YT YouTube2Text

YouTube2Text → Transcripts → 【曲博科技教室 EP325】英特爾、晶呈、創新服務、AMD、力成、Toppan、南電、景碩、欣興、宜鼎、晶化、東佑達、漢辰、營邦、天虹、康寧、崇越、補丁、台灣電鏡、牧德、聯一、聯友金屬、雍智、美光

【曲博科技教室 EP325】英特爾、晶呈、創新服務、AMD、力成、Toppan、南電、景碩、欣興、宜鼎、晶化、東佑達、漢辰、營邦、天虹、康寧、崇越、補丁、台灣電鏡、牧德、聯一、聯友金屬、雍智、美光 — Transcript

by 曲博科技教室 Dr. J Class · 5,931 words · 949 segments · language zh-TW · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Hello everyone, welcome to Bo's Tech
0:02Classroom, I'm Bo. It's Thursday night,
0:05time for our tech Q&A. If you enjoy our
0:08show, remember to like, subscribe,
0:10share, and hit the notification bell.
0:13Now today, let's start with a warning
0:15about scams. Bo’s official YouTube,
0:18Facebook, and LINE communities will
0:20never discuss individual stock prices
0:22privately with you, nor will we
0:24recommend specific stock investments.
0:26If you receive any messages from
0:28strangers asking about your stock
0:30trading or recommending stocks, those
0:32are scam groups. Please don't hesitate
0:35to report them and spread the word.
0:38Thank you for your help. First, let's
0:40take a look at the questions from last
0:42week that we haven't answered yet. The
0:47first question is from a URL provided
0:49by a viewer regarding a widely
0:51misunderstood Intel MicroLED packaging
0:53patent. It was thought to be aimed at
0:56optical communication, but it turns out
0:58to be for making CPUs emit light.
1:03Someone asked about this last week, and
1:05I've explained that I’m not sure
1:07about this "making CPUs glow" thing,
1:09but if it's just to make the CPU light
1:12up, it’s meaningless. If you just
1:20want to let the user know it's working,
1:22you don't need a special MicroLED, just
1:24one or two would suffice. A company
1:31like Intel must be conducting research
1:33and development on MicroLED
1:35communication. The reason is simple: it
1:42is an industry consensus that future
1:44processors will need to use optical
1:46communication, and using MicroLED is
1:48essentially one of the proposed
1:50solutions. Recently, something happened
1:56: the chairman of Largan Precision held
1:58an investor conference and stated that
2:01MicroLED optical communication is not
2:03feasible. He put forward several
2:05reasons for this. Following that, AUO's
2:08stock price plummeted. As a result, the
2:11chairman of AUO came out to clarify
2:13that things aren't actually like that.
2:15So, who is right and who is wrong? I
2:17will film a separate episode to explain
2:20the reasons behind this. In short,
2:23MicroLED is one of the solutions for
2:25chip-to-chip optical communication, and
2:28it certainly has its advantages. The
2:32issues raised by Largan's chairman do
2:34exist, but the technology still holds
2:36its own advantages. As for the pros and
2:40cons of MicroLED compared to regular
2:42lasers, I will make another video to
2:44compare them in detail, and I think
2:46we'll discuss this issue further in the
2:48future. Next, a viewer asked if the
2:54LADY laser drilling and gas etching
2:56technology from Jingcheng Technology
2:58and the MAN technology from Innovation
3:01Service are viable alternatives to
3:03electroplating processes. First, let me
3:07explain what LADY is. This is actually
3:14a technology proposed by Jingcheng
3:16Technology, Laser Arrow Decomposition
3:19Year, which is essentially for glass
3:21via drilling; it uses a tiny laser beam
3:23to modify the glass, followed by
3:25high-density drilling and etching with
3:27Jingcheng’s proprietary gas. As you
3:36know, Jingcheng has been pushing for
3:38dry etching, which means using gas to
3:40etch the glass, and this is indeed one
3:42such method. However, the cost of dry
3:48etching is theoretically higher than
3:50wet etching, which uses aqueous
3:52solutions. So, fundamentally speaking,
3:58this is certainly viable, but the
4:00etching process likely faces some
4:01difficulties, so I believe they are
4:04still working on it and it needs more
4:06testing, but it is theoretically
4:08feasible. As for the MAN technology
4:13from Innovation Service, it uses
4:14massive transfer of copper pillars;
4:17what does that mean? In packaging,
4:20sometimes the chip itself is quite
4:23thick. Therefore, you need to run metal
4:26interconnects from the bottom layer to
4:28the top layer. The depth, or length, of
4:33those metal interconnects could reach
4:35up to one millimeter. And one
4:40millimeter is 1,000 micrometers. As you
4:43know, for copper pillars over several
4:45hundred micrometers, using
4:47electroplating takes a very long time
4:49and the cost is quite high, so how does
4:51Innovation Service solve this? They are
4:55really a very creative company; their
4:57boss came up with the idea of using
4:59massive transfer of copper pillars. And
5:02as you know, massive transfer means you
5:04can move several thousand copper
5:06pillars at once. Of course, compared to
5:11electroplating, which processes the
5:14entire wafer or glass panel at once,
5:16massive transfer is still slower—I
5:18don't think that can be changed—but
5:20massive transfer has its own advantages
5:23. For one, the quality of the copper
5:31pillars used in massive transfer should
5:33be better than those grown by plating;
5:35and then there is the difference in
5:36speed between transfer and plating. The
5:41simple conclusion is that each has its
5:43pros and cons, so it is hard to judge
5:45which one is definitively more suitable
5:48. In short, this copper pillar massive
5:51transfer is likely still under
5:52certification, and I personally think
5:55it is feasible. Currently, everyone
5:57uses electroplating, and electroplating
5:59definitely works. It's just that if you
6:03need electricity for a long time, even
6:05though this is cheaper, it takes a long
6:07time, so both have their own pros and
6:09cons. Is mass transfer definitely
6:13faster, then? We probably need actual
6:15tests to know for sure. Regardless, I
6:18think both technologies are viable, but
6:21they each have their own challenges to
6:23overcome, so in the end, it depends on
6:25whether they can solve them. I haven't
6:28heard any further news at the moment;
6:30they are likely still working on it. A
6:35netizen asked about AMD; as you all
6:37know, it was traditionally a chip
6:39company, and their software offerings
6:41have been fewer compared to NVIDIA. You
6:48all know NVIDIA doesn't just have chips
6:51; they have switches, servers, systems,
6:53the CUDA software framework,
6:55higher-level applications and software,
6:58and even world models, but AMD does not
7:00. So, the fastest method is acquisition
7:05, which is why AMD just acquired World
7:08Labs. World Labs was founded by Fei-Fei
7:12Li, and they are also working on world
7:14models, hence the name World Labs. To
7:17be honest, since Cosmos is the world
7:19model launched by NVIDIA, and both are
7:21world models, what is the difference?
7:24There is an article here from
7:25TechOrange that I think is quite
7:27well-written; feel free to check it out
7:29if you're interested. I will post it
7:32under the video later. This part
7:36mentions the World Labs world model. It
7:41states here that there are three types
7:43of world models: the renderer, the
7:45simulator, and the third one, which is
7:47the planner. What we generally call a
7:51renderer has the main task of
7:53outputting observational results, which
7:55is generating images or videos that
7:57humans can understand. Its benchmark is
8:00the realism of the simulation. Fei-Fei
8:04Li uses text-to-video models as an
8:06example, showing that these systems can
8:09generate lifelike imagery based on
8:11prompts and can even generate
8:12interactive scenes in real-time. So,
8:17Google's Gemini and World Labs 'RTFM
8:19belong to this category; the biggest
8:22advantage of a renderer is quality—
8:24the visual experience. A simulator, on
8:30the other hand, outputs states,
8:32allowing both humans and computer
8:34programs to calculate and interact with
8:36them. The model needs to know not only
8:39what an object looks like, but also
8:41understand its geometric structure,
8:43material properties, and laws of motion
8:45. Therefore, it is more than just an
8:48image or a frame. So, fields like
8:50architectural gaming developers,
8:52autonomous driving, and robotics
8:54research require models that can
8:56reflect the physical behaviors of the
8:58world. So, WLabs launched this mble,
9:01which falls into this category, while
9:04NVIDIA is centered around Omniverse and
9:06its Cosmos world model. This is a type
9:10of parallel simulation stack, so it
9:12basically counts as a simulator. The
9:18third type is called a planner, which
9:20takes observations and goals to decide
9:22what the model should do next. For
9:27instance, if a robot sees a cup on a
9:29table, how to reach out and grab it, or
9:31how an autonomous vehicle should plan
9:34its route amid traffic—this is what
9:36we call planning. There is a comparison
9:41provided by Jin here that you can refer
9:44to, but to put it simply, NVIDIA’s
9:46Cosmos focuses more on training.
9:51Because it uses massive amounts of data
9:53to train and build this world model. As
9:56for WLabs, what they released is more
9:59about content generation and spatial
10:01understanding. 3D scene construction.
10:05So, conceptually, there are some
10:07differences, but the goal is the same:
10:09to build a world model that allows
10:11machines to understand the world,
10:13rather than just seeing it as an image.
10:15That’s the gist. That’s actually a
10:17good question; let me see if I can find
10:19some good information to explain that
10:20further. I mentioned the process last
10:22time, that it's for Fan-Out Panel-Level
10:24Packaging (FOPLP), so it doesn't use
10:26through-hole vias. That is correct.
10:28I’m not sure if I misspoke last week
10:30and they are correcting me—sometimes
10:32when I’m livestreaming, what I say
10:34doesn't match what I’m thinking.
10:36Regardless, they are correct; FO-PLP
10:39doesn't need through-hole vias, and the
10:42process is indeed for FO-PLP. However,
10:44TGV requires through-hole vias, so
10:47typically for ABF substrates—that is,
10:50wiring substrates and glass interposers
10:53—you need TGV, or Through-Glass Vias.
10:56So, it’s written here that
10:58through-glass vias are used in ABF
11:00substrates for glass-core substrates.
11:03That’s also correct, and indeed the
11:05case. Currently, there are two
11:07companies capable of mass production;
11:09one seems to be a Japanese company, T.
11:11And they mentioned that in Taiwan,
11:13Nanya PCB has already produced it and
11:16is in mass production, according to
11:18them. As for Kinsus and Unimicron, they
11:20are still working on it; I know those
11:22two are still striving for it. But
11:24I’m not entirely sure about Nanya PCB
11:25and the Japanese company having
11:27achieved mass production; I haven't
11:28specifically researched that. It is
11:30possible. Next, what manufacturers in
11:35Taiwan currently help with liquid
11:37cooling reliability testing for AI
11:39chips? What is liquid cooling
11:43reliability testing? As you all know,
11:45current AI chips use liquid cooling—
11:47which basically involves cold plates
11:49and chilled water, or cooling liquids.
11:52With these high-power electronic
11:54components generating high heat, they
11:56undergo constant thermal exchange.
11:59Under conditions of prolonged
12:01high-temperature, high-load, and
12:03cooling cycles, we must verify their
12:05reliability and lifespan—essentially,
12:07whether this chip will fail due to such
12:10stress. As for his question about which
12:13manufacturers do this, I believe it
12:15would be the testing service providers.
12:18So, it should mainly be companies like
12:20MA-tek or Integrated Service Technology
12:22. I am certain companies like these
12:24perform this type of testing, and there
12:26may be some smaller firms as well.
12:29Moving on, a netizen asked about Team
12:31Group, which is a memory module company
12:33. As you all know, because of the
12:35memory shortage, prices have surged, so
12:38companies like this have recently made
12:40a killing. Let’s take a look at just
12:42how well this company is doing. Its
12:43stock price is over 1,000. It has been
12:46falling recently. If you look at the
12:49stock price, you will see it has been
12:51on a downward trend, effectively a
12:53gradual decline from a high of nearly
12:562,000 down to 1,000 now; it has
12:57actually dropped by half. Now, let’s
13:01take a look at its income statement.
13:03You will find it quite extraordinary;
13:05they earned 16 yuan in the first half
13:07of this year alone. Plus, the second
13:10half might be at least...With an EPS of
13:13300, yet its stock price is only 1,000.
13:15In other words, its P/E ratio is only
13:18three times. This isn't unique; if you
13:20look at the P/E ratio of Hynix, it’s
13:23actually about the same. Where does the
13:25problem lie? The problem lies in the
13:27fact that memory is still widely
13:29considered a cyclical stock. Although
13:33in the first half of this year, the
13:35market was hunting for stocks and some
13:37began claiming that memory is now an AI
13:39concept stock, not a cyclical one, but
13:41...It is very clear that its stock
13:43price still reflects a cyclical stock.
13:47In other words, most investors did not
13:49buy into that narrative. Most investors
13:52haven't really categorized memory as an
13:54AI concept, which is why the P/E ratio
13:56is as absurdly low as three or four
13:59times. So, your question is, can you
14:01invest in it? I personally believe that
14:04such a low P/E ratio theoretically
14:07holds value; in other words, its value
14:09is apparent. The key still lies in the
14:13recent US midterm elections. It looks
14:17like Trump is having a tough time; I
14:19think all the media has seen this.
14:22After the election ends, what Trump’s
14:24reaction will be is actually still hard
14:26to judge. That is why you’ve noticed
14:29the stock market has been struggling to
14:30move upward recently. In other words,
14:33everyone still wants to hedge their
14:35risks first; that’s about it. However
14:38, as you know, these stocks have fallen
14:40quite a bit from their highs. Among
14:44them, there are likely companies that
14:46have been oversold, meaning their true
14:49value is starting to emerge. After all,
14:55memory is a very profitable business,
14:57but if the market considers it a
14:59cyclical stock, you need to be more
15:01conservative. In other words, don't
15:06expect too much of a rebound. Unless,
15:09of course, the demand for artificial
15:11intelligence can be sustained. But can
15:14it actually be sustained? Something
15:18happened recently: I mentioned in my
15:20short video that bonds from some U.S.
15:22cloud service providers have fallen
15:24below their issue price or par value.
15:29That is actually not good news, as it
15:31signals that market investors are
15:33starting to doubt whether these cloud
15:34providers should be borrowing money to
15:36invest in data centers. Just imagine,
15:40if the promotion of artificial
15:42intelligence doesn't meet expectations
15:44over the next year or two, it's very
15:46likely that this demand will slow down.
15:49And once demand slows, you know that a
15:51decline in investment is likely because
15:53the capital market won't have as much
15:55money for you to spend recklessly. As
15:58soon as news of slowing investment
16:00emerges, I believe stocks will
16:02definitely react. I think this is key,
16:06so judging AI demand over the next two
16:08years will be crucial. I’ve mentioned
16:11the principles for judging this many
16:13times: artificial intelligence still
16:15has many usage limitations. Can these
16:18limitations be overcome? That is the
16:20key. Next, a viewer asked about Jinhua
16:23Technology's TBF. TBF is essentially an
16:26ABF alternative; it’s a material used
16:28for IC substrates. Using this material
16:31could potentially save steps and
16:33improve yields. Right, but the problem
16:35is that substrate manufacturers aren't
16:37necessarily willing to test it. You
16:40know, I talk about Taiwan because it's
16:42too easy to buy ABF materials there,
16:44and there are no sanctions. They can
16:47buy it anywhere, so these companies
16:49have no desire to test anyone else's
16:51materials. In this regard, I think
16:53Taiwanese manufacturers are quite poor.
16:55You shouldn't be like this; you must
16:57always remember, you shouldn't have
16:59just one supplier, you shouldn't let
17:01one supplier choke you. You must
17:04allocate a production line to test
17:06other options. The issue is that
17:09because Taiwan hasn't been blockaded,
17:11it’s natural for them to just focus
17:13on making money without thinking too
17:14much about it. So if the client isn't
17:17willing to test, successfully
17:19introducing a new product will be very
17:21arduous and difficult. Now, let’s
17:25talk about Toyoda; they make sliding
17:27modules, electric cylinders, and other
17:29mechanical devices like electric
17:31grippers. And linear motor modules—
17:34these are the kinds of mechanical
17:36components used in robotics. Therefore,
17:40this company’s products certainly
17:42have a specific market, though they are
17:44under some pressure recently because
17:46manufacturers in mainland China have
17:48been rising rapidly. However, since
17:53some clients require "red-free"(
17:55non-Chinese) supply chains, I
17:57personally believe this company has a
17:59high barrier to entry and, of course, a
18:02solid market. So it is a good company,
18:05but I suggest you judge whether the P/E
18:07ratio is reasonable. I won't look at
18:09that; P/E ratios aren't my specialty,
18:11so you’ll have to judge for yourself.
18:13Moving on to Hermes-Epitek, a
18:14subsidiary of the Hermes Group. Frankly
18:16speaking, this is also a good company.
18:18As you all know, companies under the
18:20Hermes Group are basically good
18:22companies; they are very solid and work
18:24extremely hard. This company makes ion
18:26implanters. As you know, for
18:28semiconductor equipment, especially
18:30advanced or manufacturing equipment, I
18:33am very supportive of Taiwan producing
18:35its own. And you know the Hermes Group
18:39has historically worked on many
18:41high-end pieces of equipment. So, this
18:45ion implanter is something Taiwan has
18:47yet to break through in. It is
18:50commendable that this company has
18:51stepped up to build high-current ion
18:53implanters, and I believe they will
18:55certainly succeed. As you know, ion
18:58implanters are high-barrier equipment.
19:01In a semiconductor fab, the highest
19:04barrier is obviously lithography
19:06equipment. But ion implanters are
19:09likely the second or third most
19:10difficult; in other words, the level of
19:13difficulty is high. Hermes-Epitek is
19:15making such equipment, and I believe
19:17they will succeed. They might not reach
19:20advanced nodes initially, but they
19:22certainly have a chance at mature nodes
19:24. If you are looking to invest, I still
19:27believe price is the key. The price
19:29must be reasonable, because I don't
19:31think they are profitable yet. In a
19:33situation where they aren't profitable
19:35yet, assessing a reasonable price is
19:37tricky. However, I believe their
19:39products will be usable, at least in
19:41mature processes first, and then
19:43gradually move toward advanced ones. I
19:46think that’s fine, mainly because the
19:48Hermes Group is a company that focuses
19:50on diligent management. About "Yingji,"
19:54I don't need to discuss that; I don't
19:56know what that is or what it means. It
20:02does system integration for AMD and
20:04Intel servers, as well as PCB storage
20:07servers. So, this is a server company.
20:16I feel the key is that server company
20:18stocks seem to have corrected recently,
20:20though this one is relatively strong.
20:25Let's look at the income statement to
20:27get a rough idea of its current
20:29profitability. See, it was 8.66 for the
20:33first two quarters, so this year it'll
20:36probably do 15 to 20. 15 to 20, with a
20:40stock price of 600, so that's about 30
20:44times. That's about right; a P/E ratio
20:46of 30 is reasonable. You can compare it
20:51with Wistron or Quanta; these are all
20:53server system integration companies, so
20:55see if the valuation is reasonable. Is
21:00this full-color solar power a future
21:02trend? Oh, you mean the solar panels.
21:05Making them in different colors. That
21:08must be what you mean; standard solar
21:10panels are usually dark blue or black.
21:13They can look ugly when used on
21:15building exteriors and don't blend well
21:17. If they can be made in full color,
21:20it's more creative and can blend with
21:22buildings, so I think it's feasible. I
21:27wouldn't call it a trend, as it will
21:29likely be more expensive, but it is
21:31creative and certainly doable. I'll
21:37find some time to research what methods
21:39they use to make it look colorful
21:41visually. Next, a viewer asked about
21:44PTFE, which the chairman of EMC said
21:47was not feasible. I mentioned this
21:49before; PTFE is Teflon, and as you know
21:52, most are currently made with
21:54fiberglass. Fiberglass and Teflon have
21:57different properties. The main issue is
22:03that Teflon is a slippery material, and
22:05you know these substrates are stacked
22:07layer by layer, so if the intermediate
22:09material is slippery, it's not easy to
22:11laminate or bond them together. I
22:17believe there are still difficulties to
22:19overcome. However, because Teflon has
22:23its advantages—I forget the specifics
22:25, but I think it has lower dielectric
22:27constant and dielectric loss—it has
22:30other merits. Its cost or manufacturing
22:34difficulty might be its downsides. So,
22:37what I've heard is that substrates
22:39usually have 20 to 30 layers. I heard
22:43that only a few critical layers need to
22:45use Teflon, so they would switch to it
22:47there. So, for materials like PTFE
22:51Teflon, I think it will be used in some
22:54layers. The problem with PTFE Teflon
22:57lies in the P/E ratio of the
22:59manufacturing companies; look at
23:01Tai-Hong's stock price at 266, it’s
23:04actually been hyped for quite a long
23:06time. If you look at the trend,
23:09you’ll see it has corrected recently,
23:12but it had previously risen to over 300
23:14. It's a bit of an exaggeration, but
23:18look, its EPS for the first half of
23:20this year was only 1, so the full year
23:22will probably total around 3 or 4. At 3
23:26or 4 dollars with a stock price of 300,
23:29the P/E ratio is 100 times. Whether
23:32this is reasonable, that is the
23:34question. So I personally think PTFE,
23:38this Teflon material, should be viable;
23:40some technical difficulties will be
23:43gradually overcome. But the stock price
23:47is already overhyped, so you have to be
23:49very careful. The next issue is Csun,
23:54which makes semiconductor equipment,
23:56including ALD systems; these labs are
23:59mainly for joint R&D with customers.
24:02Can this be used to lock in customer
24:04production lines? Yes, exactly. Its
24:06purpose should be that too, though I
24:08don't know which customer, maybe a
24:10foundry. May I ask about the
24:12competitiveness of Csun? I personally
24:15support any company in Taiwan that
24:17manufactures equipment. Csun's main
24:24equipment includes PVD physical vapor
24:26deposition tools, ALD atomic layer
24:29deposition tools, and equipment for
24:31advanced packaging. The profit status
24:38of these companies is generally okay,
24:40so if you're looking for a job, I think
24:42it's a decent company; you could give
24:44it a try. If it's for investment, I'll
24:48say it again: you have to look at
24:49whether the P/E ratio is reasonable.
24:52But in short, it isn't necessarily easy
24:54for Taiwan to make equipment, but if
24:57they work hard, they will achieve
24:58results. Next, a viewer asks why
25:01everyone is chasing FAs? Because the
25:04gross margin is too high, let me
25:05explain. One FA, I recall, seems to
25:08have 32 optical fibers. Such a set is
25:12said to sell for 100 USD now; seeing
25:14that, you know they are really gouging
25:17customers. Previously, because demand
25:20for FAs wasn't high, there weren't many
25:22manufacturers making them. Also, old
25:24optical cables didn't have as many as
25:2632 strands. Those were for optical
25:28communications; now they are for data
25:30centers, so the number of fibers has
25:32surged, and demand has exploded. So
25:35manufacturers are definitely taking
25:37advantage, setting prices very high.
25:43But think about it, this FA, as I said,
25:45is just glass, quartz, or silicon
25:48wafers with V-grooves. Let me show you
25:54this diagram; it uses a substrate with
25:56V-grooves to hold 32 optical fibers,
25:59then pulls them out. At 100 USD apiece,
26:03tell me, wouldn't you want to start a
26:05company to make them? So I’ll bet you
26:08all that a ton of companies are jumping
26:10into this now. Just look at this
26:12year’s semiconductor expo, a bunch of
26:15people are making FOPLP. The gross
26:17margin is very high now, I know, but
26:19I’ll bet you that by 2030, this FOPLP
26:21will definitely become a red ocean
26:23market. Let’s make a bet; today is
26:252026. There are three years until 2030,
26:27and since it’s already the end of the
26:29year, that’s '27, '28, and '29. This
26:31thing will definitely become a red
26:32ocean market, and in the end, it will
26:34be a cutthroat bloodbath. We’ll wait
26:36and see; if this kind of product can
26:38sell for 100 USD each, which company
26:41wouldn’t want to get in on it?
26:43Wouldn't every company want to do it?
26:45Under these conditions, if everyone
26:47does it together, just wait and see.
26:49But as for doing it now, those who
26:50start first make the profit. In short,
26:52whoever can mass-produce first will
26:54cash in first, and those who come later
26:56will likely just end up like that. As
26:57for Corning’s glass substrate, it’s
27:00similar, but you don’t need to worry,
27:02because the FOPLP TSMC is using now is
27:04not a glass substrate. So personally, I
27:06think everyone has a chance; you
27:07can’t say that just because Corning
27:09is doing it, others can’t—it’s
27:11not like that. At the very least, what
27:14TSMC is using should be FOPLP. It’s
27:16not a glass substrate. So FOPLP should
27:19have opportunities, it’s just that
27:21the barrier to entry is limited. Since
27:23it’s so profitable, once everyone
27:25jumps in, the price will inevitably
27:27drop after a while, and it will become
27:30a red ocean market—perhaps even
27:32before 2030. Next, a netizen asked if
27:34it's difficult for Taiwan’s memory
27:36industry to catch up with South Korea.
27:38It is, because the gap is very wide. If
27:40you ask whether China has a better
27:41chance of catching up with South Korea
27:43than Taiwan, I think the chances are
27:44much higher for China. This is mainly
27:46because, in the past, Taiwan’s memory
27:48sector would make money for one year
27:50and lose for three, so they weren't
27:51very willing to invest. As a result,
27:53the memory business hasn’t made much
27:55profit, so their lack of investment has
27:57led to a technological lag. Once
27:59technology falls behind for two or
28:01three years, it becomes very difficult
28:02to catch up again. So I don't think we
28:04need to dwell on this anymore; Taiwan
28:06is likely unable to keep up. Instead,
28:08you should look at when China will
28:10catch up to South Korea; that's more
28:12relevant. Now, let’s look at the
28:14difference between CHUANYUE’s liquid
28:16cooling technology and micro-channels.
28:19The channel cover plate involves
28:21creating liquid cooling channels on the
28:23packaging cover. Currently, liquid
28:26cooling happens outside the packaging
28:28shell, so the cooling and the packaging
28:30are separate. But with micro-channels,
28:33they are integrated with the packaging.
28:36So it is mentioned here that CHUANYUE
28:39is investing in advanced heat
28:40dissipation liquid cooling technology
28:43as a channel module cooling solution,
28:45launching modules that use micro-pillar
28:47array structures to build chip
28:49micro-channel-based heat dissipation. I
28:53might need to see its actual structure.
28:56Looking at the description, it feels
28:59similar to a micro-channel lid, as it
29:01says "micro-channel" here, and a
29:03micro-channel lid is a "micro-channel
29:05lid," so I think they are similar
29:07things. That's fine, I'll look into it
29:10further to see if there are any
29:12differences, and if there are, we'll
29:14introduce them later. Nanya Technology
29:16—well, let me explain. Nanya
29:18Technology is actually a company spun
29:20off by Nanya Technology Corporation. As
29:23you know, Nanya is a wafer manufacturer
29:26, and some memory needs to be
29:28customized, which requires customized
29:30chips. For customized memory, you need
29:33a design firm, so Nanya set up Nanya
29:36Technology, which primarily focuses on
29:38such customized memory. This market is
29:41quite niche, and this company is indeed
29:43profitable. So, I personally feel that
29:46there is room for this company to exist
29:49, it’s just that, from what I recall,
29:51its price-to-earnings ratio isn't low
29:53either. The price was driven up after
29:58it went public. Look at 620—its EPS
30:02for the first half of this year is 9
30:04yuan, so let's assume it’s 20 yuan
30:07for the full year. If it's 20 yuan and
30:11the stock is 600, that’s about 30
30:13times, so it’s still fairly
30:15reasonable, just a bit on the high side
30:17. Customized memory is quite niche; it
30:20belongs to a niche market. It's not
30:23easy to scale up, and that’s the
30:25current situation. Next, a netizen
30:28asked if Intel's large-scale supply of
30:31300mm, or 12-inch, indium phosphide (
30:33InP) silicon photonics technology
30:36mass-production wafers is any different
30:39from Lumentum's indium phosphide
30:41technology. I think they are largely
30:43similar, because Intel's silicon
30:45photonics technology has always been
30:47good, and Lumentum also has silicon
30:50photonics technology, but we are
30:51talking about indium phosphide, so it
30:54refers to light-emitting or
30:55light-receiving components. A 12-inch
30:57wafer—this is the first time I've
30:59heard of this. Well, I guess they’re
31:01impressive, since most are currently
31:02six inches. Jumping straight from six
31:04inches to 12 inches is truly impressive
31:06. I don't know. As you all know,
31:07theoretically, the larger the wafer,
31:09the lower the unit cost. If they can
31:11really do this, then it certainly
31:12counts as impressive. Will there be
31:14oversupply? I think it still depends on
31:16the demand for artificial intelligence,
31:18this AI demand. If we go by the big
31:20picture they've painted, then
31:22theoretically, we don't need to worry
31:24for now, but I am very concerned about
31:26two years from now. In short, I think
31:28this year and next year will be fine.
31:30Next is this Taiwan e-microscope
31:32company, which makes electron
31:34microscopes. As I’ve always said, any
31:36Taiwanese company involved in equipment
31:38manufacturing has a very solid and
31:40diligent team, so I believe their
31:42products should be up to a certain
31:44standard, no problem there. But this
31:47company hasn't gone public yet. If
31:50you're looking for a job, I think
31:52it’s fine; it's a good company. But
31:55if it's for investment, I still think
31:57price matters. Next, a netizen asked
31:59when the demand for AI infrastructure
32:02will stop rising. That’s something we
32:04need to observe, and I think 2026 will
32:06be the key. Next year should be fine.
32:09But the year after that, who knows? We
32:12need to keep a close eye on the
32:14situation. Will Cybercab appear in
32:16Taiwan? That’s certainly possible,
32:19but let me tell you, when it comes to
32:21Cybercab, netizens keep challenging me
32:23about whether vision-only is feasible.
32:26Let me show you. This Cybercab is still
32:29being tested in the U.S. right now. And
32:34Vision-only has some blind spots in
32:36practice. These blind spots can only be
32:39managed by their current method of
32:41remote human monitoring. Whether this
32:44approach can be scaled, I don’t know.
32:47We might have to wait and see. I don't
32:50really want to talk about this anymore;
32:52let's wait until it’s been on the
32:53road for a while. If nothing happens,
32:56then I was wrong; if an accident
32:57happens, we’ll just have to wait and
32:59see. Musk once proposed letting regular
33:02car owners buy this vehicle and join a
33:04robotaxi program. Sure, but let's wait
33:08and see how it handles Taiwan's roads
33:10with that vision-only system. What will
33:13happen if an accident occurs? Actually,
33:16it’s quite simple, everyone. You all
33:19know that Cybercab is essentially just
33:21FSD. The day you see Tesla state in the
33:26owner's manual for its FSD-equipped
33:29electric vehicles that "This FSD is
33:31fully autonomous, so if an accident
33:33happens, Tesla will take responsibility
33:35," The moment they dare to write that,
33:37I will immediately apologize and admit
33:40I was wrong. I bet you they won’t
33:43dare to write it. Because they know
33:45very clearly how safe that thing
33:47actually is. They are perfectly aware
33:49of where the risks lie. Next, a netizen
33:53mentioned HVDC, power concept stocks.
33:57Currently, what's lacking is external
34:00power generation—what does "external
34:02power generation" mean? Does it mean a
34:04power shortage? That’s possible.
34:06I’m not sure what that sentence means
34:08, but a power shortage is possible;
34:10electricity is what’s missing. As for
34:12HVDC, you probably spelled it wrong.
34:14HVDC stands for High-Voltage Direct
34:16Current. Why high voltage? Because our
34:19current electricity is AC, and
34:21converting to DC leads to a lot of
34:23waste due to efficiency issues. So,
34:26sending high-voltage DC directly into
34:28data centers to be converted to low
34:31voltage will be much more efficient.
34:33When you mention US CDMOs, you are
34:35referring to biotechnology contract
34:37manufacturing, right? Contract
34:38Development and Manufacturing
34:40Organization. Are you talking about
34:42biotechnology? Biotechnology. If it is
34:47biotechnology contract manufacturing, I
34:50think using the US might be a cost
34:52issue. Given the cost factor, I think
34:55it is quite difficult. Marketech is an
34:58AI company, specifically in optical
35:00inspection, or automated optical
35:02inspection. To be honest, there are
35:06more and more AI competitors now, but I
35:08recall Marketech being the best, so
35:10they have a certain competitiveness,
35:12though competitors will continue to
35:14increase. What does Lianyi Optics do to
35:18be named by NVIDIA? Optical glass
35:26blanks, okay, as you know, optical
35:28glass is hot-pressed, so they make
35:30blanks, likely for optical glass lenses
35:32; it's the raw material before
35:34processing, where molten glass is
35:36initially molded into blanks or rough
35:39parts, then handed over to Largan or
35:41Genius for grinding or finishing. It's
35:50nothing special for companies to be
35:52named by them; I think this is very
35:54normal. Their current strategy is just
35:56for NVIDIA to name companies, and the
35:58stocks of those named companies will
36:00skyrocket. But in return, you cannot
36:02make money off NVIDIA; that's their
36:04current strategy. So being named by
36:07them is just a way to hype the stock.
36:13There are also metal recycling
36:15companies; I remember they recycle
36:17tungsten, and because prices surged,
36:19their profits surged, but after all,
36:21it's a bit cyclical. The current
36:27situation is certainly very good, but
36:29it's hard to say that metals will
36:30always be rising in price. It’s a
36:34good company; I have read their
36:36investor conference materials, and they
36:38are quite serious about their
36:39operations. Metal recycling is also a
36:43trend, but their recent stock price
36:45surge is mostly because the metals they
36:48recycle—I remember it's tungsten—
36:50have increased in price. Companies with
36:52increased profits due to rising raw
36:54material prices still need to be
36:56watched carefully, as they might not
36:58always keep rising; they recycle
37:00tungsten and cobalt, that’s right. Xu
37:03Bo recently mentioned a seminar on
37:05using AI to address mesh generation
37:07issues in structural design software;
37:10is this for Ansys? As for their
37:13solution, is it an offline AI-assisted
37:15version or a paid AI? It should be paid
37:18, and there will still be people using
37:21it because companies like ANSYS are
37:23primarily B2B. For professional design
37:27or simulation, you still can't do
37:29without tools like ICs, Synopsys, or
37:31Cadence; they're expensive, but they
37:34offer high precision. So I think the
37:37professional work should be left to the
37:39professionals. What do you think about
37:42the collaboration between Phison and
37:43Unicomp? Since Unicomp does
37:49testing-related work, and they also
37:51seem to handle probes, Pogo pins, and
37:54burn-in—using high temperature and
37:56voltage for testing—are you referring
37:58to the connector maker, Phison,
38:00collaborating with them? That’s
38:06definitely possible, because a lot of
38:08your testing would likely involve
38:10Phison. Micron thinks 4 F² is
38:13unreasonable, but it's not about being
38:15reasonable; it's about whether or not
38:17it can be done. They think the
38:20technology is near-mythical, meaning
38:22it's difficult. But the question is,
38:24just because you find it difficult
38:26doesn't mean others do. You could fail
38:28to do it and still claim you did, right
38:30? Look at TSMC's 3nm process; if you
38:33actually measure it, no part of it is
38:35truly 3nm. That’s why China is so
38:38bold in their claims. Chinese memory
38:40manufacturers are already saying they
38:42will mass-produce 4 F², so you might
38:44as well just follow suit—worry about
38:46the actual scaling later. That’s too
38:49honest; don't be that honest. There’s
38:51definitely difficulty involved, but
38:53Innotron is quite strong, and their
38:55technology is solid. However, if I
38:57recall correctly, Innotron’s P/E
38:59ratio is also very high. I won't look
39:02it up, you can check that for
39:03yourselves. Innotron is a very unique
39:06company with strong competitiveness,
39:08and their chairman is highly creative
39:10and skilled in technical implementation
39:12. So it is a good company, but as for
39:16what a reasonable P/E ratio is,
39:18that’s not my area of expertise;
39:21you’ll have to evaluate that
39:22yourselves. Designing bacteriophages
39:26using AI. Biotech is like that these
39:29days. Some people are even using AI to
39:34design viruses; frankly, that’s a bit
39:36reckless and could lead to trouble.
39:41Bacteriophages are essentially
39:43microorganisms that attack bacteria.
39:46They usually carry DNA, and unless
39:49there's a specific requirement, we
39:51don't consider on-premises solutions—
39:53yes, exactly. You can just use tokens
39:56directly. Are you talking about
39:58professional-grade design or general
40:00user needs? Well, that’s all for
40:02today’s livestream. If you have any
40:04tech questions, feel free to post them
40:06in the comments below, and we can
40:08discuss them. Thank you, everyone, and
40:10goodnight.

About this transcript

This page contains the full transcript of 【曲博科技教室 EP325】英特爾、晶呈、創新服務、AMD、力成、Toppan、南電、景碩、欣興、宜鼎、晶化、東佑達、漢辰、營邦、天虹、康寧、崇越、補丁、台灣電鏡、牧德、聯一、聯友金屬、雍智、美光 by 曲博科技教室 Dr. J Class, generated from the public captions YouTube serves with the video. The transcript has 5,931 words across 949 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

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

Free YouTube transcript tool

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