, and we talk about this
14:19often, and there’s, there’s
14:20something I like: it's when we look at
14:22the households—how do we say?—
14:24households in the United States that
14:25are equipped with AI. We're talking
14:26about around 2%of households in the
14:27United States, which are mostly a
14:28tech-savvy population, right? And
14:30>> rather avant-garde, there are 2%who pay
14:32for AI today.
14:33>> 2.2%, that's incredible and
14:35>> it's it's very little.
14:35>> I'm telling you, there is room for
14:37growth,
14:37>> there's a margin that is enormous when
14:39you look at Netflix, we're more around
14:4070%. So on one hand, we have that. On
14:42the other hand, we still have a company
14:44like Anthropic that is making 75
14:45billion in recurring revenue. I remind
14:47you that at the end of 2025 we were
14:48talking about 9 billion in recurring
14:50revenue. No company in history has
14:52added 70 billion in ARR in less than a
14:53year like that.
14:54>> So it's already a company that, even if
14:56it had its IPO today, would be worth
14:58about 30 times its ARR.
14:59>> ARR is a bit like a form of annualized
15:01revenue equivalent. It's called annual
15:04recurring revenue. Basically, it's the
15:07number of subscriptions signed today
15:10that should, over the next 12 months,
15:12we agree, annualized, represent 70
15:14billion in recurring revenue.
15:17>> That's it. With a zero growth
15:18assumption. Meaning if I stop growing
15:21now, with what I have in stock over the
15:23next 12 rolling months, I'll make 75
15:25billion or 65, but 65-75, and OpenAI
15:27officially announced 50. We know they
15:29are a bit behind, so those are roughly
15:31the figures. I thought OpenAI was
15:32monetizing better than Anthropic.
15:34>> No, rather worse. Rather worse. They
15:36have more users. They have a little bit
15:39more users. However, when you look at
15:41the details of both, Anthropic today
15:43has 92%of its paid subscriptions at 20
15:45dollars a month, and they already have
15:46almost 8%of their paid subscribers on
15:48100 dollar-a-month plans. Whereas
15:51ChatGPT has more users, but many free
15:53ones and many subscriptions also in a
15:55version that's 7-8 dollars I think,
15:57something like that per month. So they
15:59monetize rather worse. ChatGPT,
16:01>> I laugh because I wonder how one can
16:03live with a 20-dollar subscription. I I
16:05crush them all the time.
16:07>> Well, you've just seen another reserve
16:08of growth that will be able to, that
16:09will be able to explain the 2 trillion.
16:10What about this curve where, indeed,
16:12all those who are—maybe not all of
16:13them, but already within those 92%
16:15paying 20 dollars—at least half of
16:17them will quickly switch to 800 dollars
16:18. We're talking about multiplying
16:20revenue by 4 in that segment without
16:22having 1 euro of user growth. I manage
16:25licenses for our clients every single
16:27day. We have 15,000 cloud licenses at
16:28Misteria, 13,000 cloud licenses. We see
16:30among our 13,000 cloud licenses that
16:32they all started at $ 20 a month, and
16:34today 10%of my fleet is at $ 100 a
16:35month, but it cost me € 0 to move
16:37them from $ 20 to $ 100; it's the same
16:39thing. If we go back to the genesis of
16:42our conversation, the first point, many
16:44people are afraid the price of these
16:47tokens will rise because, personally,
16:49if you tell me tomorrow that Claude is
16:51doubling its price, I’d say, "Fine,
16:55I’ll go from $ 200 to $ 400, it's
16:57still worth it in the end." They
16:58shouldn't hear me say that, but they
16:59know it very well anyway.
17:01>> The thing is, and this is where
17:03antitrust comes in, they would all have
17:05to do it at the same time, because at
17:07some point, if Claude says, "It’s X
17:09price"—I was talking about harness
17:11earlier—I am now capable of switching
17:13from one AI to another during the day,
17:15or depending on the task, really
17:17on-demand. Tomorrow I can just switch
17:21to ChatGPT, and oh well, I don't care,
17:24or Mistral, as long as it's equivalent.
17:27>> That’s very true for the token price
17:29and cost, especially when we're coding
17:31and can plug in any API. So they have
17:33no interest in raising prices too high,
17:35unless it's for an exceptional model
17:37we've never seen, but for daily use,
17:38they don't really have an incentive to
17:40go very high. However, regarding the
17:41tools themselves, it's a bit less true.
17:43When I have 100 Claude licenses in a
17:45company, you don't just switch tools
17:46overnight. So if Claude becomes a
17:48little more expensive than OpenAI, it's
17:50not that easy to change. Still, the
17:52competition is so fierce between the
17:54two, and I think a company like
17:55Anthropic is capable of losing hundreds
17:57of billions of dollars before becoming
17:58profitable, just because they want to
18:00go from 2%of households equipped—and
18:01it's roughly the same for businesses—
18:03to more. I don't know if it's 2%or, for
18:05that matter, it's certainly less, but
18:07they want to go to 50%and 70%. So there
18:09is such a war between the two. They are
18:10so willing to lose money, I can't
18:12imagine them raising their prices
18:13tomorrow at all, and I think they have
18:15a market to go after as long as the
18:16company isn't doing 300 or 400 billion
18:18in revenue. We're not there yet. Today,
18:19Open Entropic announced 16 billion in
18:21actual revenue for the last quarter
18:23versus 4 billion the previous year. So,
18:25it's going fast, though. But they are
18:27still far from 500 or 400 billion
18:28dollars in revenue. So, I think we
18:31shouldn't face too big an increase in
18:33price. Besides, I should point out,
18:35regarding subscriptions, because that's
18:36still a big part of their revenue:
18:38subscriptions paid by companies have
18:39tended to decrease over the past year,
18:40rather than increase in price, notably
18:42because both are ready for an
18:43impossible war and. And it's a good
18:44thing that Google is not far off,
18:45because they are also entering this war
18:47. I don't think there are many others
18:48capable of competing with them.
18:49>> Grok.
18:50>> No, I don't think so on the enterprise
18:51side, I don't think so. We really see
18:52very little of them. They really suffer
18:54from Elon Musk's image. Now, I'm biased
18:57by French companies, as I only work in
18:58France or Europe, but Twitter and Elon
19:00Musk on one side, and Twitter, which is
19:02also quite anti-enterprise, it really
19:04pushes people away and we come back to
19:06our problem. The model is very, very
19:08good and certainly as efficient as the
19:10others and perhaps one day cheaper, but
19:11the tool, no. The tool is not a tool
19:13for the working world. It's not a tool
19:15that companies want to say, "Hey, I'll
19:16equip all my staff with Grok at € 30
19:18a month." So, I don't think they have
19:21any vocation to compete on that. In
19:23terms of the value chain, I imagine
19:25them elsewhere. It's our friends at
19:26SpaceX with Grok.
19:27>> OK. On that, well, clearly, we see it's
19:31growing, it's great. You say 75 billion
19:34, in fact, somewhere when you tell me
19:36about the mind-blowing growth rates and
19:38the growth potential behind it, that's
19:40what's important, because that's what
19:42you look at when you're an investor; I
19:44do quite a bit of start-ups and private
19:46equity. You're going to look at the
19:48revenue. Well, it's important, but not
19:50always that important. What you're
19:51looking for is revenue growth. And
19:53potentially there, what you're telling
19:55me is 2.2%of American households. So
19:57actually, you're saying the growth
19:58potential isn't inexhaustible, but it
20:00is massive. You can really go, you're
20:03saying easily 10x, 50x, plus businesses
20:06. In fact, you end up at your 75
20:09billion, times 10 is 750 billion, and
20:11times 50, well, you're past the
20:12valuation. Knowing that these things
20:14can be quite profitable. OK. Uh, I
20:16could say there isn't necessarily a
20:18bubble, but that's just revenue.
20:20>> Yeah.
20:20>> What's underneath it? Because, well,
20:22that's all very nice, but are they
20:24actually profitable? Are they still—
20:27often these are startups, we know early
20:29on they're in such an acquisition phase
20:31that before being profitable—look at
20:33Amazon, it took them 15-20 years and
20:34they only were because of AWS.
20:36Ultimately, Google took a very long
20:38time. Meta, don't even get me started.
20:40So, where do we stand on that?
20:42>> For now, it's a money pit. For now,
20:43it's a money pit. And, they are not
20:45profitable, that much is clear.
20:46>> But wait, Anthropic was talking about
20:48it and announced a few weeks ago that
20:50they had their first profitable quarter
20:51. When you look at the numbers, if you
20:54look at the 2025 figures, Anthropic
20:56officially announced 4.5 billion or so
20:58in revenue and 8 billion in losses at
21:00the end of the P&L. Watch out for
21:04anyone who saw—I saw many articles on
21:05this—because Anthropic announced a 42
21:07billion loss. Of that, 34 billion are
21:10convertible bonds. They aren't cash
21:12losses. It's just that on the income
21:14statement, it shows minus 34 billion,
21:16but it's not a cash or operational
21:18issue for the company at all. So the
21:21real operational loss is 8 billion. So
21:23that's a fact, these are companies
21:24capable of losing a lot, a whole lot
21:26more. Now, Anthropic is still claiming
21:28to be profitable this last quarter, and
21:30I believe them. I believe them, at
21:32least regarding their regular business,
21:34I would say their day-to-day activity
21:35of selling licenses.
21:37>> But you know what? I mean, it's just,
21:38you have to, I don't know, you have to
21:40believe it or not. I mean, a company
21:42that's ready to go public,
21:43>> anyway, they're going to have financial
21:45reports that are transparent
21:46>> completely.
21:47>> You can't—they can't just say they're
21:48profitable when they're not. No, but
21:50today when you analyze the numbers,
21:52they're not profitable, and when they
21:53release their official 2026 figures,
21:55they'll show us they're not profitable.
21:56The only subject is what's the role of
21:58compute costs in all this. And for me,
21:59that's the central issue. There are
22:01plenty of experts who say, "Actually,
22:03Anthropic will never be profitable
22:04because they'll never manage to sell
22:06enough licenses to cover compute costs.
22:07" That would mean the business model
22:09itself for these labs isn't profitable.
22:11Compute.
22:12>> So compute is that today, when you're
22:14Anthropic, you have to buy Nvidia chips
22:16that cost a lot of money. Training.
22:18Actually, Anthropic happens to use
22:19Amazon instead. But take OpenAI: OpenAI
22:21buys a lot of chips, uses Azure to
22:23train its models. So, high model
22:25training costs and, behind that, high
22:27inference costs. Every time you enter a
22:29prompt into ChatGPT, it runs a chip in
22:31Mississippi, more or less, and it costs
22:33them.
22:33>> So, coming back to what I was saying
22:35earlier, one buys its own engine, the
22:37other rents it.
22:37>> Yeah. OpenAI is more into renting for
22:40now with AWS, but those are still
22:41compute costs; they're just not
22:43capitalized the same way, but they're
22:44still compute costs. And the big
22:46question is: based on their current
22:48business, which in my opinion is set to
22:49stay the same for years to come, is it
22:51"I sell licenses to companies at 100
22:53€ per month per collaborator and I'll
22:54make 500k in revenue with that"? On
22:57that, today, the cost of compute should
22:58, for me, be largely sorted out. Why?
23:00Because today, when you take the Opus
23:023.5 model. If this model remains the
23:05same model and doesn't evolve over the
23:07next five years, every day that passes,
23:09there will be more users than the day
23:11before. And every day that passes, user
23:131 will make more prompts on Tuesday
23:14than on Monday, and more on Wednesday
23:16than on Tuesday. So usage is going to
23:18increase because these models are
23:19already incredibly high-performing. And
23:22so when we see today, let's take an
23:23example, if OpenAI announces, I don't
23:25know, 50 billion in revenue this year.
23:27If they announce the model makes about
23:2910 billion in profit. But if we take
23:30out the 30 billion in compute costs,
23:32well, then we're clearly in the red.
23:33Yes, but what are these compute costs
23:34for? Are they to run the current
23:36business model or are they for
23:37fundamental research to one day reach
23:39AGI, which is quite different. And
23:41today in the US, under American
23:42standards, all R&D must go into the P&L
23:44, into the income statement, which
23:45isn't the case in France. That's why
23:47I'm also looking forward to comparing
23:48it a bit with Mistral, so I don't say
23:49anything stupid, but I think in France,
23:51we don't have to put all R&D costs in
23:52the P&L. And so I could very well say,
23:54actually, out of my 30 billion in model
23:57training or model costs, okay, there
23:59are 5 that are real costs, 10 that are
24:01real costs, so I'll take those out of
24:03the earnings, it's fundamental research
24:05.
24:05>> And that's what we'll have to do.
24:07>> Meaning that you just do your accounts,
24:09right, in a company, you talk about the
24:11P&L, so you have money coming in, money
24:13going out, uh, and so research and
24:15development, well for some, as you say
24:17in the US, it's money going out, it
24:18costs and that's it. Uh, in France, you
24:21can say, "Oh no, that's outside, it's
24:23not really my core business. My
24:25business is, I sell software, I
24:27maintain software, I run it, and so on.
24:31But everything related to research and
24:32development, that doesn't count. So,
24:33even though technically you have
24:34engineers and teams working on it, it
24:36costs you money, so it should decrease
24:38your bottom line.
24:39>> That's it. So when we talk about EBITDA
24:41, for me, EBITDA is what I'll invest
24:42rather at the end, it's after my net
24:44income. So my EBITDA isn't impacted by
24:46my 10 billion in investment. Here, it
24:48turns out that's the case. So, in their
24:50P&L. So I think these are companies
24:52that have found their business model,
24:54contrary to what many say. I think they
24:56will be profitable in this business and
24:58that it is a business that can be
24:59profitable. Moreover, experts who are
25:02in favor of the profitability of these
25:04companies are banking on margins of
25:05around 15, 20, 25%. They’re not SaaS,
25:08but still, those are nice margins on
25:09huge numbers.
25:10>> I’m torn on this, and that’s where
25:11I think I can put this to you, and
25:13maybe you’ll agree, and perhaps
25:14there’s no real answer to what I’m
25:15about to say. But
25:17>> it’s the first time—I was going to
25:19say one of the first times, but it’s
25:20the first time I haven’t questioned
25:22it. Meaning if I have to put 200 bucks
25:25per person in my company, you know, if
25:27I have to put 2,000, 3,000 bucks a
25:29month into the company so that they all
25:31have a good AI, a super AI, or several
25:33good AIs for that matter, I don’t ask
25:35any questions. I know the time saved is
25:39so phenomenal, the lead we gain over
25:40those who don't have it is phenomenal.
25:42I have no issue with it. So you think
25:45to yourself," Well, this is crazy. ""
25:46Because, plus, I never stop. "Meaning,
25:49I think I took my first ChatGPT
25:50subscription about a month after the
25:53launch.
25:54>> Was it 3.1, no, or
25:56>> 3.5 in November 2022, so it must have
25:59been January 2023 or something like
26:01that, and since then I’ve never
26:03stopped paying for AI, and I usually
26:05have more than one. However, I’m not
26:10loyal at all. Meaning, as I said
26:12earlier, you know, I’m a big
26:14flip-flopper. I mean, maybe I’m an
26:16outlier, but I don’t think so. Yes,
26:18you are.
26:18>> I think you are. I think you are
26:20because you happen to be a super user.
26:22Uh, you belong to...Well, I didn’t
26:24include it here, but there is a study,
26:27or rather a post that was very
26:28interesting by a former ChatGPT
26:30co-founder that was done recently, by
26:32Andrej Karpathy on Twitter, so maybe we
26:34can find it, which explains a bit how
26:37AI users are broken down. And he
26:39actually says that today, worldwide—I
26:41don’t want to say anything foolish—
26:43but 50 to 60%have never even tested a
26:46paid AI, already 50 to 60%.
26:48>> So they have access to really average
26:50models. Then you have, uh, 10, 20, 30%
26:53who use ChatGPT with a paid license but
26:55just for research. And actually, there
26:58are several layers like that, and in
26:59the end, if you take the best models
27:01that exist in the world today, only
27:024,000 or 5,000 people have access. And
27:04so I think your example is precisely
27:06that of a super user. You’re in the
27:08first or second layer, not the 4th or
27:105th one I mentioned at the beginning.
27:11So you are one of the super users for
27:13whom it is quite simple to switch. But
27:15I want to tell you, if you had been
27:17early like that on SaaS, maybe you
27:18would have also said," I'll never in my
27:20life subscribe to Salesforce or HubSpot
27:21, it makes no sense. "" I'll create my
27:23own SaaS 10 years ago and it would be
27:25super agile, etc. "It’s a bit the
27:26same with AI; you’re more the type of
27:28profile that can switch. I see it in my
27:30company already, we have 200, just for
27:32us, 200-250 Claude licenses; it’s not
27:34that simple to switch. And that’s
27:36where you hit what all organizations
27:37are going to face, it’s, uh, you’ll
27:39have to change all the licenses,
27:40re-invoice things, recreate all my
27:42projects, recreate all my skills. And
27:43in fact, it’s not as simple as that.
27:45And that’s where the tool layer of
27:46these tools comes in. And for me,
27:47that’s what fascinates me, we talk
27:48about AI, but I think the real strength
27:50of these tools is not the AI, it’s
27:52really the tool. And they manage to
27:54create tools that lock you in. And I
27:56see it, if tomorrow I disconnect Claude
27:57from my employees 'mailboxes to put in
27:59ChatGPT, well, change management for
28:01just 200 employees will be a nightmare,
28:03and so I am still quite locked in. So I
28:05think you are indeed one of those
28:06people who digs deep, is very
28:08interested, and can switch quite
28:09quickly, and even then, if you change
28:11your 10, 20, 30 licenses, I don’t
28:12think it’s that easy. But you have it
28:14at the individual scale, yes, you’re
28:15right at the individual scale, sorry,
28:17but at the organizational scale, I
28:18think it’s much harder, and they are
28:20betting on that. Very clear. Listen,
28:23I’m working on it anyway because I
28:25also know that it’s part of the
28:26things I’ve always hated being locked
28:28into, and it has always been my
28:29obsession. I've had too many people on
28:32my podcasts here and there whose
28:33companies closed down or whatever
28:34because they were dependent on a single
28:36technology. Uh, and so I absolutely
28:38want to be independent. Uh. Now that
28:42we've seen all that, it's good because
28:44we've laid a lot of the groundwork. You
28:45see, there's one last constraint before
28:47you explain the split to me again. What
28:49is the cost behind the underlying
28:51technology and the split in the, uh, in
28:52the IR? Uh, one last question I have. I
28:55was, I did an event the other day at
28:56one of my partners called Brevo with a
28:57lot of marketing directors and so on.
28:59One of the questions that kept coming
29:01up from the marketing departments is:"
29:04Tomorrow, our costs are actually going
29:06to explode. "Meaning, if, and it's a
29:08bit like what I was saying earlier, if
29:10they multiply by 2, 5, or 10, we're
29:11basically stuck. One of them gave the
29:14example of WhatsApp and said," Well,
29:16recently at WhatsApp, they told us:' To
29:18be on WhatsApp Business, it's x 2 '. "
29:20And well, you need it so much for part
29:22of your customer management, I mean for
29:24after-sales service and customer
29:26support, that you just look at it and
29:28say," Okay, I wish it weren't, "but you
29:31complain a bit, and they look at you
29:33and say," Sorry, "and that's it.
29:36>> Is that a real risk, or does this
29:38competition, slash the open source
29:40aspect, protect us from that a bit? No,
29:42it's a real risk. It's a real risk like
29:45with any software today. The best
29:47example we have in France is talking to
29:48real estate agents about what's
29:50happening between SeLoger and Le
29:51Boncoin today, who go to the real
29:53estate agents every year and say," It's
29:54+15, +20, + 30%every time you post an
29:56ad on my site. "And they all find
29:58themselves really, really under that
30:00rule, and so they are basically being
30:02choked by it. So it's the same thing,
30:03it's the same thing. We have that risk.
30:05Now, what I tell myself is that this
30:07risk won't happen within the next 10
30:09years. We already have 10 years to have
30:10fun. We've already been able to have a
30:11lot of fun with all this for 3 years.
30:12We still have 10 years to transform our
30:13companies. And when that moment comes,
30:15in other words, we won't be able to say
30:17we didn't know. We won't be able to say
30:19," Oh shoot, we haven't started our AI
30:20transition yet, it's still a bit too
30:22early for us. "So, I personally think
30:24that a company that trains 100%of its
30:25staff today, that really takes the
30:26subject seriously and says," I'm going
30:28to provide € 100 licenses to everyone
30:29and make sure my employees understand
30:31how to use AI, "I think if you do all
30:32that, by tomorrow you've already
30:33significantly increased your margins,
30:35which is a major point. And so tomorrow
30:37, even if these companies come and
30:38squeeze us a bit, raising their prices,
30:40>> I think the risk will be quite limited
30:42for companies. So there you have it, we
30:43saw it with Microsoft. Microsoft is
30:45also a good example for me. Sometimes I
30:46have IT directors telling me," Ah yes,
30:48but anyway, they're going to do like
30:49Microsoft. "What does" do like
30:50Microsoft "mean? The license is € 20
30:52a month at Microsoft. It's nothing.
30:53Considering everything we do with it.
30:54We pay € 20 a month for something we
30:56might have bought for 7, 8, 10 dollars
30:58or euros a month 20 years ago. Well, if
31:01Claude goes from € 25 to € 50 a
31:03month in 10 or 20 years, I think it's
31:05more likely to look like that. I think
31:07they want to become a Microsoft-sized
31:08giant,
31:09>> and so they're going to model
31:10themselves on what Microsoft does. And
31:11if they do like Microsoft, I'm buying
31:13in immediately because look at
31:14Microsoft's valuation. I mean, the
31:16number of times I might have said, or
31:19did say—and I stand by it—that
31:21Microsoft is worth nothing anymore.
31:24Google comes along with its cloud,
31:25Google Docs, I mean Google spreadsheets
31:27, it's so much better, and then, in
31:29fact, not at all, and it goes on, and
31:30then Copilot, it's really the Microsoft
31:32phoenix, uh...Windows was totally
31:36trashed. Well, no, they're still here,
31:38even if Windows Phone didn't succeed.
31:39Yeah, but they're still here. Plus,
31:41they're getting into corporate
31:43environments everywhere.
31:44>> That's clear. And yet, I still think
31:46that regarding the Office suite, you
31:47know, when you think we might say it's
31:49the biggest monopoly, there's Google,
31:50but in France, for instance, Google
31:51represents nothing in terms of
31:53workspace. They have the biggest
31:56monopoly, they can do pretty much
31:57whatever they want, yet the major
31:59competition they are currently facing
32:00head-on is Claude Slides, Claude Docs,
32:02and Claude spreadsheets. And that is
32:04still going to hurt them a lot. I’m
32:06not saying it will kill them, but it
32:08won’t exactly push prices up. So
32:11today, there are really quite few SaaS
32:12companies that have truly, uh, held
32:14their clients by the throat by just
32:15increasing prices over and over. In any
32:18case, for the very big ones, there is
32:18so much competition that I personally
32:19don’t think they’ll be in that
32:20position. They can, they will be able
32:21to do it a little bit. The risk of this
32:23actually damaging businesses, I don’t
32:24really believe in that much.
32:26>> Very funny. I’m receiving at the same
32:28time a...So what I also like about La
32:30Martingale and so on is the, uh,
32:32contradiction. We’ll get to the split
32:34, but I was looking for data on what
32:36you were saying about Actually, I
32:39wanted to look for data on Microsoft
32:41regarding what I was saying about the
32:42stock trend, and then I see an email
32:43from Frédéric Fillou. Does that ring
32:45a bell or not?
32:45>> No. Occasional, a guy I adore. And
32:47subscribe to his newsletter, it’s
32:49really very cool. I’ll put the link
32:50in. And he puts the terrifying numbers
32:52of the upcoming crash. So you see, he
32:54tells me he has the numbers. So here it
32:56is, with charts and numbers, the state
32:58of the AI bubble. For now, the house of
33:00cards continues to defy gravity, but a
33:01correction seems inevitable. Uh, for
33:03months, I’ve been accumulating
33:04figures, charts, and perspectives on
33:06the inflation of the artificial
33:07intelligence bubble. So I’ll put the
33:09link for you, uh, but it’s
33:11interesting to see also, uh, everyone
33:12should go and form their own opinion.
33:15If, if we had a, a crystal ball, uh, uh
33:18, we would be better. So how would we
33:20be unbeatable? Sorry? But we have a
33:23little La Martingale crystal ball. How
33:24do we break down 100 dollars of revenue
33:29, uh, from AI? Who is taking the money
33:32from that?
33:33>> Yeah. Well, I would do it more on the
33:35capex first, because regarding
33:36investments, or when we hear OpenAI is
33:38investing 100 dollars in AI, what does
33:40that actually mean? Especially because
33:42there is the question of sovereignty
33:43behind all of that. When Mistral or
33:44OpenAI announces 100 billion in
33:45investment, come on, I’m tempted to
33:47say, at the end of the day, we don’t
33:48really care what that means. When
33:49Mistral tells us," We're putting 100
33:50billion in AI investments in France, "
33:51well, what does that actually mean? And
33:54here, there's a huge, huge issue, which
33:56is that today, when 100 billion dollars
33:58are invested in AI, 50 billion—so 50%
33:59of those investments—go to chip
34:01makers. And so, when we look at the
34:04details, it's Nvidia on one side, and
34:06all the memory chips. Now, I'll say
34:09Nvidia, but the competitors are going
34:10to be American, right? It's going to be
34:12the chips.
34:13>> Actually, I'd say that Nvidia's biggest
34:15competitor today is perhaps Google with
34:16its TPUs.
34:17>> Yeah, practically Google. Amazon with
34:19its Trainium, OpenAI—though it's
34:21still small for now—has just released
34:23Jalapeno, their rival to Nvidia's Maia;
34:25and Microsoft has also released chips.
34:27So, all of that is American. Then you
34:29have the second layer, because for me,
34:30investing in AI means breaking down all
34:32the layers to really see what's going
34:33on. So, if we take those 50 billion in
34:35the chip layer, there's the Nvidia
34:37layer. So, think of Nvidia as those who
34:39create the chips to train AIs and those
34:41who create the chips for AI tool
34:42inference. When it's when I talk to it
34:44and it answers me. That is Nvidia's
34:46major capability. Although we could
34:47even break down Nvidia further because
34:49within it there is ASML, there are a
34:51huge number of sub-layers in Nvidia,
34:53but perhaps we shouldn't go that far.
34:55>> No, no, but we could. It's true that
34:57for ASML, to give an order of magnitude
34:59, when 100 billion dollars go into AI
35:01investments, about 2 billion go to ASML
35:02. ASML is the company that sells the
35:04machines to Nvidia to be able to etch
35:06the chips.
35:06>> And they're Dutch. Yeah.
35:07>> Yeah. So, they are European, and the
35:09big winners behind all this, as you've
35:11seen on the Korean stock market, are
35:12them too. Now, be careful, the Korean
35:14stock market is currently falling quite
35:16a bit, but you know that this year,
35:17among the four or five companies—
35:18actually, I think among the four
35:20companies worldwide that will have the
35:21highest net profits—there are two
35:22Koreans.
35:23>> Samsung.
35:24>> Samsung and SK...SK Hynix.
35:28>> And so, we're talking about—let's not
35:29say anything stupid—maybe 150 billion
35:31in net profit, something like that.
35:32It's unprecedented. And them, why?
35:34Because they cre...
35:35>> 150 billion in net income. Listen to
35:37this. 150 billion in net income. I
35:39don't even know if there's a French
35:41company that makes 150 billion in
35:43revenue.
35:44>> Like LVMH. LVMH must
35:46>> Total might
35:47>> I'll check. Yeah.
35:49>> LVMH must make 150. Yeah, maybe around
35:51that or 80, I don't remember. But
35:53anyway, it's Yeah, in any case it's
35:55equivalent or a little bit
35:56>> LVMH 80, you see. So the net income
35:59you're giving me is twice Yeah, that's
36:01it
36:01>> the revenue of LVMH.
36:03>> Exactly. And I don't think Total makes
36:06much more than 100 billion in revenue.
36:08So it's definitely huge. And
36:10>> if 200 billion for Total. OK. Yeah.
36:12Yeah. 171 billion euros
36:14>> with very small margins.
36:14>> Yeah. Yeah. Very small margins.
36:16>> Yeah. So, so, it's certain that they
36:18are becoming very strong, and
36:19especially what we call memory. It's
36:21that they, in fact, to put it simply,
36:23when you give ChatGPT a prompt and it
36:25remembers everything you told it before
36:27, that's not an Nvidia chip running,
36:28it's an SK or Samsung, or US Micron
36:30memory chip, which I think is worth
36:32almost a trillion, yet nobody knows
36:34them. So there's 50 billion going into
36:36the chips, so for sure when we invest
36:38in AI, we ask: do we invest in the
36:40chips? Is it overvalued? Hence the
36:42crash, actually. The little crash of
36:43our Korean friends over the last few
36:45days, few weeks there, it's going a bit
36:47less well, but it's not just that.
36:48There's 15 dollars going into the whole
36:50network, so that's a bit more technical
36:51, I'd say, but 20 dollars into
36:53electricity. So then out of 100, out of
36:54100 billion, we have 20 billion sorry.
36:56So 20%in electricity, like what we were
36:58saying about Schneider. Schneider is
37:00good but today the equivalent that
37:01would go to a Schneider out of 100
37:03billion in investments in France is 7.
37:05>> Out of the 20 billion in electricity,
37:07basically 15 to 20 billion in
37:08electricity, only 5 to 7 would be
37:10Schneider's core business. And as for
37:12Schneider's precise core business, I
37:13won't go into the details because it
37:14would eventually go over my head. I'm
37:16not an electricity expert but those are
37:17the orders of magnitude.
37:19>> It's it's but it's still a lot because,
37:21I mean, they are an important player on
37:23the global stage. Uh, and then, what we
37:27need to look at is more like scarcity.
37:30Because we're starting to see at some
37:31point, whether it's electricity—well,
37:33especially electricity—regardless of
37:35where it comes from, nuclear or fossil
37:37fuels, uh, if demand is growing, well,
37:39today we’re maxed out; we're using
37:41all the electricity we can, so.
37:43>> Completely, and actually I think it's
37:45who says that if if the whole world
37:46used ChatGPT as much as an average user
37:48, we wouldn't have enough electricity
37:50on Earth today, really. So there are
37:52those issues. That's why he's investing
37:55heavily in nuclear fusion, which for
37:56him would be the only way, not just for
37:58the world to use AI like today, but
38:00rather for everyone to use AI
38:01everywhere, that's what it would take.
38:04So yes, that's not bad. Now, we can
38:06also say at Schneider, when when 100
38:08billion dollars are invested in France,
38:10France captures, let's say, the 7
38:11billion from Schneider; we can put it
38:13that way. In any case, Europe, uh, a
38:16bit of building, 7 billion more, but we
38:18struggle to get 10-15%, more than 10-15
38:21%of investment in French companies, at
38:23least money that would be spent at
38:25French firms. When 100 billion is
38:27invested in the US, 90%of all that goes
38:29to American companies. And that,
38:31>> we'll provide links for those listening
38:32who would like more info on this.
38:34you've prepared for us. There, I have a
38:36visual right in front of me. We'll
38:37include links to your visual if you
38:38don't mind. And for those watching on
38:40YouTube, it's edited in. In any case,
38:42what needs to be said, you can go back
38:43to the other one. Uh, what just needs
38:46to be said, to keep in your head, if
38:48you are uh listening, running, or
38:50driving, first off, uh, if you're
38:52running, run a little bit faster for 1
38:54minute, humor me, and then we'll
38:56continue. If you're driving, don't
38:59drive faster, there's no point. Take it
39:00easy, and don't yell at your neighbor,
39:02it's no use. Just visualize it: out of
39:06every 100 invested, spending is 50%on
39:09chips, 15%on networking, 20%on
39:12electricity. and 7.5 and 7.5 on cooling
39:16and facilities. So there you go, those
39:18are the orders of magnitude. Chips are
39:20half of it. What creates a bit of a
39:23phenomenon that's heavily criticized in
39:26the United States is that Nvidia often
39:28invests massively in its clients,
39:30saying," Hey, listen, I'll give you 8
39:33billion,
39:35>> but it would be good if those 8 billion
39:36were used to buy Nvidia chips. "Which
39:39is why we talk about a bubble there,
39:40too, because they are often willing to
39:42pay a high price, thinking:" Well, at
39:44worst, it's 8 billion, they'll buy
39:46chips from me afterwards, it'll fund
39:47them to fund me, and in the end, if it
39:49doesn't work out, I'll have recovered
39:51the money. "Anyway.
39:53>> Yes, that's for sure. I always ask
39:54myself the question, and I don't have
39:55the answer to it. But I still always
39:57wonder: is this the sign of a potential
39:59bubble? Or is it precisely the sign of
40:01a market that is so intertwined that no
40:03one has any interest in killing anyone,
40:05and no one has an interest in seeing
40:06Nvidia fall, nor Google fall, nor
40:08Doctolib fall, since Nvidia recently
40:10invested quite a bit of money in
40:11Doctolib? So
40:13>> it's always hard to identify and
40:15analyze these links between all these
40:16companies. I actually think it
40:18solidifies things. So, in all of this,
40:20Nvidia has certainly made some
40:22investments that are a bit crazy
40:23because they have so much money—in
40:24this case, 100 to 150 billion in net
40:26income this year. So, in any case, in
40:28'26 it should arrive. So it's certain
40:29that among those, they might make a few
40:31blunders. But, still, it actually
40:34reassures me. I think everything is too
40:37intertwined for a giant to fall, and
40:39competition is a good thing. The fact
40:42that Nvidia is being competed with by
40:44all the big tech firms, even with
40:45Microsoft on the chip side, means they
40:47have to reinvent themselves, they'll
40:49have to cut their prices a bit, and so
40:51on. So
40:52>> they were too alone indeed, you're
40:54right about that.
40:55>> that uh and they're not so alone
40:56anymore, while there are others we
40:58don't talk about at all, and that's
40:59that's very interesting. You were
41:00talking about ASML earlier. ASML, for
41:02that matter, a much more existential
41:04danger, or TSMC. TSMC, which is the one
41:06that's going to etch the chips. Nvidia
41:08chips are, or sorry, the Taiwanese one.
41:11>> Taiwanese, that's it. TSMC, Taiwanese,
41:12they etch the chips. For them, actually
41:14, there's no competitor in the world.
41:16ASML, those who make the machine that
41:18that etches, has no competitor. I can
41:21even go further. Zeiss, which is a
41:22German player that's worth maybe only 5
41:24billion, 10 billion, it's small, well,
41:26quote-unquote small compared to the
41:27others. They are the ones who uh
41:29manufacture the lenses that go on ASML
41:31machines that go to TSMC to etch the
41:33Nvidia chip. And they, it's the same,
41:35it's a big monopoly. And so we often
41:37don't think about these firms, but the
41:38risk is perhaps much more behind them
41:40than behind Nvidia, which if tomorrow
41:41it produces a little less, honestly,
41:43there will still be competition behind
41:44it.
41:46>> Knowing that without going into but uh
41:47much further, I was talking about it
41:49recently with Nils and Guidour, I don't
41:51know if you know him, who set up uh
41:53Kyutai, it's uh this French AI that
41:54does mostly audio. OK. and uh they all
42:00develop on competing AIs. What's crazy
42:02about that is that, in fact, to develop
42:05your latest cloud models, you use
42:07ChatGPT and vice versa. So in fact, uh
42:10if I think about that, when you can,
42:12you talk about all these things that
42:14are intertwined, it's somewhere you say
42:16, no, but wait, if I want to make an
42:17ASML, a TSMC, etc., maybe by searching
42:19well with AI for a long time, an
42:21investment. No, but that's industrial,
42:23it's different. But I mean your
42:24software more and more, your research,
42:27in fact, they are so high-performing,
42:29these programs, that in fact it's an
42:31endless race uh forward uh so
42:33inevitably uh which burns through a lot
42:36, but it really is endless. It's
42:38>> yeah, that's clear. On the software
42:40side, it's endless and the big problem
42:41is going to be the hardware, it's going
42:42to be, are we managing to, it's exactly
42:43what happened at Apple. I don't know if
42:45you've read the book Apple in China,
42:47which is fascinating—my favorite book
42:49on how Apple became Apple thanks to
42:50China and how China became China thanks
42:52to Apple, which is really incredible,
42:54written by the guy from the Financial
42:56Times who followed Apple for 10 years,
42:58I think, and he explains it.
43:00>> He explains this topic very well. What
43:02happened a bit when Apple invested,
43:04invested so much money in China, was
43:06thinking: actually, we put devel, some,
43:09some Chinese engineers, we put Apple
43:11engineers in China, which meant that
43:13China developed a lot of skills. And in
43:15fact, he says, the big problem, when we
43:17wanted to bring them back to the United
43:18States, we wanted to create factories
43:20—it was Foxconn at the time—we
43:21wanted to compete a bit with Foxconn
43:22and create that in the US, and the
43:24problem was, quote unquote, the" hard "
43:25side. We didn't have the skills, we
43:27didn't have the means to build
43:28factories in 3 weeks, we didn't have
43:29all that. And so indeed, while they
43:31will be able to wage war on software
43:33thanks to plenty of AI and so on, as
43:35for the hardware, recreating an ASML
43:37would actually take decades and huge
43:39capex. And we don't really know how
43:40we're going to manage to do that. That
43:41is a risk. Beyond that, the Chinese are
43:43doing it and we don't talk about it
43:44much today, but they are very far ahead
43:45on all these subjects. We
43:47>> must not let up on this. Uh okay, so
43:49we've understood it link by link: 50%
43:51for chips, 20%for electricity, 15%for
43:53the network, and 7.5%for cooling and
43:55buildings. So the impact on
43:57construction is also crazy. Uh and now,
43:59let's zoom in a bit on the chips, if
44:03you will. Uh so the chips, we were
44:05talking about them earlier. Uh what,
44:08what, what do we have in there? Well,
44:10there it is, exactly. So out of the 50%
44:12that go to chips, you still have 25%
44:14that go to—25 points that go to
44:15Nvidia. So Nvidia is half of the chip
44:18market today. That's about 15 billion
44:20dollars. We are talking in billions:
44:21out of the 50 billion that would be
44:22developed, deployed on this project,
44:23there are 100 billion.
44:24>> It's roughly, you know, it means that
44:27if I had to summarize it broadly, uh,
44:29Nvidia is 25%of the AI market.
44:32>> Yeah. Of what's being inves...
44:34>> In any case, 50%of what...25%of what's
44:35going to be invested in AI, well,
44:37assuming a large share will go to
44:39Nvidia out of that 25%...out of that 25
44:41%of...of all of it. So, yeah, it's...
44:43it's quite...it's quite enormous. But
44:45once again, there are competitors, but
44:46still, they have a...an edge that...
44:47that is quite strong. And the same for
44:49memory chips, it's memory chips. And...
44:50and one might ask: is there a bubble?
44:52Is Nvidia really worth 5 trillion? Is
44:54...is Samsung, I don't know what it is
44:56today, 1.5 trillion, maybe 2 trillion,
44:58is that their true valuation? And there
45:00, for once, I would rather say:"
45:02Investors, be very cautious, "in the
45:04sense that, objectively, no one knows
45:06if Nvidia's market is headed toward 10
45:09trillion or rather stagnating around 5
45:11because Google will become big enough
45:13or OpenAI will become big enough. It's
45:15so opaque that on that point I always
45:18say be quite cautious. What is certain
45:20is that the market continues to be
45:21enormous. For Nvidia or SK Hynix or
45:23Samsung, direct competitors won't
45:25appear right away. I think it's much
45:27better to look at companies like ASML.
45:29I believe that for a company like ASML,
45:31the probability that a company like
45:32that sees its market shrink, I don't
45:34see how that could happen. So an ASML
45:36might only be 2%of the 100 billion
45:38invested per project, but it's a
45:39company that has no competitors. So
45:42look, I think when you look at this
45:43whole chip ecosystem, look at who among
45:44them has no competitors. Conversely,
45:47TSMC. TSMC is also a company that is
45:49worth several trillions or hundreds of
45:51billions on the stock market. Well, for
45:53TSMC, there are risks, and it's one of
45:54the big topics among all these links in
45:56the chain, it's the China risk. It's
45:58the risk of China invading Taiwan,
46:00which isn't zero, and I'm not enough of
46:01a geopolitical expert to quantify the
46:03risk, but it exists.
46:04>> Yeah, it's...it's present when...when
46:06it exists. It is always present in the
46:08market. Exactly.
46:09>> To give you ASML, I didn't even know it
46:11was listed. I had never thought about
46:12it, you see. So it's worth, uh, it's
46:14worth 168 uh dollars at the moment I'm
46:19speaking to you. Uh, it was, in October
46:232022, 434, to give you a low point for
46:26the dates, 379 even in October 2022. Uh
46:31, it really started to go up in 2025
46:34because we went, uh, from 2025, August
46:372025, 72 basically 6,700 in August 2025
46:43to, uh, you see, 1600, 1800 now.
46:46>> And you, but is it already priced in
46:48that it's unique and that no one can
46:50compete with...
46:52>> Yes, it's already priced in. Now,
46:54what's priced in is whether we will
46:55have...What isn't priced in yet is
46:57whether we'll need two, three, or five
46:58times more, given the adoption I keep
47:00bringing up. We will sell ASML machines
47:02as long as there are more AI users in
47:04companies tomorrow than today and
47:05yesterday, and that is the case. And so
47:07, I think this market is very deep,
47:09even though they aren't allowed to sell
47:10machines in China. So even by cutting
47:12themselves off entirely from the
47:14Chinese market, which wants to master
47:16the whole supply chain on its own, they
47:18are, well, they are very big and they
47:19are quite alone. So I think that, and
47:22I'm talking about ASML, but you could
47:23break it down, we could even go further
47:25into chips and break them down into 10,
47:2715, 20 micro-markets, it doesn't really
47:28change much. However, be much more
47:30careful with an SK Hynix that might
47:32come out of the woodwork with memory
47:34chips and, from one year to the next,
47:35could potentially be cut in half. That
47:37wouldn't surprise me because they have
47:38competitors. They have Samsung as a
47:39competitor, they have plenty of others,
47:41but there's TSMC that doesn't have many
47:42competitors, there's ASML that doesn't
47:43have many competitors. Applied
47:45Materials, same thing on a very
47:46specific subject regarding chips. So,
47:48uh, I think that is
47:50>> I remind you that neither Martin nor I
47:52are wealth management advisors or
47:54financial investment advisors. We are
47:57not at all giving you, uh, advice.
47:59We're telling you to look, uh, well,
48:02we've been talking to you all this time
48:04about many players, Nvidia, SK Hynix,
48:07Foxconn. Now, Foxconn, I'm not sure,
48:09you see, you might be able to invest
48:10via Taiwan, but it's still in China,
48:11really.
48:12>> Yeah, that's what I think.
48:13>> So Applied Materials, ASML, we've
48:15talked about them a lot here. And so,
48:18are there any you see as completely
48:21underrated, underpriced,
48:22under-the-radar, or not really? They
48:24are already
48:25>> No, not really. They're all—I was
48:26about to say, there's a bubble, I don't
48:28like that word, but there's been a huge
48:30, huge wave for chips anyway. So they
48:32have all risen a lot. No, at the same
48:34time, I don't see any where I could say
48:35that's really underpriced, it's just
48:36nonsense. But not the other way around
48:38either. So no, for now, it's quite hard
48:40to say. No one predicted the Korean
48:42drop. It’s going to be quite
48:43interesting to analyze, though. The
48:45first people who came to give me their
48:46analysis on what was happening in the
48:48Korean stock market, they’re telling
48:49me," Look, it’s diving, you should
48:50have expected it, "and so on. Three
48:51months ago, we were talking about SK
48:52Hynix and Samsung, and we weren't
48:53talking at all about a potential drop
48:55like this.
48:55>> But didn't a drop already happen a few
48:57months ago? Doesn't Korea yo-yo a bit?
49:00Well, in any case, over the last 18
49:02months, it really exploded upward. And
49:04so, does this drop date back to 6
49:05months ago or more like 3 months ago?
49:07I’d say more like 3 months, honestly,
49:09and we’re talking about -30%or -40%,
49:10if I’m not mistaken, on the Korean
49:12index. So, there you go, that was also
49:15hard to predict. Now, demand is only
49:17going to increase. So if you follow
49:19this, everyone will do their own
49:20analysis, but demand is only going to
49:22increase, and very, very strongly, and
49:24all those who have no competitors will,
49:26in my opinion, be in higher and higher
49:27demand. So there are indeed, within
49:29this value chain, a few players who are
49:30affected by all this, by these criteria
49:32.
49:32>> What else do we want to look at today,
49:33Martin? What did you bring for me?
49:34Because I have one or two questions. Uh
49:37>> notably, I’d like to look into
49:39something with you, uh, yes. Well, it's
49:42kind of the idea of SaaS
49:45>> and the side effects
49:47>> because beyond SaaS, we talked a bit
49:48here about the" SaaS Apocalypse, "right
49:50? It's the fact of realizing that I'm
49:52actually developing a lot of software
49:53that I wouldn't have developed
49:55otherwise. So that’s part of my 200
49:56dollars per person per month because I
49:59tell myself that, ultimately, we have
50:01so much more than just AI. We’ve
50:03developed our own software suites.
50:05Besides, I could turn off my AIs and
50:06most of those programs would keep
50:08running.
50:09>> So I have less dependency on others. So
50:11there were companies that had to...BCM.
50:14Actually, you can see it, there have
50:17been some major shifts recently among
50:20insurers in the United States because
50:22Muse was capable—Muse being that Meta
50:25model—of analyzing all your insurance
50:28policies and telling you," Watch out,
50:31you're paying way too much and you're
50:34poorly covered. "And obviously, the
50:37ones you pay too much for and that
50:38cover you poorly are generally the ones
50:39making the most profit, because they
50:41sell you a lousy product at a very high
50:42price and don't cover you well. So when
50:44you go to ask them, when you have an
50:45accident, they'll say," Well, first off
50:47, you have a—what do you call it? ""
50:49The thing you pay when you... ""
50:51>> A deductible, a super high deductible,
50:53and on top of that, the max I'll
50:54reimburse you is 20,000, even if your
50:56damage is worth 200, you see. "So, you
50:58end up paying a lot for something that
50:59covers you poorly, and for the
51:00insurance, well, Muse will tell you,"
51:02This is trash, I'm canceling your
51:03subscription and signing you up for
51:05something that costs less and covers
51:06you better. "
51:07>> Yeah.
51:07>> And so we’ve had a lot of insurers
51:08who have lowered their...
51:09>> That’s for sure. AI will make access
51:11to information really easy for everyone
51:13. That was already somewhat the case
51:15with the internet, but now it’s
51:16faster and more detailed. So access to
51:17information becomes simple. Access to
51:18negotiation becomes simple too, because
51:20understanding something doesn't mean
51:21succeeding in negotiating it well with
51:23your provider. So that’s definitely
51:24going to be very powerful. Now, it’s
51:26also going to, quote-unquote, set
51:28things straight. It's a question of
51:29whether a company should make money
51:30because there’s an information
51:31asymmetry, or should it make money
51:32because it provides value? I think
51:33it’s more the latter. So all the
51:35companies that were making money
51:36because they took advantage of
51:37information asymmetry, typically an
51:39insurer in this case, they are going to
51:40be very, very challenged. We can even
51:42go much more micro. A notary who wasn't
51:44doing their job well. Now they have a
51:45client coming in front of them saying,"
51:47I prepared for the meeting with GPT. ""
51:48Here is what it tells me. "" Here is
51:49what you’re going to propose, and
51:50it’s up to you to choose between the
51:51last two scenarios. "Well, if I’m not
51:53well-prepared as a notary or as a
51:54doctor—and tomorrow everyone will be
51:55affected—it’s certain that it will
51:57drive all prices down. And
51:59>> tomorrow, for my own lawyers and
52:00doctors, I’m sorry, but for the last
52:01six months, I’ve been showing up
52:03saying, well...
52:03>> Exactly. So
52:05>> that, for sure, is the strategy for my
52:06knee surgery, that’s certain. Yeah,
52:07that’s it. And and it’s going to
52:09spread. So it’s certain that it will
52:10really challenge companies. And later,
52:11it’s exactly what happened with SaaS,
52:12the fact is that when there was
52:14>> so SaaS, sorry, before we lose everyone
52:16, explain it to me.
52:17>> SaaS is software you pay for, so it's"
52:19Software as a Service. "
52:21>> So" software "is the program, and" as a
52:22service "means you pay a subscription
52:24to access that software. We
52:25>> use what kind of SaaS?
52:27>> Salesforce. Salesforce, that’s for
52:29people who are in business. Uh, finally
52:31, a—well, Spotify is a bit different,
52:33but it’s a SaaS; it’s software, but
52:34you pay a subscription and you have to
52:36pay artists behind it. So they still
52:38have a barrier to entry, which is that
52:39you can’t just redevelop your own
52:41>> No. Yeah, they still have rights, in my
52:42opinion, a bit like that for music
52:43where they share the rights, so it’s
52:45different. No, but something like
52:46Aircall, it’s used for making calls.
52:48Aircall is a pure SaaS app. An idea of
52:50a SaaS somewhere that you can more or
52:52less replace, but you have coaching
52:54SaaS or even somewhere a little, you
52:56know Runkeeper or whatever you call it,
53:01except for the community part.
53:03>> The community part cannot be replaced.
53:05>> And actually, your part of putting in
53:07your GPS data, connecting it to your
53:08Garmin or Strava, that can be developed
53:10in a single day now.
53:11>> Oh yeah, absolutely; we have a
53:13consultant staffing tool, and well,
53:14there’s nothing simpler—it’s just
53:16binary rules, just binary rules.
53:17There’s no community, there’s no
53:19specific proprietary data, the company
53:21and us, the young ones. Well, it’s
53:23certain that for all those pieces of
53:24software, it’s very, very hard to
53:25survive. And I dreamed of that for so
53:27many years, and it costs a lot to
53:28develop when you build it yourself.
53:30There you go, but we are we are
53:31handling it right,
53:32>> It's going much faster. So there was
53:33this SaaS apocalypse that started with
53:35the Claude plugins, really. They were
53:36the ones who launched all that, saying
53:38we're going to connect to all the tools
53:39, and so you'll be able to have a
53:41single interface from Claude and talk
53:42to all your tools. And we saw a
53:44trillion dollars in valuation go up in
53:46smoke in just a few days, I think. So
53:48everyone thought," It's the end of SaaS
53:50, we won't pay for subscriptions
53:51anymore, we'll recreate everything
53:52in-house. "And I'm revisiting that
53:54subject a bit. Well, I've always been
53:56cautious, and I think we were moving a
53:58bit too fast on the SaaS apocalypse.
54:00Personally, I think it's really going
54:01to split into two. Yes, the SaaS
54:02companies like those you mentioned or I
54:04mentioned that don't bring much value
54:05to customers beyond just creating a
54:07nice interface, they are dead. They'll
54:09never survive. But on the other hand,
54:11and we see it in the stock market with
54:13Salesforce, HubSpot, and also
54:15ServiceNow, all these companies that
54:16hold their customers' data, that have
54:18an architecture that's useful, et
54:20cetera, I think they have bright days
54:22ahead, and we have the case. Right now,
54:25I'm looking a lot at a company called
54:27Veeva, Veeva Industries, something like
54:29that, or Veeva, Veeva Systems. It's a
54:32spin-off originally from Salesforce, so
54:34it's a CRM for pharma companies. Well,
54:36Veeva is a company that lost 40%of its
54:37valuation on the stock market with the
54:39SaaS apocalypse and has recovered 100%
54:40—well, it gained over 100%in the last
54:426 months, or almost + 85%or something
54:43like that. And Salesforce is the same;
54:46Salesforce dropped 30-40%in recent
54:48months, and now they're up 30%because
54:50we realize it's not so simple to
54:52replace giants like that. So I really
54:54believe in what we call the single
54:56interface, meaning you will have a
54:57single interface.
54:58>> Veeva, sorry, to get the date...
55:00November 11, 2025, $ 300. Let's see,
55:04I'll try to find the low. June 22, 2026
55:08, $ 153. So, roughly 150; so 300 to 150
55:11, cut in half, bam. And now today it's
55:15313, so we are beyond its previous max.
55:19So we're back to 317, that's the 5-year
55:22high.
55:23>> We're talking about a high that was
55:24back in 2020. So it's a company too.
55:27After that, you have to know how to
55:28bounce back. That's the question. You
55:29see what I mean? Well, I personally
55:31think there was a bit of market
55:32hysteria. Everyone got scared, everyone
55:34wanted to pull out. I mean, remember
55:36that a large portion of the people
55:38investing and divesting from these
55:39companies are not people who use AI on
55:41a daily basis. So there is still a bit
55:42of a misunderstanding of the subject.
55:44Salesforce has made up for it all today
55:45. ServiceNow has almost made up for its
55:47entire drop too. So if you're
55:48specialized—all these companies that
55:49are specialized, like Veeva, they only
55:51do biotech.
55:51>> If you are ultra-specialized, if you
55:53have the client's data, if you know how
55:54to integrate well with Claude, that's
55:56the big Salesforce topic. I think it's
55:57a fascinating use case. Salesforce. At
55:59first, they said," We're going to
56:01create our own AI called Agentforce. "
56:02Overall, it didn't work. It was kind of
56:05lame. Everyone was saying," Actually,
56:06let's recode Salesforce, "and so on.
56:07Then Marc Benioff, one of the greatest
56:09entrepreneurs of our time, came along
56:10and said," Actually, we understand the
56:12issue; what we're going to do instead
56:13is create an app that integrates into
56:15Claude. "And today, they've created—
56:17in fact, we have the visual—but today
56:19, you can use and open Salesforce
56:21inside Claude. And so they thought, if
56:24the unique interface becomes Claude or
56:26ChatGPT, if that’s the case, if
56:28it’s through this interface that all
56:30users pass, our clients must be able to
56:32simply query and use Salesforce. I'm
56:35going to—I'm going to give you a
56:36thing, and it's true that it's a
56:37question we had. I was telling you
56:39about this event last week with the
56:41marketing director of Air France, for
56:42example, and I was saying," But because
56:44there are people today who are too
56:46attached to interfaces, "and actually
56:47many people in the room, not all, but
56:49they were saying," Actually, my job has
56:51always been to build interfaces. ""
56:53>> I don't want to do all those things you
56:56call APIs, connectors, whatever. "I say
56:58," But wait, you know, as an Air France
56:59customer, I don't want to have to deal
57:01with your website. "I don't care,
57:02actually. I'll say, look at my calendar
57:05, look at my stuff, I need to take a
57:06round trip to New York on such and such
57:08date. Uh, ideally, I'd like to find a
57:11business class ticket that isn't too
57:13expensive if possible. Find me the best
57:15options, suggest them within three days
57:17or three hours, it doesn't matter. And
57:20then, uh, I'll have my stuff with my
57:22discount card or my it's so magical, I
57:27don't want to use their their thing
57:28where I have to go back and forth 12
57:29times and
57:30>> That's clear. And those who fight
57:31against that, to me, they're they're
57:32dead. And there are many, a Sopra
57:33Steria today that make a lot, a lot of
57:34software, they're fighting against that
57:36. They don't even have an API, and that
57:38for me today still holds up a bit, but
57:39I'm not at all sure it will last long.
57:41Whereas all the inter all the tools
57:43like Salesforce, like Canva, when we go
57:45on ChatGPT today, we can open Canva in
57:47ChatGPT because Canva said we're not
57:49going to fight this. The employee has
57:51their memory in ChatGPT, they have
57:52their projects in ChatGPT. Now they're
57:54starting to have their chat, their
57:56meeting notes. Well, if they can open
57:57Canva directly in their tool, but if
57:58that's the case, what interest do I
58:00have in unsubscribing from Canva? None.
58:02But I want to give a quick shout-out to
58:03one of my sponsors on my other podcast,
58:04Génération Self, which is called
58:05Qonto. Uh, who allows me to connect uh
58:08Qonto and I think to myself, when will
58:10traditional banks allow me to do that?
58:12Today, if I tell Qonto, open, make me a
58:17quote
58:19>> for Martin Pavanello Mysteria at €
58:211500 for such and such service, or send
58:24them an invoice, or turn this quote
58:26into an invoice, or create a new
58:28temporary card for such an employee, or
58:31uh what did I do? Give me a summary of
58:34my marketing expenses.
58:35>> Yeah.
58:36>> Uh, over the last 6 months, and then I
58:38see it with graphs, things, and stuff;
58:39I can even have it animated in every
58:41direction. Yeah.
58:42>> But it's just so crazy.
58:44>> It's crazy, but I think you'd prefer to
58:45do it even more on ChatGPT.
58:47>> But I do it, no, I do it on I do it on
58:49Claude. Yeah, using a connector. Yeah,
58:51it works on CL GVT. There you go, and
58:53there you go. And so I'm saying this
58:54because I was recording, and I never do
58:56this, but I don't care, I'm not saying
58:57it because it's my sponsor, I'm not
58:58paid to do this. But if you want to try
59:01what I'm talking about, use Do it
59:02Europe. You'll get three free words,
59:04and you'll thank me. But in this case,
59:06uh, you say," But
59:08>> do you think traditional banks will
59:10release this in the next two years? "
59:12>> No, I don't think so.
59:13>> No, no, I don't think so. I think
59:14they're too afraid, they're going to
59:15get hit hard, actually. That might be
59:16where SaaS will take over.
59:18>> And yes, actually, it's not really SaaS
59:20that's going to be challenged here,
59:21it's the legacy company. But yes, all
59:23the legacy companies that refuse to
59:24open up like this are really going to
59:25suffer. Pennylane opened up, Doctrine
59:27is in the process of opening up.
59:28Everything, I think they are all
59:29opening up. Everyone is opening up,
59:31thinking," Either I fight against the
59:32GPT interface, or I know the match is
59:34already lost. "So I'm going to give
59:36what I can give. My proprietary data,
59:39my access. Salesforce is working a lot
59:40on agents.
59:41>> legacy ones are doing it anyway. Maybe
59:43I am, I saw this morning that Carrefour
59:44had an MCP, so maybe I can do my
59:46shopping at Carrefour, and I wouldn't
59:47find that stupid at all because it's
59:49really the most annoying thing in life
59:50too.
59:50>> Exactly. And conversely, if you're
59:52choosing between Carrefour and Franprix
59:54and only Carrefour has its MCP, I'm
59:55sure you'll go to Carrefour more. No,
59:57not because you prefer Carrefour, but
59:58it's so convenient to have it on
59:59ChatGPT, and everyone is saying that.
1:00:01It's so convenient to have HubSpot in
1:00:02Claude, so the SaaS companies going
1:00:04against that, I think it's risky. Those
1:00:06who follow this wave, in my opinion,
1:00:08there's not too much risk of being
1:00:09disrupted, and companies won't want to
1:00:11...
1:00:12>> This is super interesting, what we're
1:00:13saying, because it...You see, by
1:00:15talking like this, you say," Well,
1:00:16maybe tech companies are more
1:00:17tech-savvy and will follow this a bit
1:00:19more, "or some tech companies—not all
1:00:20—and look at which ones are opening
1:00:22up or not and will go for it, and those
1:00:23will be the ones that move the fastest.
1:00:26Because afterwards, the issue is that
1:00:28you could also say to yourself,
1:00:29conversely, at Air France, it doesn't
1:00:33work exactly like that. Because, I mean
1:00:34, I’m going to...if Air France is
1:00:36already selling via Booking and other
1:00:38platforms. So, they’re not the only
1:00:39ones distributing; it’s in their
1:00:41interest to do it themselves as Air
1:00:42France. On the other hand, you might
1:00:44tell yourself, I’m going to create
1:00:45other dependencies. I’m going to
1:00:46shoot myself in the foot because I no
1:00:48longer have an interface, because, at
1:00:49some point, why wouldn't Claude move up
1:00:51the chain and say:" Yeah, but you know
1:00:52what? Everything else you do,
1:00:54Salesforce, I don’t care, I’m going
1:00:55to redo it myself. "And there are quite
1:00:57a few companies that suspect or see
1:00:59that this is indeed happening.
1:01:01>> Which is the big Office risk, the
1:01:03Office suite risk, or will we leave it
1:01:05one day? I started using Claude Slides;
1:01:07I can’t make a PowerPoint anymore,
1:01:08it’s impossible. Even a PowerPoint
1:01:10with Claude, it’s slower, it’s not
1:01:13as well done, Claude Slides is
1:01:15brilliant. And so yes, yes, they are
1:01:17going to face that risk. Now, I still
1:01:20think that when they have the data, and
1:01:22conversely, you see an opposite point
1:01:24that we saw which is great, well, we
1:01:25are moving out of our tools.
1:01:29>> Because we realized that everything is
1:01:30on our SharePoint, and actually, since
1:01:31we have a single interface that is our
1:01:32cloud, it will organize my SharePoint
1:01:33perfectly. It is highly structured; it
1:01:36will very easily find things in
1:01:37SharePoint and modify them. And my base
1:01:39, you don’t know SharePoint, what is
1:01:40it? It’s
1:01:41>> SharePoint is the, the cloud, it’s
1:01:43the Google Drive of Microsoft.
1:01:46>> OK, got it.
1:01:46>> So it’s really the same, except that
1:01:48a lot of things I used to do on Notion,
1:01:50I don’t need anymore because since I
1:01:51talk to my Claude, and it’s
1:01:52well-connected to SharePoint, that’s
1:01:54enough for me. And for us, we’re
1:01:55talking about a company like ours,
1:01:56that’s tens of thousands of euros a
1:01:57year with Notion that we’re going to
1:01:58cut. So indeed, there is a real issue
1:02:00that it will refocus on far fewer
1:02:02players, that’s for sure, but players
1:02:04that will be increasingly connected to
1:02:06your AI because the single interface, I
1:02:08believe in that 1000%.
1:02:10>> I agree, but especially the interface;
1:02:12I mean, when you go further—I was
1:02:15talking about voice earlier—I don't
1:02:17use a keyboard anymore, I only talk to
1:02:20my AI. So, at some point, it really is
1:02:23Jarvis.
1:02:23>> Yeah, that’s Claude’s problem, by
1:02:25the way. They lose a lot, they really
1:02:26lose out on that. Their mic isn't
1:02:29nearly as competitive, and I keep
1:02:30saying that the winner in AI might be
1:02:32the one with the best mic; you wouldn't
1:02:34imagine, it's GPT—Nil explained this
1:02:36to me—but they have, well, I use
1:02:38Whisper too, but there’s a technology
1:02:40behind it, a voice technology
1:02:44>> that's pretty insane, an incredible
1:02:46test. You know Lou, the wonderful
1:02:48producer of La Matinale? We were
1:02:50preparing for a meeting together the
1:02:52other day, and often, voice features
1:02:55use speech-to-text, hm hm
1:02:57>> to speech, you understand? So first
1:02:58they listen to what you say, it gets
1:03:01transformed into text, analyzed, and so
1:03:03on. And there, it actually recognized
1:03:05Lou and Mathieu.
1:03:07>> There were three of us in the meeting,
1:03:08well, the two of us. Plus, that means
1:03:10we were using a voice version that only
1:03:12did speech-to-speech, it didn't
1:03:13necessarily go through text anymore.
1:03:16>> And now, we're reaching things that are
1:03:17becoming really, really
1:03:19>> uh, yeah. for me, that really do.
1:03:23>> Yeah, it’s mind-blowing, but it's
1:03:24still crazy that Claude hasn't managed
1:03:26to implement this mic yet. For us, it
1:03:28comes out at 100%very quickly. I don't
1:03:32know, I don't know how long they're
1:03:33going to struggle with this. Copilot is
1:03:35the same, by the way. Copilot doesn't
1:03:37have a good mic, and that slows down
1:03:38the adoption of generative AI in
1:03:39companies. Not having a good mic, and I
1:03:41think there are lots of side effects
1:03:43like that, right, that accelerate or
1:03:45slow down AI development in a company.
1:03:47Very clear, Martin. I think we’ve
1:03:51covered a lot of ground. Yeah.
1:03:52>> Yeah. Thanks for all of that, in any
1:03:54case. How did you get started in all of
1:03:57this 3 years ago?
1:04:00>> I started by using ChatGPT day and
1:04:02night starting exactly on November 30,
1:04:042022, and it fascinated me. I wasn't
1:04:06from that field at all. It fascinated
1:04:08me, and as I used it, and even today,
1:04:10I'm talking about 1 to 1.5 hours of
1:04:12monitoring every day for the past 3
1:04:14years, every single day. I spend an
1:04:15hour a day keeping up to date. The goal
1:04:17isn't to see every new thing, but you
1:04:18find yourself asking more and more
1:04:20questions. Oh, look, data centers are
1:04:22falling on the stock market. Oh, look,
1:04:23the Korean stock market, and it makes
1:04:24you dig into these topics. So, I
1:04:25started using it a lot, and then I
1:04:27thought, wow, actually everyone is
1:04:29going to need to do what I'm doing and
1:04:30learn how to use it. Hence MisterIA,
1:04:32which is behind
1:04:33>> So you started selling services, right?
1:04:36First solo, and then with other people.
1:04:38Is that right?
1:04:38>> Yeah. No, I really like the newsletter
1:04:40medium. I've had many newsletters on
1:04:42many subjects in my life, and in March
1:04:44'23, I decided to do a newsletter about
1:04:46AI called MisterIA. And so, for 6
1:04:47months, it was just a newsletter that
1:04:49quickly climbed to 20,000 or 30,000
1:04:51subscribers, because people were quite
1:04:52interested. And it was after 6 months,
1:04:54so in September/October '23, that I
1:04:55thought:" Actually, every day people
1:04:57ask me how to get trained in AI, so
1:04:58let's go, I should be the one training
1:04:59them. "And so, I hired two or three
1:05:01employees immediately, and from all
1:05:02that, MisterIA was born, where we are
1:05:04200 today.
1:05:04>> Fantastic, did you keep the newsletter
1:05:06going?
1:05:06>> Yeah, every week, still. We have five
1:05:08of them now, even. We do a lot by
1:05:09industry: health, finance, public
1:05:11sector, law, generalist.
1:05:12>> Very good, well, we'll find them at
1:05:14misteria.com.
1:05:15>> com misteria.com misteria.com. I'll put
1:05:17the links there. In my newsletter, you
1:05:19see, it turns out that for a few weeks
1:05:21now, I've been doing the prompt of the
1:05:22week, and everyone loves my prompts of
1:05:24the week. Very practical for many areas
1:05:26, etcetera.
1:05:27>> Uh, very cool. If someone comes from me
1:05:32>> uh
1:05:33>> and contacts MisterIA, what can they
1:05:34ask you for? What can they get as a
1:05:36>> can ask for a small discount, a little,
1:05:38a little promo code.
1:05:39>> OK. For what? For an assignment.
1:05:41>> For everything. A consulting assignment
1:05:43.
1:05:43>> You say" la Martin Gal ".
1:05:44>> Exactly. With" la Martin Gal ", you get
1:05:46a 10%discount on our consulting
1:05:47assignments and our training days. On
1:05:49licenses, it's a little bit harder, but
1:05:50then again, we can negotiate on
1:05:51licenses because we are a big, big
1:05:53reseller, or not reseller, but license
1:05:54integrator.
1:05:54>> OK. Well, go ahead, say you're coming
1:05:57from me, and negotiate. If they aren't
1:05:59nice enough, you let me know and then I
1:06:01I'll put I'll put...
1:06:02>> You know these promo codes work well
1:06:04because I am also a listener of your
1:06:06other podcasts and thanks to you I used
1:06:08the, quote-unquote," Moura "promo code.
1:06:12Following that magnificent podcast I
1:06:13watched several times. With Christophe,
1:06:15I think.
1:06:15>> Christophe Cout.
1:06:16>> And so I used it once and he told me it
1:06:18wasn't at all the first time someone
1:06:19used it at...
1:06:20>> No, I think last year, it really made
1:06:21his year. In 2025, quite seriously, he
1:06:23said to me:" It's...it's... "No, no,
1:06:25but actually, that's also why I
1:06:26negotiate, quote-unquote, it's that I
1:06:28don't take anything for it. It makes
1:06:30everyone happy, it makes you happy, if
1:06:31it works it will make people happy and
1:06:33...and there you go. I think it's worth
1:06:35it, in any case, to go see Martin or
1:06:37others if you have them in your circle.
1:06:39Just give yourself a little kick, a
1:06:42little push to go test all that out. I
1:06:45hope we weren't too jargon-heavy today,
1:06:46but I don't think so. I tried, you saw
1:06:48it, to stay grounded and explain
1:06:50everything. Do you have a little"
1:06:52Martingale "of your own, a little
1:06:53investment guilty pleasure?
1:06:55>> I don't invest much. Oh, yeah?
1:06:57>> Yeah, for now that's my, no that's my
1:06:59...
1:07:00>> That's my somewhat negative side, and
1:07:01especially because, particularly with
1:07:02AI, people are always telling me,"
1:07:03Where are you investing in AI? "And I
1:07:05say: I have 50%of my company, 40%...I'm
1:07:07already so exposed to AI that, if I
1:07:09invest in the stock market, I
1:07:11personally invest more in companies
1:07:13that build highways, airports, and
1:07:14electricity. Because with AI, I'm
1:07:16already quite, quite exposed personally
1:07:18and directly through my company. So
1:07:20there you go, that's more my guilty
1:07:22pleasure, what I watch a lot, and I
1:07:23know you're a, a guy who, at any rate,
1:07:25appreciates China. I'm a big fan of
1:07:27Chinese stocks, Chinese companies, so I
1:07:29look at that a lot.
1:07:30>> Very nice, listen. Yeah, and then later
1:07:32, you know, the things we could talk
1:07:33about but don't have the time for right
1:07:35now is the impact of AI in particular.
1:07:36For me, that makes me work a lot on
1:07:39offices and many things we'll see in
1:07:41the coming years. But there is also
1:07:43probably some political impact in
1:07:45France in parallel. So, uh, a lot of
1:07:47things, in any case, that make us ask
1:07:48plenty of questions that are
1:07:50interesting, fascinating. Thank you
1:07:51Martin, we can follow you on LinkedIn
1:07:53if we want to. Martin Pavanello, Mister
1:07:55AI, pretty much everywhere on social
1:07:57media. Subscribe to one of your
1:07:59newsletters and thank you for your time
1:08:00, it was fascinating.
1:08:01>> See you later. Ciao.

About this transcript

This page contains the full transcript of Anthropic à 2 000 milliards : opportunité du siècle ou bulle IA ? - Martin Pavanello by La Martingale, generated from the public captions YouTube serves with the video. The transcript has 15,155 words across 2,278 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 choose a caption track available on the video.

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, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.