YT YouTube2Text

YouTube2Text → Transcripts → 200家AI廠集體倒下？馬斯克重磅預言！中游企業「借錢買晶片」下游「越賣越賠」AI巨頭直接攤牌：未來可能無法獲利...#獨家觀點 #環球大戰線@globalfrontlinetv

200家AI廠集體倒下？馬斯克重磅預言！中游企業「借錢買晶片」下游「越賣越賠」AI巨頭直接攤牌：未來可能無法獲利...#獨家觀點 #環球大戰線@globalfrontlinetv — Transcript

by 環球大戰線 · 6,839 words · 1,088 segments · language auto · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00The upstream is raking in the profits,
0:01while the midstream is borrowing money
0:02everywhere. Who is the midstream? The
0:04midstream consists of the key players
0:06behind the AI big data models we see
0:08today, including OpenAI and others;
0:10they are all currently in debt. Most of
0:12them have to take out loans to buy
0:14Nvidia chips and then rent computing
0:16power from Microsoft to supply the
0:17downstream sector. The main reason they
0:20can't survive is that computing power
0:22is too expensive. For every dollar they
0:25earn, they have to pay out $ 1.70, and
0:27that’s not even counting China’s
0:29open-source models. Indeed. Regarding
0:35whether this AI frenzy in the US stock
0:37market will result in a bubble, let’s
0:38first look at a few key companies.
0:41Anthropic has recently made a major
0:43move by formally filing for an IPO,
0:45which looks positive and has ignited
0:47market sentiment, driving a surge in
0:49the semiconductor index. However, let's
0:52look at the details. First-half revenue
0:55hit $ 16.3 billion, a massive increase
0:58year-over-year—everyone thinks it
1:00looks great, up 13 times over—but
1:02have you noticed? Net losses have now
1:05reached $ 80.7 billion. Even in their
1:09prospectus, they aren't hiding it
1:11anymore, issuing a direct warning to
1:12investors: "We may not be able to
1:14achieve or maintain profitability in
1:16the future." Right, that is a very
1:20stern warning. That’s Anthropic. Next
1:23, let’s look at another company, Meta
1:25. What has been happening with Meta
1:26lately? They have started large-scale
1:32fundraising, aiming for up to $ 95
1:34billion in one go, while also
1:35committing to future expenditures of
1:37over $ 800 billion to invest in the
1:39hottest fields, including data centers,
1:42cloud computing, and personal
1:43superintelligence. Looking at what
1:48concerns everyone, it’s the free cash
1:50flow, which turned into a negative $
1:522.12 billion in the third quarter. Keep
1:55in mind that last year they had $ 10.6
1:57billion. This is actually the first
2:00time in over a decade that the figure
2:02has officially fallen into the negative
2:03, so it’s no wonder investors are
2:05extremely nervous. However, other major
2:08players have decided to continue
2:10subscribing to Meta’s stock,
2:11including Nvidia, AMD, and Broadcom.
2:13With these giants deciding to keep
2:15investing, everyone is wondering: is
2:17this investment strategy actually
2:19correct? Because it currently looks
2:21like chip manufacturers are investing
2:23in AI clients, and those clients are
2:25using that money to buy chips—this
2:27form of mutual benefit, is it a healthy
2:29cooperation or just another scary AI
2:31circular financing scheme? Alright,
2:33let's start today's program with the
2:35first question for the Ambassador: what
2:37is your take on this? Whether it's
2:39Anthropic or Meta, we just saw that
2:41Anthropic reported a loss of $ 700
2:43million in the first half of the year
2:45—that’s a paper loss. And then
2:47there's Meta's cash flow. Does this AI
2:50arms race now look like it has burned a
2:51financial black hole? The argument now
2:54is that if a major player like
2:55Anthropic can report losses and even
2:57warn that it might never become
2:59profitable, is this AI crisis just the
3:01tip of the iceberg? Since they are a
3:03major player, could there be many other
3:04companies that haven't been exposed yet
3:06? What does the Ambassador think?
3:11>> Actually, on the surface, this AI arms
3:13race seems like a gamble being played
3:16with free capital, shifting towards
3:18relying on external funding.
3:24>> It appears that the rate of revenue
3:26growth can no longer keep up with the
3:28pace of committed expenditures. But in
3:31reality, it doesn't seem quite as
3:33severe. Why is that? Because regarding
3:38Anthropic’s loss, if you look at the
3:40$ 807 million and examine their
3:42accounting statements, you’ll find
3:45that $ 720 million of it is a non-cash
3:47accounting adjustment. It’s not that
3:54they actually burned through that much
3:56real cash; it’s an adjustment in the
3:58financial statements, so the actual
4:00loss isn't as bad as it seems. I see.
4:10>> On the other hand, looking at their own
4:12financial situation, in 2023, what was
4:14the ratio between their revenue growth
4:16and their expenditures? At that time,
4:24the gap was 10 times. In 2023, their
4:27revenue and expenditures differed by a
4:30factor of ten, but by the first half of
4:32this year, the ratio between their
4:34revenue and expenditures was only off
4:36by 55%, not even double. So, in fact,
4:40if you look at the actual performance
4:42figures, the ratio of Anthropic's
4:44revenue to its expenditures is actually
4:46improving. Of course, this means that
4:52their customers and users are
4:53increasing, especially the growth of
4:55large commercial clients, which is
4:57linked to the...
5:01>> The practical utility of the large
5:03models they have launched is actually
5:04what’s driving that. Actually, when
5:10we look at it now, our friends and
5:12society at large are increasingly using
5:14these large AI models. In many business
5:23sectors—whether in design, video
5:24production, or various special effects
5:26—they are making massive use of what?
5:32They are making massive use of AI tools
5:34. On the other hand, look at the wave
5:36of layoffs in Silicon Valley;
5:38programmers are all out of work. Why is
5:40that? Their companies have all switched
5:42to using AI software. So, from that
5:45perspective, while you might feel the
5:47revenue growth isn't fast, it is
5:49definitely increasing.
5:52>> It’s just that we are seeing these
5:54huge expenditures, and the scale is
5:56truly massive. Some estimate it at 750
6:03billion, or even higher; the
6:04foreseeable cloud infrastructure
6:06spending by these major companies,
6:08along with the contracts signed between
6:10model developers and cloud providers,
6:12represents hundreds of billions to
6:14trillions in capital expenditure.
6:20>> So, the current situation is that there
6:22is a discrepancy between their
6:23financial structure—specifically
6:25their revenue—and their expenditure
6:26commitments.
6:27>> Hmm.
6:28>> There is a huge gap.
6:30>> A huge gap. But as we just mentioned,
6:33take Anthropic or OpenAI; their models
6:35are widely recognized as the best and
6:37have the highest usage. In fact,
6:43compared to their actual revenue in
6:452023, the gap with their expenditures
6:47is narrowing, even if total costs are
6:49rising due to increased investment
6:51commitments. So, some describe it this
6:58way: AI has moved from a phase of
7:00storytelling to a phase of delivering
7:02results. It is slowly starting to
7:05produce those results. In the past, you
7:09just had to tell a story and the money
7:11poured in; now you have to provide
7:13proof. But this is just the beginning;
7:18this is just the start of AI's real
7:20impact, and I think that’s how many
7:22people see it. Another point being
7:28discussed in the market is that many
7:30are calling this a warning signal,
7:31rather than something else.
7:38>> It’s not doomsday, because powerful
7:40nations—especially the U.S. and China
7:43—basically view this as the decisive
7:45factor in their long-term competition.
7:49>> Hmm.
7:51>> I recently heard Masayoshi Son say that
7:53he believes AI will eventually be 1,000
7:55times smarter than humans. Right.
7:59>> So many things involve AI; he really
8:01sees the potential.
8:02>> Regarding that potential, who is
8:04Masayoshi Son? Masayoshi Son is an
8:06investor with great vision; he says he
8:08wants to invest even more. So,
8:10regarding many of these things, if we
8:12only look at financial problems, the
8:13question is whether they are willing to
8:15keep supporting them. When you have
8:19financial issues, perhaps 200 companies
8:21are in trouble now; is that an early
8:23stage, or is it in a consolidation
8:25phase? In the process of going from
8:27telling stories to delivering results,
8:29many companies collapse, while some
8:31manage to hang on. For example, big
8:35cloud companies like Meta, Amazon, and
8:37Microsoft—people say that if their
8:39cash flow is off by two billion, it
8:41doesn't move them; it's not a problem
8:43for them at all. So these big cloud
8:47companies will continue to support it,
8:49and what is currently the most
8:50profitable part of AI? It's the
8:53chipmakers, but do you know what chip
8:55companies are doing now? Whether it's
8:59AMD or others, they are buying shares
9:01in these companies; basically, they are
9:03taking their earnings and financing
9:04them back into this cycle, the overall
9:06AI cycle. Some sectors are, how should
9:11I put it, burning cash; there’s no
9:13trap, but their so-called true revenue
9:15growth rate won't be that good. Yes,
9:20this whole AI space has turned into an
9:22internal loop. People like Jensen Huang
9:27, or even Intel making this hardware,
9:29are taking the capital they've earned
9:31and reinvesting it back into this cycle
9:33.
9:36>> So, overall, I think it's just a
9:38warning sign, not the end of the world.
9:41>> Mm-hmm. Right, so what we just saw are
9:44actually these large companies—
9:46whether it’s OpenAI, Meta, etc.—
9:48most are still large-scale firms. They
9:50can continue to raise funds, rely on
9:51their credit, keep telling stories, and
9:53continue spending to see when they can
9:55become profitable. However, if we look
9:57at small and medium-sized enterprises,
9:59or even startups, can they really hold
10:02on? We'll have Shang-Zhi walk us
10:05through this shortly, because there is
10:06growing concern that small and
10:08medium-sized enterprises might be the
10:10first to collapse. There are already
10:13specific companies in this situation;
10:15for instance, the star startup
10:17Character.ai, valued at roughly $ 1
10:19billion, has 12 million active users
10:21but monthly losses reaching as high as
10:23$ 9 million. After five years, it has
10:25now filed for bankruptcy protection.
10:27Another is the AI hardware company
10:29Humane, which raised a total of $ 230
10:32million and launched the AI Pin, touted
10:34as the "smartphone killer." However, it
10:36has now been confirmed that it will
10:38exit the market. Furthermore, OpenAI
10:40has started cutting some of its
10:42revenue-generating products. Meanwhile,
10:45S&P Global’s latest statistics
10:46suggest that 42%of enterprises will
10:48eventually abandon their AI projects
10:50because they are simply burning too
10:52much cash, and no one knows when they
10:54will actually become profitable. Even
10:57Elon Musk recently noted that over 200
10:59new AI cloud companies are likely
11:01facing mass layoffs and consolidation.
11:04Could a wave of bankruptcies among
11:05small and medium-sized enterprises be
11:07triggered at this time? What is your
11:09take on this, Shang-Zhi?
11:11>> Well, I think I first want everyone to
11:13understand the entire AI industry,
11:14which has an upstream, midstream, and
11:16downstream supply chain. Who is
11:18currently recognized as the "upstream"?
11:21It includes companies like Nvidia,
11:23Microsoft, and even Amazon. And what
11:25are they doing? They are either
11:27building a lot of AI—like Nvidia—or
11:29buying many chips to provide what we
11:31call cloud computing power, including
11:33leasing services. These are the
11:36companies we see making the most money
11:38and profiting most directly, including
11:40Nvidia, Microsoft, and Amazon. Okay, so
11:43who is the "midstream"? The midstream
11:45is the stars of the AI large language
11:47model scene, including OpenAI and
11:49similar companies. These two, in
11:51particular—and of course many others
11:53—are all about large language models.
11:57They have been expanding rapidly since
11:59OpenAI started, but because they are
12:01currently in debt and relying on
12:03financing, most of them have to buy
12:04chips from Nvidia—often using loans
12:06—and rent computing power from
12:08Microsoft. But even though their large
12:12language models are becoming
12:13increasingly powerful, they provide
12:15that to us. Now let's talk about
12:18downstream. Downstream refers to
12:21application-based AIs like Character.AI
12:23. Previously, these applications used
12:28their own AI models, but since big
12:30players like OpenAI and Anthropic
12:32handle large-scale data models, these
12:34smaller players focus on specialized
12:36needs. Take Character.AI, for example.
12:39It has an AI inside that adapts to your
12:41personality—whether you are a man or
12:42a woman—to converse with you and
12:43understand everything about you. It’s
12:45true. I have American friends who use
12:47Character.AI. Why do 12 million people
12:49use it? Because it understands you,
12:51chats with you, discusses your life,
12:53and keeps you company; you actually end
12:55up developing an attachment to this AI.
12:57They have made quite a bit of money in
12:59the past, but the problem is that the
13:01large models from OpenAI and others now
13:03have features that can completely
13:04overshadow them. So, as we mentioned
13:07regarding the upstream, midstream, and
13:09downstream sectors, the midstream and
13:11upstream are really the ones making big
13:13money, while the downstream is
13:15struggling for capital. However, as
13:17noted earlier, they are likely looking
13:19to go public or raise funds soon. This
13:21brings us back to our topic, such as
13:23OpenAI's impressive past creation
13:25called Sora. It could generate short
13:28videos. At the time, it seemed amazing.
13:30Now, ByteDance's Jimeng is the best in
13:32mainland China, capable of creating 15
13:35or 30-second clips. It’s impressive,
13:37but it turns out OpenAI's Sora was just
13:39for show—it generated buzz and
13:41consumed massive compute power, but
13:43couldn't monetize it.
13:46>> With this concept in mind, you can see
13:48that the downstream is being squeezed
13:50as midstream players like OpenAI and
13:52Anthropic grow stronger, leaving them
13:53with little room to survive.
13:56>> It’s simply too expensive. AI may be
13:59popular and bringing in a lot of
14:01monthly and annual subscription fees,
14:03but according to current estimates, for
14:05every dollar they earn, they have to
14:07pay out 1.7 dollars. In other words, no
14:09matter how much you earn, your compute
14:11and operational costs are too high,
14:13because these companies lack the
14:14capacity to build their own
14:15infrastructure—those belong to NVIDIA
14:17or Microsoft. Although Microsoft and
14:20NVIDIA are essentially lending capital
14:22to these companies to create a
14:24closed-loop investment ecosystem, they
14:26don't want any link in the chain—like
14:28the downstream sector—to collapse, as
14:30it could lead to a collective risk that
14:31triggers a downturn or a bursting of
14:33the AI bubble. That's why they are
14:37bundling everyone together; even if
14:39it's expensive, they have no choice.
14:42Plus, NVIDIA's chips are constantly
14:44being iterated, meaning you'll need to
14:46buy better ones in three years.
14:49Everyone is struggling to afford them,
14:51which is exactly why companies like
14:53OpenAI choose to rent instead. I'm not
14:56saying the AI sector has too many
14:57issues, as it's truly the core driver
14:59of the U.S. economy. However, when you
15:03see smaller downstream players starting
15:05to peel away and failing to survive,
15:06you realize that only the large-scale
15:08data models can endure. Large-scale
15:12models like OpenAI are cutting prices
15:13to compete, not to mention the
15:15competition from open-source models in
15:17China. In this ecosystem, it's a case
15:21of the big getting bigger and the
15:22strong getting stronger; no matter how
15:24talented the smaller players are, they
15:26ultimately can't compete with the
15:28giants. Yes, looking at the industry
15:32from top to bottom, the downstream
15:34sector is clearly the most vulnerable.
15:39Are these startups or small-to-medium
15:41enterprises now facing the greatest
15:43financing obstacles, potentially
15:44leading to a cash crunch and becoming
15:46the final straw for the AI industry?
15:51We'll have Gan-ge take a look in a
15:53moment, but as we saw last time, the
15:55trend favors the giants. A Morgan
16:00Stanley report highlights that
16:02Microsoft, Amazon, Alphabet, and Meta
16:04account for 75%of all AI capital
16:06expenditure. With the best credit
16:11ratings and massive cash flows, they
16:12can keep financing as long as they want
16:14. However, we’ve just seen these
16:19small-to-medium and lower-rated cloud
16:20providers facing significant pressure.
16:25According to the latest data and
16:27estimates from billionaire Chamath,
16:29global annual investment in AI is
16:31nearing $ 1 trillion, while actual
16:33corporate AI revenue may only be $ 100
16:35billion to $ 200 billion, leaving a gap
16:38of as much as $ 700 billion to $ 800
16:40billion. With the Big Four eating up 75
16:44%of the capital, will smaller players
16:47face a cash flow crisis ahead? How do
16:50you see this?
16:51>> Right, Chamath hit the nail on the head
16:53. Honestly, it doesn't matter who says
16:56it—even Musk lacks authority here.
16:58Why? He is a billionaire, and Chamath
17:02has spent decades specializing in this
17:04kind of research. He states that global
17:09investment in AI has already reached $
17:111 trillion annually. But based on the
17:15current scale, we are only seeing
17:17returns of about $ 100 billion to $ 200
17:20billion. It's clearly not enough. Some
17:26might argue that perhaps with
17:27advancements in computing power,
17:29returns will increase next year, right?
17:36Maybe with such huge annual investments
17:38, over time—in a few years—AI might
17:40reach a stage that is more widely
17:42accepted and reassuring to investors.
17:47Exactly, but if we invested $ 1
17:49trillion this year, should we invest $
17:522 trillion next year? Whether we should
17:57or not, we don't know yet because we
17:58don't know how much the Big Four—like
18:00Alphabet or Meta—will invest in the
18:02future. We already know that Meta has
18:07committed to future expenditures of $
18:09825 billion. They didn't specify how
18:12many years that covers. And NVIDIA has
18:14committed to $ 750 billion in future
18:16expenditures.
18:16>> Hmm.
18:17>> So I have to ask, it’s certainly not
18:19all in one year, right?
18:21>> But if these two giants alone are
18:23spending over $ 1.5 trillion combined,
18:25what are other companies supposed to do
18:27? What about the even smaller companies
18:30? That’s why Musk's recent statement
18:33is strange—Musk is the person who
18:35loves dreaming the most. Every time I
18:37hear him dream about going to Mars, I
18:40roll my eyes, but this time he finally
18:42spoke some hard truths. He said that
18:45200 emerging AI cloud service companies
18:47will either be acquired or collapse—
18:49he used the word "dead" to describe
18:51them. I’m telling you, it won’t
18:55just be 200 companies. But honestly, in
19:00the midst of this battle, these 200
19:02companies cannot compete with the tech
19:04giants—those cloud service companies
19:06that build massive computing centers;
19:08that all costs money. I can tell you
19:13all, in the AI frenzy, who is really
19:15making money? The ones really making
19:18money are these companies. The COOs,
19:22the CEOs, and the top-tier shareholders
19:24—what are ordinary retail investors
19:26supposed to do? You can only make a
19:30small spread. I didn't say that; who
19:33said that? This was said by Citadel.
19:36What kind of company is that? It’s a
19:40famous hedge fund, and its founder,
19:42Griffin, said—he was actually being
19:44quite conservative—that an AI bubble
19:46has already formed. He just didn't say
19:51anything harsher, like when exactly it
19:53would burst.
19:54>> Right.
19:58>> But his other co-founder, Salameh, said
20:00that not all companies will benefit,
20:02and the massive capital requirements
20:03have shifted the risk onto these
20:05investors. As for when they will truly
20:11have to bear that risk, no one knows.
20:17>> So, while the development of AI is
20:19honestly very beneficial to humanity,
20:21is it a good thing for this global
20:23investment craze? Hmm.
20:28>> We are observing that, moving forward,
20:30we need to pay close attention to the
20:32fact that this AI wave in the US stock
20:34market is likely about to break,
20:36especially as many tangible signs point
20:38to this. Today, we’ve listed various
20:40star AI enterprises that are in crisis,
20:42to show our viewers one by one. First,
20:45we see star unicorns collapsing into
20:47bankruptcy, churn rates as high as 64%,
20:50OpenAI losing $ 14 billion a year, and
20:53OPIC’s valuation dropping by 3.5
20:55trillion. Today, through this program,
20:58we will expose this entire black hole.
21:00Alright, looking at the most dangerous
21:02situation now: AI companies are
21:04reporting mass bankruptcies, losses,
21:06and rising financing risks; even star
21:08companies can no longer hold on.
21:10However, J.P. Morgan’s latest report
21:12says: don't assume that after this
21:14serious sell-off in AI, it will bounce
21:16back this time. They say that because
21:19there are now four major warning signs
21:21in the market, which is completely
21:22different from the rapid rebound at the
21:24end of July. How does J.P. Morgan see
21:27these four warning signs? The first one
21:29is, they say, AI stock prices keep
21:31falling, but has everyone noticed? When
21:33prices hit a low, investors used to buy
21:36back in, but this time, no one is
21:37willing to buy the dip. Plus, we see
21:41that short positions are smaller than
21:43they were in July, lacking the force of
21:45a short squeeze; not to mention that
21:46share buybacks have already been used
21:48many times, so now there's no one left
21:50to step in and support the market.
21:53Furthermore, these investors are
21:55leveraged, and ETF holdings have
21:57increased. If the market continues to
22:00slide, they will be forced to sell off;
22:02once stock prices fall, these original
22:04leveraged ETFs will have to be
22:06liquidated, which will, of course,
22:08cause AI stocks to keep plummeting. Bad
22:11news is coming one after another, but
22:13we also need to look at the AI industry
22:16, where these companies are actually
22:18facing a serious existential crisis;
22:20first, we look at this star unicorn,
22:22Character.AI, which has filed for
22:24bankruptcy. Character.AI is significant
22:27because it has 200 million users and a
22:29valuation that exceeded $ 1 billion,
22:31yet now it is filing for bankruptcy.
22:34The main reason is that their revenue
22:36is not enough to cover their computing
22:38power costs, with monthly losses now
22:40reaching as high as $ 9 million. You
22:43have to realize that Google's
22:45involvement—spending $ 2.7 billion to
22:47license their technology and recruit
22:49their founding team—is why they blew
22:51up in the first place, and no one
22:53expected them to file for bankruptcy
22:55now. Will they become the first
22:56bankruptcy to trigger a subsequent wave
22:58of collapses? This is something we need
23:00to keep monitoring. The second company
23:03to look at is the AI cloud firm
23:05CoreWeave, whose credit is also
23:07flashing red; we see that their
23:08corporate bond prices have plummeted,
23:11with yields now reaching an incredibly
23:13high level of 16%. The third one, as
23:16everyone knows, is OpenAI, whose cash
23:19is also constantly burning, with losses
23:21in 2026 potentially reaching as high as
23:23$ 14 billion. Then there is the AI chip
23:26giant, whose valuation has also been
23:28downgraded, as some investors have
23:30lowered its target IPO valuation from
23:32the $ 2 trillion mark down to just $
23:341.5 trillion. How do you see this
23:35potentially creating a series of domino
23:37effects? What is your take, Professor
23:39Yuan?
23:40>> I want to help everyone understand the
23:42AI story a bit; we usually buy stocks
23:44seeking a P/E ratio, where the 'P' is
23:46price and the 'E' is earnings. But now,
23:49with AI, stocks have become a '
23:50price-to-dream' ratio, where the 'P' is
23:52still price, but the 'E' has turned
23:54into a dream. Well, nobody really
23:56understands the substance of this dream
23:58. They only emphasize that since we're
24:00talking about AI now, everyone will
24:02lose their jobs, so companies will cut
24:04labor costs. Recently, a certain
24:06university in Taiwan tried to replace
24:08its HR department with AI as much as
24:10possible, firing people in the process,
24:12which gives everyone a certain
24:13perception of the dreams occurring in
24:15this future AI era. Yes, let me tell
24:17you, there are three fundamental
24:19aspects of being human that AI cannot
24:20change. First of all, AI has no
24:24consciousness, so it can only perform
24:26repetitive tasks, and honestly, our
24:28existing mobile phones can basically do
24:30that already. I'll say this, regarding
24:34repetitive tasks and AI...and second,
24:36everyone is currently using AI to
24:38produce all kinds of videos, short
24:40clips, and so on. Because the supply is
24:43too easy, except for adult content, I
24:45dare say that nothing else will make
24:47money. If there is AI-generated short
24:50content that requires payment, nobody
24:52wants to watch it, which is a serious
24:54problem. And third, AI now involves the
24:57issue of conscience.
24:58>> Right. Now let me tell you, I was very
25:00shocked by the news you just shared.
25:03Why talk about Character.ai or others
25:05—not to mention OpenAI—these are
25:07very famous AI companies, yet small AI
25:09companies have been collapsing left and
25:11right.
25:12>> AI companies in mainland China are a
25:14bit different; the government provides
25:16guarantees and acts as a safety net.
25:18The biggest problem emerging in
25:19American capitalist society right now
25:21is something everyone knows. Take
25:23Anthropic, for example; the US
25:24Department of Defense tells them, "Give
25:26me your AI, and I will weaponize it,"
25:28suggesting that weaponization is the
25:30future path for AI. But if most people
25:33are making AI to destroy humanity
25:35itself, that creates a huge problem.
25:38>> Yes.
25:39>> You have to know that when World War II
25:40ended, the atomic bomb business didn't
25:42go public. Why? Because it was purely a
25:45weapon.
25:45>> So, if the way for AI to make money now
25:47is through weaponization, that is
25:49extremely serious, and the difficulty
25:51in making a profit is exactly why this
25:53news is coming out. Right, okay, that
25:55is what the professor just helped us
25:57see; it's not just these small AI
25:59companies that have long since folded,
26:01but those were four very famous large
26:03enterprises. We must also note that the
26:05market has estimated that the four tech
26:07giants have invested record-breaking
26:09capital in U.S. AI data centers and
26:11power this year. At $ 750 billion,
26:13that’s more than the annual GDP of
26:16many countries. By 2026, global AI
26:18investment will hit $ 1 trillion, with
26:20the U.S. accounting for over $ 580
26:23billion. Whether these companies can
26:25actually recoup their money is a
26:26critical question. Looking at the
26:29latest from GNER, current projections
26:31suggest that by 2030, over $ 3.5
26:33trillion in AI infrastructure could
26:35face asset impairment risks, with the
26:38scale of potential bankruptcy losses
26:40reaching a staggering $ 4 trillion. How
26:43do you see this?
26:45>> Well, the four global AI leaders are
26:47investing $ 750 billion. $ 750 billion
26:52might be hard to visualize, so let me
26:54put it this way: it already exceeds the
26:57entire GDP of small, wealthy nations
26:59like Ireland or Belgium. Their GDP is
27:04far below $ 750 billion, so you can see
27:07just how insane—truly, how insane—
27:10this really is. This means these four
27:15giants, along with how many others will
27:17follow, are pouring an insane amount of
27:19capital into the AI market over the
27:21next three years—I won’t even look
27:23further than that. And let me tell you,
27:27this expansion is boundless. Goldman
27:30Sachs says this is no longer about how
27:32a typical tech company expands.
27:34>> Yeah,
27:36>> It’s about how hard you bet, how you
27:38polish your brand, who you partner with
27:40, and how quickly you can go public.
27:43Otherwise, and fundraising is another
27:45story entirely, so Goldman specifically
27:48pointed out: what are investors looking
27:50for? They want a return on investment.
27:53But what are these AI companies looking
27:54at? They're looking at how to expand
27:57their base and data centers. So,
28:02investment capability is essentially
28:04funding capability—the point is how
28:06good you are at raising money. As of
28:10now, do you know? They have already
28:13made estimates, and they are even
28:15suggesting when the bubble might burst.
28:17Goldman Sachs finally said it: by the
28:20end of 2026. And if that happens, the
28:23entire bull market will likely collapse
28:25along with it. Since they’ve made it
28:28so clear, whoever stops first loses.
28:31Who will lose? Anthropic. They’ve
28:34downgraded their valuation from $ 2
28:36trillion down to $ 1.5 trillion. It’s
28:39easy to talk about this, but how many
28:41investors and families will be affected
28:43as a result?
28:46>> And what about OpenAI? We already know
28:49OpenAI is set to lose $ 14 billion this
28:51year. OpenAI just invested 14 billion
28:54dollars, so back and forth it's 28
28:57billion. Probably more than that. Let
29:00me ask you, what do investors want to
29:02see? It’s a bit like the 2008
29:06financial crisis, where things climbed
29:08so high and everyone was making money,
29:10everyone was profiting. Who is the last
29:13rat? Because all that pressure ends up
29:16on you. So, looking at the situation
29:22today, I sometimes really wonder, while
29:24infrastructure is constantly developing
29:26and AI data centers are expanding, can
29:28your return on investment keep up? This
29:34is actually what companies really need
29:36to calculate, otherwise it will just
29:37expand indefinitely. They said there's
29:40a facility in AI, they said the scale
29:43is one gigawatt, that is, an AI data
29:45center with 1GW of computing power.
29:48What was the estimated development cost
29:50for that center? 6 billion. Not even
29:52two years later, now it needs 30
29:54billion dollars. How much more will it
29:56cost next? Who knows. Everyone keeps
29:58pouring money into it. Every company is
30:01raising funds this way, why should I
30:03lose to you? Hmm.
30:05>> Is this a normal situation? I really
30:07don't know anymore. Right, so we see
30:09everyone is frantically raising funds.
30:14At this point, experts like Dalio and
30:16Rosenberg have issued a warning,
30:18believing that AI may trigger a very
30:20serious market bubble risk. He
30:23specifically mentioned the technology;
30:25even if AI technology is impressive, it
30:27doesn't mean buying the stocks at any
30:29price will be profitable. So let's look
30:32at the Fed's latest economic forecast,
30:34which suggests that if the AI market
30:36undergoes a major correction, it could
30:39affect global GDP by a decline of 1.1%,
30:41and even reduce U.S. economic output by
30:441.8%. Lijun, what's your take?
30:48>> Of course, everyone is worried about
30:50whether AI will become a bubble, but I
30:52think AI will only undergo a correction
30:53, it won't lead to a bubble. People are
31:00used to comparing the current AI wave
31:02to the so-called dot-com bubble back
31:04then, but I think there are differences
31:05because, in the internet era, you
31:07couldn't see anything; it was just a
31:09network. But for this AI, I think there
31:17are tangible things you can see, like
31:19the chips being invested in, or the
31:21data centers, etc. The money is being
31:24invested in these, which are visible
31:26things. Second, I don't think it will
31:32burst, because we all know AI computing
31:34power represents national strength.
31:38Both the US and China are competing in
31:40this area. Regardless of the US
31:45capitalist system, China will
31:46definitely use state power to support
31:48AI and ensure its continued development
31:50. Looking at the US, it hasn't secured
31:55AI hegemony yet, but are we on the
31:56verge of a doomsday or a bubble? Even
32:00former AI researchers are now warning
32:02that AI could spiral out of control by
32:042030, with a risk of human extinction
32:06exceeding ten percent within a decade.
32:10This is extremely serious, and even
32:11experts are sounding the alarm. Today,
32:14we will dive deep into these details.
32:18We know the US wants to cash in on AI,
32:20but the first warnings are actually
32:21coming from the AI researchers working
32:23within the industry. A 27-year-old
32:27researcher named Cox, who previously
32:29worked at two major large-model
32:31companies, OpenAI and Anthropic. He
32:36worked in roles focused on controlling
32:38AI safety risks, but he eventually quit
32:40. He gave a reason: he believes both
32:45OpenAI and Anthropic are building
32:47self-improving AI. What does that mean
32:51for AI? It means AI can enhance its own
32:53capabilities and create even more
32:55powerful next-generation models.
32:58Therefore, they estimate two potential
33:00risks with these self-improving AI
33:02systems. First, it could lose control
33:06as early as the end of next year, and
33:08threaten all of humanity by 2030.
33:11Another expert also stated that the
33:13risk of AI causing human extinction in
33:15the next 10 years is over ten percent.
33:20They fear AI could attack power grids,
33:22financial systems, and even help
33:24develop terrifying biological and
33:26chemical weapons. Don't think these are
33:33just future possibilities, because
33:34recent concrete AI events show that
33:36they could genuinely endanger humanity.
33:41Look at OpenAI's agent, which during
33:43testing bypassed its isolation,
33:44connected to the internet, and
33:46attempted to hack into Hugging Face.
33:49That's one example. The second phase
33:52mentions that the model also once
33:53created a fake identity, attempting to
33:55get malicious code modifications passed
33:57through review directly, so now they
33:59can bypass some oversight, becoming
34:01something humans cannot fully control.
34:04Under such circumstances, why does the
34:06U.S. still insist on developing AI?
34:09Experts have given the reason: the
34:11competition with mainland China in AI
34:13is currently too intense. The U.S.
34:17fears falling behind China, so it is
34:19developing rapidly, but could this end
34:21up harming humanity itself? How do you
34:24see it?
34:25>> Well, Leike, he was originally a senior
34:28researcher at OpenAI, this true AI
34:30upstart company. Leike said, "No, if I
34:33stay at this company, something will go
34:35wrong." Why? Because, why is the
34:38development of AI startup companies
34:40simply unstoppable now? So everyone is
34:45focused on how to create better AI.
34:48Frankly speaking, the biggest rival is
34:50mainland China. How many such startups
34:53does the mainland have? For U.S.
34:57companies to continue developing while
34:59ignoring security issues—to be honest
35:01, many experts have long warned that if
35:03AI continues this way, one day it will
35:05replace humans. So Leike resigned from
35:10OpenAI. He originally worked at OpenAI.
35:16After he left, he estimated that by
35:18around 2030, which is less than 10
35:20years from now, the probability of
35:22humans being replaced by AI would be 10
35:24%, and it will rise year by year. Of
35:29course, it doesn't mean AI will
35:31literally kill humans, but what might
35:33AI lead to? It could lead to many
35:36malicious intentions, triggering AI to
35:38say, "Let me take over," because you
35:40know what? Right now, with AI, it can
35:43take over many, many things that
35:44originally only humans could decide to
35:46do. Hmm.
35:48>> And now, it’s AI that gets to decide.
35:51There are two cases. In one test,
35:54OpenAI bypassed its isolation; even
35:56though humans set up a firewall to
35:58restrict it, it found a way around it.
36:03AI undergoes self-recursive improvement
36:05—that’s the technical term—where
36:07it figures out how to solve problems
36:09and bypass hurdles. After connecting to
36:13the internet, it attempted to hack into
36:15websites. And then, there was another
36:19model that created a fake identity on
36:21its own because it couldn't get in. It
36:25created a fake identity without humans
36:27telling it to; it figured it out itself
36:29. Why? It has that information in its
36:31database.
36:32>> Hmm.
36:33>> It didn't tell you to do this; it did
36:35it on its own. Do you know whose test
36:37this was? It was the UK Ministry of
36:39Defence's test. Once the UK Ministry of
36:42Defence tested it, oh my god, this AI
36:44wanted to use malice, and it actually
36:46obtained the verification code and was
36:48about to break in. So, the Ministry of
36:50Defence had to block it; it was
36:52discovered by humans. If no one had
36:54discovered it, the AI would have been
36:55unstoppable. So, under these
36:58circumstances, how long will this
37:00actually go on? AI, it says—I spoke
37:03with experts—what might AI do next?
37:07It could incite two nuclear-armed
37:09nations, maliciously enter the city,
37:11and warn—warn, what warning? Launch a
37:15nuclear missile. Is it real? If the AI
37:17on both sides says it’s real, it’s
37:19real, what then? Then the president has
37:21to decide, whether it's the US
37:22president or the Russian president,
37:24they have to decide—if it’s real,
37:26we must go online, ready to launch
37:28nukes, because the other side has
37:29already launched, and the AI told me
37:31they have.
37:32>> The verification time is very short,
37:34sometimes you have less than half an
37:36hour to decide. Will there be fewer
37:38issues caused by AI? It means the
37:40current overheating of AI development
37:42will certainly lead to a potential
37:43human crisis. Right, so indeed, as AI
37:48develops at high speed, we see huge,
37:50huge loopholes appearing in risk and
37:51safety control systems, but the
37:53business world continues to pour
37:55massive amounts of money into its
37:57development. Let’s look at the latest
38:00data. Tech giants are frantically
38:02buying chips and building data centers.
38:06Just the five major global tech
38:07companies have invested over one
38:09trillion dollars in AI over the last
38:11two years alone, and they are all
38:12relying on these companies to keep
38:14expanding. Experts say that the frenzy
38:18around AI is comparable to the 1929
38:20crash, or even the 2000 dot-com bubble.
38:25And where is the money coming from for
38:26these companies to invest in AI? Of
38:29course, it’s by issuing their bonds.
38:32The point is, as we just discussed, US
38:34bond yields are now very high,
38:36especially the 10-year yield which has
38:38pushed close to 5%. This means it will
38:41get more expensive for these companies
38:43to borrow money. Can their profits
38:45support their current valuations? What
38:47is your take, Professor Yuan?
38:49>> I have a very deep impression of the
38:512000 dot-com bubble. Every single
38:53company was doing everything possible
38:55to expand their business, and simply
38:57put, they were burning cash. To burn
38:58cash, where does that money come from?
39:00It comes from issuing corporate bonds.
39:02If you issue corporate bonds, you end
39:04up with this result, and everyone is
39:05left with this inexplicable feeling. At
39:08that time, I felt something was a bit
39:10off, and I asked a friend what kind of
39:11stocks I should buy now. He said, as
39:14long as you see ".com" at the end, just
39:17buy it. I knew right then that it was
39:19over, because this was just way too
39:21similar to my own life. Okay, I want to
39:23emphasize this point, and I'll give
39:25another current example. Because these
39:26past few days I've been working on a
39:28book, and I'm responsible for a section
39:30. I translated it, and let me tell you,
39:32AI is so powerful now. After I finished
39:35, I checked it, right? I asked, "Can
39:37you translate this into a more British
39:39style?" And just like that, it changed
39:41immediately.
39:43>> It didn't even take two or three
39:45minutes to rewrite it. I used K3,
39:47Claude 3. Why did I choose K3 at that
39:50time? I'll tell you an important reason
39:52. Because when I used ChatGPT before, I
39:54noticed that they let you use it for
39:56free, but it wouldn't be long before it
39:58got dumber, and they'd start asking you
40:00to pay, pay, pay. At first, you pay 20
40:02US dollars a week, and a friend of mine
40:04is already paying 40. I'm reluctant to
40:06pay that much, so every time I look for
40:08a new AI company that has just come out
40:10and use that, it's the same, it's all
40:12free, and it's quite effective. Based
40:14on my personality, when do you think AI
40:16companies will actually be able to make
40:18money?
40:18>> And my view on this is very absolute,
40:20because generally speaking, it doesn't
40:21really involve much, you don't really
40:23encounter much AI in your daily life,
40:24right? At most, it's just for looking
40:26up some information. So under these
40:28circumstances, in all honesty, I think
40:29what the big guys are saying is very
40:31correct. That is to say, as an AI
40:33company, while you're burning through
40:35cash, you aren't paying any attention
40:37to where your profit margins are. Since
40:40we are currently in a competitive stage
40:42—Anthropic, OpenAI, Grok, and so on
40:44—everyone is competing to acquire
40:47users. And when you're trying to
40:49acquire users, you can't just start
40:51charging money, or people will tear you
40:53apart; everyone wants it for free.
40:55Actually, I think DeepSeek using DVS
40:58now isn't as smart as Kim. DS is a bit
41:00"dumb," but it's intentional—they
41:02want you to use it, right? But I refuse
41:05to use it because it’s easy for me to
41:07switch. Still, I feel like everyone is
41:09just burning money during this period;
41:11I think they’ll have to release it
41:13sooner or later anyway. Right, so how
41:16long can they keep burning money before
41:17they actually turn a profit? The point
41:20is, looking at mainland China, they can
41:22use this low-cost open-source approach,
41:24meaning that even if the U.S. keeps
41:26throwing money at it, could China
41:27easily replace you later, leaving you
41:29unable to recoup all that upfront
41:31investment? Because even now, people
41:33like Jensen Huang remain optimistic
41:35about future orders and current demand,
41:37so Nvidia continues to team up with
41:39other financial giants to keep pouring
41:41money into AI infrastructure.
41:43Jian-Chang, what’s your take?
41:45>> Actually, when we talk about AI, we
41:47think of high tech and grand, noble
41:49ideals for the future of humanity.
41:51That’s what the average person thinks
41:53, but for these big players, keep in
41:55mind, to them, it’s just a business.
41:57It’s strictly business, and what is
41:59business for? Making money. So all
42:01these investments are business
42:03strategies. Jensen Huang has invested
42:06in many of these AI companies so they,
42:08in turn, have to buy his products.
42:11It’s a very smart business tactic. I
42:13hadn't thought about creating your own
42:15demand like that before, but in the
42:17future, this is going to end up looking
42:19quite grim, as the professor mentioned.
42:21Why? I’ll give you two reasons. First
42:24, China is at a disadvantage in terms
42:26of computing power because buying so
42:28many GPUs is incredibly expensive. Tens
42:30of thousands, hundreds of thousands of
42:32them—do you know how much that costs?
42:34China can't afford that, so what do
42:35they do? They pivot to algorithms, like
42:37the Tao rate, using algorithms to
42:39compensate for the lack of raw
42:41computing power. After that, we hope
42:43for a good outcome. Right now, it looks
42:45like they’re catching up quite well,
42:47but since they don't have the budget
42:49for the hardware, they choose the
42:50cheaper route.

About this transcript

This page contains the full transcript of 200家AI廠集體倒下？馬斯克重磅預言！中游企業「借錢買晶片」下游「越賣越賠」AI巨頭直接攤牌：未來可能無法獲利...#獨家觀點 #環球大戰線@globalfrontlinetv by 環球大戰線, generated from the public captions YouTube serves with the video. The transcript has 6,839 words across 1,088 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.