AIを導入したのに生産性が落ちる…今、企業で起きているヤバい現象とは【マイキー佐野 投資学】 — Transcript
Full transcript
- 0:00Usually, it should go down before you
- 0:01even ride it. That's exactly why. So,
- 0:04anyone who says their productivity went
- 0:06up is probably not using AI correctly.
- 0:11Hi, this is Yuki. I'm shooting for
- 0:13YouTube again today.
- 0:14Let's get started. Let's do it.
- 0:15This time, even though we’ve already
- 0:17talked a lot about semiconductors and
- 0:20AI—I mean, you gave me dozens of
- 0:22companies before—Yes.
- 0:24I’m not asking for new company names,
- 0:26but I want to shift perspectives and
- 0:28have you tell me the points in the AI
- 0:30industry that we should keep an eye on.
- 0:33Distortion.
- 0:34Distortion?
- 0:35In a word, distortion.
- 0:37Yes.
- 0:38So, to explain what that means, things
- 0:40called intelligence—like marginal
- 0:41costs or token prices—These are
- 0:43dropping like crazy, right? Conversely,
- 0:45there’s the physical infrastructure
- 0:47that supports that foundation, right?
- 0:48Data centers, semiconductors, and so on
- 0:50—the input costs are skyrocketing,
- 0:52creating an inflationary state. So, you
- 0:55understand what’s happening: buying
- 0:57expensive equipment to create cheap
- 0:59products.
- 0:59Expensive equipment.
- 1:00Introducing it.
- 1:01Introducing it.
- 1:02And making cheap products; that’s
- 1:04what it is.
- 1:05And that’s the token price. Even
- 1:06though costs are rising more and more,
- 1:09the cost per token is starting to drop
- 1:11significantly. In traditional economic
- 1:13theory, technological evolution lowers
- 1:15overall production costs. And
- 1:17ultimately, production costs are
- 1:19reduced, leading to lower prices for
- 1:21products or services. But it's common
- 1:24for production costs to fall below the
- 1:26selling price.
- 1:27As production costs fall below the
- 1:29selling price, profits just keep
- 1:31growing, right? Right, that’s the
- 1:33fundamental, traditional theory of
- 1:35economics. So, as I said, while
- 1:37technological evolution lowers overall
- 1:40production costs, the prices of
- 1:42products and services eventually fall.
- 1:45But what’s happening now is the
- 1:46opposite: the selling price is dropping
- 1:49lower than production costs, or to put
- 1:51it another way, production costs are
- 1:53rising while prices fall, creating a
- 1:55structure that is squeezing profits,
- 1:57which is the current reality. To
- 2:02explain this more academically, when we
- 2:04talk about the shrinking profit margins
- 2:06in the AI ecosystem, it’s easiest to
- 2:08start by dividing it into upstream and
- 2:11downstream. Upstream refers to
- 2:13infrastructure and semiconductor
- 2:15companies. Downstream refers to model
- 2:18development and the provision of APIs.
- 2:20Cutting-edge semiconductors,
- 2:21Yeah.
- 2:22The HBM market and physical production
- 2:24capacity are reaching their limits. And
- 2:27another factor is that the technical
- 2:29intensity has created a structurally
- 2:31undersupplied state. In fact, it seems
- 2:33there’s a shortage of memory as well.
- 2:36NVIDIA itself, as I mentioned, has
- 2:38already hit 290 billion just in orders,
- 2:41but what are memory firms like SK
- 2:43projecting for the future? They’ve
- 2:47signed long-term contracts with data
- 2:49center clients. In other words, 70%of
- 2:52the maximum supply is already secured
- 2:54elsewhere, so the outlook is that the
- 2:56memory shortage will continue. It’s
- 2:58going to be tight. Samsung and other
- 3:01foundries are raising their prices by
- 3:03up to 15%, and NVIDIA is doing the same
- 3:06with a 15%hike. This was in the news
- 3:08recently, too. So, as you can see, the
- 3:10upstream sector overwhelmingly holds
- 3:13the pricing power. That is the current
- 3:15situation. NVIDIA’s recent earnings
- 3:17were quite bullish, though. Actually,
- 3:19they need 110 billion dollars—
- 3:21including memory and various other
- 3:24parts—to utilize their future GPUs.
- 3:26This figure went from 100 billion to
- 3:29290 billion in just one year. Meanwhile
- 3:31, the downstream side—the ones being
- 3:34supplied—are being charged these
- 3:36incredibly high prices by the suppliers
- 3:38. For instance, look at China's
- 3:40DeepSeek going open source; it’s
- 3:42actually causing prices to drop, right?
- 3:44This is, well, part of China's strategy
- 3:46. I mean, China itself is being heavily
- 3:49restricted by the U.S. in terms of
- 3:51capital investment. While companies
- 3:53like Huawei and SMIC are emerging, by
- 3:55going open source, they’re preventing
- 3:57U.S. firms from collecting subscription
- 3:59fees. And have you heard of the new one
- 4:01that came out recently, Qwen-Max?
- 4:02Another new one just appeared. Well, a
- 4:04lot of them are popping up. Yes.
- 4:06Yes. This is causing further price
- 4:10destruction, and downstream, companies
- 4:12have no pricing power because customers
- 4:14already hold the leverage.
- 4:16They can just choose the cheaper option
- 4:18. I mean, I'm using a 3,000 yen service
- 4:20myself.
- 4:21So, the customer has all the bargaining
- 4:23power; if they want to find a service,
- 4:25they can find countless options for any
- 4:27price point.
- 4:28That's why buyers are in a position to
- 4:29choose at their own convenience.
- 4:31It’s that kind of distortion, isn't
- 4:32it?
- 4:32Exactly. Yes. Yes. Yes.
- 4:35So, we’re seeing a state of both
- 4:36inflation and deflation.
- 4:39This is happening within a single
- 4:40industry, meaning it's getting harder
- 4:42to pull a profit, which is why
- 4:43companies are being grilled by
- 4:45investors and the market. In this
- 4:49situation, it’s becoming very
- 4:50difficult for development firms and
- 4:52infrastructure providers to justify
- 4:54their position, and the reality is that
- 4:55their profit structures are being
- 4:57squeezed. That’s why you call it a
- 4:59distortion.
- 5:01But there are both bulls and bears when
- 5:02it comes to this current AI trend. For
- 5:06example, the bulls argue that as the
- 5:08price of tokens drops, demand will
- 5:10explode, ultimately increasing total
- 5:12revenue. We covered this on YouTube
- 5:14before. Jevons paradox.
- 5:15Ah, yes, we did.
- 5:17The steam engine one, right?
- 5:18Yes. Exactly. We talked about that, and
- 5:21it's essentially an argument for
- 5:22volume-driven growth. I mean, these
- 5:24companies are paying massive amounts of
- 5:26capital to hyperscalers and for AI,
- 5:27right? It's seen as strong leadership,
- 5:29but when you actually examine corporate
- 5:31financial data at the micro level,
- 5:33there’s a significant disconnect
- 5:35between investment and profitability.
- 5:40There are these AI advisory firms like
- 5:42Mackey Research and Spek Research. They
- 5:46conducted a quantitative analysis of
- 5:48919 quarterly earnings reports from the
- 5:51top 60 major US financial institutions
- 5:54over the past three years. And what
- 5:57actually happened? In about 80%of those
- 6:01earnings calls—four out of five—AI
- 6:04was mentioned. However, more than 50%of
- 6:07the discussion was about rising
- 6:09implementation costs and increased
- 6:12CAPEX. So, in the entire study, only
- 6:15two companies were able to quantify and
- 6:18clearly state the actual monetary
- 6:20returns generated by AI implementation.
- 6:23Basically, there's no profit coming in.
- 6:25The reality is that for executives,
- 6:27marketing, and PR, the only thing being
- 6:30pushed is a story about dramatic
- 6:32productivity gains from AI.
- 6:34Yes. Yes. Yes.
- 6:36So, when you look at the gap between
- 6:38the actual financial situation and the
- 6:40earnings calls, the conclusion is that
- 6:42there is almost no quantitative proof
- 6:44of profitability so far. Actually, this
- 6:49is something I predicted on YouTube two
- 6:51or three years ago.
- 6:53I'll do my best to find it. Yes.
- 6:55Please do. It’s something called the
- 6:58Solow Paradox.
- 6:59Oh, the paradox, yes.
- 7:03It was proposed in 1987 by Robert Solow
- 7:06, a Nobel laureate in economics, who
- 7:08pointed out the economic mystery that "
- 7:10computers are everywhere except in the
- 7:12productivity statistics." That is the
- 7:16Solow Paradox. This PC and internet
- 7:19revolution, it actually takes ten years
- 7:21for the productivity to show up in the
- 7:23numbers. I've talked about this on
- 7:26YouTube many times. This is the content
- 7:27I mentioned earlier. The expectation is
- 7:30that by introducing computers, business
- 7:32efficiency will rise and social
- 7:34productivity will grow rapidly. But the
- 7:38reality is, despite huge investments,
- 7:40national economic growth and corporate
- 7:42productivity remain mostly flat. This
- 7:45is the current situation. The
- 7:48background to this is that even if new
- 7:50technology spreads, it just becomes a
- 7:52cost center unless workflows and
- 7:54organizational structures evolve. I
- 7:57explained this on YouTube before, right
- 7:59?
- 8:00In fact, the situation where AI
- 8:02adoption actually increases labor hours
- 8:04in Silicon Valley is exactly that
- 8:06background. What's happening with AI
- 8:09investment now is exactly this Solow
- 8:11Paradox, and what we're actually seeing
- 8:13is FOMO. There's the fear of missing
- 8:16out, the anxiety that if you fall
- 8:18behind others, you might eventually
- 8:20lose or be unable to keep up. So, there
- 8:24are many investors who think people are
- 8:26investing without calculating clear
- 8:28returns or cost-effectiveness. It's
- 8:35just a scramble; it's the idea that AI
- 8:37costs are being wasted as defensive
- 8:39spending to steal customers or market
- 8:41share from rivals rather than
- 8:42increasing overall wealth, which is the
- 8:44IT productivity paradox. There is
- 8:53something called the IT productivity
- 8:55paradox, where simply introducing
- 8:56convenient AI or IT without
- 8:58fundamentally changing workflows or
- 9:00organizations leads to failure. That is
- 9:05what they call the IT productivity
- 9:07paradox. This is exactly what is
- 9:08happening right now. So, does that mean
- 9:10AI isn't improving productivity? Well,
- 9:13I’ll explain why it’s not a waste.
- 9:17This is what some call the "intangible
- 9:19asset time lag," referring to what
- 9:21happens during the initial adoption
- 9:23phase. Everyone is using AI, but more
- 9:27importantly, organizations need to
- 9:29fundamentally change how they operate.
- 9:33It’s a gap or time lag created by how
- 9:35much intangible capital was built up
- 9:36versus how much is actually integrated
- 9:38into the organization. This is
- 9:41something that happens naturally. It
- 9:43really does. This is exactly what
- 9:45happened during the IT boom, too. Even
- 9:47if you adopt IT, you won't maximize its
- 9:49full potential unless you fundamentally
- 9:51restructure your operations and
- 9:52organization. The same thing is
- 9:56happening with AI now, which brings us
- 9:58to the "Productivity J-Curve."
- 10:01There’s a theory called the
- 10:02Productivity J-Curve, which is a key
- 10:04concept in economic productivity
- 10:06analysis. Here’s the interesting part
- 10:08. It explains a dynamic process where
- 10:12Total Factor Productivity (TFP) and
- 10:14corporate profitability initially
- 10:16decline before eventually rising. As I
- 10:20mentioned, in Silicon Valley, companies
- 10:23saw their productivity drop after
- 10:25adopting AI. In fact, their labor hours
- 10:27actually increased. In Japanese
- 10:30companies, people mistakenly think AI
- 10:32just makes their tasks easier, but in
- 10:35reality, using AI creates more
- 10:36challenges, hypotheses, and testing,
- 10:38which ultimately increases labor hours
- 10:41—this is exactly the Productivity
- 10:43J-Curve. This concept was proposed by
- 10:47three American economists. This is a
- 10:49theory developed by a Stanford
- 10:51professor, Danny Rock from the Wharton
- 10:52School at the University of
- 10:54Pennsylvania, and Chayberson from the
- 10:56University of Chicago. As I explained
- 10:58regarding the difference between
- 11:00Japanese companies and AI firms,
- 11:01startup companies are seeing their
- 11:03volume increase, so if you'd like to
- 11:05know more, please check out my YouTube
- 11:06channel. Yes. I explain it there.
- 11:10During this early adoption phase,
- 11:12companies need to allocate capital and
- 11:13labor not only to direct hardware and
- 11:15software costs but also to intangible
- 11:17assets that are difficult to measure.
- 11:21That’s why it takes time to build the
- 11:23foundation. So, when it comes to
- 11:25redesigning business processes to adapt
- 11:27operations to AI-supported workflows,
- 11:29structural changes take time, which is
- 11:31why there’s a period where
- 11:33productivity drops. Regarding machine
- 11:37learning, there is the human capital
- 11:39aspect; this needs to be retrained, and
- 11:41it takes time for employees to actually
- 11:43acquire the skills to use AI tools
- 11:45effectively. Also, with data management
- 11:49infrastructure and system updates, you
- 11:51need to build data pipelines and design
- 11:53organizational structures, so a
- 11:55temporary drop is only natural. This is
- 11:58the productivity curve. Conversely,
- 12:00anyone saying their productivity went
- 12:01up is probably not using AI correctly.
- 12:03That's right. Yes. Yes. Building these
- 12:06things requires labor hours and
- 12:08resources. These are recorded as
- 12:11expenses.
- 12:12But you are essentially building
- 12:14intangible assets. These are not
- 12:16recorded on the balance sheet. This is
- 12:18a situation that happens very often.
- 12:20The point is, this is standard. And
- 12:22there is also empirical evidence of
- 12:24this J-curve based on microdata. To put
- 12:27it simply, there is a government agency
- 12:29under the Department of Commerce called
- 12:31the Census Bureau that recently
- 12:32analyzed statistical data on the U.S.
- 12:34population and economy. This data was
- 12:37actually used to conduct industrial AI
- 12:39demonstration experiments in 2025. And
- 12:42what was the result? It showed that
- 12:44initial productivity actually dropped
- 12:47by 44%. So, it turned out that the
- 12:49sites where AI was introduced
- 12:50experienced that deep "J-curve" dip
- 12:52just as the theory predicted. There is
- 12:55plenty of proven data on how to change
- 12:57conventional work rules, or on the
- 12:59confusion in the field, such as time
- 13:01spent on system adjustments or
- 13:02inventory increasing because processes
- 13:04weren't syncing up. So, productivity is
- 13:08actually decreasing, and this is
- 13:09something confirmed by studies at
- 13:11Stanford and the University of Toronto;
- 13:13introducing AI can actually lower
- 13:15productivity. If used correctly, most
- 13:17of the time—though not in every case.
- 13:19But Japanese people tend to be a bit
- 13:21naive, thinking, "My company is
- 13:22different." They claim, "Our
- 13:24productivity is going up," but they are
- 13:25likely wrong in most cases and probably
- 13:27need to change their organizational
- 13:29structure, or—
- 13:29Wait, we weren't talking about that,
- 13:31were we? Please don't go off on a
- 13:32tangent. Sorry for the trash talk. Yeah
- 13:34, yeah, yeah, there are a lot of people
- 13:35like that.
- 13:36Well, well, well, well.
- 13:37Research was published by universities
- 13:40like Stanford, Oxford, and MIT on
- 13:41something called the "Cognitive
- 13:43Offloading Model 2026." What this means
- 13:45is that by offloading thinking tasks
- 13:47and decision-making to AI, labor time
- 13:49actually decreases. But I always say
- 13:51this, and I use AI extensively. Labor
- 13:54time actually increases. So, labor time
- 13:56goes up. This means we're in a state of
- 13:58constantly delegating cognition to AI.
- 14:01And as a result, the expertise and
- 14:02field intuition that humans have
- 14:04cultivated over many years begin to
- 14:06fade. That's why AI on its own doesn't
- 14:08create value; it doesn't show its true
- 14:10potential unless combined with human
- 14:12judgment. In fact, human skills
- 14:14themselves begin to rust. So, the
- 14:16conclusion is that if you can't fully
- 14:18utilize the potential of AI, the time
- 14:20it takes to climb out of the bottom of
- 14:22the J-curve or see a sharp rise in
- 14:24productivity will be delayed. But those
- 14:26with an even more naive outlook say
- 14:28that doesn't concern them, because now
- 14:30they can easily create reports. For
- 14:32those whose working hours dropped from
- 14:34eight to four because reports became
- 14:36easy to make, they won't even fall into
- 14:38a J-curve to begin with; they'll just
- 14:40end up in a state of blissful ignorance
- 14:42.
- 14:42I mean, that really is just losing to
- 14:44AI. That is,
- 14:45Only people who aren't thinking clearly
- 14:46would say something like that.
- 14:47It’s true that a temporary 44%drop in
- 14:50productivity due to operational
- 14:51confusion on the ground is unavoidable.
- 14:54However, the fact that working hours
- 14:55are actually decreasing just means
- 14:57people are getting better at slacking
- 14:59off, which means we are facing a
- 15:01degradation in judgment standards. To
- 15:05summarize this, at the micro level,
- 15:07investment returns are becoming delayed
- 15:09, while upstream, massive investment
- 15:12continues, and they hold bargaining
- 15:14power, effectively influencing
- 15:15financial and credit markets, which
- 15:17explains the current state. In the
- 15:22stock market, there is an AI frenzy,
- 15:24but in contrast, bonds and credit
- 15:26default swaps (CDS) are rising, showing
- 15:29growing alarm over debt financing for
- 15:31AI infrastructure. With news of
- 15:35Broadcom's chip development and
- 15:37negotiations for 60 billion in
- 15:38infrastructure funding, leveraged
- 15:40financing has increased, causing CDS
- 15:43credit spreads to widen. As a result,
- 15:45risk premiums have risen. The CDS
- 15:47spread for Broadcom has actually risen
- 15:5080 points since the beginning of the
- 15:52year. The same thing happened with
- 15:54NVIDIA, you see. Before the 500 billion
- 15:56funding was even announced, the CDS
- 15:58spread shot up. NVIDIA itself, that is.
- 16:01So, this is a market analysis report
- 16:03from Goldman Sachs. While EPS estimates
- 16:06have been revised upward, the P/E
- 16:08ratios for AI-related stocks have
- 16:10actually been trending downward
- 16:12recently. They are seeing a downward
- 16:14trend lately. In other words, the
- 16:17market isn't just looking at
- 16:19speculation or raw revenue anymore; the
- 16:21core issue has become who can generate
- 16:23cash flow and when. Goldman Sachs
- 16:27argues that this is the shift currently
- 16:29taking place. It goes back to that
- 16:31J-curve I mentioned earlier.
- 16:33Productivity initially drops in the
- 16:35short term, but it tends to rise
- 16:37afterward. If the right processes are
- 16:39in place, this is exactly what happened
- 16:41during the previous AI and IT eras.
- 16:43With that in mind, we need to consider
- 16:44how to view this current divergence. I
- 16:46think looking at this data will give
- 16:48you a clear perspective. By the way,
- 16:49it’s a fact that the wider this
- 16:51divergence becomes, the harder it is to
- 16:52recover. Yeah.
- 16:53Absolutely.
- 16:54Yes.
- 16:55However, as I mentioned earlier,
- 16:56there's Jevons paradox—if this
- 16:58spreads across the board, the situation
- 17:00could change entirely. Just keep that
- 17:02in mind. Right. Anyway, I hope this was
- 17:04helpful for everyone today. We talked
- 17:06about the AI divergence, and if you
- 17:08want more detailed information, please
- 17:10be sure to read the full report. It's
- 17:11an official publication, so I highly
- 17:13encourage you to get your hands on that
- 17:15report. Also, if you found today’s
- 17:17video interesting, educational, or
- 17:20helpful, please subscribe, like,
- 17:22comment, and share it with others.
- 17:24We talked about NVIDIA’s earnings
- 17:26during today's shoot, right?
- 17:27It’s been uploaded. Yeah.
- 17:29Right. It's totally fine, but after
- 17:31hearing about the outlook through 2028
- 17:33and now this divergence, I honestly
- 17:35don't know how I should be looking at
- 17:37the AI industry anymore.
- 17:39See what I mean?
- 17:39Right, right, right, right, right. You
- 17:41see, there's a next one coming.
- 17:43Wait, there's more?
- 17:44Yes. I’ve already told my business
- 17:46school students to look into this
- 17:48months ago, but honestly,
- 17:50I’d be a bad person if I didn’t
- 17:51talk about it on YouTube.
- 17:53Eventually.
- 17:54Yes.
- 17:54I will talk about it.
- 17:55Oh, I’ll look forward to that.
- 17:56Listening to it, I couldn't help but
- 17:58wonder, what's the point of this?
- 17:59Yes. Well, there’s actually a real
- 18:00point to it.
- 18:01Yeah.
- 18:02No, but I use AI daily, and I do wonder
- 18:05if my money is even helping to cover
- 18:07those massive capital investments. I
- 18:10really do think about things like that.
- 18:12I mean, well, honestly, I really wonder
- 18:14if they can ever recoup the costs. I
- 18:16mean, there are tons of people only
- 18:18paying 3,000 yen, right?
- 18:20Especially with the wealth gap widening
- 18:21like this, and in places like China,
- 18:23it’s cheap and they can use it freely
- 18:25.
- 18:25Yes. The meaning of these 500 billion
- 18:27yen figures is just becoming harder and
- 18:29harder to understand. Well, but that
- 18:30part is, well, it’s interesting to
- 18:32hear about. Yes. So,
- 18:34the question is, what’s the next step
- 18:36? And if you look here, you can see how
- 18:38that industry and the companies
- 18:39themselves can recover their profits.
- 18:41If you ask what this is,
- 18:42Oh, that’s where it leads.
- 18:44Exactly.
- 18:45Ah, another new genre has appeared.
- 18:47Yes. This content,
- 18:49Yes.
- 18:50I was looking forward to seeing when it
- 18:51would come up.
- 18:52I’ve been checking YouTube for almost
- 18:54two years, but nobody is talking about
- 18:55it.
- 18:56Yeah, they don't talk about it.
- 18:57So I thought I’d stop talking about
- 18:59it too.
- 18:59I mean, everyone, I know, so I’m not
- 19:01going to tell you guys.
- 19:03I did talk about it at the business
- 19:04school, though. Yes. Ha.
About this transcript
This page contains the full transcript of AIを導入したのに生産性が落ちる…今、企業で起きているヤバい現象とは【マイキー佐野 投資学】 by マイキーの非道徳な社会学, generated from the public captions YouTube serves with the video. The transcript has 3,095 words across 552 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.