YT YouTube2Text

YouTube2Text → Transcripts → 월가 10개 리포트 종합, 반박까지 20분 총정리. 꼭 보셔야합니다

월가 10개 리포트 종합, 반박까지 20분 총정리. 꼭 보셔야합니다 — Transcript

by 월텍남 - 월스트리트 테크남 · 3,831 words · 603 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00They say 80%of revenue in the AI
0:01ecosystem is just returned investment
0:03capital. If you calculate backwards from
0:06$ 1.5 trillion, you have to earn $ 6
0:08trillion a year from AI. What has
0:10changed between the Nvidia from a month
0:12ago, when Nasdaq hit its peak, and the
0:14Nvidia of today? Absolutely nothing has
0:16changed.
0:20First, let's look at the argument that
0:22AI is a bubble in terms of revenue.
0:26Bain, a global consulting firm, pointed
0:29out that while the AI market should
0:31reach $ 6 trillion annually, it's
0:33currently only around $ 1.8 trillion,
0:35but if you look closer at this—They say
0:38that AI facility investments will amount
0:40to about $ 1.5 trillion per year until
0:422031.
0:46In the case of cloud services, you
0:47typically need to multiply your capital
0:49expenditure by four to estimate the
0:51revenue.
0:53So, if you calculate backwards from this
0:56$ 1.5 trillion, you have to earn $ 6
0:58trillion a year from AI. When you talk
1:01about $ 6 trillion, that's roughly 8,000
1:03trillion won. Can this revenue actually
1:05be generated?
1:07They are pointing out that it probably
1:08won't be easy. Based on current
1:10standards: Consumer AI is between $ 200
1:13billion and $ 400 billion. And for
1:15corporate productivity, about $ 1
1:17trillion. If that's $ 1 trillion, where
1:19did the remaining $ 4.2 trillion go? How
1:22can you possibly earn that much? That is
1:24a question that could be asked.
1:26Ultimately, they are questioning whether
1:28AI can actually sell that much.
1:31TS Lombard also raises another doubt. To
1:34recoup investments, AI revenue needs to
1:36increase by more than 10 times what it
1:37is now. The required annual revenue by
1:392030 is between $ 2.3 trillion and $ 3.5
1:42trillion. This estimate is a bit more
1:43modest. Since it's $ 3 trillion, which
1:45is about 4,000 trillion won in revenue,
1:47but current corporate revenue is only
1:49about $ 200 billion. So, you need almost
1:5110 times more than that. We have about
1:53four years left until 2030.
1:56They are asking if a 10-fold increase is
1:58possible within those four years. For
1:59this to be possible, if it doubles every
2:01year, two to the power of three is eight
2:03, so it would grow eightfold, right? It
2:04needs to grow like this, but this $ 2
2:06trillion scale is quite massive. I
2:08understand the U.S. dining-out market is
2:10about $ 2 to $ 3 trillion a year, so
2:12it's a number that requires consumers to
2:14spend as much on AI as they do on food.
2:16How much do you spend on food each month
2:17? To put it simply, while there are
2:19variations, you spend about 300,000 won
2:21a month, right? Critics might argue that
2:23since everyone is expected to spend this
2:25much on AI, but many people don't even
2:26spend 30,000 won a month, wouldn't this
2:28be difficult? If I were to give just one
2:30counterargument here, I believe this 6
2:32trillion dollar figure is achievable.
2:36I personally think the 1.5 trillion
2:38dollar investment figure is a bit of an
2:40aggressive estimate, and I believe the
2:42concept of how much value AI creates is
2:44something that must be continuously
2:46verified through economy-wide GDP
2:48productivity or other indicators.
2:51And now for the second criticism. The
2:52circularity controversy, or what we call
2:54circular financing. According to the
2:56Bank for International Settlements, 55%
2:58of investment in AI companies comes from
3:00other AI companies. I'm sure you've all
3:02seen this graph.
3:04OpenAI receives funding and support from
3:06NVIDIA and purchase commitments from
3:08Microsoft, and then OpenAI buys NVIDIA
3:10products—it’s an interconnected
3:12relationship, and while only six
3:14companies are shown here, it is actually
3:16a structure where dozens of companies
3:18are intertwined.
3:20Also, 46%of AI investment funds consist
3:22mostly of relationships where they buy
3:24and sell goods among themselves. Are you
3:26just inflating your revenue? There is
3:28criticism that if you look at revenue
3:29through consolidated financial
3:31statements and remove the related
3:32duplicate transactions, the actual
3:33revenue would be less than half. That
3:35does have some validity. Looking at even
3:37more recent reports, 80%of the revenue
3:39in the AI ecosystem is actually recycled
3:41investment capital. Who do you think
3:43provided this 80%? Most of it is being
3:45invested by hyperscalers.
3:48Hyperscalers are investing 80%of the
3:50total, and the only ones currently
3:52making money from AI are these
3:54hyperscalers and frontier AI research
3:56labs.
3:58So, the issue that there is very little
4:00revenue coming from outside the
4:02ecosystem is certainly a valid point for
4:04now.
4:06But this could also be viewed as the
4:07building phase for AI, which is why
4:09significant profits aren't being
4:10generated yet.
4:14Ultimately, this verification can only
4:16be properly addressed once AI is fully
4:18monetized, becomes a service that people
4:20actually use, and generates meaningful
4:22revenue from that usage.
4:26Now for the third weakness.
4:28Isn't AI being built on debt? Recently,
4:30Big Tech's capital expenditures have
4:32caught up with their operating cash flow
4:33.
4:34The market is worried because free cash
4:36flow is turning negative. Capital
4:38expenditures are currently increasing at
4:40a rate of 70%per year. This is the graph
4:42for capital expenditures, and this is
4:43the graph for operating cash flow. You
4:45can see right away that a crossover
4:46point would be created in the third
4:48quarter, right? In fact, that result
4:50appeared in the third quarter.
4:52Alphabet and Amazon both began to show
4:54negative free cash flow.
4:56Oracle has been negative for a long time
4:58, and it's highly likely that Microsoft
5:00and Meta will soon see negative free
5:02cash flow as well.
5:04And how long will this last? It will
5:06likely continue until 2027 because there
5:08is a time lag in laying out data center
5:10infrastructure. Free cash flow is
5:12expected to turn positive again only by
5:15the second half of 2028.
5:16This means we will see Big Tech
5:18companies continuing to spend money for
5:20nearly a year and a half. In a way, this
5:22is something the market would be
5:23concerned about. Hey, can you really
5:25afford to spend that much money? And if
5:27you have no cash left, how will you
5:29manage capital expenditures? This is
5:30emerging as the core question. What you
5:32are looking at now is a projection by
5:34Morgan Stanley for 2027. The investment
5:37amount is $ 1,583 billion. It was
5:40roughly $ 800 to $ 900 billion in 2026,
5:42so it almost doubles. And for 2028? It
5:45increases by another 20%from there.
5:48AI investment is truly exceeding
5:50expectations; they say they will invest
5:52this much, but it always turns out to be
5:54more.
5:55AI investment keeps growing larger than
5:57expected in this way.
5:59A major cause for this is that the surge
6:02in memory prices is eating up almost all
6:04the increase in AI CAPEX, so I
6:05personally suspect that it won't be easy
6:07for memory stock prices to rise.
6:10Now, cash capital expenditure is $ 1.388
6:13trillion. However, $ 200 billion is
6:15recorded here as leases. $ 200 billion
6:17is about 250 trillion won. Because this
6:20is recorded as a lease, it is not
6:21captured as capital expenditure on the
6:23cash flow statement. In other words, it
6:25is not recorded as AI CAPEX. It is often
6:28simply recorded as an operating expense
6:30.
6:31But the problem here is that the reason
6:33for creating these leases is mostly to
6:35make debt look smaller.
6:37We call that an SPV.
6:39Simply put, you can understand it as a
6:41paper company.
6:43If you borrow through an
6:44off-balance-sheet SPV, you can hide the
6:46debt you hold.
6:48And you can also reduce AI CAPEX more
6:50than expected. However, a downside to
6:52creating a paper company like an SPV is
6:54that the company ends up paying more
6:56interest. What does this mean? Let's say
6:58I am Meta. And imagine I partner with a
7:00financial firm called Blackstone to
7:02create a paper company called Mota. It
7:04effectively becomes a subsidiary of Meta
7:06. Essentially, Meta can almost entirely
7:08control this SPV, which borrows money to
7:11buy a massive amount of GPUs. In
7:12exchange, Meta guarantees the credit,
7:14saying, "I'm going to use these GPUs, so
7:16you back them." In return, they help
7:18them secure debt at a lower cost. By
7:19saying, "I'll guarantee it," they've
7:21been pulling in huge amounts of
7:22financing this way. However, when Meta
7:24issues corporate bonds, they can borrow
7:26at a very low rate. They can borrow at 6
7:28%. But this SPV isn't Meta, is it?
7:32Since it's not Meta, a credit spread is
7:34added, leading to about 7%interest—
7:37effectively paying 1%more. We call this
7:39a credit spread; usually, a company with
7:41nothing couldn't borrow such large sums
7:44. But didn't Meta already say they would
7:45use it all? There are credit agreements
7:47and credit enhancements in place, and
7:49even that wasn't enough. Lately, NVIDIA
7:51and Broadcom are stepping in, saying, "
7:53Hey, it's alright.""Since you're buying
7:55our chips, keep spending," and they even
7:56provide seller financing. That’s how
7:58this 7%interest rate came to be. Now,
8:01the problem for these cloud providers is
8:03that this 1%difference is enormous. It's
8:05huge. Raising operating margins by 1%is
8:07already tough, so can they really afford
8:091%more in interest? That’s not going to
8:11be easy.
8:12Instead of that, they just decide to
8:14take it on themselves, and that's the
8:16difference; ultimately, whether it’s
8:18Broadcom, Alphabet, Meta, or Oracle, it
8:20seems more likely they will borrow
8:22indirectly via SPVs rather than issuing
8:24corporate debt directly. Looking at it
8:26now, SPV issuance is larger than direct
8:28corporate issuance. And here, regarding
8:30the private credit I mentioned earlier,
8:32since it's private and we don't know it
8:34well, unlike public funds, we can't be
8:36sure exactly what’s happening behind the
8:38scenes.
8:39That's why private credit-related stocks
8:41have fallen sharply recently, and these
8:43hidden debts could continue to emerge.
8:45This is what people are pointing out as
8:47well. In fact, if market interest rates
8:48hadn't jumped this much, there wouldn't
8:50have been a major problem. Originally,
8:52people expected the Fed to cut interest
8:53rates. Just a year ago, the consensus
8:55was that rates would be cut this year.
8:57But what happened? They ended up hiking
8:58rates, and they might do more in the
9:00future. So, while we expected a -1%
9:02change, we're currently seeing + 1%in
9:04rates. This 2%difference has started to
9:07act as a massive burden on building data
9:09centers.
9:10So, what kind of problem arose? The AI
9:12high-yield spread alone has surged to a
9:15staggering 377bp. What this high-yield
9:17spread actually means is...It’s the
9:19Treasury yield for the same duration,
9:21for example. We mentioned that the
9:23credit spread is the difference when you
9:25subtract the 10-year corporate bond
9:26yield of Nebius. This spread is exactly
9:29that, and the difference is as high as
9:313.77%. In other words, if the government
9:34bond yield is 5%, Nebius’s 10-year
9:36corporate bond has to pay 8.77%in
9:38interest. As you can see, this is rising
9:40to an all-time high. Non-AI companies
9:42usually only pay around 200 bps, so I
9:45believe distrust in AI has truly reached
9:47its peak. Even in the case of CoreWeave
9:50, take their 2032 maturity unsecured
9:52bond—the key here is that it's unsecured
9:54. You should view this as being almost
9:56identical to an unsecured personal loan
9:58, and this rate was once 15%. This 15%
10:00rate refers to new issuances; the actual
10:03interest capital cost CoreWeave is
10:05currently paying is......around 8.7%.
10:09It is significantly lower than 15%.
10:11That’s because they borrowed a lot of it
10:13back in the day. Nevertheless, the CDS—
10:15Credit Default Swap—which acts as
10:17insurance against CoreWeave going
10:19bankrupt, has risen to nearly 1,100 bps
10:22. At this level, you can multiply by
10:24five to estimate the probability of
10:25bankruptcy within five years, which is
10:27seen as nearly 55%. A 55%chance of
10:29bankruptcy within five years. A 5-year
10:32CDS is essentially insurance I bought to
10:34hedge against that risk.
10:35In reality, hyperscalers have such solid
10:38core businesses that they likely won't
10:40go bankrupt. However, the argument for
10:42AI bears is that many AI firms might
10:44have a tough time.
10:46And now for the fourth point: there is
10:48an argument that AI is not actually
10:49needed in the field. Gartner states that
10:51generative AI will experience a "Trough
10:54of Disillusionment" throughout 2026.
10:56There is something called the "Hype
10:57Cycle." You have probably heard of it;
10:59it's a process where expectations for
11:01new technology skyrocket, deflate, and
11:03then rise again. It starts with an
11:04innovation that everyone gets incredibly
11:06excited about. Then it hits a peak, and
11:08enters what is called the "Trough of
11:10Disillusionment." After that, it leads
11:12to enlightenment and reaches a plateau;
11:14most technological advancements follow
11:16this path. And we have reached the
11:18Trough of Disillusionment among those
11:20stages. Do you see how the AI hype is
11:21falling now? At first, AI was incredibly
11:24hot. People were saying, "Wow, AI can do
11:26this?""It draws Ghibli-style art,
11:27answers my questions, and creates images
11:29!"—people thought it was fascinating,
11:30but now that excitement is cooling off
11:32rapidly. Some might say, "So what? What
11:34does that even mean?" and interest drops
11:36sharply, but I’m not so sure. I think
11:38that perspective makes little sense,
11:40because in fields like software
11:41development and professional services,
11:43we are reaching a point where work is
11:45impossible without AI.
11:47From the perspective of an ordinary
11:48person, I thought that could be the case
11:50. Because whether it's GPT-5, GPT-5.5,
11:53or GPT-6, they all answer my questions
11:56perfectly well. The difference in
11:58utility felt by an individual might not
11:59be that significant.
12:01So, I understand why those who don't use
12:04AI agents much might be quite
12:05pessimistic about this AI.
12:07They might think, "I tried it and it
12:09wasn't a big deal, do we really need to
12:10keep developing this?""I'm already using
12:12it just fine as it is, right?" They
12:13could think like that. That is entirely
12:15possible. That's why there might be some
12:18hype cycle, and while some wonder if it
12:20will drop off, I see it more as a
12:21potential temporary stagnation in
12:23technological progress. We call this the
12:25"chasm," don't we? Even with electric
12:28vehicles, the chasm period is quite long
12:30, and there is clearly a time gap
12:31between early adopters and the
12:33mainstream market. Because when new
12:35technology emerges, there aren't as many
12:37people who adopt it immediately as you'd
12:38think. So, it is predicted that we will
12:40get stuck in this chasm. They are saying
12:42that the year 2026 is particularly risky
12:45. In this case, personal AI agents like
12:47Meta's Muse need to play a big role in
12:49changing public perception.
12:52I've actually used Meta Muse, and for
12:54example, when I asked it to find and
12:56cancel subscriptions I'm paying for but
12:58not actually watching, it really created
13:00an incredible economic effect.
13:02"Wow, once you try this, how did I ever
13:05live without an AI agent?"
13:06I use AI agents for all of my work, and
13:08AI even helps with more than half of the
13:11process of planning and creating these
13:12videos.
13:15Perhaps not many people feel that level
13:17of utility yet, but as AI agents
13:18continue to evolve and we reach an era
13:20where it’s just a simple click away, I
13:22believe we will cross over this chasm of
13:24disillusionment.
13:26And it also faces a lot of criticism
13:27like this. That although companies have
13:30introduced custom AI, only 5%actually
13:32reach the operational stage. So I looked
13:34into it, and that’s a result from MIT,
13:36but it’s from about a year ago.
13:39It’s a bit difficult to say if there are
13:41truly any companies making a profit with
13:43this AI.
13:44It’s true that there aren't many cases
13:46where AI is significantly driving up
13:48revenue.
13:49On the other hand, what is clearly
13:50making a difference is that AI is
13:52evidently reducing the costs being spent
13:54.
13:55I once surveyed four banks—Morgan
13:57Stanley, JPMorgan, Bank of America, and
13:59Citibank—and all four banks have seen
14:02their revenue increase by about 40%over
14:04the last four years. Yet, it is said
14:06that the headcount has only increased by
14:081%over four years. What do you think
14:10this means? In the end, while they
14:12didn't actually fire people, AI is
14:14indirectly replacing humans, reducing
14:16labor costs and ultimately leading to an
14:18increase in margins.
14:20Not every company is like this. Most Big
14:22Tech firms or large corporations are
14:24utilizing AI effectively to
14:25significantly cut down these costs. The
14:27research results are still quite varied
14:29.
14:30Some say it's bad, others say it's good
14:32, but as someone who can clearly feel
14:34that expenditures are being reduced, I'm
14:36inclined to side with this view for now
14:38.
14:39And they are also receiving this kind of
14:40criticism.
14:42There is talk that the reason spending
14:44remains flat is because token prices are
14:46falling due to model competition.
14:49If you look at the actual data, token
14:51prices have dropped by 67%over the past
14:53five months.
14:55It means the cost of running AI is
14:57becoming increasingly cheaper. However,
14:59total spending has increased by 122%.
15:03So, in the end, Jevons paradox—where
15:05lower token prices lead to higher total
15:07demand—is clearly manifesting, and the
15:09problem arises when this paradox no
15:12longer holds while token prices keep
15:14falling.
15:15As GPUs develop and systems become more
15:17sophisticated, token prices fall, but if
15:19demand doesn't increase proportionally,
15:21that could become a real problem. I am
15:23also leaning towards the AI bull case,
15:25but I definitely consider this part a
15:27risk. What if token prices drop crazily
15:30but the demand just isn't there?
15:32My concern is that only 1%of AI
15:34enthusiasts are using AI like crazy,
15:36creating demand only among them.
15:40The remaining 10%or the median 50%don't
15:42really use AI that much. Looking at it,
15:45even using Chinese models is the same.
15:47If this happens, it won't be excess
15:48demand but excess supply, causing GPU
15:50prices to crash and the GPU
15:52infrastructure built on debt could be
15:54completely ruined. I understand exactly
15:56what that means, but as of now, They say
15:58it's still in the early adoption phase
16:00and we don't know if it will expand, but
16:02I think I know. I believe it will
16:04eventually have to work out. In
16:05conclusion, the biggest topic of debate
16:07is whether computing power will be
16:09insufficient or surplus by 2027. If AI
16:10demand is as high as expected and prices
16:12don't fall, that's fortunate, but what
16:14if there's little demand? Only 80%of
16:16what was expected? In that case, we
16:18would really be in big trouble, wouldn't
16:20we? At that point, GPU prices could
16:22become worthless. Then, those numerous
16:24value chains would collapse all at once
16:25, the price of GPUs held as collateral
16:27for bonds would crash, and other
16:29financial products backed by those bonds
16:31would crumble, leading to a situation
16:32almost identical to the subprime
16:34mortgage crisis. However, speaking from
16:36the current field perspective, the
16:38demand for AI computing is still
16:39incredibly, incredibly large. If you
16:41open up 1 gigawatt of power grid right
16:43now, there would be over 100 teams
16:45lining up and pulling tickets to get
16:47that 1 gigawatt. That’s how immense the
16:49demand for AI is right now, and you know
16:51Nvidia's gaming graphics cards, right?
16:53There is a graphics card called the RTX
16:555090. When this first came out, it cost
16:57about 3 million won, and even then,
16:59people were saying it was expensive. It
17:00was originally supposed to sell for 2
17:02million won, but it went to 3 million,
17:03then 5 million, and now it’s finally
17:04starting to sell for 10 million won, or
17:0612 million won in China. The price keeps
17:08going up. In terms of the stage of AI
17:09development, it is actually at a very
17:11early stage. It's a technology that has
17:13only been out for two or three years.
17:15Therefore, I believe the market is still
17:17wide open.
17:18In fact, regarding big tech clouds,
17:20revenue for Microsoft grew by 48%,
17:23Google by 103%, and AWS by 39%.
17:26Originally. The growth rate for this
17:28cloud was about 20 to 30%, but they
17:30suddenly started hitting the accelerator
17:32to full speed, and now it keeps
17:34accelerating from 40%up to 100%. It
17:36should be slowing down, but it is
17:37actually going back up like this. And
17:39does this continue only for this year?
17:40No, it doesn't. The order backlog has
17:42already piled up to over 1.8 trillion
17:44dollars, more than 2,000 trillion won.
17:45So, we see this as generating sufficient
17:47revenue through next year and the year
17:48after as well. And now, for big tech,
17:50the forward P/E ratios have fallen
17:52significantly. For companies like Nvidia
17:54and Broadcom, the forward P/E is
17:56currently around 13 to 15 times, and big
17:58tech is at 18 times, right? This is
18:00below the 10-year average of 20 to 23
18:02times. The S&P tech stock average was 23
18:04times, but it is currently trading well
18:06below that.
18:08You might wonder if this can really be
18:10called a bubble, right? A bubble has to
18:11grow to be a bubble; but now people are
18:13pointing at something that hasn't even
18:14formed yet, hitting the air and asking,
18:16"Isn't this a bubble?""Aren't we popping
18:18this?" It seems like they keep trying to
18:19pop it. The cause, as I mentioned in my
18:21video, is that as funds liquidated and
18:23reduced their leverage positions in
18:25large quantities, the largest leverage
18:27liquidation since 2017 occurred. Because
18:29of the current volatility, when AI
18:30stocks plummet and prices fall. What do
18:32people do? They look for a reason.
18:34People ask, "Hey, why is this falling?"
18:36"Isn't it an AI bubble?" There is a
18:38tendency to artificially create a
18:39narrative about an AI bubble just to
18:41stir up fear. It really just depends on
18:43what reasons people decide to attach to
18:44it. What has actually changed in the
18:46company Nvidia from when it hit Nasdaq
18:47highs a month ago compared to now?
18:49Nothing at all has changed. Jensen Huang
18:51is still out there working hard and
18:52traveling the world. Did Samsung
18:53Electronics 'performance change
18:55dramatically in just a month? That's not
18:56the case. Samsung Electronics is working
18:58hard to make good memory chips and is
18:59exporting them very well. But as soon as
19:01the stock price falls, people suddenly
19:03come out and start arguing that AI was
19:05actually a bubble, but I'm not so sure.
19:06Actually, I believe every single logic
19:08I've presented in my videos can be
19:10refuted. Of course, it's possible that
19:12I've fallen into confirmation bias. I
19:14acknowledge that to some extent, but
19:16unless there is a lot of truly reliable
19:18data showing AI is a bubble, I think
19:20it’s an exaggeration to claim it's a
19:22bubble based on the data we have now.
19:25I keep suspecting that someone might be
19:27spreading this narrative repeatedly to
19:29short the market and profit from it.
19:32I am continuously posting these
19:34materials for free in my Telegram
19:35channel, along with reports on U.S.
19:37stock analysis, company analysis, and
19:39industry analysis, so please check the
19:41link in the pinned comment below. Also,
19:43hitting the subscribe and like buttons
19:45is a huge help to the creator. I would
19:47really appreciate it.

About this transcript

This page contains the full transcript of 월가 10개 리포트 종합, 반박까지 20분 총정리. 꼭 보셔야합니다 by 월텍남 - 월스트리트 테크남, generated from the public captions YouTube serves with the video. The transcript has 3,831 words across 603 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.