0:00.000–0:01.420
enWelcome to Taiwan Era.
0:01.580–0:03.540
enHello everyone, I'm Chong Cheng Hua.
0:03.840–0:08.180
enWe saw the Philadelphia Semiconductor Index undergo a rather sharp correction from late
0:08.180–0:09.140
enJune to mid-July.
0:09.260–0:13.280
enBut don't worry, everyone says this isn't an AI bubble, just a cleanup of funds and
0:13.280–0:13.840
endeleveraging.
0:14.060–0:15.140
enHas the deleveraging ended?
0:15.340–0:16.360
enWhat about the fundamentals?
0:16.660–0:18.620
enHave there been any new changes in the fundamentals?
0:18.980–0:22.080
enDo AI models from China pose a threat to USI developers?
0:22.720–0:24.620
enIs the competition intensifying?
0:24.820–0:29.500
enToday, I am very pleased to invite Junong Capital CEO, CEO Shen Jun, to analyze this
0:29.500–0:29.880
enfor us.
0:30.000–0:33.120
enHello,CEO,BrotherChong,Hello to all the viewers
0:33.120–0:37.840
enRight,looking back,from late June to mid July,the SOX index corrected by about 13%,
0:37.840–0:40.400
enwhile some funds flowed into the so-called magnificent seven,
0:40.400–0:42.880
enwith their average stock price rising by 4.5%.
0:42.880–0:45.840
enTesla and Alphabet released their earnings reports this week,
0:45.840–0:48.840
enso let's carefully examine this so-called i-logic.
0:48.840–0:50.200
enWe are seeing a shift.
0:50.200–0:53.600
enIn the past,we focused on capital expenditure and scale.
0:53.600–0:56.480
enThe more you invested,the more the stock rose.
0:56.480–0:59.600
enWill we be more stringent in examining these AI earnings now?
1:00.000–1:03.720
zh在此,AI有所造成的異常在我們的實現
1:03.720–1:07.680
zhGoogle的 CEO曾經說
1:07.680–1:11.400
zhAI曾經說1%的人持有99%的貨幣
1:11.400–1:16.320
zh這不是我們的貨幣
1:16.320–1:19.040
zh它決定了一個非常強烈的貨幣
1:19.040–1:21.920
zh也會引起大增加和貨幣
1:21.920–1:23.360
zh貨幣貨幣貨幣貨幣貨幣貨幣貨幣
1:23.360–1:26.360
zh我們看到了如此重新的貨幣
1:26.360–1:27.640
zh最近的貨幣貨幣貨幣
1:27.640–1:32.040
zh為什麼之前的強烈賞出了,甚至於是一種股市場,
1:32.040–1:35.000
enwhile the M7 stocks that weren't rising began to catch up?
1:36.040–1:39.400
enIt's simply because the gap between the winners and losers was too large.
1:40.760–1:46.360
enYou can see it yourself. By mid-year, as of June 31st, the stocks had surged by nearly 80%.
1:48.120–1:52.920
enAmong all US stock indices, except for small cap stocks which might have risen by 20%.
1:52.920–1:57.960
enOthers like the NASDAQ were actually up less than 20%, but the stocks alone rose by 80%.
1:57.960–2:00.960
zh所以,市場感受到太多重
2:00.960–2:01.960
zh更多
2:01.960–2:03.960
zh有個例子,有個例子
2:03.960–2:05.960
zh有個例子,是有個例子
2:05.960–2:08.960
zh我們可能會有四年來的屁股
2:08.960–2:10.960
zh但他們有四年來的屁股
2:10.960–2:12.960
zh他們說有150,000
2:12.960–2:15.960
zh你能夠借到1.5萬
2:15.960–2:17.960
zh是那些稍稅
2:17.960–2:19.960
zh這情況有所在的
2:19.960–2:21.960
zh有關影響的影響
2:21.960–2:23.960
zh從我們的影片
2:23.960–2:24.960
zh從一位的主意
2:24.960–2:29.880
enthe key focus is the extreme distortion of resources which will not be resolved in the future
2:29.880–2:31.260
enlet me share a number with you
2:31.260–2:36.600
enwe all know that csps are expected to spend 700 billion dollars or 800 billion dollars this year
2:36.600–2:37.400
enthat's the book
2:37.400–2:41.980
enlet me mention that this year the semiconductor sector including tsmc
2:41.980–2:46.380
enase and memory combined will earn nearly 700 billion dollars
2:46.380–2:49.980
enoh so that's it it's moved from here to there
2:49.980–2:54.360
enand this is exactly what causes extreme volatility in stock prices
2:54.360–2:57.860
enRight,so coming back to the question,is there an AI bubble?
2:57.860–3:00.300
enWe just mentioned whether AI would lead to a bubble.
3:00.300–3:04.360
enI want to share that in the AI space,some are working hard for their money,like the
3:04.360–3:06.460
enhumanoid robots sector.
3:06.460–3:11.160
enOr even those we consider winners,like OpenAI,they are working hard for their money.
3:11.160–3:15.520
enThey haven't made a profit yet,they might be losing $10 billion or even $20 billion
3:15.520–3:16.520
enevery quarter.
3:16.520–3:20.120
enThey are taking huge risks,and the humanoid robotics companies they pull along are also
3:20.120–3:23.620
entaking huge risks,yet their profit margins remain relatively fixed.
3:23.620–3:24.120
zh嗯
3:24.120–3:28.120
enconversely for those selling shovels wow their profits keep piling up
3:28.120–3:34.220
eneven the director of sk hinnick said excuse me i feel this isn't quite normal we're making too much money could it be
3:34.220–3:37.820
enthey even said the prices are a bit high and they should probably be a bit healthier
3:37.820–3:42.720
enif you read the whole thing he sounds a bit guilty wondering how they could be earning so much
3:42.720–3:45.520
enhe didn't actually see a downturn coming yet
3:45.520–3:53.520
enso from this perspective we really need to be mentally prepared for the investment market knowing that it will never be balanced
3:53.520–3:58.220
enWhen extreme stock price volatility occurs, sometimes people get cautious, causing the
3:58.220–4:01.840
enprice to rebalance, but in the long run, it will still lead to polarization.
4:02.980–4:08.200
enYes, but when we look back at the memory industry, we see they are continuously expanding capacity.
4:08.660–4:12.960
enWe know that the demand for memory remains very high, but it will certainly take time
4:12.960–4:17.060
enfor the supply side to catch up. I just want to remind everyone, because many ghost stories
4:17.060–4:21.140
enare surfacing right now, as memory manufacturers everywhere are expanding plants, whether it's
4:21.140–4:25.300
zh在香港,臺灣,或在中國,包括YMTC和CXMT,
4:25.300–4:26.300
zh都在增加了
4:26.300–4:27.300
zh人們可能會想想
4:27.300–4:28.860
zh如果您撥開工程
4:28.860–4:31.220
zh決定,它們將在增加了什麼我們的名義
4:31.220–4:32.260
zh是一種連續的連續
4:32.260–4:33.700
zh它們現在變成了一種連續
4:33.700–4:35.380
zh在最後一種連續的連續
4:35.380–4:36.660
zh至於2030,
4:36.660–4:38.620
enthere will still be a very significant shortage.
4:38.860–4:40.220
enThis can help strengthen confidence,
4:40.560–4:42.340
enconfirming that if the memory sector corrects,
4:42.560–4:45.880
enit is perfectly fine in terms of both stock price and industry fundamentals.
4:46.520–4:50.640
enLike recently, Sammy, who used to love to talk down the market and so on,
4:51.040–4:54.840
ensuddenly came out and said he believes memory will remain in shortage all the way through.
4:55.460–4:57.340
enHis perspective, or for example,
4:57.720–5:02.500
enlike even when you look at TSMC's earnings call mentioning envious gross margins and so on,
5:02.880–5:04.120
enactually tells you one thing,
5:04.120–5:09.440
enwhy some people will tell you decisively that memory demand will be excellent and will continue to be in short supply
5:09.440–5:12.380
enyou might think if the price gets too high i just won't buy
5:12.380–5:17.560
enthe conventional investment logic for business has always been if it goes up i won't use it or i'll just stop
5:17.560–5:19.300
enso how could demand possibly be limitless
5:19.300–5:23.900
enwhat i want to share with everyone that they are ignoring or why they see what others are ignoring
5:23.900–5:29.460
enis that we now see tsmc's capex at 60 billion this year maybe 80 billion next year
5:29.460–5:33.320
enor nvidia's gpus and amd's gpus doing great
5:33.320–5:37.040
zh我們當然忘記 that for every GPU or CPU produced
5:37.040–5:38.620
zh它 must be paired with a unit
5:38.620–5:40.060
ena fixed unit of memory
5:40.060–5:42.580
zh它是不 sold separately
5:42.580–5:44.340
enmemory is not sold alone
5:44.340–5:46.640
zh所以我們的 AI future becomes clear
5:46.640–5:48.760
enwith so many GPUs coming out
5:48.760–5:49.980
zh如果我們不存在的話
5:49.980–5:50.000
zh我們不存在的話
5:50.000–5:51.380
zh我們不存在的話
5:51.380–5:54.400
zh所以我們看它從 TSMC's perspective
5:54.400–5:56.120
zh從 the semiconductor perspective
5:56.120–5:58.320
zh我們的 demand for memory 4 years from now
5:58.320–6:00.480
enshould be at least 4 times what it is today
6:00.480–6:00.880
zh嗯
6:00.880–6:01.880
zh在最少
6:01.880–6:02.880
enYes
6:02.880–6:05.880
enAnd this premise is assuming my GPU volume hasn't even increased
6:05.880–6:08.880
enIf I produce 100 units of GPUs this year and 100 units 4 years from now
6:08.880–6:11.880
enAlthough there is migration and they might be more powerful in 4 years
6:11.880–6:15.880
enIf it's still 100 units, the memory I need to pair with them will still be 4 times as much
6:15.880–6:21.880
enUnless I downgrade, for instance, fine, I'm driving a Bentley, but I don't want air conditioning
6:21.880–6:24.880
enI don't want ABS brakes, then your performance will definitely suffer
6:24.880–6:28.880
enSo later on you will keep hearing people talk about downgrading specs
6:28.880–6:30.880
zh你可能會認為 memory是不好的,
6:30.880–6:33.380
zh但不,因為有沒有的,
6:33.380–6:38.880
zh你會認為大家是在中國的增長和產品的增長。
6:38.880–6:42.880
zh如果看來,產品是在增長的增長並且在2028,
6:42.880–6:44.880
zh或者我們叫它,四年以後,
6:44.880–6:45.880
zh加上了它的增長。
6:45.880–6:46.880
zh嗯嗯。
6:46.880–6:48.880
zh我們就說,
6:48.880–6:50.880
zh在這四次,
6:50.880–6:51.880
zh所以,
6:51.880–6:52.880
zh在這一年,
6:52.880–6:54.880
zh四次。
6:54.880–6:57.880
zh如果我們的資料設計是在四年以後,
6:57.880–7:03.740
enit is now, wouldn't TSMC be finished? Wouldn't Nvidia be finished? That's impossible.
7:05.020–7:10.340
enWe're all optimistic that GPUs will keep trending upward, so investors who are worried about memory,
7:10.880–7:15.280
enif you are truly worried that memory will crash, then you should be like Michael Burry and not buy
7:15.280–7:22.040
enany semiconductors at all. You can't say you're bullish on TSMC but bearish on memory. Let's get
7:22.040–7:26.700
enthis logic straight. You can't have it both ways. If it looks like memory is going to collapse in
7:26.700–7:29.700
zh未來,這些世界都會崩潰。
7:29.700–7:32.700
zh當然,我們不可能會否有機會,
7:32.700–7:35.700
zh我們不可能會否會發展開四年
7:35.700–7:37.700
zh和世界的世界返回到大陸,
7:37.700–7:39.700
zh讓 AI後來的確定。
7:39.700–7:43.700
zh我們不可能會否會發生。
7:43.700–7:46.700
zh但你必須理解釋這個 logic,
7:46.700–7:48.700
zh如果 AI還會升幅度的數量,
7:48.700–7:50.700
zh如果數量的數量會升幅,
7:50.700–7:52.700
zh就會增加了。
7:53.700–7:55.700
zh這就是這些人所說的,
7:55.700–7:58.460
zh不只是給你一個路線或是領導的
7:59.212–8:01.652
zh而我們剛才提到的
8:01.652–8:02.212
zh《病》
8:02.212–8:04.892
zh我們還沒有說了
8:04.892–8:05.772
zh我們還沒有說了
8:05.772–8:07.332
zh《軟體》
8:07.332–8:08.532
zh《軟體》
8:08.532–8:08.692
zh《軟體》
8:08.692–8:09.172
zh《軟體》
8:09.172–8:09.892
zh《軟體》
8:09.892–8:11.012
zh《軟體》
8:11.012–8:11.892
zh《軟體》
8:11.892–8:12.932
zh《軟體》
8:12.932–8:14.132
zh《軟體》
8:14.212–8:22.652
enAs we've mentioned before, TSMC estimates that future applications for AI will require 8 times the current computing power of these, brains.
8:24.212–8:28.612
enYes, the memory demand is simply unimaginable, it's terrifying.
8:29.272–8:33.932
enTo be honest, in our reports from visiting Samsung and Hinex, even they can't calculate.
8:34.692–8:38.072
enExactly how much is needed, they don't know how to negotiate with clients.
8:38.072–8:40.672
zh如果他們太高,他們會負敗的交易,
8:40.672–8:42.112
zh但在實際上,他們會負敗的貨幫助,
8:42.112–8:44.112
zh他們會負敗的貨幫助,
8:44.112–8:46.392
zh他們會負敗的貨幫助。
8:46.392–8:48.192
zh當然,他們會負敗的貨幫助,
8:48.192–8:50.992
zh他們會負敗的貨幫助,
8:50.992–8:52.752
zh但根據上一個討論的評論,
8:52.752–8:54.632
zh他們會負敗的貨幫助。
8:54.632–8:56.352
zh另一個重點的評論,
8:56.352–8:59.312
zh包括了一種同時的評論,
8:59.312–9:00.592
zh我們看到最近,
9:00.592–9:02.192
zh人們覺得這些關係的關係
9:02.192–9:04.312
zh是被高傳的高傳領的高傳領的
9:04.312–9:09.672
enand their ratings are increasing so it's a situation where if gpus increase racks increase
9:09.672–9:15.272
enand servers increase including from ai from communications passive components and pcbs
9:15.272–9:20.392
enall demand is moving upward if we talk about computing power as we just mentioned cpus and
9:20.392–9:26.632
engpus are paired with memory they are a set in the future like the hvdc power supply for servers
9:26.632–9:32.472
enit is paired with passive components passive components are used for power so if you believe
9:32.472–9:34.472
zh如果您相信,您的需要提供力量的力量,
9:34.472–9:36.472
zh您的支援力量是否有力量的力量,
9:36.472–9:38.472
zh您的支援力量是否有力量的力量。
9:38.472–9:41.472
zh如果您相信,您的支援力量是否有力量的力量。
9:41.472–9:44.472
zh如果您相信,您的支援力量是否有力量的力量。
9:44.472–9:46.472
zh您相信,您的支援力量是否有力量的力量。
9:46.472–9:49.472
zh您相信,您的支援力量是否有力量的力量。
9:49.472–9:52.472
zh所以,您的支援力量是否有力量的力量。
9:52.472–9:54.472
zh所以,您的支援力量是否有力量的力量。
9:54.472–9:56.472
zh您的支援力量是否有力量的力量。
9:56.472–9:58.472
zh您的支援力量是否有力量的力量。
9:58.472–10:00.472
zh所以,您的支援力量是否有力量的力量。
10:00.472–10:30.472
zh優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優�
10:30.472–10:35.392
enbut shortages are normal because after all this massive demand, the supply simply cannot keep up.
10:35.652–10:38.632
enNow, as the first wave pulls back and finishes consolidating,
10:39.032–10:43.292
enonce the actual shortages begin, the stock price will follow the logic of supply shortages.
10:43.632–10:45.392
enSo, what should its valuation be?
10:45.572–10:49.252
enActually, in due time, the capital market will give it a reasonable valuation.
10:49.252–10:53.612
enOkay, so looking at it from an economic perspective or from the fundamentals of the industry,
10:53.972–10:55.512
enmemory is currently in short supply.
10:55.672–10:59.512
enThe supply side is severely lacking, and this will likely continue for a long time.
10:59.512–11:02.592
zh但是,有一個不斷的經濟發生在這個業務中的
11:02.592–11:05.652
zh美國的美國人說是中國人的美國人
11:05.652–11:08.472
zh像 Samsung是傳續於美國人的美國人
11:08.472–11:09.872
zh而是太多貨
11:09.872–11:12.312
zh的供貨力是完全在美國人的手段
11:12.312–11:14.952
zh而且是中國人的美國人現在太高
11:14.952–11:17.312
zh所以政府是購買的貨帶
11:17.312–11:18.652
zh你認為這個人的這個?
11:18.992–11:21.472
zh我認為這個人認為他們認為貨帶的貨帶
11:21.472–11:24.092
zh但在實際上,我們的認為我們的課計
11:24.092–11:27.272
zh難以為他們的貨帶貨帶的貨帶
11:27.272–11:31.312
enWith US companies, they can ask for equity, like they did with Intel or others.
11:31.592–11:35.532
enBut for the US government to take equity in a Korean company, that's a bit difficult.
11:35.872–11:38.912
enWe haven't seen them take an equity stake in TSMC, right?
11:39.072–11:43.472
enHowever, they could follow the TSMC model, meaning if you don't come and set up a plant here.
11:43.912–11:47.252
enIf you don't come here to build what they call a US-based memory facility,
11:47.672–11:49.312
enbecause everyone knows there's a shortage,
11:49.752–11:53.412
enit's only natural for them to want these companies to expand their capacity here.
11:53.412–11:58.632
enhonestly if these companies don't face such overwhelming demand pushing them
11:58.632–12:04.552
enthey really aren't willing to expand because for memory manufacturers they know one thing
12:04.552–12:09.732
eni don't need to expand my plants when the time comes and you have a shortage i'll just double
12:09.732–12:14.512
enthe price and i'll make the same amount of money or even more and i can do that for a long time
12:14.512–12:19.792
enof course they want a healthy market so their customers don't go under but for them over
12:19.792–12:49.792
zh肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料�
12:49.792–12:50.792
enKIMI-K3
12:50.792–12:55.632
enThis was actually one of the factors that led to a market correction last Friday.
12:56.372–13:01.612
enPeople thought, okay, last month the White House decided that Anthropix Claude 3.5 was
13:01.612–13:04.592
entoo powerful, so it restricted access for non-Americans.
13:05.412–13:11.132
enAnd then, shortly after, China released Quen 2.5, which scored nearly the same as 3.5,
13:11.132–13:13.732
enbut it's actually open source and launched immediately.
13:14.692–13:19.732
enHow big of a shock is this to the current USI model landscape, and has market competition
13:19.732–13:25.232
zh其實, everyone is just looking at the final evaluation results.
13:25.632–13:28.272
enIt has 7 metrics, and Quen won 6 of them.
13:29.232–13:32.592
enAnd one look tells us that Chinese people are really good at taking tests.
13:32.892–13:36.812
enWhat we are best at is test-taking, whether it's the SATs or our joint entrance exams,
13:37.132–13:39.072
enwe eventually just master the question bank.
13:40.132–13:44.352
enEssentially, they looked at those 7 performance metrics and tuned their model to be perfectly
13:44.352–13:47.572
ensuited for those, whether for AI agents or image generation.
13:47.572–13:50.972
enYou can actually see how their models work
13:50.972–13:55.052
enCurrently, whether it's Anthropic or OpenAI, they are all very powerful
13:55.052–13:59.692
enSo powerful, in fact, that sometimes they become powerful enough to be stupid
13:59.692–14:04.852
enFor example, if you use Gemini in a coffee shop
14:04.852–14:07.592
enIt starts off strong with great proposals and estimates
14:07.592–14:11.092
enBut after a month, it gets slow because every time it makes a decision
14:11.092–14:13.372
enIt reconsiders every previous decision
14:13.372–14:17.712
enyou can imagine how that piles up
14:17.712–14:20.012
enby the second month
14:20.012–14:21.272
enit starts making errors
14:21.272–14:23.972
enlike ordering thousands of rubber gloves for a cafe
14:23.972–14:27.932
enit starts ordering carrots
14:27.932–14:29.112
enwhen there isn't even a kitchen
14:29.112–14:30.092
enso in the end
14:30.092–14:31.172
enit just crashes
14:31.172–14:34.372
enthe problem with AI now
14:34.372–14:35.872
enis that while it might be powerful
14:35.872–14:37.572
eneven single task powerful
14:37.572–14:40.512
eni have to tune it to handle these 7 specific metrics
14:40.512–14:42.792
enso you have to train it specifically for that
14:43.372–14:46.492
zh看起來,不像 DeepSeq
14:46.492–14:52.852
enDeepSeq is like a needle, a small model that had a huge impact when it came out.
14:54.412–14:55.432
enWhat does that mean?
14:56.012–15:00.152
enIt means I don't need to be that powerful, and I don't need that much computing power.
15:00.952–15:05.912
enBut with Quen, I have massive computing power, yet I perform better than you, and the way
15:05.912–15:09.912
enI perform better is by tuning it specifically to handle the performance you're looking for.
15:10.432–15:14.152
enComing back to the point, even though it performs well, people don't need to worry
15:14.152–15:16.792
zh太多了因為它還需要大量的空氣力
15:16.792–15:18.952
zh然後它將會很快的空氣力
15:18.952–15:21.112
zh有一個東西人都沒有認為
15:21.112–15:22.072
zh它的價值
15:22.072–15:25.192
zh它的價值是一樣的價值
15:25.192–15:28.632
zh它是0.95%和1.07%
15:28.632–15:31.592
zh但安thropic是超費的
15:31.592–15:32.392
en$3
15:32.392–15:33.912
zh但看起來
15:33.912–15:35.752
zh是 OpenAI really doing worse?
15:35.752–15:37.992
zh它是非常好
15:37.992–15:38.792
zh的開放
15:38.792–15:40.592
zh為
15:40.592–15:42.552
zh因為人們說它是費的
15:42.552–15:44.312
enare enterprises willing to pay?
15:44.752–15:47.512
enthis is what we need to go back and discuss regarding AI
15:48.232–15:49.912
enit's not just about performing well
15:50.792–15:52.872
enor generating pretty pictures or answers
15:54.072–15:54.872
enif that were all
15:54.872–15:57.752
enit would just be for us non-paying or low-paying users
15:58.819–15:59.819
zh而是我們的不費用
15:59.819–16:01.819
zh原因是因為它做得到的事情
16:01.819–16:03.819
zh它是做得到的
16:03.819–16:05.819
zh它是需要的需要
16:05.819–16:07.819
zh像是需要的
16:07.819–16:09.819
zh如果用它是用它
16:09.819–16:11.819
zh或是關於你的研究
16:11.819–16:13.819
zh你不能夠做得到的
16:13.819–16:21.819
zh你會看到,甚至是很貴的,而且它的效果不太好,為什麼它是很高的?
16:23.819–16:32.819
zh你可能說,這些人們不懂得用 Open Source,但我認為不覺得他們的獨自貴。
16:32.819–17:02.819
zh他們應該能夠實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實
17:02.819–17:04.259
zh是在那裡的事,
17:04.259–17:05.579
zh之後K was released,
17:05.579–17:08.119
zh它的機器 immediately hit the limit.
17:08.119–17:09.919
zh它 seems unable to allow new users
17:09.919–17:11.759
zh去上網上用了新的新版本來使用了新版本來的
17:11.759–17:16.319
zh因為它的背景已經不足了。
17:16.319–17:18.719
zh所以,它不是在這裡的底色
17:18.719–17:21.059
zh它可以用 AI models on lighter hardware.
17:21.059–17:24.219
zh它們的額外 demand是其實大量的。
17:24.219–17:27.919
zh所以,是這其實好事?
17:27.919–17:29.919
zh是這其實好事?
17:29.919–17:31.119
zh有更多的影響,
17:31.119–17:32.859
zh是更多的更多的更多
17:32.859–17:34.259
zh為什麼是更多的更多?
17:35.619–17:37.159
zh因為我們剛剛說
17:37.159–17:38.559
zh如果是個體驗的問題
17:38.559–17:40.159
zh而你的模特兒是有點困難
17:40.159–17:42.359
zh而你使用它在不足的地方
17:42.359–17:43.419
zh但實際上的使用者
17:43.419–17:45.459
zh不會讓我能夠享受到更多的效果
17:45.459–17:46.419
zh而我不會讓我能夠
17:46.419–17:47.719
zh稍為的更多的效果
17:49.319–17:51.259
zh如果我知道這些資料
17:51.259–17:52.259
zh在正確的地方
17:52.259–17:54.259
zh要有更多的執行
17:54.259–17:55.579
zh是當然好
17:55.579–17:56.379
zh那些人會有盡量的
17:56.379–17:58.159
zh所以你能看到
17:58.159–18:28.159
zh這次的壞壞壞在售貨物店裡的壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞�
18:28.159–18:32.159
enthe real concern might be when we go back to examine OpenAI
18:32.159–18:34.159
enit becomes a situation where, sure,
18:34.159–18:38.159
enCLOD's benchmark scores might not be significantly higher than GPT-4's
18:38.159–18:42.159
enbut they are able to make much more money from enterprise clients
18:42.159–18:44.159
enExactly
18:44.159–18:48.159
enAs we mentioned, OpenAI's costs and efficiency are similar to CLOD's
18:48.159–18:50.159
enso CLOD 3 might not hurt Anthropic,
18:50.159–18:54.159
enbut it could definitely hit OpenAI
18:54.159–19:00.159
zh有甚至今的說話,如果有一個AI瀏規, OpenAI會是說的一種類似的一種類似的。
19:01.159–19:05.159
zh所以,我們看來的責任 OpenAI在中國的責任?
19:06.159–19:10.159
zh如果它不影響 Anthropik,它會直接射擊 OpenAI,
19:10.159–19:13.159
zh尤其是跟 Oracle和 Cloudflare的背景。
19:14.159–19:16.159
zh你覺得這個動作?
19:18.159–19:20.159
zh其實,我們不相信AI會流動,
19:20.159–19:25.159
zh但最弱的原因是在《AI》和《Oracle》中的主要是《OpenAI》。
19:30.859–19:32.159
zh你能看到了解的目標
19:32.159–19:35.359
zh像它的借貨倫、購乎率、購乎率、
19:35.359–19:37.959
zh或是在購買購買貨。
19:37.959–19:39.959
zh這些是否有任何的影響。
19:45.859–19:49.819
zh只要是,《OpenAI》需要找到一種商品的商品作為
19:50.159–19:53.919
zh所以,看著他們在做的
19:53.919–19:55.399
zh他們要進行進行的
19:55.399–19:57.399
zh他們想要做一個話題
19:57.399–19:58.399
zh他們要做一個話題
19:58.399–19:59.399
zh他們要做一些其他東西
19:59.399–20:02.399
zh他們要做到販賣物的物品
20:02.399–20:04.399
zh我們會把這個做法
20:04.399–20:05.399
zh很清楚
20:05.399–20:08.399
zh他們真的不能夠保持
20:08.399–20:10.399
zhunless美國給他們
20:10.399–20:13.399
zh他們的公司不可能會有問題
20:15.399–20:17.399
zh所以我們會看 closely
20:17.399–20:18.279
zh因為他們的商品
20:18.279–20:19.779
zh如果這個認識的認識是沒有發展,
20:19.779–20:20.279
zh在這個中處的認識是沒有發展,
20:20.279–20:21.779
zh在這個中處的認識是沒有發展,
20:21.779–20:23.779
zh在這個中處的認識是沒有發展,
20:23.779–20:25.779
zh在這個中處的認識是沒有發展,
20:25.779–20:27.779
zh在這個中處的認識是沒有發展,
20:27.779–20:29.779
zh在這個中處的認識是沒有發展,
20:29.779–20:31.779
zh在這個中處的認識是沒有發展,
20:31.779–20:33.779
zh在這個中處的認識是沒有發展,
20:33.779–20:35.779
zh在這個中處的認識是沒有發展,
20:35.779–20:37.779
zh在這個中處的認識是沒有發展,
20:37.779–20:39.779
zh如果他們開始發展,
20:39.779–20:40.779
zh如果他們發展,
20:40.779–20:41.779
zh他們發展,他們發展,
20:41.779–20:42.779
zh他們發展,他們發展,
20:42.779–21:12.779
zh如果能否購買貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨
21:12.779–21:16.619
zh在美國的美國,我們覺得我們的貨幣是否會有效果,
21:16.619–21:19.979
zh我們的其他人,不一定會有效果。
21:19.979–21:25.019
zh其實,我們覺得我們的貨幣未來的未來,
21:25.019–21:28.139
zh現在,當然,我們的貨幣未來的貨幣未來,
21:28.139–21:29.179
zh那是確實。
21:29.179–21:31.979
zh現在,如果你貨幣,你貨幣。
21:31.979–21:35.819
zh而且,你更多貨幣,像更多貨幣,
21:35.819–21:39.019
zh像TSMC,他們不能算貨幣,
21:39.019–21:40.719
enEventually, they will make money.
21:41.879–21:47.479
enYes, it's hard-earned money, and they are truly carrying human hope forward, like OpenAI and Anthropic.
21:48.419–21:50.939
enThey have to figure out how to make money themselves.
21:51.879–21:54.219
enWhy do I say the real winners haven't appeared?
21:54.979–21:57.079
enBecause it's about the application side.
21:58.099–22:02.339
enBecause once this AI, brain, is smart enough, robots will emerge.
22:03.119–22:06.339
enAI pets will emerge, AI autonomous vehicles will emerge.
22:06.339–22:09.899
enyou might be able to use what others have already spent so much money on
22:09.899–22:12.219
endeveloped and made ready
22:12.219–22:14.419
enand i just need to pay a reasonable price
22:14.419–22:16.499
eni can integrate it into my system
22:16.499–22:18.219
enit's a bit like apple now
22:18.219–22:20.339
enapple chose gemini to integrate
22:20.339–22:22.099
enif it's smart enough
22:22.099–22:24.339
enit might just make their iphone a huge hit
22:24.339–22:28.319
enthese things could happen on the application side in the future
22:28.319–22:31.539
enin the application sector
22:31.539–22:34.419
enthose who can seize the initiative and hit that sweet spot
22:34.419–22:36.139
enwill likely be the first wave winners
22:36.139–22:39.979
zh但便是企業的,在一旦會發生,
22:39.979–22:41.979
zh在一旦會發生的一流程。
22:43.179–22:45.859
zh這就是為了我們已經看著當中的AI模式
22:45.859–22:46.979
zh是完全適合的。
22:50.659–22:52.019
zh像是在手機上,
22:52.019–22:53.699
zh它們已經在8-10年來
22:53.699–22:55.219
zh讓它們發生了一些
22:55.219–22:55.979
zh手機,
22:55.979–22:56.699
zh手機,
22:56.699–22:56.859
zh手機,
22:56.859–22:58.219
zh最後手機。
23:01.419–23:01.899
zh現在,
23:01.899–23:03.979
zh它們都已經在三年。
23:03.979–23:12.879
zh一切都要做到,每公司都要推出了,就像每公司都要推出了,像每公司都要推出了,像 Inventech的手法上,在那天
23:12.879–23:16.759
enBenQ smartphones and the rest all died off
23:16.759–23:17.379
enWhy?
23:17.839–23:19.259
enPeople won't buy from everyone
23:19.259–23:20.899
enIt will be the same in the future
23:20.899–23:23.179
enThere might be a hundred brands of robots
23:23.179–23:25.139
enThat's when a real bubble forms
23:25.139–23:26.679
enYou'll need all sorts of parts
23:26.679–23:28.259
enAnd there will be massive shortages
23:28.259–23:29.219
enBut in the end
23:29.219–23:31.739
enThere will be a pile of robots that no one buys
23:31.739–23:34.519
enThat's exactly where the real bubble lies
23:34.519–23:38.119
enSo, before those applications truly emerge
23:38.119–23:40.839
enThat's why many say we are in the early stages of AI
23:40.839–23:42.379
enLike the top of the third inning
23:42.879–23:45.199
zh那是他們所說的,OK?
23:45.199–23:46.599
zh很多 pessimists might think
23:46.599–23:52.139
enyou're lying to me again, trying to trick me every day, saying it's only the third inning when stock prices have already risen like this.
23:52.259–23:55.519
enThey think that because they see the massive scale of the applications coming later.
23:55.719–23:56.659
enYes, exactly.
23:56.919–24:00.439
enSo we see that at this point in time, it is indeed still the top of the third inning.
24:00.599–24:03.179
enFor those of us selling shovels,there's no need to worry.
24:03.359–24:07.839
enOnce the leverage is removed and funds are rebalanced,buying the dip is a very good entry point.
24:07.839–24:12.499
enThank you very much to the CEO for joining us for this interview,and thank you to our viewers for watching.
24:12.699–24:13.819
enTaiwan's Grand Era
24:16.599–24:32.888
zh請訂閱 訂閱 轉發 打賞支援明鏡與點點欄目
0:00.000–0:01.420
Welcome to Taiwan Era.
0:01.580–0:03.540
Hello everyone, I'm Chong Cheng Hua.
0:03.840–0:08.180
We saw the Philadelphia Semiconductor Index undergo a rather sharp correction from late
0:08.180–0:09.140
June to mid-July.
0:09.260–0:13.280
But don't worry, everyone says this isn't an AI bubble, just a cleanup of funds and
0:13.280–0:13.840
deleveraging.
0:14.060–0:15.140
Has the deleveraging ended?
0:15.340–0:16.360
What about the fundamentals?
0:16.660–0:18.620
Have there been any new changes in the fundamentals?
0:18.980–0:22.080
Do AI models from China pose a threat to USI developers?
0:22.720–0:24.620
Is the competition intensifying?
0:24.820–0:29.500
Today, I am very pleased to invite Junong Capital CEO, CEO Shen Jun, to analyze this
0:29.500–0:29.880
for us.
0:30.000–0:33.120
Hello,CEO,BrotherChong,Hello to all the viewers
0:33.120–0:37.840
Right,looking back,from late June to mid July,the SOX index corrected by about 13%,
0:37.840–0:40.400
while some funds flowed into the so-called magnificent seven,
0:40.400–0:42.880
with their average stock price rising by 4.5%.
0:42.880–0:45.840
Tesla and Alphabet released their earnings reports this week,
0:45.840–0:48.840
so let's carefully examine this so-called i-logic.
0:48.840–0:50.200
We are seeing a shift.
0:50.200–0:53.600
In the past,we focused on capital expenditure and scale.
0:53.600–0:56.480
The more you invested,the more the stock rose.
0:56.480–0:59.600
Will we be more stringent in examining these AI earnings now?
1:00.000–1:03.720
在此,AI有所造成的異常在我們的實現
1:03.720–1:07.680
Google的 CEO曾經說
1:07.680–1:11.400
AI曾經說1%的人持有99%的貨幣
1:11.400–1:16.320
這不是我們的貨幣
1:16.320–1:19.040
它決定了一個非常強烈的貨幣
1:19.040–1:21.920
也會引起大增加和貨幣
1:21.920–1:23.360
貨幣貨幣貨幣貨幣貨幣貨幣貨幣
1:23.360–1:26.360
我們看到了如此重新的貨幣
1:26.360–1:27.640
最近的貨幣貨幣貨幣
1:27.640–1:32.040
為什麼之前的強烈賞出了,甚至於是一種股市場,
1:32.040–1:35.000
while the M7 stocks that weren't rising began to catch up?
1:36.040–1:39.400
It's simply because the gap between the winners and losers was too large.
1:40.760–1:46.360
You can see it yourself. By mid-year, as of June 31st, the stocks had surged by nearly 80%.
1:48.120–1:52.920
Among all US stock indices, except for small cap stocks which might have risen by 20%.
1:52.920–1:57.960
Others like the NASDAQ were actually up less than 20%, but the stocks alone rose by 80%.
1:57.960–2:00.960
所以,市場感受到太多重
2:00.960–2:01.960
更多
2:01.960–2:03.960
有個例子,有個例子
2:03.960–2:05.960
有個例子,是有個例子
2:05.960–2:08.960
我們可能會有四年來的屁股
2:08.960–2:10.960
但他們有四年來的屁股
2:10.960–2:12.960
他們說有150,000
2:12.960–2:15.960
你能夠借到1.5萬
2:15.960–2:17.960
是那些稍稅
2:17.960–2:19.960
這情況有所在的
2:19.960–2:21.960
有關影響的影響
2:21.960–2:23.960
從我們的影片
2:23.960–2:24.960
從一位的主意
2:24.960–2:29.880
the key focus is the extreme distortion of resources which will not be resolved in the future
2:29.880–2:31.260
let me share a number with you
2:31.260–2:36.600
we all know that csps are expected to spend 700 billion dollars or 800 billion dollars this year
2:36.600–2:37.400
that's the book
2:37.400–2:41.980
let me mention that this year the semiconductor sector including tsmc
2:41.980–2:46.380
ase and memory combined will earn nearly 700 billion dollars
2:46.380–2:49.980
oh so that's it it's moved from here to there
2:49.980–2:54.360
and this is exactly what causes extreme volatility in stock prices
2:54.360–2:57.860
Right,so coming back to the question,is there an AI bubble?
2:57.860–3:00.300
We just mentioned whether AI would lead to a bubble.
3:00.300–3:04.360
I want to share that in the AI space,some are working hard for their money,like the
3:04.360–3:06.460
humanoid robots sector.
3:06.460–3:11.160
Or even those we consider winners,like OpenAI,they are working hard for their money.
3:11.160–3:15.520
They haven't made a profit yet,they might be losing $10 billion or even $20 billion
3:15.520–3:16.520
every quarter.
3:16.520–3:20.120
They are taking huge risks,and the humanoid robotics companies they pull along are also
3:20.120–3:23.620
taking huge risks,yet their profit margins remain relatively fixed.
3:23.620–3:24.120
嗯
3:24.120–3:28.120
conversely for those selling shovels wow their profits keep piling up
3:28.120–3:34.220
even the director of sk hinnick said excuse me i feel this isn't quite normal we're making too much money could it be
3:34.220–3:37.820
they even said the prices are a bit high and they should probably be a bit healthier
3:37.820–3:42.720
if you read the whole thing he sounds a bit guilty wondering how they could be earning so much
3:42.720–3:45.520
he didn't actually see a downturn coming yet
3:45.520–3:53.520
so from this perspective we really need to be mentally prepared for the investment market knowing that it will never be balanced
3:53.520–3:58.220
When extreme stock price volatility occurs, sometimes people get cautious, causing the
3:58.220–4:01.840
price to rebalance, but in the long run, it will still lead to polarization.
4:02.980–4:08.200
Yes, but when we look back at the memory industry, we see they are continuously expanding capacity.
4:08.660–4:12.960
We know that the demand for memory remains very high, but it will certainly take time
4:12.960–4:17.060
for the supply side to catch up. I just want to remind everyone, because many ghost stories
4:17.060–4:21.140
are surfacing right now, as memory manufacturers everywhere are expanding plants, whether it's
4:21.140–4:25.300
在香港,臺灣,或在中國,包括YMTC和CXMT,
4:25.300–4:26.300
都在增加了
4:26.300–4:27.300
人們可能會想想
4:27.300–4:28.860
如果您撥開工程
4:28.860–4:31.220
決定,它們將在增加了什麼我們的名義
4:31.220–4:32.260
是一種連續的連續
4:32.260–4:33.700
它們現在變成了一種連續
4:33.700–4:35.380
在最後一種連續的連續
4:35.380–4:36.660
至於2030,
4:36.660–4:38.620
there will still be a very significant shortage.
4:38.860–4:40.220
This can help strengthen confidence,
4:40.560–4:42.340
confirming that if the memory sector corrects,
4:42.560–4:45.880
it is perfectly fine in terms of both stock price and industry fundamentals.
4:46.520–4:50.640
Like recently, Sammy, who used to love to talk down the market and so on,
4:51.040–4:54.840
suddenly came out and said he believes memory will remain in shortage all the way through.
4:55.460–4:57.340
His perspective, or for example,
4:57.720–5:02.500
like even when you look at TSMC's earnings call mentioning envious gross margins and so on,
5:02.880–5:04.120
actually tells you one thing,
5:04.120–5:09.440
why some people will tell you decisively that memory demand will be excellent and will continue to be in short supply
5:09.440–5:12.380
you might think if the price gets too high i just won't buy
5:12.380–5:17.560
the conventional investment logic for business has always been if it goes up i won't use it or i'll just stop
5:17.560–5:19.300
so how could demand possibly be limitless
5:19.300–5:23.900
what i want to share with everyone that they are ignoring or why they see what others are ignoring
5:23.900–5:29.460
is that we now see tsmc's capex at 60 billion this year maybe 80 billion next year
5:29.460–5:33.320
or nvidia's gpus and amd's gpus doing great
5:33.320–5:37.040
我們當然忘記 that for every GPU or CPU produced
5:37.040–5:38.620
它 must be paired with a unit
5:38.620–5:40.060
a fixed unit of memory
5:40.060–5:42.580
它是不 sold separately
5:42.580–5:44.340
memory is not sold alone
5:44.340–5:46.640
所以我們的 AI future becomes clear
5:46.640–5:48.760
with so many GPUs coming out
5:48.760–5:49.980
如果我們不存在的話
5:49.980–5:50.000
我們不存在的話
5:50.000–5:51.380
我們不存在的話
5:51.380–5:54.400
所以我們看它從 TSMC's perspective
5:54.400–5:56.120
從 the semiconductor perspective
5:56.120–5:58.320
我們的 demand for memory 4 years from now
5:58.320–6:00.480
should be at least 4 times what it is today
6:00.480–6:00.880
嗯
6:00.880–6:01.880
在最少
6:01.880–6:02.880
Yes
6:02.880–6:05.880
And this premise is assuming my GPU volume hasn't even increased
6:05.880–6:08.880
If I produce 100 units of GPUs this year and 100 units 4 years from now
6:08.880–6:11.880
Although there is migration and they might be more powerful in 4 years
6:11.880–6:15.880
If it's still 100 units, the memory I need to pair with them will still be 4 times as much
6:15.880–6:21.880
Unless I downgrade, for instance, fine, I'm driving a Bentley, but I don't want air conditioning
6:21.880–6:24.880
I don't want ABS brakes, then your performance will definitely suffer
6:24.880–6:28.880
So later on you will keep hearing people talk about downgrading specs
6:28.880–6:30.880
你可能會認為 memory是不好的,
6:30.880–6:33.380
但不,因為有沒有的,
6:33.380–6:38.880
你會認為大家是在中國的增長和產品的增長。
6:38.880–6:42.880
如果看來,產品是在增長的增長並且在2028,
6:42.880–6:44.880
或者我們叫它,四年以後,
6:44.880–6:45.880
加上了它的增長。
6:45.880–6:46.880
嗯嗯。
6:46.880–6:48.880
我們就說,
6:48.880–6:50.880
在這四次,
6:50.880–6:51.880
所以,
6:51.880–6:52.880
在這一年,
6:52.880–6:54.880
四次。
6:54.880–6:57.880
如果我們的資料設計是在四年以後,
6:57.880–7:03.740
it is now, wouldn't TSMC be finished? Wouldn't Nvidia be finished? That's impossible.
7:05.020–7:10.340
We're all optimistic that GPUs will keep trending upward, so investors who are worried about memory,
7:10.880–7:15.280
if you are truly worried that memory will crash, then you should be like Michael Burry and not buy
7:15.280–7:22.040
any semiconductors at all. You can't say you're bullish on TSMC but bearish on memory. Let's get
7:22.040–7:26.700
this logic straight. You can't have it both ways. If it looks like memory is going to collapse in
7:26.700–7:29.700
未來,這些世界都會崩潰。
7:29.700–7:32.700
當然,我們不可能會否有機會,
7:32.700–7:35.700
我們不可能會否會發展開四年
7:35.700–7:37.700
和世界的世界返回到大陸,
7:37.700–7:39.700
讓 AI後來的確定。
7:39.700–7:43.700
我們不可能會否會發生。
7:43.700–7:46.700
但你必須理解釋這個 logic,
7:46.700–7:48.700
如果 AI還會升幅度的數量,
7:48.700–7:50.700
如果數量的數量會升幅,
7:50.700–7:52.700
就會增加了。
7:53.700–7:55.700
這就是這些人所說的,
7:55.700–7:58.460
不只是給你一個路線或是領導的
7:59.212–8:01.652
而我們剛才提到的
8:01.652–8:02.212
《病》
8:02.212–8:04.892
我們還沒有說了
8:04.892–8:05.772
我們還沒有說了
8:05.772–8:07.332
《軟體》
8:07.332–8:08.532
《軟體》
8:08.532–8:08.692
《軟體》
8:08.692–8:09.172
《軟體》
8:09.172–8:09.892
《軟體》
8:09.892–8:11.012
《軟體》
8:11.012–8:11.892
《軟體》
8:11.892–8:12.932
《軟體》
8:12.932–8:14.132
《軟體》
8:14.212–8:22.652
As we've mentioned before, TSMC estimates that future applications for AI will require 8 times the current computing power of these, brains.
8:24.212–8:28.612
Yes, the memory demand is simply unimaginable, it's terrifying.
8:29.272–8:33.932
To be honest, in our reports from visiting Samsung and Hinex, even they can't calculate.
8:34.692–8:38.072
Exactly how much is needed, they don't know how to negotiate with clients.
8:38.072–8:40.672
如果他們太高,他們會負敗的交易,
8:40.672–8:42.112
但在實際上,他們會負敗的貨幫助,
8:42.112–8:44.112
他們會負敗的貨幫助,
8:44.112–8:46.392
他們會負敗的貨幫助。
8:46.392–8:48.192
當然,他們會負敗的貨幫助,
8:48.192–8:50.992
他們會負敗的貨幫助,
8:50.992–8:52.752
但根據上一個討論的評論,
8:52.752–8:54.632
他們會負敗的貨幫助。
8:54.632–8:56.352
另一個重點的評論,
8:56.352–8:59.312
包括了一種同時的評論,
8:59.312–9:00.592
我們看到最近,
9:00.592–9:02.192
人們覺得這些關係的關係
9:02.192–9:04.312
是被高傳的高傳領的高傳領的
9:04.312–9:09.672
and their ratings are increasing so it's a situation where if gpus increase racks increase
9:09.672–9:15.272
and servers increase including from ai from communications passive components and pcbs
9:15.272–9:20.392
all demand is moving upward if we talk about computing power as we just mentioned cpus and
9:20.392–9:26.632
gpus are paired with memory they are a set in the future like the hvdc power supply for servers
9:26.632–9:32.472
it is paired with passive components passive components are used for power so if you believe
9:32.472–9:34.472
如果您相信,您的需要提供力量的力量,
9:34.472–9:36.472
您的支援力量是否有力量的力量,
9:36.472–9:38.472
您的支援力量是否有力量的力量。
9:38.472–9:41.472
如果您相信,您的支援力量是否有力量的力量。
9:41.472–9:44.472
如果您相信,您的支援力量是否有力量的力量。
9:44.472–9:46.472
您相信,您的支援力量是否有力量的力量。
9:46.472–9:49.472
您相信,您的支援力量是否有力量的力量。
9:49.472–9:52.472
所以,您的支援力量是否有力量的力量。
9:52.472–9:54.472
所以,您的支援力量是否有力量的力量。
9:54.472–9:56.472
您的支援力量是否有力量的力量。
9:56.472–9:58.472
您的支援力量是否有力量的力量。
9:58.472–10:00.472
所以,您的支援力量是否有力量的力量。
10:00.472–10:30.472
優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優�
10:30.472–10:35.392
but shortages are normal because after all this massive demand, the supply simply cannot keep up.
10:35.652–10:38.632
Now, as the first wave pulls back and finishes consolidating,
10:39.032–10:43.292
once the actual shortages begin, the stock price will follow the logic of supply shortages.
10:43.632–10:45.392
So, what should its valuation be?
10:45.572–10:49.252
Actually, in due time, the capital market will give it a reasonable valuation.
10:49.252–10:53.612
Okay, so looking at it from an economic perspective or from the fundamentals of the industry,
10:53.972–10:55.512
memory is currently in short supply.
10:55.672–10:59.512
The supply side is severely lacking, and this will likely continue for a long time.
10:59.512–11:02.592
但是,有一個不斷的經濟發生在這個業務中的
11:02.592–11:05.652
美國的美國人說是中國人的美國人
11:05.652–11:08.472
像 Samsung是傳續於美國人的美國人
11:08.472–11:09.872
而是太多貨
11:09.872–11:12.312
的供貨力是完全在美國人的手段
11:12.312–11:14.952
而且是中國人的美國人現在太高
11:14.952–11:17.312
所以政府是購買的貨帶
11:17.312–11:18.652
你認為這個人的這個?
11:18.992–11:21.472
我認為這個人認為他們認為貨帶的貨帶
11:21.472–11:24.092
但在實際上,我們的認為我們的課計
11:24.092–11:27.272
難以為他們的貨帶貨帶的貨帶
11:27.272–11:31.312
With US companies, they can ask for equity, like they did with Intel or others.
11:31.592–11:35.532
But for the US government to take equity in a Korean company, that's a bit difficult.
11:35.872–11:38.912
We haven't seen them take an equity stake in TSMC, right?
11:39.072–11:43.472
However, they could follow the TSMC model, meaning if you don't come and set up a plant here.
11:43.912–11:47.252
If you don't come here to build what they call a US-based memory facility,
11:47.672–11:49.312
because everyone knows there's a shortage,
11:49.752–11:53.412
it's only natural for them to want these companies to expand their capacity here.
11:53.412–11:58.632
honestly if these companies don't face such overwhelming demand pushing them
11:58.632–12:04.552
they really aren't willing to expand because for memory manufacturers they know one thing
12:04.552–12:09.732
i don't need to expand my plants when the time comes and you have a shortage i'll just double
12:09.732–12:14.512
the price and i'll make the same amount of money or even more and i can do that for a long time
12:14.512–12:19.792
of course they want a healthy market so their customers don't go under but for them over
12:19.792–12:49.792
肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料�
12:49.792–12:50.792
KIMI-K3
12:50.792–12:55.632
This was actually one of the factors that led to a market correction last Friday.
12:56.372–13:01.612
People thought, okay, last month the White House decided that Anthropix Claude 3.5 was
13:01.612–13:04.592
too powerful, so it restricted access for non-Americans.
13:05.412–13:11.132
And then, shortly after, China released Quen 2.5, which scored nearly the same as 3.5,
13:11.132–13:13.732
but it's actually open source and launched immediately.
13:14.692–13:19.732
How big of a shock is this to the current USI model landscape, and has market competition
13:19.732–13:25.232
其實, everyone is just looking at the final evaluation results.
13:25.632–13:28.272
It has 7 metrics, and Quen won 6 of them.
13:29.232–13:32.592
And one look tells us that Chinese people are really good at taking tests.
13:32.892–13:36.812
What we are best at is test-taking, whether it's the SATs or our joint entrance exams,
13:37.132–13:39.072
we eventually just master the question bank.
13:40.132–13:44.352
Essentially, they looked at those 7 performance metrics and tuned their model to be perfectly
13:44.352–13:47.572
suited for those, whether for AI agents or image generation.
13:47.572–13:50.972
You can actually see how their models work
13:50.972–13:55.052
Currently, whether it's Anthropic or OpenAI, they are all very powerful
13:55.052–13:59.692
So powerful, in fact, that sometimes they become powerful enough to be stupid
13:59.692–14:04.852
For example, if you use Gemini in a coffee shop
14:04.852–14:07.592
It starts off strong with great proposals and estimates
14:07.592–14:11.092
But after a month, it gets slow because every time it makes a decision
14:11.092–14:13.372
It reconsiders every previous decision
14:13.372–14:17.712
you can imagine how that piles up
14:17.712–14:20.012
by the second month
14:20.012–14:21.272
it starts making errors
14:21.272–14:23.972
like ordering thousands of rubber gloves for a cafe
14:23.972–14:27.932
it starts ordering carrots
14:27.932–14:29.112
when there isn't even a kitchen
14:29.112–14:30.092
so in the end
14:30.092–14:31.172
it just crashes
14:31.172–14:34.372
the problem with AI now
14:34.372–14:35.872
is that while it might be powerful
14:35.872–14:37.572
even single task powerful
14:37.572–14:40.512
i have to tune it to handle these 7 specific metrics
14:40.512–14:42.792
so you have to train it specifically for that
14:43.372–14:46.492
看起來,不像 DeepSeq
14:46.492–14:52.852
DeepSeq is like a needle, a small model that had a huge impact when it came out.
14:54.412–14:55.432
What does that mean?
14:56.012–15:00.152
It means I don't need to be that powerful, and I don't need that much computing power.
15:00.952–15:05.912
But with Quen, I have massive computing power, yet I perform better than you, and the way
15:05.912–15:09.912
I perform better is by tuning it specifically to handle the performance you're looking for.
15:10.432–15:14.152
Coming back to the point, even though it performs well, people don't need to worry
15:14.152–15:16.792
太多了因為它還需要大量的空氣力
15:16.792–15:18.952
然後它將會很快的空氣力
15:18.952–15:21.112
有一個東西人都沒有認為
15:21.112–15:22.072
它的價值
15:22.072–15:25.192
它的價值是一樣的價值
15:25.192–15:28.632
它是0.95%和1.07%
15:28.632–15:31.592
但安thropic是超費的
15:31.592–15:32.392
$3
15:32.392–15:33.912
但看起來
15:33.912–15:35.752
是 OpenAI really doing worse?
15:35.752–15:37.992
它是非常好
15:37.992–15:38.792
的開放
15:38.792–15:40.592
為
15:40.592–15:42.552
因為人們說它是費的
15:42.552–15:44.312
are enterprises willing to pay?
15:44.752–15:47.512
this is what we need to go back and discuss regarding AI
15:48.232–15:49.912
it's not just about performing well
15:50.792–15:52.872
or generating pretty pictures or answers
15:54.072–15:54.872
if that were all
15:54.872–15:57.752
it would just be for us non-paying or low-paying users
15:58.819–15:59.819
而是我們的不費用
15:59.819–16:01.819
原因是因為它做得到的事情
16:01.819–16:03.819
它是做得到的
16:03.819–16:05.819
它是需要的需要
16:05.819–16:07.819
像是需要的
16:07.819–16:09.819
如果用它是用它
16:09.819–16:11.819
或是關於你的研究
16:11.819–16:13.819
你不能夠做得到的
16:13.819–16:21.819
你會看到,甚至是很貴的,而且它的效果不太好,為什麼它是很高的?
16:23.819–16:32.819
你可能說,這些人們不懂得用 Open Source,但我認為不覺得他們的獨自貴。
16:32.819–17:02.819
他們應該能夠實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實
17:02.819–17:04.259
是在那裡的事,
17:04.259–17:05.579
之後K was released,
17:05.579–17:08.119
它的機器 immediately hit the limit.
17:08.119–17:09.919
它 seems unable to allow new users
17:09.919–17:11.759
去上網上用了新的新版本來使用了新版本來的
17:11.759–17:16.319
因為它的背景已經不足了。
17:16.319–17:18.719
所以,它不是在這裡的底色
17:18.719–17:21.059
它可以用 AI models on lighter hardware.
17:21.059–17:24.219
它們的額外 demand是其實大量的。
17:24.219–17:27.919
所以,是這其實好事?
17:27.919–17:29.919
是這其實好事?
17:29.919–17:31.119
有更多的影響,
17:31.119–17:32.859
是更多的更多的更多
17:32.859–17:34.259
為什麼是更多的更多?
17:35.619–17:37.159
因為我們剛剛說
17:37.159–17:38.559
如果是個體驗的問題
17:38.559–17:40.159
而你的模特兒是有點困難
17:40.159–17:42.359
而你使用它在不足的地方
17:42.359–17:43.419
但實際上的使用者
17:43.419–17:45.459
不會讓我能夠享受到更多的效果
17:45.459–17:46.419
而我不會讓我能夠
17:46.419–17:47.719
稍為的更多的效果
17:49.319–17:51.259
如果我知道這些資料
17:51.259–17:52.259
在正確的地方
17:52.259–17:54.259
要有更多的執行
17:54.259–17:55.579
是當然好
17:55.579–17:56.379
那些人會有盡量的
17:56.379–17:58.159
所以你能看到
17:58.159–18:28.159
這次的壞壞壞在售貨物店裡的壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞�
18:28.159–18:32.159
the real concern might be when we go back to examine OpenAI
18:32.159–18:34.159
it becomes a situation where, sure,
18:34.159–18:38.159
CLOD's benchmark scores might not be significantly higher than GPT-4's
18:38.159–18:42.159
but they are able to make much more money from enterprise clients
18:42.159–18:44.159
Exactly
18:44.159–18:48.159
As we mentioned, OpenAI's costs and efficiency are similar to CLOD's
18:48.159–18:50.159
so CLOD 3 might not hurt Anthropic,
18:50.159–18:54.159
but it could definitely hit OpenAI
18:54.159–19:00.159
有甚至今的說話,如果有一個AI瀏規, OpenAI會是說的一種類似的一種類似的。
19:01.159–19:05.159
所以,我們看來的責任 OpenAI在中國的責任?
19:06.159–19:10.159
如果它不影響 Anthropik,它會直接射擊 OpenAI,
19:10.159–19:13.159
尤其是跟 Oracle和 Cloudflare的背景。
19:14.159–19:16.159
你覺得這個動作?
19:18.159–19:20.159
其實,我們不相信AI會流動,
19:20.159–19:25.159
但最弱的原因是在《AI》和《Oracle》中的主要是《OpenAI》。
19:30.859–19:32.159
你能看到了解的目標
19:32.159–19:35.359
像它的借貨倫、購乎率、購乎率、
19:35.359–19:37.959
或是在購買購買貨。
19:37.959–19:39.959
這些是否有任何的影響。
19:45.859–19:49.819
只要是,《OpenAI》需要找到一種商品的商品作為
19:50.159–19:53.919
所以,看著他們在做的
19:53.919–19:55.399
他們要進行進行的
19:55.399–19:57.399
他們想要做一個話題
19:57.399–19:58.399
他們要做一個話題
19:58.399–19:59.399
他們要做一些其他東西
19:59.399–20:02.399
他們要做到販賣物的物品
20:02.399–20:04.399
我們會把這個做法
20:04.399–20:05.399
很清楚
20:05.399–20:08.399
他們真的不能夠保持
20:08.399–20:10.399
unless美國給他們
20:10.399–20:13.399
他們的公司不可能會有問題
20:15.399–20:17.399
所以我們會看 closely
20:17.399–20:18.279
因為他們的商品
20:18.279–20:19.779
如果這個認識的認識是沒有發展,
20:19.779–20:20.279
在這個中處的認識是沒有發展,
20:20.279–20:21.779
在這個中處的認識是沒有發展,
20:21.779–20:23.779
在這個中處的認識是沒有發展,
20:23.779–20:25.779
在這個中處的認識是沒有發展,
20:25.779–20:27.779
在這個中處的認識是沒有發展,
20:27.779–20:29.779
在這個中處的認識是沒有發展,
20:29.779–20:31.779
在這個中處的認識是沒有發展,
20:31.779–20:33.779
在這個中處的認識是沒有發展,
20:33.779–20:35.779
在這個中處的認識是沒有發展,
20:35.779–20:37.779
在這個中處的認識是沒有發展,
20:37.779–20:39.779
如果他們開始發展,
20:39.779–20:40.779
如果他們發展,
20:40.779–20:41.779
他們發展,他們發展,
20:41.779–20:42.779
他們發展,他們發展,
20:42.779–21:12.779
如果能否購買貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨
21:12.779–21:16.619
在美國的美國,我們覺得我們的貨幣是否會有效果,
21:16.619–21:19.979
我們的其他人,不一定會有效果。
21:19.979–21:25.019
其實,我們覺得我們的貨幣未來的未來,
21:25.019–21:28.139
現在,當然,我們的貨幣未來的貨幣未來,
21:28.139–21:29.179
那是確實。
21:29.179–21:31.979
現在,如果你貨幣,你貨幣。
21:31.979–21:35.819
而且,你更多貨幣,像更多貨幣,
21:35.819–21:39.019
像TSMC,他們不能算貨幣,
21:39.019–21:40.719
Eventually, they will make money.
21:41.879–21:47.479
Yes, it's hard-earned money, and they are truly carrying human hope forward, like OpenAI and Anthropic.
21:48.419–21:50.939
They have to figure out how to make money themselves.
21:51.879–21:54.219
Why do I say the real winners haven't appeared?
21:54.979–21:57.079
Because it's about the application side.
21:58.099–22:02.339
Because once this AI, brain, is smart enough, robots will emerge.
22:03.119–22:06.339
AI pets will emerge, AI autonomous vehicles will emerge.
22:06.339–22:09.899
you might be able to use what others have already spent so much money on
22:09.899–22:12.219
developed and made ready
22:12.219–22:14.419
and i just need to pay a reasonable price
22:14.419–22:16.499
i can integrate it into my system
22:16.499–22:18.219
it's a bit like apple now
22:18.219–22:20.339
apple chose gemini to integrate
22:20.339–22:22.099
if it's smart enough
22:22.099–22:24.339
it might just make their iphone a huge hit
22:24.339–22:28.319
these things could happen on the application side in the future
22:28.319–22:31.539
in the application sector
22:31.539–22:34.419
those who can seize the initiative and hit that sweet spot
22:34.419–22:36.139
will likely be the first wave winners
22:36.139–22:39.979
但便是企業的,在一旦會發生,
22:39.979–22:41.979
在一旦會發生的一流程。
22:43.179–22:45.859
這就是為了我們已經看著當中的AI模式
22:45.859–22:46.979
是完全適合的。
22:50.659–22:52.019
像是在手機上,
22:52.019–22:53.699
它們已經在8-10年來
22:53.699–22:55.219
讓它們發生了一些
22:55.219–22:55.979
手機,
22:55.979–22:56.699
手機,
22:56.699–22:56.859
手機,
22:56.859–22:58.219
最後手機。
23:01.419–23:01.899
現在,
23:01.899–23:03.979
它們都已經在三年。
23:03.979–23:12.879
一切都要做到,每公司都要推出了,就像每公司都要推出了,像每公司都要推出了,像 Inventech的手法上,在那天
23:12.879–23:16.759
BenQ smartphones and the rest all died off
23:16.759–23:17.379
Why?
23:17.839–23:19.259
People won't buy from everyone
23:19.259–23:20.899
It will be the same in the future
23:20.899–23:23.179
There might be a hundred brands of robots
23:23.179–23:25.139
That's when a real bubble forms
23:25.139–23:26.679
You'll need all sorts of parts
23:26.679–23:28.259
And there will be massive shortages
23:28.259–23:29.219
But in the end
23:29.219–23:31.739
There will be a pile of robots that no one buys
23:31.739–23:34.519
That's exactly where the real bubble lies
23:34.519–23:38.119
So, before those applications truly emerge
23:38.119–23:40.839
That's why many say we are in the early stages of AI
23:40.839–23:42.379
Like the top of the third inning
23:42.879–23:45.199
那是他們所說的,OK?
23:45.199–23:46.599
很多 pessimists might think
23:46.599–23:52.139
you're lying to me again, trying to trick me every day, saying it's only the third inning when stock prices have already risen like this.
23:52.259–23:55.519
They think that because they see the massive scale of the applications coming later.
23:55.719–23:56.659
Yes, exactly.
23:56.919–24:00.439
So we see that at this point in time, it is indeed still the top of the third inning.
24:00.599–24:03.179
For those of us selling shovels,there's no need to worry.
24:03.359–24:07.839
Once the leverage is removed and funds are rebalanced,buying the dip is a very good entry point.
24:07.839–24:12.499
Thank you very much to the CEO for joining us for this interview,and thank you to our viewers for watching.
24:12.699–24:13.819
Taiwan's Grand Era
24:16.599–24:32.888
請訂閱 訂閱 轉發 打賞支援明鏡與點點欄目
0:00.000–0:01.420
enWelcome to Taiwan Era.
Welcome to Taiwan Era.
0:01.580–0:03.540
enHello everyone, I'm Chong Cheng Hua.
Hello everyone, I'm Chong Cheng Hua.
0:03.840–0:08.180
enWe saw the Philadelphia Semiconductor Index undergo a rather sharp correction from late
We saw the Philadelphia Semiconductor Index undergo a rather sharp correction from late
0:08.180–0:09.140
enJune to mid-July.
June to mid-July.
0:09.260–0:13.280
enBut don't worry, everyone says this isn't an AI bubble, just a cleanup of funds and
But don't worry, everyone says this isn't an AI bubble, just a cleanup of funds and
0:13.280–0:13.840
endeleveraging.
deleveraging.
0:14.060–0:15.140
enHas the deleveraging ended?
Has the deleveraging ended?
0:15.340–0:16.360
enWhat about the fundamentals?
What about the fundamentals?
0:16.660–0:18.620
enHave there been any new changes in the fundamentals?
Have there been any new changes in the fundamentals?
0:18.980–0:22.080
enDo AI models from China pose a threat to USI developers?
Do AI models from China pose a threat to USI developers?
0:22.720–0:24.620
enIs the competition intensifying?
Is the competition intensifying?
0:24.820–0:29.500
enToday, I am very pleased to invite Junong Capital CEO, CEO Shen Jun, to analyze this
Today, I am very pleased to invite Junong Capital CEO, CEO Shen Jun, to analyze this
0:29.500–0:29.880
enfor us.
for us.
0:30.000–0:33.120
enHello,CEO,BrotherChong,Hello to all the viewers
Hello,CEO,BrotherChong,Hello to all the viewers
0:33.120–0:37.840
enRight,looking back,from late June to mid July,the SOX index corrected by about 13%,
Right,looking back,from late June to mid July,the SOX index corrected by about 13%,
0:37.840–0:40.400
enwhile some funds flowed into the so-called magnificent seven,
while some funds flowed into the so-called magnificent seven,
0:40.400–0:42.880
enwith their average stock price rising by 4.5%.
with their average stock price rising by 4.5%.
0:42.880–0:45.840
enTesla and Alphabet released their earnings reports this week,
Tesla and Alphabet released their earnings reports this week,
0:45.840–0:48.840
enso let's carefully examine this so-called i-logic.
so let's carefully examine this so-called i-logic.
0:48.840–0:50.200
enWe are seeing a shift.
We are seeing a shift.
0:50.200–0:53.600
enIn the past,we focused on capital expenditure and scale.
In the past,we focused on capital expenditure and scale.
0:53.600–0:56.480
enThe more you invested,the more the stock rose.
The more you invested,the more the stock rose.
0:56.480–0:59.600
enWill we be more stringent in examining these AI earnings now?
Will we be more stringent in examining these AI earnings now?
1:00.000–1:03.720
zh在此,AI有所造成的異常在我們的實現
在此,AI有所造成的異常在我們的實現
1:03.720–1:07.680
zhGoogle的 CEO曾經說
Google的 CEO曾經說
1:07.680–1:11.400
zhAI曾經說1%的人持有99%的貨幣
AI曾經說1%的人持有99%的貨幣
1:11.400–1:16.320
zh這不是我們的貨幣
這不是我們的貨幣
1:16.320–1:19.040
zh它決定了一個非常強烈的貨幣
它決定了一個非常強烈的貨幣
1:19.040–1:21.920
zh也會引起大增加和貨幣
也會引起大增加和貨幣
1:21.920–1:23.360
zh貨幣貨幣貨幣貨幣貨幣貨幣貨幣
貨幣貨幣貨幣貨幣貨幣貨幣貨幣
1:23.360–1:26.360
zh我們看到了如此重新的貨幣
我們看到了如此重新的貨幣
1:26.360–1:27.640
zh最近的貨幣貨幣貨幣
最近的貨幣貨幣貨幣
1:27.640–1:32.040
zh為什麼之前的強烈賞出了,甚至於是一種股市場,
為什麼之前的強烈賞出了,甚至於是一種股市場,
1:32.040–1:35.000
enwhile the M7 stocks that weren't rising began to catch up?
while the M7 stocks that weren't rising began to catch up?
1:36.040–1:39.400
enIt's simply because the gap between the winners and losers was too large.
It's simply because the gap between the winners and losers was too large.
1:40.760–1:46.360
enYou can see it yourself. By mid-year, as of June 31st, the stocks had surged by nearly 80%.
You can see it yourself. By mid-year, as of June 31st, the stocks had surged by nearly 80%.
1:48.120–1:52.920
enAmong all US stock indices, except for small cap stocks which might have risen by 20%.
Among all US stock indices, except for small cap stocks which might have risen by 20%.
1:52.920–1:57.960
enOthers like the NASDAQ were actually up less than 20%, but the stocks alone rose by 80%.
Others like the NASDAQ were actually up less than 20%, but the stocks alone rose by 80%.
1:57.960–2:00.960
zh所以,市場感受到太多重
所以,市場感受到太多重
2:00.960–2:01.960
zh更多
更多
2:01.960–2:03.960
zh有個例子,有個例子
有個例子,有個例子
2:03.960–2:05.960
zh有個例子,是有個例子
有個例子,是有個例子
2:05.960–2:08.960
zh我們可能會有四年來的屁股
我們可能會有四年來的屁股
2:08.960–2:10.960
zh但他們有四年來的屁股
但他們有四年來的屁股
2:10.960–2:12.960
zh他們說有150,000
他們說有150,000
2:12.960–2:15.960
zh你能夠借到1.5萬
你能夠借到1.5萬
2:15.960–2:17.960
zh是那些稍稅
是那些稍稅
2:17.960–2:19.960
zh這情況有所在的
這情況有所在的
2:19.960–2:21.960
zh有關影響的影響
有關影響的影響
2:21.960–2:23.960
zh從我們的影片
從我們的影片
2:23.960–2:24.960
zh從一位的主意
從一位的主意
2:24.960–2:29.880
enthe key focus is the extreme distortion of resources which will not be resolved in the future
the key focus is the extreme distortion of resources which will not be resolved in the future
2:29.880–2:31.260
enlet me share a number with you
let me share a number with you
2:31.260–2:36.600
enwe all know that csps are expected to spend 700 billion dollars or 800 billion dollars this year
we all know that csps are expected to spend 700 billion dollars or 800 billion dollars this year
2:36.600–2:37.400
enthat's the book
that's the book
2:37.400–2:41.980
enlet me mention that this year the semiconductor sector including tsmc
let me mention that this year the semiconductor sector including tsmc
2:41.980–2:46.380
enase and memory combined will earn nearly 700 billion dollars
ase and memory combined will earn nearly 700 billion dollars
2:46.380–2:49.980
enoh so that's it it's moved from here to there
oh so that's it it's moved from here to there
2:49.980–2:54.360
enand this is exactly what causes extreme volatility in stock prices
and this is exactly what causes extreme volatility in stock prices
2:54.360–2:57.860
enRight,so coming back to the question,is there an AI bubble?
Right,so coming back to the question,is there an AI bubble?
2:57.860–3:00.300
enWe just mentioned whether AI would lead to a bubble.
We just mentioned whether AI would lead to a bubble.
3:00.300–3:04.360
enI want to share that in the AI space,some are working hard for their money,like the
I want to share that in the AI space,some are working hard for their money,like the
3:04.360–3:06.460
enhumanoid robots sector.
humanoid robots sector.
3:06.460–3:11.160
enOr even those we consider winners,like OpenAI,they are working hard for their money.
Or even those we consider winners,like OpenAI,they are working hard for their money.
3:11.160–3:15.520
enThey haven't made a profit yet,they might be losing $10 billion or even $20 billion
They haven't made a profit yet,they might be losing $10 billion or even $20 billion
3:15.520–3:16.520
enevery quarter.
every quarter.
3:16.520–3:20.120
enThey are taking huge risks,and the humanoid robotics companies they pull along are also
They are taking huge risks,and the humanoid robotics companies they pull along are also
3:20.120–3:23.620
entaking huge risks,yet their profit margins remain relatively fixed.
taking huge risks,yet their profit margins remain relatively fixed.
3:23.620–3:24.120
zh嗯
嗯
3:24.120–3:28.120
enconversely for those selling shovels wow their profits keep piling up
conversely for those selling shovels wow their profits keep piling up
3:28.120–3:34.220
eneven the director of sk hinnick said excuse me i feel this isn't quite normal we're making too much money could it be
even the director of sk hinnick said excuse me i feel this isn't quite normal we're making too much money could it be
3:34.220–3:37.820
enthey even said the prices are a bit high and they should probably be a bit healthier
they even said the prices are a bit high and they should probably be a bit healthier
3:37.820–3:42.720
enif you read the whole thing he sounds a bit guilty wondering how they could be earning so much
if you read the whole thing he sounds a bit guilty wondering how they could be earning so much
3:42.720–3:45.520
enhe didn't actually see a downturn coming yet
he didn't actually see a downturn coming yet
3:45.520–3:53.520
enso from this perspective we really need to be mentally prepared for the investment market knowing that it will never be balanced
so from this perspective we really need to be mentally prepared for the investment market knowing that it will never be balanced
3:53.520–3:58.220
enWhen extreme stock price volatility occurs, sometimes people get cautious, causing the
When extreme stock price volatility occurs, sometimes people get cautious, causing the
3:58.220–4:01.840
enprice to rebalance, but in the long run, it will still lead to polarization.
price to rebalance, but in the long run, it will still lead to polarization.
4:02.980–4:08.200
enYes, but when we look back at the memory industry, we see they are continuously expanding capacity.
Yes, but when we look back at the memory industry, we see they are continuously expanding capacity.
4:08.660–4:12.960
enWe know that the demand for memory remains very high, but it will certainly take time
We know that the demand for memory remains very high, but it will certainly take time
4:12.960–4:17.060
enfor the supply side to catch up. I just want to remind everyone, because many ghost stories
for the supply side to catch up. I just want to remind everyone, because many ghost stories
4:17.060–4:21.140
enare surfacing right now, as memory manufacturers everywhere are expanding plants, whether it's
are surfacing right now, as memory manufacturers everywhere are expanding plants, whether it's
4:21.140–4:25.300
zh在香港,臺灣,或在中國,包括YMTC和CXMT,
在香港,臺灣,或在中國,包括YMTC和CXMT,
4:25.300–4:26.300
zh都在增加了
都在增加了
4:26.300–4:27.300
zh人們可能會想想
人們可能會想想
4:27.300–4:28.860
zh如果您撥開工程
如果您撥開工程
4:28.860–4:31.220
zh決定,它們將在增加了什麼我們的名義
決定,它們將在增加了什麼我們的名義
4:31.220–4:32.260
zh是一種連續的連續
是一種連續的連續
4:32.260–4:33.700
zh它們現在變成了一種連續
它們現在變成了一種連續
4:33.700–4:35.380
zh在最後一種連續的連續
在最後一種連續的連續
4:35.380–4:36.660
zh至於2030,
至於2030,
4:36.660–4:38.620
enthere will still be a very significant shortage.
there will still be a very significant shortage.
4:38.860–4:40.220
enThis can help strengthen confidence,
This can help strengthen confidence,
4:40.560–4:42.340
enconfirming that if the memory sector corrects,
confirming that if the memory sector corrects,
4:42.560–4:45.880
enit is perfectly fine in terms of both stock price and industry fundamentals.
it is perfectly fine in terms of both stock price and industry fundamentals.
4:46.520–4:50.640
enLike recently, Sammy, who used to love to talk down the market and so on,
Like recently, Sammy, who used to love to talk down the market and so on,
4:51.040–4:54.840
ensuddenly came out and said he believes memory will remain in shortage all the way through.
suddenly came out and said he believes memory will remain in shortage all the way through.
4:55.460–4:57.340
enHis perspective, or for example,
His perspective, or for example,
4:57.720–5:02.500
enlike even when you look at TSMC's earnings call mentioning envious gross margins and so on,
like even when you look at TSMC's earnings call mentioning envious gross margins and so on,
5:02.880–5:04.120
enactually tells you one thing,
actually tells you one thing,
5:04.120–5:09.440
enwhy some people will tell you decisively that memory demand will be excellent and will continue to be in short supply
why some people will tell you decisively that memory demand will be excellent and will continue to be in short supply
5:09.440–5:12.380
enyou might think if the price gets too high i just won't buy
you might think if the price gets too high i just won't buy
5:12.380–5:17.560
enthe conventional investment logic for business has always been if it goes up i won't use it or i'll just stop
the conventional investment logic for business has always been if it goes up i won't use it or i'll just stop
5:17.560–5:19.300
enso how could demand possibly be limitless
so how could demand possibly be limitless
5:19.300–5:23.900
enwhat i want to share with everyone that they are ignoring or why they see what others are ignoring
what i want to share with everyone that they are ignoring or why they see what others are ignoring
5:23.900–5:29.460
enis that we now see tsmc's capex at 60 billion this year maybe 80 billion next year
is that we now see tsmc's capex at 60 billion this year maybe 80 billion next year
5:29.460–5:33.320
enor nvidia's gpus and amd's gpus doing great
or nvidia's gpus and amd's gpus doing great
5:33.320–5:37.040
zh我們當然忘記 that for every GPU or CPU produced
我們當然忘記 that for every GPU or CPU produced
5:37.040–5:38.620
zh它 must be paired with a unit
它 must be paired with a unit
5:38.620–5:40.060
ena fixed unit of memory
a fixed unit of memory
5:40.060–5:42.580
zh它是不 sold separately
它是不 sold separately
5:42.580–5:44.340
enmemory is not sold alone
memory is not sold alone
5:44.340–5:46.640
zh所以我們的 AI future becomes clear
所以我們的 AI future becomes clear
5:46.640–5:48.760
enwith so many GPUs coming out
with so many GPUs coming out
5:48.760–5:49.980
zh如果我們不存在的話
如果我們不存在的話
5:49.980–5:50.000
zh我們不存在的話
我們不存在的話
5:50.000–5:51.380
zh我們不存在的話
我們不存在的話
5:51.380–5:54.400
zh所以我們看它從 TSMC's perspective
所以我們看它從 TSMC's perspective
5:54.400–5:56.120
zh從 the semiconductor perspective
從 the semiconductor perspective
5:56.120–5:58.320
zh我們的 demand for memory 4 years from now
我們的 demand for memory 4 years from now
5:58.320–6:00.480
enshould be at least 4 times what it is today
should be at least 4 times what it is today
6:00.480–6:00.880
zh嗯
嗯
6:00.880–6:01.880
zh在最少
在最少
6:01.880–6:02.880
enYes
Yes
6:02.880–6:05.880
enAnd this premise is assuming my GPU volume hasn't even increased
And this premise is assuming my GPU volume hasn't even increased
6:05.880–6:08.880
enIf I produce 100 units of GPUs this year and 100 units 4 years from now
If I produce 100 units of GPUs this year and 100 units 4 years from now
6:08.880–6:11.880
enAlthough there is migration and they might be more powerful in 4 years
Although there is migration and they might be more powerful in 4 years
6:11.880–6:15.880
enIf it's still 100 units, the memory I need to pair with them will still be 4 times as much
If it's still 100 units, the memory I need to pair with them will still be 4 times as much
6:15.880–6:21.880
enUnless I downgrade, for instance, fine, I'm driving a Bentley, but I don't want air conditioning
Unless I downgrade, for instance, fine, I'm driving a Bentley, but I don't want air conditioning
6:21.880–6:24.880
enI don't want ABS brakes, then your performance will definitely suffer
I don't want ABS brakes, then your performance will definitely suffer
6:24.880–6:28.880
enSo later on you will keep hearing people talk about downgrading specs
So later on you will keep hearing people talk about downgrading specs
6:28.880–6:30.880
zh你可能會認為 memory是不好的,
你可能會認為 memory是不好的,
6:30.880–6:33.380
zh但不,因為有沒有的,
但不,因為有沒有的,
6:33.380–6:38.880
zh你會認為大家是在中國的增長和產品的增長。
你會認為大家是在中國的增長和產品的增長。
6:38.880–6:42.880
zh如果看來,產品是在增長的增長並且在2028,
如果看來,產品是在增長的增長並且在2028,
6:42.880–6:44.880
zh或者我們叫它,四年以後,
或者我們叫它,四年以後,
6:44.880–6:45.880
zh加上了它的增長。
加上了它的增長。
6:45.880–6:46.880
zh嗯嗯。
嗯嗯。
6:46.880–6:48.880
zh我們就說,
我們就說,
6:48.880–6:50.880
zh在這四次,
在這四次,
6:50.880–6:51.880
zh所以,
所以,
6:51.880–6:52.880
zh在這一年,
在這一年,
6:52.880–6:54.880
zh四次。
四次。
6:54.880–6:57.880
zh如果我們的資料設計是在四年以後,
如果我們的資料設計是在四年以後,
6:57.880–7:03.740
enit is now, wouldn't TSMC be finished? Wouldn't Nvidia be finished? That's impossible.
it is now, wouldn't TSMC be finished? Wouldn't Nvidia be finished? That's impossible.
7:05.020–7:10.340
enWe're all optimistic that GPUs will keep trending upward, so investors who are worried about memory,
We're all optimistic that GPUs will keep trending upward, so investors who are worried about memory,
7:10.880–7:15.280
enif you are truly worried that memory will crash, then you should be like Michael Burry and not buy
if you are truly worried that memory will crash, then you should be like Michael Burry and not buy
7:15.280–7:22.040
enany semiconductors at all. You can't say you're bullish on TSMC but bearish on memory. Let's get
any semiconductors at all. You can't say you're bullish on TSMC but bearish on memory. Let's get
7:22.040–7:26.700
enthis logic straight. You can't have it both ways. If it looks like memory is going to collapse in
this logic straight. You can't have it both ways. If it looks like memory is going to collapse in
7:26.700–7:29.700
zh未來,這些世界都會崩潰。
未來,這些世界都會崩潰。
7:29.700–7:32.700
zh當然,我們不可能會否有機會,
當然,我們不可能會否有機會,
7:32.700–7:35.700
zh我們不可能會否會發展開四年
我們不可能會否會發展開四年
7:35.700–7:37.700
zh和世界的世界返回到大陸,
和世界的世界返回到大陸,
7:37.700–7:39.700
zh讓 AI後來的確定。
讓 AI後來的確定。
7:39.700–7:43.700
zh我們不可能會否會發生。
我們不可能會否會發生。
7:43.700–7:46.700
zh但你必須理解釋這個 logic,
但你必須理解釋這個 logic,
7:46.700–7:48.700
zh如果 AI還會升幅度的數量,
如果 AI還會升幅度的數量,
7:48.700–7:50.700
zh如果數量的數量會升幅,
如果數量的數量會升幅,
7:50.700–7:52.700
zh就會增加了。
就會增加了。
7:53.700–7:55.700
zh這就是這些人所說的,
這就是這些人所說的,
7:55.700–7:58.460
zh不只是給你一個路線或是領導的
不只是給你一個路線或是領導的
7:59.212–8:01.652
zh而我們剛才提到的
而我們剛才提到的
8:01.652–8:02.212
zh《病》
《病》
8:02.212–8:04.892
zh我們還沒有說了
我們還沒有說了
8:04.892–8:05.772
zh我們還沒有說了
我們還沒有說了
8:05.772–8:07.332
zh《軟體》
《軟體》
8:07.332–8:08.532
zh《軟體》
《軟體》
8:08.532–8:08.692
zh《軟體》
《軟體》
8:08.692–8:09.172
zh《軟體》
《軟體》
8:09.172–8:09.892
zh《軟體》
《軟體》
8:09.892–8:11.012
zh《軟體》
《軟體》
8:11.012–8:11.892
zh《軟體》
《軟體》
8:11.892–8:12.932
zh《軟體》
《軟體》
8:12.932–8:14.132
zh《軟體》
《軟體》
8:14.212–8:22.652
enAs we've mentioned before, TSMC estimates that future applications for AI will require 8 times the current computing power of these, brains.
As we've mentioned before, TSMC estimates that future applications for AI will require 8 times the current computing power of these, brains.
8:24.212–8:28.612
enYes, the memory demand is simply unimaginable, it's terrifying.
Yes, the memory demand is simply unimaginable, it's terrifying.
8:29.272–8:33.932
enTo be honest, in our reports from visiting Samsung and Hinex, even they can't calculate.
To be honest, in our reports from visiting Samsung and Hinex, even they can't calculate.
8:34.692–8:38.072
enExactly how much is needed, they don't know how to negotiate with clients.
Exactly how much is needed, they don't know how to negotiate with clients.
8:38.072–8:40.672
zh如果他們太高,他們會負敗的交易,
如果他們太高,他們會負敗的交易,
8:40.672–8:42.112
zh但在實際上,他們會負敗的貨幫助,
但在實際上,他們會負敗的貨幫助,
8:42.112–8:44.112
zh他們會負敗的貨幫助,
他們會負敗的貨幫助,
8:44.112–8:46.392
zh他們會負敗的貨幫助。
他們會負敗的貨幫助。
8:46.392–8:48.192
zh當然,他們會負敗的貨幫助,
當然,他們會負敗的貨幫助,
8:48.192–8:50.992
zh他們會負敗的貨幫助,
他們會負敗的貨幫助,
8:50.992–8:52.752
zh但根據上一個討論的評論,
但根據上一個討論的評論,
8:52.752–8:54.632
zh他們會負敗的貨幫助。
他們會負敗的貨幫助。
8:54.632–8:56.352
zh另一個重點的評論,
另一個重點的評論,
8:56.352–8:59.312
zh包括了一種同時的評論,
包括了一種同時的評論,
8:59.312–9:00.592
zh我們看到最近,
我們看到最近,
9:00.592–9:02.192
zh人們覺得這些關係的關係
人們覺得這些關係的關係
9:02.192–9:04.312
zh是被高傳的高傳領的高傳領的
是被高傳的高傳領的高傳領的
9:04.312–9:09.672
enand their ratings are increasing so it's a situation where if gpus increase racks increase
and their ratings are increasing so it's a situation where if gpus increase racks increase
9:09.672–9:15.272
enand servers increase including from ai from communications passive components and pcbs
and servers increase including from ai from communications passive components and pcbs
9:15.272–9:20.392
enall demand is moving upward if we talk about computing power as we just mentioned cpus and
all demand is moving upward if we talk about computing power as we just mentioned cpus and
9:20.392–9:26.632
engpus are paired with memory they are a set in the future like the hvdc power supply for servers
gpus are paired with memory they are a set in the future like the hvdc power supply for servers
9:26.632–9:32.472
enit is paired with passive components passive components are used for power so if you believe
it is paired with passive components passive components are used for power so if you believe
9:32.472–9:34.472
zh如果您相信,您的需要提供力量的力量,
如果您相信,您的需要提供力量的力量,
9:34.472–9:36.472
zh您的支援力量是否有力量的力量,
您的支援力量是否有力量的力量,
9:36.472–9:38.472
zh您的支援力量是否有力量的力量。
您的支援力量是否有力量的力量。
9:38.472–9:41.472
zh如果您相信,您的支援力量是否有力量的力量。
如果您相信,您的支援力量是否有力量的力量。
9:41.472–9:44.472
zh如果您相信,您的支援力量是否有力量的力量。
如果您相信,您的支援力量是否有力量的力量。
9:44.472–9:46.472
zh您相信,您的支援力量是否有力量的力量。
您相信,您的支援力量是否有力量的力量。
9:46.472–9:49.472
zh您相信,您的支援力量是否有力量的力量。
您相信,您的支援力量是否有力量的力量。
9:49.472–9:52.472
zh所以,您的支援力量是否有力量的力量。
所以,您的支援力量是否有力量的力量。
9:52.472–9:54.472
zh所以,您的支援力量是否有力量的力量。
所以,您的支援力量是否有力量的力量。
9:54.472–9:56.472
zh您的支援力量是否有力量的力量。
您的支援力量是否有力量的力量。
9:56.472–9:58.472
zh您的支援力量是否有力量的力量。
您的支援力量是否有力量的力量。
9:58.472–10:00.472
zh所以,您的支援力量是否有力量的力量。
所以,您的支援力量是否有力量的力量。
10:00.472–10:30.472
zh優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優�
優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優優�
10:30.472–10:35.392
enbut shortages are normal because after all this massive demand, the supply simply cannot keep up.
but shortages are normal because after all this massive demand, the supply simply cannot keep up.
10:35.652–10:38.632
enNow, as the first wave pulls back and finishes consolidating,
Now, as the first wave pulls back and finishes consolidating,
10:39.032–10:43.292
enonce the actual shortages begin, the stock price will follow the logic of supply shortages.
once the actual shortages begin, the stock price will follow the logic of supply shortages.
10:43.632–10:45.392
enSo, what should its valuation be?
So, what should its valuation be?
10:45.572–10:49.252
enActually, in due time, the capital market will give it a reasonable valuation.
Actually, in due time, the capital market will give it a reasonable valuation.
10:49.252–10:53.612
enOkay, so looking at it from an economic perspective or from the fundamentals of the industry,
Okay, so looking at it from an economic perspective or from the fundamentals of the industry,
10:53.972–10:55.512
enmemory is currently in short supply.
memory is currently in short supply.
10:55.672–10:59.512
enThe supply side is severely lacking, and this will likely continue for a long time.
The supply side is severely lacking, and this will likely continue for a long time.
10:59.512–11:02.592
zh但是,有一個不斷的經濟發生在這個業務中的
但是,有一個不斷的經濟發生在這個業務中的
11:02.592–11:05.652
zh美國的美國人說是中國人的美國人
美國的美國人說是中國人的美國人
11:05.652–11:08.472
zh像 Samsung是傳續於美國人的美國人
像 Samsung是傳續於美國人的美國人
11:08.472–11:09.872
zh而是太多貨
而是太多貨
11:09.872–11:12.312
zh的供貨力是完全在美國人的手段
的供貨力是完全在美國人的手段
11:12.312–11:14.952
zh而且是中國人的美國人現在太高
而且是中國人的美國人現在太高
11:14.952–11:17.312
zh所以政府是購買的貨帶
所以政府是購買的貨帶
11:17.312–11:18.652
zh你認為這個人的這個?
你認為這個人的這個?
11:18.992–11:21.472
zh我認為這個人認為他們認為貨帶的貨帶
我認為這個人認為他們認為貨帶的貨帶
11:21.472–11:24.092
zh但在實際上,我們的認為我們的課計
但在實際上,我們的認為我們的課計
11:24.092–11:27.272
zh難以為他們的貨帶貨帶的貨帶
難以為他們的貨帶貨帶的貨帶
11:27.272–11:31.312
enWith US companies, they can ask for equity, like they did with Intel or others.
With US companies, they can ask for equity, like they did with Intel or others.
11:31.592–11:35.532
enBut for the US government to take equity in a Korean company, that's a bit difficult.
But for the US government to take equity in a Korean company, that's a bit difficult.
11:35.872–11:38.912
enWe haven't seen them take an equity stake in TSMC, right?
We haven't seen them take an equity stake in TSMC, right?
11:39.072–11:43.472
enHowever, they could follow the TSMC model, meaning if you don't come and set up a plant here.
However, they could follow the TSMC model, meaning if you don't come and set up a plant here.
11:43.912–11:47.252
enIf you don't come here to build what they call a US-based memory facility,
If you don't come here to build what they call a US-based memory facility,
11:47.672–11:49.312
enbecause everyone knows there's a shortage,
because everyone knows there's a shortage,
11:49.752–11:53.412
enit's only natural for them to want these companies to expand their capacity here.
it's only natural for them to want these companies to expand their capacity here.
11:53.412–11:58.632
enhonestly if these companies don't face such overwhelming demand pushing them
honestly if these companies don't face such overwhelming demand pushing them
11:58.632–12:04.552
enthey really aren't willing to expand because for memory manufacturers they know one thing
they really aren't willing to expand because for memory manufacturers they know one thing
12:04.552–12:09.732
eni don't need to expand my plants when the time comes and you have a shortage i'll just double
i don't need to expand my plants when the time comes and you have a shortage i'll just double
12:09.732–12:14.512
enthe price and i'll make the same amount of money or even more and i can do that for a long time
the price and i'll make the same amount of money or even more and i can do that for a long time
12:14.512–12:19.792
enof course they want a healthy market so their customers don't go under but for them over
of course they want a healthy market so their customers don't go under but for them over
12:19.792–12:49.792
zh肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料�
肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料肥料�
12:49.792–12:50.792
enKIMI-K3
KIMI-K3
12:50.792–12:55.632
enThis was actually one of the factors that led to a market correction last Friday.
This was actually one of the factors that led to a market correction last Friday.
12:56.372–13:01.612
enPeople thought, okay, last month the White House decided that Anthropix Claude 3.5 was
People thought, okay, last month the White House decided that Anthropix Claude 3.5 was
13:01.612–13:04.592
entoo powerful, so it restricted access for non-Americans.
too powerful, so it restricted access for non-Americans.
13:05.412–13:11.132
enAnd then, shortly after, China released Quen 2.5, which scored nearly the same as 3.5,
And then, shortly after, China released Quen 2.5, which scored nearly the same as 3.5,
13:11.132–13:13.732
enbut it's actually open source and launched immediately.
but it's actually open source and launched immediately.
13:14.692–13:19.732
enHow big of a shock is this to the current USI model landscape, and has market competition
How big of a shock is this to the current USI model landscape, and has market competition
13:19.732–13:25.232
zh其實, everyone is just looking at the final evaluation results.
其實, everyone is just looking at the final evaluation results.
13:25.632–13:28.272
enIt has 7 metrics, and Quen won 6 of them.
It has 7 metrics, and Quen won 6 of them.
13:29.232–13:32.592
enAnd one look tells us that Chinese people are really good at taking tests.
And one look tells us that Chinese people are really good at taking tests.
13:32.892–13:36.812
enWhat we are best at is test-taking, whether it's the SATs or our joint entrance exams,
What we are best at is test-taking, whether it's the SATs or our joint entrance exams,
13:37.132–13:39.072
enwe eventually just master the question bank.
we eventually just master the question bank.
13:40.132–13:44.352
enEssentially, they looked at those 7 performance metrics and tuned their model to be perfectly
Essentially, they looked at those 7 performance metrics and tuned their model to be perfectly
13:44.352–13:47.572
ensuited for those, whether for AI agents or image generation.
suited for those, whether for AI agents or image generation.
13:47.572–13:50.972
enYou can actually see how their models work
You can actually see how their models work
13:50.972–13:55.052
enCurrently, whether it's Anthropic or OpenAI, they are all very powerful
Currently, whether it's Anthropic or OpenAI, they are all very powerful
13:55.052–13:59.692
enSo powerful, in fact, that sometimes they become powerful enough to be stupid
So powerful, in fact, that sometimes they become powerful enough to be stupid
13:59.692–14:04.852
enFor example, if you use Gemini in a coffee shop
For example, if you use Gemini in a coffee shop
14:04.852–14:07.592
enIt starts off strong with great proposals and estimates
It starts off strong with great proposals and estimates
14:07.592–14:11.092
enBut after a month, it gets slow because every time it makes a decision
But after a month, it gets slow because every time it makes a decision
14:11.092–14:13.372
enIt reconsiders every previous decision
It reconsiders every previous decision
14:13.372–14:17.712
enyou can imagine how that piles up
you can imagine how that piles up
14:17.712–14:20.012
enby the second month
by the second month
14:20.012–14:21.272
enit starts making errors
it starts making errors
14:21.272–14:23.972
enlike ordering thousands of rubber gloves for a cafe
like ordering thousands of rubber gloves for a cafe
14:23.972–14:27.932
enit starts ordering carrots
it starts ordering carrots
14:27.932–14:29.112
enwhen there isn't even a kitchen
when there isn't even a kitchen
14:29.112–14:30.092
enso in the end
so in the end
14:30.092–14:31.172
enit just crashes
it just crashes
14:31.172–14:34.372
enthe problem with AI now
the problem with AI now
14:34.372–14:35.872
enis that while it might be powerful
is that while it might be powerful
14:35.872–14:37.572
eneven single task powerful
even single task powerful
14:37.572–14:40.512
eni have to tune it to handle these 7 specific metrics
i have to tune it to handle these 7 specific metrics
14:40.512–14:42.792
enso you have to train it specifically for that
so you have to train it specifically for that
14:43.372–14:46.492
zh看起來,不像 DeepSeq
看起來,不像 DeepSeq
14:46.492–14:52.852
enDeepSeq is like a needle, a small model that had a huge impact when it came out.
DeepSeq is like a needle, a small model that had a huge impact when it came out.
14:54.412–14:55.432
enWhat does that mean?
What does that mean?
14:56.012–15:00.152
enIt means I don't need to be that powerful, and I don't need that much computing power.
It means I don't need to be that powerful, and I don't need that much computing power.
15:00.952–15:05.912
enBut with Quen, I have massive computing power, yet I perform better than you, and the way
But with Quen, I have massive computing power, yet I perform better than you, and the way
15:05.912–15:09.912
enI perform better is by tuning it specifically to handle the performance you're looking for.
I perform better is by tuning it specifically to handle the performance you're looking for.
15:10.432–15:14.152
enComing back to the point, even though it performs well, people don't need to worry
Coming back to the point, even though it performs well, people don't need to worry
15:14.152–15:16.792
zh太多了因為它還需要大量的空氣力
太多了因為它還需要大量的空氣力
15:16.792–15:18.952
zh然後它將會很快的空氣力
然後它將會很快的空氣力
15:18.952–15:21.112
zh有一個東西人都沒有認為
有一個東西人都沒有認為
15:21.112–15:22.072
zh它的價值
它的價值
15:22.072–15:25.192
zh它的價值是一樣的價值
它的價值是一樣的價值
15:25.192–15:28.632
zh它是0.95%和1.07%
它是0.95%和1.07%
15:28.632–15:31.592
zh但安thropic是超費的
但安thropic是超費的
15:31.592–15:32.392
en$3
$3
15:32.392–15:33.912
zh但看起來
但看起來
15:33.912–15:35.752
zh是 OpenAI really doing worse?
是 OpenAI really doing worse?
15:35.752–15:37.992
zh它是非常好
它是非常好
15:37.992–15:38.792
zh的開放
的開放
15:38.792–15:40.592
zh為
為
15:40.592–15:42.552
zh因為人們說它是費的
因為人們說它是費的
15:42.552–15:44.312
enare enterprises willing to pay?
are enterprises willing to pay?
15:44.752–15:47.512
enthis is what we need to go back and discuss regarding AI
this is what we need to go back and discuss regarding AI
15:48.232–15:49.912
enit's not just about performing well
it's not just about performing well
15:50.792–15:52.872
enor generating pretty pictures or answers
or generating pretty pictures or answers
15:54.072–15:54.872
enif that were all
if that were all
15:54.872–15:57.752
enit would just be for us non-paying or low-paying users
it would just be for us non-paying or low-paying users
15:58.819–15:59.819
zh而是我們的不費用
而是我們的不費用
15:59.819–16:01.819
zh原因是因為它做得到的事情
原因是因為它做得到的事情
16:01.819–16:03.819
zh它是做得到的
它是做得到的
16:03.819–16:05.819
zh它是需要的需要
它是需要的需要
16:05.819–16:07.819
zh像是需要的
像是需要的
16:07.819–16:09.819
zh如果用它是用它
如果用它是用它
16:09.819–16:11.819
zh或是關於你的研究
或是關於你的研究
16:11.819–16:13.819
zh你不能夠做得到的
你不能夠做得到的
16:13.819–16:21.819
zh你會看到,甚至是很貴的,而且它的效果不太好,為什麼它是很高的?
你會看到,甚至是很貴的,而且它的效果不太好,為什麼它是很高的?
16:23.819–16:32.819
zh你可能說,這些人們不懂得用 Open Source,但我認為不覺得他們的獨自貴。
你可能說,這些人們不懂得用 Open Source,但我認為不覺得他們的獨自貴。
16:32.819–17:02.819
zh他們應該能夠實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實
他們應該能夠實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實現實
17:02.819–17:04.259
zh是在那裡的事,
是在那裡的事,
17:04.259–17:05.579
zh之後K was released,
之後K was released,
17:05.579–17:08.119
zh它的機器 immediately hit the limit.
它的機器 immediately hit the limit.
17:08.119–17:09.919
zh它 seems unable to allow new users
它 seems unable to allow new users
17:09.919–17:11.759
zh去上網上用了新的新版本來使用了新版本來的
去上網上用了新的新版本來使用了新版本來的
17:11.759–17:16.319
zh因為它的背景已經不足了。
因為它的背景已經不足了。
17:16.319–17:18.719
zh所以,它不是在這裡的底色
所以,它不是在這裡的底色
17:18.719–17:21.059
zh它可以用 AI models on lighter hardware.
它可以用 AI models on lighter hardware.
17:21.059–17:24.219
zh它們的額外 demand是其實大量的。
它們的額外 demand是其實大量的。
17:24.219–17:27.919
zh所以,是這其實好事?
所以,是這其實好事?
17:27.919–17:29.919
zh是這其實好事?
是這其實好事?
17:29.919–17:31.119
zh有更多的影響,
有更多的影響,
17:31.119–17:32.859
zh是更多的更多的更多
是更多的更多的更多
17:32.859–17:34.259
zh為什麼是更多的更多?
為什麼是更多的更多?
17:35.619–17:37.159
zh因為我們剛剛說
因為我們剛剛說
17:37.159–17:38.559
zh如果是個體驗的問題
如果是個體驗的問題
17:38.559–17:40.159
zh而你的模特兒是有點困難
而你的模特兒是有點困難
17:40.159–17:42.359
zh而你使用它在不足的地方
而你使用它在不足的地方
17:42.359–17:43.419
zh但實際上的使用者
但實際上的使用者
17:43.419–17:45.459
zh不會讓我能夠享受到更多的效果
不會讓我能夠享受到更多的效果
17:45.459–17:46.419
zh而我不會讓我能夠
而我不會讓我能夠
17:46.419–17:47.719
zh稍為的更多的效果
稍為的更多的效果
17:49.319–17:51.259
zh如果我知道這些資料
如果我知道這些資料
17:51.259–17:52.259
zh在正確的地方
在正確的地方
17:52.259–17:54.259
zh要有更多的執行
要有更多的執行
17:54.259–17:55.579
zh是當然好
是當然好
17:55.579–17:56.379
zh那些人會有盡量的
那些人會有盡量的
17:56.379–17:58.159
zh所以你能看到
所以你能看到
17:58.159–18:28.159
zh這次的壞壞壞在售貨物店裡的壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞�
這次的壞壞壞在售貨物店裡的壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞壞�
18:28.159–18:32.159
enthe real concern might be when we go back to examine OpenAI
the real concern might be when we go back to examine OpenAI
18:32.159–18:34.159
enit becomes a situation where, sure,
it becomes a situation where, sure,
18:34.159–18:38.159
enCLOD's benchmark scores might not be significantly higher than GPT-4's
CLOD's benchmark scores might not be significantly higher than GPT-4's
18:38.159–18:42.159
enbut they are able to make much more money from enterprise clients
but they are able to make much more money from enterprise clients
18:42.159–18:44.159
enExactly
Exactly
18:44.159–18:48.159
enAs we mentioned, OpenAI's costs and efficiency are similar to CLOD's
As we mentioned, OpenAI's costs and efficiency are similar to CLOD's
18:48.159–18:50.159
enso CLOD 3 might not hurt Anthropic,
so CLOD 3 might not hurt Anthropic,
18:50.159–18:54.159
enbut it could definitely hit OpenAI
but it could definitely hit OpenAI
18:54.159–19:00.159
zh有甚至今的說話,如果有一個AI瀏規, OpenAI會是說的一種類似的一種類似的。
有甚至今的說話,如果有一個AI瀏規, OpenAI會是說的一種類似的一種類似的。
19:01.159–19:05.159
zh所以,我們看來的責任 OpenAI在中國的責任?
所以,我們看來的責任 OpenAI在中國的責任?
19:06.159–19:10.159
zh如果它不影響 Anthropik,它會直接射擊 OpenAI,
如果它不影響 Anthropik,它會直接射擊 OpenAI,
19:10.159–19:13.159
zh尤其是跟 Oracle和 Cloudflare的背景。
尤其是跟 Oracle和 Cloudflare的背景。
19:14.159–19:16.159
zh你覺得這個動作?
你覺得這個動作?
19:18.159–19:20.159
zh其實,我們不相信AI會流動,
其實,我們不相信AI會流動,
19:20.159–19:25.159
zh但最弱的原因是在《AI》和《Oracle》中的主要是《OpenAI》。
但最弱的原因是在《AI》和《Oracle》中的主要是《OpenAI》。
19:30.859–19:32.159
zh你能看到了解的目標
你能看到了解的目標
19:32.159–19:35.359
zh像它的借貨倫、購乎率、購乎率、
像它的借貨倫、購乎率、購乎率、
19:35.359–19:37.959
zh或是在購買購買貨。
或是在購買購買貨。
19:37.959–19:39.959
zh這些是否有任何的影響。
這些是否有任何的影響。
19:45.859–19:49.819
zh只要是,《OpenAI》需要找到一種商品的商品作為
只要是,《OpenAI》需要找到一種商品的商品作為
19:50.159–19:53.919
zh所以,看著他們在做的
所以,看著他們在做的
19:53.919–19:55.399
zh他們要進行進行的
他們要進行進行的
19:55.399–19:57.399
zh他們想要做一個話題
他們想要做一個話題
19:57.399–19:58.399
zh他們要做一個話題
他們要做一個話題
19:58.399–19:59.399
zh他們要做一些其他東西
他們要做一些其他東西
19:59.399–20:02.399
zh他們要做到販賣物的物品
他們要做到販賣物的物品
20:02.399–20:04.399
zh我們會把這個做法
我們會把這個做法
20:04.399–20:05.399
zh很清楚
很清楚
20:05.399–20:08.399
zh他們真的不能夠保持
他們真的不能夠保持
20:08.399–20:10.399
zhunless美國給他們
unless美國給他們
20:10.399–20:13.399
zh他們的公司不可能會有問題
他們的公司不可能會有問題
20:15.399–20:17.399
zh所以我們會看 closely
所以我們會看 closely
20:17.399–20:18.279
zh因為他們的商品
因為他們的商品
20:18.279–20:19.779
zh如果這個認識的認識是沒有發展,
如果這個認識的認識是沒有發展,
20:19.779–20:20.279
zh在這個中處的認識是沒有發展,
在這個中處的認識是沒有發展,
20:20.279–20:21.779
zh在這個中處的認識是沒有發展,
在這個中處的認識是沒有發展,
20:21.779–20:23.779
zh在這個中處的認識是沒有發展,
在這個中處的認識是沒有發展,
20:23.779–20:25.779
zh在這個中處的認識是沒有發展,
在這個中處的認識是沒有發展,
20:25.779–20:27.779
zh在這個中處的認識是沒有發展,
在這個中處的認識是沒有發展,
20:27.779–20:29.779
zh在這個中處的認識是沒有發展,
在這個中處的認識是沒有發展,
20:29.779–20:31.779
zh在這個中處的認識是沒有發展,
在這個中處的認識是沒有發展,
20:31.779–20:33.779
zh在這個中處的認識是沒有發展,
在這個中處的認識是沒有發展,
20:33.779–20:35.779
zh在這個中處的認識是沒有發展,
在這個中處的認識是沒有發展,
20:35.779–20:37.779
zh在這個中處的認識是沒有發展,
在這個中處的認識是沒有發展,
20:37.779–20:39.779
zh如果他們開始發展,
如果他們開始發展,
20:39.779–20:40.779
zh如果他們發展,
如果他們發展,
20:40.779–20:41.779
zh他們發展,他們發展,
他們發展,他們發展,
20:41.779–20:42.779
zh他們發展,他們發展,
他們發展,他們發展,
20:42.779–21:12.779
zh如果能否購買貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨
如果能否購買貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨貨
21:12.779–21:16.619
zh在美國的美國,我們覺得我們的貨幣是否會有效果,
在美國的美國,我們覺得我們的貨幣是否會有效果,
21:16.619–21:19.979
zh我們的其他人,不一定會有效果。
我們的其他人,不一定會有效果。
21:19.979–21:25.019
zh其實,我們覺得我們的貨幣未來的未來,
其實,我們覺得我們的貨幣未來的未來,
21:25.019–21:28.139
zh現在,當然,我們的貨幣未來的貨幣未來,
現在,當然,我們的貨幣未來的貨幣未來,
21:28.139–21:29.179
zh那是確實。
那是確實。
21:29.179–21:31.979
zh現在,如果你貨幣,你貨幣。
現在,如果你貨幣,你貨幣。
21:31.979–21:35.819
zh而且,你更多貨幣,像更多貨幣,
而且,你更多貨幣,像更多貨幣,
21:35.819–21:39.019
zh像TSMC,他們不能算貨幣,
像TSMC,他們不能算貨幣,
21:39.019–21:40.719
enEventually, they will make money.
Eventually, they will make money.
21:41.879–21:47.479
enYes, it's hard-earned money, and they are truly carrying human hope forward, like OpenAI and Anthropic.
Yes, it's hard-earned money, and they are truly carrying human hope forward, like OpenAI and Anthropic.
21:48.419–21:50.939
enThey have to figure out how to make money themselves.
They have to figure out how to make money themselves.
21:51.879–21:54.219
enWhy do I say the real winners haven't appeared?
Why do I say the real winners haven't appeared?
21:54.979–21:57.079
enBecause it's about the application side.
Because it's about the application side.
21:58.099–22:02.339
enBecause once this AI, brain, is smart enough, robots will emerge.
Because once this AI, brain, is smart enough, robots will emerge.
22:03.119–22:06.339
enAI pets will emerge, AI autonomous vehicles will emerge.
AI pets will emerge, AI autonomous vehicles will emerge.
22:06.339–22:09.899
enyou might be able to use what others have already spent so much money on
you might be able to use what others have already spent so much money on
22:09.899–22:12.219
endeveloped and made ready
developed and made ready
22:12.219–22:14.419
enand i just need to pay a reasonable price
and i just need to pay a reasonable price
22:14.419–22:16.499
eni can integrate it into my system
i can integrate it into my system
22:16.499–22:18.219
enit's a bit like apple now
it's a bit like apple now
22:18.219–22:20.339
enapple chose gemini to integrate
apple chose gemini to integrate
22:20.339–22:22.099
enif it's smart enough
if it's smart enough
22:22.099–22:24.339
enit might just make their iphone a huge hit
it might just make their iphone a huge hit
22:24.339–22:28.319
enthese things could happen on the application side in the future
these things could happen on the application side in the future
22:28.319–22:31.539
enin the application sector
in the application sector
22:31.539–22:34.419
enthose who can seize the initiative and hit that sweet spot
those who can seize the initiative and hit that sweet spot
22:34.419–22:36.139
enwill likely be the first wave winners
will likely be the first wave winners
22:36.139–22:39.979
zh但便是企業的,在一旦會發生,
但便是企業的,在一旦會發生,
22:39.979–22:41.979
zh在一旦會發生的一流程。
在一旦會發生的一流程。
22:43.179–22:45.859
zh這就是為了我們已經看著當中的AI模式
這就是為了我們已經看著當中的AI模式
22:45.859–22:46.979
zh是完全適合的。
是完全適合的。
22:50.659–22:52.019
zh像是在手機上,
像是在手機上,
22:52.019–22:53.699
zh它們已經在8-10年來
它們已經在8-10年來
22:53.699–22:55.219
zh讓它們發生了一些
讓它們發生了一些
22:55.219–22:55.979
zh手機,
手機,
22:55.979–22:56.699
zh手機,
手機,
22:56.699–22:56.859
zh手機,
手機,
22:56.859–22:58.219
zh最後手機。
最後手機。
23:01.419–23:01.899
zh現在,
現在,
23:01.899–23:03.979
zh它們都已經在三年。
它們都已經在三年。
23:03.979–23:12.879
zh一切都要做到,每公司都要推出了,就像每公司都要推出了,像每公司都要推出了,像 Inventech的手法上,在那天
一切都要做到,每公司都要推出了,就像每公司都要推出了,像每公司都要推出了,像 Inventech的手法上,在那天
23:12.879–23:16.759
enBenQ smartphones and the rest all died off
BenQ smartphones and the rest all died off
23:16.759–23:17.379
enWhy?
Why?
23:17.839–23:19.259
enPeople won't buy from everyone
People won't buy from everyone
23:19.259–23:20.899
enIt will be the same in the future
It will be the same in the future
23:20.899–23:23.179
enThere might be a hundred brands of robots
There might be a hundred brands of robots
23:23.179–23:25.139
enThat's when a real bubble forms
That's when a real bubble forms
23:25.139–23:26.679
enYou'll need all sorts of parts
You'll need all sorts of parts
23:26.679–23:28.259
enAnd there will be massive shortages
And there will be massive shortages
23:28.259–23:29.219
enBut in the end
But in the end
23:29.219–23:31.739
enThere will be a pile of robots that no one buys
There will be a pile of robots that no one buys
23:31.739–23:34.519
enThat's exactly where the real bubble lies
That's exactly where the real bubble lies
23:34.519–23:38.119
enSo, before those applications truly emerge
So, before those applications truly emerge
23:38.119–23:40.839
enThat's why many say we are in the early stages of AI
That's why many say we are in the early stages of AI
23:40.839–23:42.379
enLike the top of the third inning
Like the top of the third inning
23:42.879–23:45.199
zh那是他們所說的,OK?
那是他們所說的,OK?
23:45.199–23:46.599
zh很多 pessimists might think
很多 pessimists might think
23:46.599–23:52.139
enyou're lying to me again, trying to trick me every day, saying it's only the third inning when stock prices have already risen like this.
you're lying to me again, trying to trick me every day, saying it's only the third inning when stock prices have already risen like this.
23:52.259–23:55.519
enThey think that because they see the massive scale of the applications coming later.
They think that because they see the massive scale of the applications coming later.
23:55.719–23:56.659
enYes, exactly.
Yes, exactly.
23:56.919–24:00.439
enSo we see that at this point in time, it is indeed still the top of the third inning.
So we see that at this point in time, it is indeed still the top of the third inning.
24:00.599–24:03.179
enFor those of us selling shovels,there's no need to worry.
For those of us selling shovels,there's no need to worry.
24:03.359–24:07.839
enOnce the leverage is removed and funds are rebalanced,buying the dip is a very good entry point.
Once the leverage is removed and funds are rebalanced,buying the dip is a very good entry point.
24:07.839–24:12.499
enThank you very much to the CEO for joining us for this interview,and thank you to our viewers for watching.
Thank you very much to the CEO for joining us for this interview,and thank you to our viewers for watching.
24:12.699–24:13.819
enTaiwan's Grand Era
Taiwan's Grand Era
24:16.599–24:32.888
zh請訂閱 訂閱 轉發 打賞支援明鏡與點點欄目
請訂閱 訂閱 轉發 打賞支援明鏡與點點欄目
影片筆記：全球AI股價震盪 基本面沒有鬆動只是修正? AI算力激記憶體需求 下個超級周期將爆發?Kimi K3橫空出世 掀美中AI爭霸戰?｜張振驊ft.鈞弘資本創辦人 沈萬鈞｜台灣大時代｜2026.07.25｜
一句話總結
影片探討AI產業當前處於「第三局上半」的早期階段，指出雖然應用端（如人形機器人）尚未爆發導致泡沫疑慮，但半導體基礎設施（特別是記憶體）因供需結構性短缺而具備長期支撐；同時分析中美AI模型競爭現狀，強調應用層成熟後將引發硬體整合與潛在泡沫，建議投資者關注資金重新平衡後的逢低買入機會。
核心重點
- 市場修正與資源扭曲： SOX指數修正及資金流向「Magnificent Seven」反映資源極度扭曲。目前AI應用端（OpenAI、人形機器人）仍在虧損，但賣「鏟子」的半導體廠商利潤豐厚，導致股價波動與偏離。
- 記憶體供需結構性短缺： GPU/CPU與記憶體為綁定銷售的固定單位（Fixed Unit）。隨著TSMC與Nvidia/AMD產能擴張，未來四年記憶體需求將至少為目前的4倍。目前短缺為結構性問題，供給無法即時跟上，看好記憶體長期短缺帶來的股價支撐。
- 中美AI模型競爭現狀：
- 中國發布的 Quen 2.5 在評分上接近美國 Anthropic 的 Claude 3.5，且為開源。中國模型擅長針對特定指標優化（刷題）。
- 美國大模型（Gemini、OpenAI）雖強大但可能因過度複雜導致效率下降。
- DeepSeq 等小型模型顯示出「不需極大算力即可產生巨大影響」的趨勢。
- 商業價值與效能的錯位： 模型效能優異不代表企業願意付費。OpenAI 的價值在於企業客戶的付費意願。CLOD（疑為Claude）的基準測試分數可能不顯著高於 GPT-4，但能從企業客戶賺取更多利潤。
- 應用層爆發與泡沫預警： 真正的爆發點在於應用層（AI機器人、AI寵物、自動駕駛汽車）。當應用層成熟時，可能出現類似手機產業的泡沫與整合現象（品牌消失、零件短缺、堆積無人購買的產品）。
- 投資階段論： 目前處於AI早期階段，比喻為棒球比賽的「第三局上半」（top of the third inning）。對於賣鏟子的人（基礎設施/模型提供商）無需擔心；當槓桿移除且資金重新平衡時，逢低買入（buying the dip）是好的切入點。
詳細大綱
1. 市場回顧與 AI 泡沫辯證
- SOX指數修正與資金流向： 針對 Philadelphia Semiconductor Index (SOX) 指數修正及資金流向 Magnificent Seven (M7)，指出當前市場存在資源極度扭曲。
- 應用端 vs. 基礎設施：
- AI 應用端（如 OpenAI、人形機器人）仍在虧損狀態。
- 賣「鏟子」的半導體廠商（Nvidia, AMD, TSMC等）利潤豐厚。
- 這種利潤與虧損的對比導致股價波動與極端偏離。
2. 半導體與記憶體產業基本面
- 產能與營收對比： 討論 TSMC、Ase（封測廠）、記憶體廠商的營收與資本支出（csps，疑為CapEx）對比。
- 綁定銷售邏輯： GPU/CPU 與記憶體是綁定銷售的固定單位（Fixed Unit）。
- 需求預測與供給瓶頸：
- 隨著 TSMC 與 Nvidia/AMD 產能擴張，未來四年記憶體需求將至少為目前的 4 倍。
- 目前記憶體短缺是結構性問題，供給無法即時跟上。
- 看好記憶體長期短缺帶來的股價支撐。
- 廠商動態： 提及韓國記憶體廠商（Samsung）與美國政府的關係及擴產意願。
3. AI 模型競爭與技術趨勢
- 中美模型對比：
- 中國發布的 Quen 2.5 在評分上接近美國 Anthropic 的 Claude 3.5，且為開源。
- 中國模型擅長針對特定指標進行優化（刷題）。
- 美國大模型（Gemini、OpenAI）雖強大但可能因過度複雜導致效率下降。
- 小型模型趨勢： DeepSeq 等小型模型顯示出「不需極大算力即可產生巨大影響」的趨勢。
- Kimi K3： 提及 Kimi-K3 橫空出世，掀美中AI爭霸戰（具體細節需查證）。
4. 商業模式與應用層爆發
- 效能與付費意願的錯位：
- Quen 擁有大量運算能力且效能優異，但需考慮「空氣力」（疑點）與成本。
- 效能好不代表企業願意付費；OpenAI 的價值在於企業客戶的付費意願。
- CLOD（疑為Claude）的基準測試分數可能不顯著高於 GPT-4，但能從企業客戶賺取更多利潤。
- CLOD 3 可能不會傷害 Anthropic，但可能衝擊 OpenAI。
- OpenAI 的競爭壓力：
- 面臨來自 CLOD 等競爭對手的壓力，特別是結合 Oracle 和 Cloudflare 背景的情況。
- OpenAI 需要找到商品化路徑，否則若無美國支持，公司可能面臨問題。
- 講者認為 AI 不會流動（疑點），但 OpenAI 在 AI 和 Oracle 中的角色關鍵。
- 應用層爆發點：
- 當 AI 大腦足夠聰明時，機器人、AI 寵物、自動駕駛汽車將出現。
- 應用層整合類似 Apple 選擇 Gemini 整合進 iPhone 的模式。
- 應用層是未來第一波贏家的關鍵，需抓住主動權並擊中甜蜜點。
- 產業泡沫與歷史借鏡：
- 手機產業歷史：Inventech、BenQ 等品牌消失，因為消費者不會購買所有品牌。
- 未來機器人產業可能出現上百個品牌，導致真實泡沫形成與零件短缺。
- 最終結果可能是堆積無人購買的機器人。
5. 市場階段論與投資建議
- 當前階段： 目前處於 AI 早期階段，被比喻為「第三局上半」（top of the third inning）。
- 觀點對比： 悲觀者認為股價已漲高，但講者認為應用規模龐大，目前仍屬早期。
- 投資策略：
- 對於賣鏟子的人（基礎設施提供商）無需擔心。
- 當槓桿移除且資金重新平衡時，逢低買入（buying the dip）是好的切入點。
工具 / 模型 / 名詞整理
- 指數/市場：
- Philadelphia Semiconductor Index (SOX)
- Magnificent Seven (M7)
- NASDAQ
- Small Cap Stocks
- 公司/企業：
- Google (Alphabet)
- Tesla
- Junong Capital (君龍資本/鈞弘資本)
- OpenAI
- TSMC (台積電)
- Ase (安世/封測廠)
- Samsung (三星)
- SK Hynix (海力士)
- YMTC (長江存儲)
- CXMT (長鑫存儲)
- Nvidia
- AMD
- Intel
- Anthropic
- DeepSeq (疑點：需查證，可能為 DeepSeek 或其他模型名稱)
- Oracle
- Cloudflare
- Apple
- Inventech (品牌名稱，疑點)
- BenQ (品牌名稱)
- AI 模型/產品：
- Quen 2.5 (疑點：聽寫可能為 Qwen 2.5)
- Claude 3.5 (Anthropic)
- CLOD / CLOD 3 (疑點：聽似 Claude)
- GPT-4
- Gemini (Google)
- Kimi-K3 (提及於修正段落前)
- Humanoid Robots (人形機器人)
- GPUs, CPUs
- HVDC Power Supply (伺服器高壓直流電源)
- Passive Components (零組件/無源元件)
- PCBs (印刷電路板)
- iPhone
操作流程整理
- 產業觀察流程：
- 監控 SOX 指數修正與資金流向 Magnificent Seven。
- 分析 AI 應用端（OpenAI、人形機器人）的虧損狀況與半導體賣鏟者（Nvidia, TSMC）的利潤對比。
- 評估 GPU/CPU 產能擴張對記憶體需求的影響（未來四年4倍增長預測）。
- 比較中美 AI 模型（Quen 2.5 vs. Claude 3.5）的效能與開源策略。
- 關注小型模型（DeepSeq）帶來的算力效率啟示。
- 觀察應用層（機器人、自動駕駛）的成熟度與整合趨勢（如 Apple 整合 Gemini）。
- 判斷市場階段（第三局上半），並在資金重新平衡時尋找逢低買入機會。
值得注意的限制或風險
- 應用端虧損： AI 應用端（如人形機器人、OpenAI）目前仍在虧損，若應用爆發延遲，可能影響整體產業節奏。
- 記憶體供給瓶頸： 雖然看好記憶體短缺帶來的股價支撐，但供給無法即時跟上需求（4倍增長）可能導致價格波動或延遲產品發布。
- 模型效率問題： 美國大模型（Gemini、OpenAI）可能因過度複雜導致效率下降，而中國模型擅長針對指標優化（刷題），兩者競爭格局存在不確定性。
- 應用層泡沫風險： 未來機器人產業可能出現上百個品牌，導致真實泡沫形成、零件短缺，最終堆積無人購買的機器人。
- OpenAI 的競爭壓力： 面臨來自 CLOD 等競爭對手的壓力，特別是結合 Oracle 和 Cloudflare 背景的情況，若無法找到商品化路徑且無美國支持，公司可能面臨問題。
- 市場資源扭曲： 當前市場存在資源極度扭曲，股價波動與偏離可能反映非基本面因素。
逐字稿辨識疑點
- Quen 2.5： 逐字稿為 "Quen"，通常指代阿里雲的 "Qwen" 系列，但依規則標為疑點。
- DeepSeq： 逐字稿為 "DeepSeq"，在 AI 領域較常見為 "DeepSeek"，但依規則標為疑點。
- KIMI-K3： 逐字稿提及 "KIMI-K3"，Kimi 為月之暗面產品，但 "K3" 具體型號需查證。
- SOX / sox： 指代費城半導體指數，大小寫混用。
- csps： 逐字稿為 "csps"，結合上下文 "spend 700 billion dollars"，疑為 "CapEx" (資本支出) 或 "Cloud Service Providers" 之聽寫錯誤，但依規則標為疑點。
- envious gross margins： 逐字稿為 "envious"，結合 TSMC earnings call 語境，疑為 "enviable" (令人羨慕的) 之聽寫錯誤，但依規則標為疑點。
- sk hinnick： 逐字稿為 "sk hinnick"，應指 "SK Hynix"，但依規則標為疑點。
- Anthix： 逐字稿為 "Anthix"，應指 "Anthropic"，但依規則標為疑點。
- 4 years from now / 四年後的屁股： 逐字稿中出現 "四年來的屁股"，語意不通，疑為 "四年後的..." 之聽寫錯誤，但依規則標為疑點。
- 負敗的貨幫助： 多次出現 "負敗的貨幫助"，語意不明，疑為 "失敗的貨物/幫助" 或特定術語聽寫錯誤，但依規則標為疑點。
- 肥料肥料...： 大段重複 "肥料"，疑為聽寫錯誤或語音識別故障，無法確定原意。
- 1%的人持有99%的貨幣： 語意不明，可能為 "算力"、"數據" 或 "貨幣" 之聽寫錯誤，但依規則標為疑點。
- 6月31日： 日期錯誤，6月只有30天，但依規則標為疑點。
- 空氣力： 逐字稿中多次出現「大量的空氣力」、「很快的空氣力」、「它的價值是一樣的價值...但安thropic是超費的...它是非常好 的開放...因為人們說它是費的」。此處語意不通，疑似聽寫錯誤，可能原意為「算力」、「成本」或「效率」相關詞彙，但無法確定具體所指。
- 安thropic： 應為 Anthropic，但逐字稿拼寫為「安thropic」。
- CLOD： 逐字稿中多次出現「CLOD's benchmark scores」、「CLOD 3」，根據上下文對比 GPT-4 與 Anthropic，疑點為「Claude」的聽寫錯誤，但依規則標為疑點。
- 射擊： 逐字稿提到「它會直接射擊 OpenAI」，疑點為「衝擊」或「打擊」的聽寫錯誤。
- 借貨倫、購乎率： 逐字稿提到「像它的借貨倫、購乎率、購乎率、或是在購買購買貨」，語意不明，疑點為財務或商業術語的聽寫錯誤（如「借貸率」、「購買率」等）。
- 流動： 逐字稿提到「我們不相信AI會流動」，疑點為「流動」在此處語意不明，可能為「流動性」或其他詞彙的誤聽。
- 第三局上半： 逐字稿提到「top of the third inning」，中文譯為「第三局上半」，此為棒球術語，用於比喻市場階段。
- 賣鏟子的人： 逐字稿提到「For those of us selling shovels」，中文譯為「賣鏟子的人」，此為投資隱喻（淘金熱中賣鏟子的人）。
- 明鏡與點點欄目： 逐字稿結尾提到「打賞支援明鏡與點點欄目」，疑點為節目或頻道名稱。
可延伸追問
- Quen 2.5 與 Qwen 2.5 的具體差異： 若 Quen 為 Qwen 之誤，其開源策略對全球 AI 生態系的具體影響為何？
- DeepSeq 的技術架構： DeepSeq 作為小型模型，其技術架構如何實現「不需極大算力即可產生巨大影響」？
- 記憶體供應鏈細節： 未來四年記憶體需求增長 4 倍的具體驅動因素為何？三星、海力士、長江存儲、長鑫存儲的擴產進度與競爭格局如何？
- OpenAI 與 Oracle/Cloudflare 的合作細節： OpenAI 與 Oracle 和 Cloudflare 的結合具體如何影響其商業模式與競爭力？
- 應用層泡沫的具體指標： 如何判斷 AI 機器人或自動駕駛汽車產業是否已進入泡沫階段？有哪些先行指標？
- 資金重新平衡的時機： 目前市場處於「第三局上半」，資金重新平衡的具體觸發條件為何？投資者應關注哪些信號？
尚未產生學習筆記
請在 Telegram 指令最後加上「學習」，例如：videonote 網址 英文 雙語 學習