AI正走向严重供给过剩:顶级对冲基金经理 David Orr 的预测
- Orr在6月卖掉了所有AI股票,因为他预计算力供给增速将超过可落地的需求。 风险在于短期内“智能过剩”:资本开支将持续到2028年,而实体应用需要数年才能铺开。如果供给过剩在12–18个月内出现,6月可能就是高点;如果要等4年,AI股票可能还会先涨100%。
- Google、Amazon和Microsoft仍是Orr首选的AI投资标的,尽管他已暂时离场。 他提到一些资本回报率达35%的项目,约一半产能已签约;除非软件客户短期内大批倒闭,否则“它们的下行风险很小”。如果算力稀缺持续10年,这笔投资会格外有吸引力。
- 用AI替代昂贵的人力,并不意味着软件利润能够长久,也不保证算力需求会大幅增长。 Orr预计,5年内常规会计工作将几乎免费;竞争者会复制这些工具,消费者则会享受到成本节省。拥有持久护城河的公司可能变成“印钞机”,竞争激烈的企业则会把改善的利润率让出去。
- 科学突破可能推翻供给过剩论,但Orr目前无法确认需求是否会出现。 将药物实验扩大100倍,需要实验室、受试者和新数据;智能可能因此等待现实世界跟上。纯粹的智力突破则可能推高算力价格:“我真的不知道。”
- Orr认为,相比针对数据中心的政治限制,税收和发债对市场构成的风险更明确。 2030年前后企业税上调的概率若为30–50%,或足以解释标普500指数约4%的跌幅,甚至可能是6–7%。Anthropic若IPO,还会进一步吸收流动性:“又是一记重击。”
- AI威胁的不只是软件公司的护城河,还包括搜索广告,甚至社交媒体。 Orr怀疑,一旦竞争模型足够好,普通搜索还能否支撑Google式的“广告墙”。Meta的平台越来越偏向娱乐,5年内可能沦为大宗商品式的同质化服务;不过他明确表示,这只是偏悲观的情景。
- Orr的应对方式是控制敞口,而不是大举做空。 他的基金过去30天贝塔略为负值,但目标是接近零。AI有真实的看多逻辑,他预计自己“至少有40%的时候会错”:“我就先不参与了。”
1. 持仓过于拥挤,让这轮抛售难以判断
- Orr认为,AI股票从6月起就容易下跌:太多投资者被套在高位,也没有新的保证金买家接盘。“市场只是在等着下跌。”他描述一些公司的杠杆达到6–10倍;日经指数跌5%、日本小盘股却上涨,这更像是被迫平仓,而非基本面被有序重估。
- Wiethe问,这种压力如何传导到Orr的10位投资组合经理。Orr说,其他7位经理的规模都很小,即使判断失误,也不会对基金造成实质影响;最差的结果看起来约为8%。Mike和Rodrigo赚了不少钱。基金整体贝塔过去30天略为负值,但Orr不确定是否该相信这个数字,目标是让贝塔维持在零附近。
- 与2022年的比较有其局限:当时Orr做空的是“非常差”的公司,而AI是“基本面真实存在的事物”,确实有看涨的情景。他暂时不打算抄底,但认为自己可能会买入半导体ETF做一笔短线交易。
2. 供给过剩何时出现,比AI的长期前景更关键
- 在行情大部分上涨阶段,Orr一直看多,直到6月卖光所有AI股票。“智能永远不嫌多”,但其他瓶颈可能让算力暂时过剩。如果一年或一年半后供给过剩,6月可能就是高点;如果要等4年,AI股票可能还会再涨100%。
- 他的依据是:资本开支规模惊人,且将持续到2028年;需求却集中在仍在摸索如何整合AI的软件公司。客户最终可能会说:“我们其实不需要这个”,然后转售算力。他提到一些小型运营商,算力报价从9月的每小时12美元降至约4美元。这可能很重要,但“非常难以判断”:小供应商的产能不能直接与大型合约相提并论。
- Wiethe提到,在拥挤的“做多硬件、做空软件”交易平仓后,IGV与SOXX、SMH一同下跌。Orr不认同背后的基本面相关性:“这些叙事其实说不通。”他另有一个跨越多年的担忧:缺乏强大护城河的软件,无论硬件支出如何,终究都会变得脆弱。
- Orr首选Google、Amazon和Microsoft等超大规模云服务商,但他预期抛售压力将至,因此已离场。他提到一些资本回报率达35%、约一半产能已签约的项目。除非软件客户短期内大批破产,否则下行空间似乎有限;如果算力稀缺持续10年,这些投资的回报将极为可观,甚至可能抵消它们其他业务的疲软。
3. 税收和流动性,比德州催化因素更重要
- Orr认为,企业税上调可能解释大盘走弱。如果民主党在2028年选举中表现出色,税改或许会在2030年前后落地。若其概率为30–50%,标普500指数跌约4%算是合理:“也许应该跌6%或7%,我不知道。”
- 假设政府阻止95%的新项目,现有产能所有者会受益。但Orr怀疑德州选举新闻会带来多大变化:电厂本来就是瓶颈,项目也可以迁往别处。除了税收,他认为政治因素总体影响有限,不过更强硬的反垄断执法可能压低股价。
- Wiethe提出Anthropic可能在年底前IPO。Orr认为这取决于IPO是否成行,并指出G7发债,尤其是债券市场发债,已经在吸收流动性。再来一笔规模巨大的发行,只会增加看空理由。
- Orr认为Google最可能成为反垄断目标,但AI威胁其垄断利润,早在他2023年的信中就已提及。Maps和YouTube各有优势;搜索则面临其他模型的竞争,这些模型无需堆出“广告墙”,也能给出普通的餐厅推荐。他预计搜索业务终会衰退,但不会像有线电视那样每年萎缩约10%。
4. 消除工作所需的算力,可能远低于预期
- Orr的出发点是,很多数字产品并不需要特别高的智能。Windows 7当时已经“非常好”;后续版本说明,有用功能可能很快就会触及上限。同样,他预计常规办公室工作,包括会计,将在5年内变得几乎免费,而不是支撑无止境的算力支出。
- 医学研究可能消耗大量资源,但每个有希望的计算方向,最终仍需要实验和新数据。实验量要扩大100倍,就意味着要在多年间建设实验室。Orr强调:“没人能确定事情会如何发展”;他预计瓶颈在于现实世界的执行,而非智能。
- Wiethe追问基础科学领域的突破,包括固态电池:AI能否重现那些主要靠思考实现的突破?Orr承认,如果智能足够强,确实可能做到,并大幅推高算力价格。Wiethe还提到癌症治疗的叙事,为Anthropic 1.5万亿美元的估值提供支撑。Orr希望这些应用能成真,但他说:“我真的不知道。”
- Wiethe以Progressive Insurance为例,追问谁能拿走劳动力成本节省的收益。Orr的基准判断是,多数办公室岗位会消失。受规模或网络效应保护的企业可以裁员,变成“印钞机”;竞争激烈的企业则会失去官僚体系带来的壁垒,把节省下来的成本让给客户。他认为,这两种结果都未必会消耗足够多的算力,支撑超大规模云服务商的资本开支。
5. 免费终端折射出软件定价面临的威胁
- Wiethe问,垂直领域AI能否按结果而非使用量收费。Orr承认,这种模式“几年内可能看起来很有吸引力”,但预计更强大的AI会在5年内复制简单方案。他的基金已经搭建了一个风险系统和终端:“可能不如专业产品好,但我更喜欢它。”
- Orr设想,用户连接IBKR账户后,就能自动获得投资组合风险分析;也许再过1、2年,几乎不需要任何设置。他预计会免费提供这个产品,用户或许只需自行承担token成本,同时通过产品标志获得营销曝光。Wiethe追问Orr的做法能否推广开来;Orr说,可能还有十几个人会做出类似工具并免费提供。
- Bloomberg过去投入多年清洗数据;Orr认为,AI处理旧年报后,这项工作将面临威胁。他押注,几年内自己会更信任AI生成的数据。面对每位用户每年25,000–35,000美元的传统定价,他设想会有十几个人提供免费替代品。Wiethe则指出,机构间的IB聊天、券商接入和已记录的交易,仍有价值,不只是可互换的数据。
6. 更充裕的供给,无法解决劳动力和平台之争
- Orr希望被替代的人才去改善实体环境,而不是优化广告。他举了一个具体例子:一位合伙人在国家公园旁建酒店,请建筑师让建筑与周边景观相融合,但项目因利率上升而推迟。他的愿景是,打造更多经过用心设计的宜居空间。
- Wiethe质疑转型过程:标准化住房本来就是为了压低成本,被替代的员工可能会拉低工资;职业身份也不会自动转化为手艺人的技能。Orr仍设想,数百万人能找到有趣的工作,生产人们看重的东西,并获得相对于其他选择而言不错的收入。Wiethe称之为“美好的愿景”,但对于大规模失业后的转型路径仍不信服。
- 谈到Meta时,Wiethe区分了用户主动消费社交内容与搜索结果令人沮丧这两种体验。Orr担心平台“每年都变得没那么社交”:“现在它只是媒体。”失去过去那种社交依赖后,平台可能在5年内变得同质化。他也明确承认另一种看涨情景:用户持续使用平台,AI增强的内容越来越有吸引力。
- Wiethe认为,创作者与受众之间的网络效应、无止境滚动的信息流,都是强大壁垒;他也提出,小众服务可能会吸引用户离开。Orr承认社交媒体有网络效应,但认为Spotify的推荐算法比TikTok或Instagram更好;后两者推送的内容容易让人上瘾,用户却未必喜欢。他猜测,AI可能学会一个人的偏好,并做出更好的推荐。
完整逐字稿
AI will be in abundance. We’ll have too much intelligence, and it will take years for the physical world to catch up. You can never have too much intelligence, but in the short term you’ll run into all sorts of other bottlenecks and have too much computing power.
In fact, apart from software development companies, no one in particular buys this capacity. Will software companies really need even more than we already have? It’s only a matter of time before they say, “Oh, we don’t really need this.” Then you’ll start reselling your capacity, and once that starts, the price will drop to a tiny fraction of its original price.
Welcome to the show, “Other People’s Money.” With me today is David Orr, founder of Militia Capital, a hedge fund he has run for several years and which has had fantastic returns: 48% per annum compounded. David, you’ve grown it to over $500 million. Thank you for coming on the show.
Thank you for inviting me.
1. AI Trade Unwinds
We’re in difficult circumstances. We’re recording a week after the US midterm elections, and the market has reacted quite violently to the results. Some of the ongoing moves began even before the election results were announced. I’m curious what you think about the rollback in what was the most popular trade of the year: long-term investments in AI infrastructure, data centers, and semiconductors.
Yes, as you said, it’s been declining to some extent since June. Several of the best traders I know believed June was the peak of market euphoria. I think the market was just looking for an excuse for AI stocks to fall sharply in price.
That limited the potential for growth to some extent, with too many people stuck at these high prices. It was a kind of downward trend. There were no new margin buyers. The market was just waiting to fall. I’ve written about this several times. I expected it.
No one really knows why a certain thing will cause a fall. I don’t think anyone knows why the news about the governor of Texas was probably the most important. Although I don’t think that’s the real reason.
Both in June and this time, there’s been a lot of talk about fund blowups and forced position closures. When a sell-off like this happens, how do you distinguish fundamental reasons from asset liquidation?
These companies operate with leverage of 6 to 10, so some trades just fall through. Strange things happen: the Nikkei fell 5%, and Japanese small-cap companies suddenly rose. This is clearly someone draining the account because of excessive leverage. There are many cases like that, and they cause a chain reaction.
So everyone who was long semiconductors with leverage or short semiconductors effectively got crushed this week. I tried to stay away to keep myself safe. I came out on top, and Mike and Rodrigo made a good trade. They both made good money.
You now have 10 portfolio managers. When you think the market has peaked, do you tell them? How do you make sure actions in one of your divisions don’t cause you problems?
The other seven managers are so small that even if they do something stupid, it won’t affect the fund. But it was nice to see the results from each of them. It’s good to finally see a stress test like this. In fact, none of them had a serious drawdown. The worst result seems to be minus 8%. Considering this move, 8% isn’t that much.
Some of the guys just went flat, so I didn’t see any problems. I try to reassure them by saying, “This is a specific market environment. Try not to worry too much.” Everyone’s risk control was pretty good, so things are going according to plan.
Do you have any idea whether the decline will continue?
I’m not sure. I’m not buying the drawdown yet. I’ll probably buy a semiconductor ETF, at least for the short term.
I was a big optimist about AI for most of the way up, but in June I sold all my AI stocks because I think AI is going to be oversupplied. We’ll have too much intelligence, and it will take years for the physical world to catch up. In the long run, I think it will work. You can never have too much intelligence, but in the short term you’ll run into other bottlenecks and have excess computing power.
Will this happen in, say, a year or a year and a half? If so, June must have been the peak. The market always looks a little ahead. But if it happens in 4 years, perhaps AI shares will rise another 100% before then.
We just lived in a world where everyone thought we lacked computing power. How did we go from a deficit mindset to a surplus mindset so quickly?
I think it was just a matter of positioning, and the price had already adjusted the narrative to that. That’s all that happened.
2. CapEx Drives Glut
I don’t think anything fundamental has suddenly changed. So what fundamental indicators are people pointing to now to say we’re in for a glut? If you were to argue that the surplus would become apparent in a year or two, what would you point to now?
Look at the huge capital expenditure figures for the next 2 years, actually up to 2028. They’re simply colossal. Apart from software companies, no one is actually buying this capacity.
Will software companies really need even more than we already have? I think they’re learning to use it. They’re learning to integrate it, and when it works, what we have now may seem excessive.
I just know that as long as they continue to make such huge capital expenditures, it’s only a matter of time before software companies say, “We don’t really need this.” Then you’ll start reselling your capacity, and once that starts, the price will drop to a tiny fraction.
I saw a few headlines. This was the most interesting, and not enough people were talking about it in October. A few people were saying, “Oh, I have excess capacity.” For example, in September I was offered computing at $12 an hour, but now suddenly no one is offering that price anymore. Now it’s more like $4 an hour.
To me, that’s probably a key indicator of whether there’s going to be a huge crash, but it’s also very ambiguous.
Most of the big contracts, of course, are with these huge companies. When people said, “I have this spare capacity,” do you mean very small data centers, like private data centers or single operators?
Yes, smaller companies. Maybe it isn’t the case for a large company, but for a smaller one it’s a huge hassle, and it’s not worth setting all this up at all. So it’s not a perfect comparison.
3. Software Correlation
What about the software component of this trade? Interestingly, in the first half of the year, people were betting on semiconductors and AI stocks going up, and software going down. Every time semiconductors went up, software went down.
Then people seemed to realize this and started playing both directions because they were supposedly uncorrelated, and software started to show some strength. We see IGV falling along with SOXX and SMH. Do you think this “long AI, short software” correlation has finally broken down, and has this trade become too popular?
Yes, I think they’re just getting too popular. It’s kind of random. These narratives don’t really make any sense, do they?
Fundamentally, in my opinion, betting against most software companies is right if you look a few years ahead. It’s hard to imagine how they wouldn’t be in full flight unless they have a good moat. So why would that have to be correlated with AI hardware? It doesn’t make sense anyway. It’s just, “Oh, this thing is doomed anyway,” if I’m right about that.
Those who have a wide moat will probably win. So it doesn’t really matter. Why does capital spending on AI matter one way or the other? They don’t have [a moat].
4. Hyperscalers Advantage
So what fundamental trades do you see, then?
My favorite bet on AI is simply to play the rise of hyperscalers like Google, Amazon, and Microsoft. But I predicted they would probably sell off, so I got out for a while. Maybe I’ll come back to them at some point.
Is this based on the idea that slowing down capital spending actually helps them because they’ve funded everything up front and they’ll be the ones with the computing power?
Yes, it was a subtle thing. After the governor’s victory, they fell in premarket trading. I almost bought the dip for a bargain, but got scared. The next day they jumped up, and then the whole AI trade collapsed the day after that. They jumped the next day because of what you’re talking about.
It’s good for hyperscalers if, hypothetically, the government blocks 95% of all new projects. Then they own this truly unique, great asset.
But I don’t think the government, even with this story about the governor of Texas, really changes anything. She ran a good campaign and all that, but the real bottleneck was just the power plants. This should have been severely limited in any case. You can always build them in another state or another location, and I don’t think that news mattered at all.
The reason I like hyperscalers is the way they’ve structured the contracts. Essentially, they’re projects with a 35% return on capital, and they’ve contracted about half of the capacity they’re building. If you do the math, their downside risk is pretty small. Even if there’s a pretty nasty downturn, Amazon, Google, and Microsoft won’t lose a lot of money in a bad scenario unless software companies go bankrupt in very large numbers and very quickly.
In a good outcome, if I’m wrong, let’s assume computing power remains in short supply and we don’t have enough intelligence for 10 years. Then you have a 35% return on capital investment, which is just incredible.
It’s clear that the market is now a machine for creating narratives. It’s incredibly driven by this. What needs to happen for the narrative to change back?
I think it’s just a matter of positioning. Who is the marginal buyer now? Who is inspired by AI now? I’m not sure who it is. Where is the big money here? I just don’t see it happening anytime soon.
We had weakness in AI trading the day before this, but the broad market generally held up. I’d say yesterday was really the first time that weakness spread to the rest of the market and we saw a real sell-off at the index level. Do you think this was a purely technical phenomenon, or did people have concerns about other market sectors besides AI?
Well, the Democrats performed really well. Let’s say they do well in ’28; they could, for example, raise corporate tax rates. Perhaps the market has realized this, and that would be a valid reason for the market to fall 4% right now. It would have a completely logical fundamental justification.
Perhaps the market has thought about this, given that the S&P 500 is down only 4%. For such a strong Democratic showing, the tax threat seems entirely justified. How much might it cost? If you just repealed Trump’s tax breaks, how would that affect valuations? You said we fell by about 4%. Somehow that’s not enough.
Yes, it would be postponed to the future. Let’s say it happens sometime in 2030. If it happened, I’d say 4% wouldn’t be enough. But let’s say there’s a 30–50% chance that it will happen. Then 4% seems quite acceptable. Maybe it should be 6% or 7%. I don’t know, but this is about the right move.
We’ve already covered most of this path. As for the immediate catalysts, Anthropic’s IPO is expected to be priced sometime before the end of the year. Its valuation has already dropped a bit. What happens if the price is significantly reduced?
I think there are already a lot of these huge liquidity suckers in the market. Let’s say the G7 is issuing debt at a crazy pace right now, especially in the bond market. They’re sucking up all the liquidity they can. And then there’s Anthropic. Maybe they’ll come to market, maybe they won’t. That would be another case like this. So, purely for liquidity reasons, you have another reason to be quite pessimistic. And if Anthropic hits the market, it will be another blow.
You said you don’t think the threat of Democrats closing data centers is as great as the market thinks. You think it has more to do with oversupply, which was already potentially obvious. You believe the impact of tax increases is a serious factor that should be considered. Are there any other political moves as a result of this election that people should be thinking about now?
Apart from taxes, I don’t think politics matters much at all. Most Democrats are pretty moderate. The entire American system generally works quite well. It has stood the test of time. So that’s a key indicator. There may be some little things besides that, but they won’t really affect anything.
5. Antitrust And Search
You start seeing memes joking about Silicon Valley and how people are going to get out of there. They all pretty actively supported the Trump administration at the beginning of his second term, and now you start seeing jokes about them turning left. Do you think there will be any retaliation against the tech sector for what many see as its flirtation with the Trump administration?
Maybe they’ll start taking more antitrust actions. This is long overdue. I think that would be great for the world, but probably pretty bad for stocks. Interestingly, these companies seem to be self-destructing. As I wrote in my letter in 2023, AI has become a major threat to their monopoly profits. I think that’s exactly right, but it has nothing to do with politics.
Which big tech company do you think is most at risk from antitrust laws?
Probably only Google. They have so many lines of business.
Yes, but they could always go after more of them.
Yes, I think that’s true. They just went through search and advertising, which looks like it might be on the decline.
I don’t know if it was a downturn.
I don’t think it was a decline. The growth rate has just slowed down significantly, so it’s not a recession yet. I think we jumped to conclusions, but it certainly looks like I’m using AI for a lot of my searches now. This has affected my search behavior, and I don’t know if getting ads in my AI results would influence which models I use. We haven’t crossed that Rubicon yet to see how people will react. Who knows? Maybe Google can figure out how to monetize this.
Google Maps is pretty good for them. And YouTube probably benefits a lot from AI. Isn’t that right? You can create a lot more content, videos by one creator that are great and all. They also have the advantage of scale in that business.
But regarding search, I think AI is a hindrance. Google won’t die in an instant, but this business will start to decline at some point. I don’t think it’s going to reach the level of cable TV, which is falling by about 10% per year. I don’t think it will be that bad, but people are creatures of habit. They continue doing the same thing even if it’s not the best option, and Google can integrate its AI into search so people don’t leave. But they’ll no longer be able to get by simply with these walls of advertising.
I don’t think anyone will succeed at that. I don’t think any AI will be able to do it. Let’s say there’s the best model for the most difficult tasks, like programming, but it comes at a high price. They don’t make money from advertising. But the average person might want to find a good restaurant within a five-minute walk. In a few years, all AIs will be good enough to do that. They’re unlikely to be able to show annoying ads or do what Google Search did, which actually degraded quality by making people search longer. I don’t think anyone will be able to do something like that in the future.
How do you relate this to your idea that hyperscalers are winners?
It’s just that this part of their business has a very favorable risk-reward ratio because of how they structured these deals. If I’m wrong and intelligence is in short supply for the next decade, they’ll invest these hundreds of billions of dollars at a 35% return on capital. That’s simply incredible. So that alone can overshadow any bearish arguments about their other businesses.
6. Physical Bottlenecks
What do you think has brought us to this point of scarcity? Is it the efficiency of the models, and the fact that we don’t need advanced models for most tasks because they don’t require an IQ of 300? The fact that they can be automated, and that we don’t actually do that much complex work?
Yes, that’s right. Even programming, you know… I don’t really understand what people spend money on in some products, but let’s take Windows as an example. Windows 7 was just really good, and Windows hasn’t gotten any better since then because it was already perfect. You couldn’t figure out how to make it better, but they keep releasing new versions of Windows. I don’t know exactly why.
I feel like that applies to a lot of things: Instagram or anything else. They’re not high-tech products, like Netflix. These are quite simple businesses. They don’t require much superintelligence. I think intelligence could be used more for things like medical research. But the physical limitation will arise much faster, because you need a test subject or something, right? How do you scale the number of subjects by 100 times?
How much computing power is needed to develop drugs? If we believe we can cure cancer and all other serious diseases with AI, how many resources could that consume?
It may consume a lot, but you probably need to collect data. No one knows for sure how this will work, but I think the AI will say, “Okay, this path looks interesting. Let’s start exploring it.” You’ll need a team of people, robots, or someone else to conduct experiments, do research, and collect data. Then the AI will say, “Okay, now we need to conduct the next series of studies, get new data, and so on.” I think the possibilities are endless.
You can just keep solving it using more and more resources. But when I look at it more broadly and ask, “If this increases the number of experiments by a factor of 100, how long would it take to physically build all these labs?” It would take many years. So I think that’s most likely to be the bottleneck.
Let’s remove the physical barriers. Could drug development—or, let’s call them basic science problems—reach a new level? We still don’t have a good solid-state battery, for example. What if AI helped us design one? There are other complex scientific problems that AI could potentially solve. Do you think technological progress, as a source of demand for computing power, is large enough to offset the excess—or perceived excess—supply?
I think that outcome is possible. The way people invent things requires a long-term accumulation of knowledge, and it’s usually done by more than one person. You could say that Claude Shannon, or someone like him, was like that to some extent. But take Einstein, for example. He had only one or two narrow themes that he developed, and he was kind of a loner intellectual.
How many more people like that will there be? We haven’t really had people like this for almost 100 years, right? For people with such exceptional intelligence, who could sit down, think about problems in their heads, and solve them, progress effectively stopped maybe 70 or 80 years ago. I don’t want to give an exact number.
7. Demand Narratives Shift
Perhaps if computers become much smarter than they were, progress will resume, and the price of computing could skyrocket. But if things go wrong, I think you’ll run into a bottleneck instead. A lot of people have invested a lot of money in making this process continue: all the banks, big investment managers, Silicon Valley. This question is almost existential for the entire industry. They need some kind of big breakthrough.
You can see it in the language people use. They’re starting to talk about cancer treatment as the reason why Anthropic is worth $1.5 trillion. It’s becoming more about results than demand for labor replacement. They need to find more demand. They’re probably saying this because they understand what I said earlier: software companies themselves are unlikely to buy everything. They’re the main customers now, so they need a new big customer. They need to sell this idea somehow.
I hope this works. I really hope so. That would be great. But will it work? I really don’t know.
8. White Collar Automation
What about office workers? Do you think that’s a sufficient source of demand if it were actually implemented?
No, it’s all too simple. Given how AI is developing, it’s safe to assume that all of this will be reduced to zero in five years, and the cost of accounting or something like that will become almost free and require almost no computing power. We could replace a huge portion of human labor, and no one would benefit except the shareholders of existing companies.
I don’t even think the shareholders would benefit much. It’s a pretty crappy business, right? Let’s say you do corporate taxes or personal taxes. Take a complex personal tax return. It’s quite complicated. You need a good accountant, a personal accountant. But it won’t be difficult for a machine, and if it’s not difficult for one machine, they’ll probably make ten of them. The price will be practically zero. They won’t make any profit from creating this system, but they’ll probably create it. So I think the only ones who will win are the consumers.
Well, I guess I’m talking more about a company like Progressive Insurance, a large company where a lot of intellectual work is done. A large part of it is administrative. You’re dealing with a huge bureaucratic mess, and insurance departments exist at the state level. There are companies with huge revenues but also huge production costs and small profits. Because of the nature of insurance, you could argue that the margin they earn could be manageable if we could create some level of efficiency.
That’s exactly what I’m talking about: a large national or transnational corporation where a significant part of the costs goes to labor. How relevant is this story?
I think most of these office jobs will disappear. Say you have a business with a wide moat, and for some reason people can’t copy it, maybe because of scale or network effects. Perhaps they’ll have to lower their prices a little. But since they had to lay off almost all their employees, they get a kind of printing press for money. That’s a bullish forecast for some of these companies.
But in terms of actual AI use, I don’t think it will consume much computing power. The calculations they use won’t really affect the capital expenditure situation of hyperscalers. My base case is that all these jobs will disappear.
9. Sarmaya LENS ETF
If it’s a competitive industry, then the companies are really in trouble, because their entire business used to be based on bureaucracy, as you say. And now they don’t even have that. Some companies may be able to implement AI, improve margins, and become a kind of virtual money-printing machine. But in industries with more competition, it turns into another race for survival. Everyone will improve margins, and the old argument applies: “Your margin is my market share.”
Yes. It’s happened in other industries too. Think of what Amazon did to retail in general, or the railroads. I’m sure they ruined the lives of many businesses. Those businesses went out of business. There was also this funny satirical website where cavemen tried to ban the Ferris wheel because they said it would kill so many jobs.
It’s funny, but I think it’s a good thing. Most people don’t even like doing accounting. When we talk about that kind of work, it’s definitely not our highest calling.
This type of labor we talk about as replaceable raises the question of whether it will cause social problems, because all these people will suddenly become artists. We would all be creating masterpieces if we weren’t so busy filling out reports.
10. The Impact on Skilled Labor
But that’s the thing with artists: it’s kind of like Price’s law. Only a tiny percentage of people are even remotely interested. I’ve always liked the furniture example. We live in a supposedly super-rich society today, but everyone buys cheap furniture that has to be assembled out of a box, and no one has put any effort into creating it. It’s just factory junk. Maybe more people will start making quality furniture. I think there are a lot of potential professions people can pursue.
So you think these are crafts, more than anything else?
Yes. I think this is an important direction.
So shall we return to the workshops?
I hope so, if you’re lucky.
No, that’s good. I just think it’s a complicated, very complicated question. What do you think about vertical AI and the idea that companies will be able to set prices based on results? You could create a specialized tool for an industry, and it wouldn’t become a mass-market product. It wouldn’t be usage-based, at least not like it is now, and you’d be able to deliver those cost savings to a specific company or industry. Do you think it will be possible to have that kind of price advantage when providing AI solutions to large corporations?
No. Maybe for a few years it will seem like a good business. But I’m assuming it’s all so simple, and you just have to combine strange things, people, time, and everything else. I don’t think AI is advanced enough yet. But in five years, I’m sure.
Do you mean general AI, or the idea of special tools for specific solutions?
11. Software Moats and Deflation
Yes. I’ll give an example from my own business. We hired a few people this year. We’re already creating tools. We built a risk system, our own kind of risk system, and it works very well. Maybe it’s not as good as its professional counterparts, but I like it better.
We also created our own terminal. There are some obstacles to making this a generally available product. But I think I’ll probably make it free, or people might pay for their tokens if they have their own AI interface. Otherwise, I think I’ll just give it away for free. I believe it should be free. Maybe the price will be very low at first, or even a little higher, but then I’ll aggressively lower it because I know others will do the same.
Here’s my idea. Let’s say we have this risk terminal, a risk system, a kind of universal modern portfolio product that you can simply connect your IBKR account to. In a year or two, AI might become so good that you won’t need to do anything at all. It will just do everything automatically.
Wonderful. So you just run this thing, it pulls up your portfolio, and says, “This is the risk you’re taking,” and so on. From my perspective, we can use it as a marketing tool. At least every time someone shares their risk report or talks about the company, hopefully our logo will be somewhere on it. You can just show at a glance how much you’re constantly lagging behind the market.
Yes, it’s possible.
You should just give up and give your money to David.
The question of data access remains. Can you even get all the data you need for the terminal? I’m sure it costs money, and you have to maintain it.
Yes. On data, let’s take annual reports, for example. Getting those reports used to be the easiest part of the job, accounting for only a few percent of the work. But Bloomberg’s former moat was that they spent years hiring people to clean up and correct the data.
Now, let’s say we take all the old annual reports from these companies, going back, hopefully, 100 years. That would be cool. And you know what? We’ll just let artificial intelligence process them. I bet that in a few years I’ll trust AI numbers more than data from the guys at Bloomberg. So now we have a product that costs a fraction of the price and is better.
So, David, what do you think about the fact that you’re perhaps someone with a high level of initiative, and that your company has unique experience creating things like this, finding reports, and processing data?
Yes, but I’m saying I’m going to give it away for free. If this works, people who used to charge for similar services for other companies were asking for $25,000 or $35,000 per user per year. Let’s say my tool is better and free. As I said, not everyone has to be so proactive about creating it and collecting the data. There are probably a dozen other people like me who will do it for free, too.
This sounds incredibly deflationary. It’s the same society of total automation and affluence that technofuturists have been talking about for a long time. Is that really how you see the world?
12. Back to Physical World
When it comes to software and information services, yes. But it would allow many smart or hardworking people to do something else in the real, physical world. Take housing construction, for example. There are many rather careless developers. They build very mediocre houses that aren’t well thought out and aren’t really places where people dream of living. They’re quite nice, modern, and trendy, but maybe we’ll get to the point where people build and carefully design wonderful places to live. That would be good.
There’s a shortage of really good houses to live in right now. That’s good, right?
Yes, that’s good. But despite the shortage of skilled labor, I’d say the cost of materials still plays a big role. And if this suddenly creates more demand, does that make sense? Who will buy all these houses?
Housing is a good example because we already build houses, right? They’re already being built, but they aren’t very well thought out. You can visit most residential complexes. There are better developers and worse ones, right? If you go to the worst developers, you can easily imagine how it is: every house looks almost identical, and the tree stands in exactly the same place. It’s not exactly a pleasant place to live. It works; it’s perfectly acceptable, but perhaps such places shouldn’t exist anymore. Maybe they can be improved or changed.
I’d argue that it might have been deliberate. It’s not that no one thought about it; everything was well thought out to build as many houses as possible with the least effort.
Yes, it’s simple. That’s true. I don’t see—I mean, if you’re saying there’ll be so much displacement of labor that suddenly everyone will be doing this, it seems like their labor won’t carry much weight because there’ll be a huge surplus of labor. People would really have to love working with their hands to want to do it.
I don’t think you’ll have to physically build it, will you? You’ll design it or contemplate it. That’s an improvement. For example, one of my partners has a very interesting hotel project. He couldn’t build it because interest rates had risen too high, but he bought the land. It was very thoughtfully designed and is right next to the national park. He found architects who would make it fit into the landscape and look visually appealing and beautiful.
I think there’s room for that. And it’s not that he’s going to—in fact, I think the number of people doing the physical work will also decrease. For example, all the drivers will disappear.
You could argue that we don’t build such beautiful, unique things anymore because people can’t afford them. Everyone would like that, right? If you go into the ultra-premium segment, even ultra-elite real estate is starting to become formulaic. But at least theoretically, in the ultra-premium segment, you can hire an architect, commission an individual project, and get something truly beautiful and wonderful to live in.
That doesn’t mean it doesn’t exist. It’s just too expensive, and I don’t see why it has to be. It’s too expensive because the supply is too small, right? That’s how I see it. You don’t have enough people doing that kind of work. Instead, everyone is working on—take Google’s search business, for example. All the smartest people are trying to figure out how to violate users’ privacy as much as possible and make search worse and more annoying so that people search even more. How can we insert another ad? Those are exactly the kinds of problems these people are working on today. Our best and brightest are working on this. They all work for the Magnificent Seven, expanding those companies’ business models.
I don’t think that’s good for anyone. Take the big law firms, for example. No one likes working in large law firms. They all hate their jobs. They’re only there because they make a lot of money, right? But what if all those people started doing something truly useful?
Big law is a good example because it’s just complete rot. But my point is that there’s a class component here. If you’re from a certain upper class, you’re supposedly expected to know your worth and your place in society and get that kind of job. If that disappears, if working at a big law firm is no longer a viable way to assert your status, I don’t necessarily think carpentry will suddenly replace it. Maybe to some extent. If you look at many tech company founders, they have these kinds of passion projects, like opening a surf shop or something like that. Charlie Munger did this, right? He designed all sorts of weird things, and he didn’t even do it for the money.
What I’m saying is that people will actually enjoy doing it, and it can generate a pretty decent income compared with other opportunities. It won’t just be them, but all their peers, right? Millions of people will build things that others want to use all day, things they’ll love and want to be surrounded by.
13. Social Media Moats Debate
That’s a beautiful vision, David. I’m not sure I see the path the same way you do if we all suddenly lose our jobs. You mentioned what Google is doing by making search worse, and how artificial intelligence is cannibalizing it. Of the entire Magnificent Seven, Meta’s advertising business seems to be doing pretty well. You could argue that people are addicted to their phones and social media, and that this has negative externalities for the individual. But in general, I’d say people choose to spend more time on their phones. No one wants to search for something and not get a good result. Whether or not it’s addictive, people choose to spend more time on their phones. They choose to interact more often with sponsored content and all these kinds of ads on social networks. Do you consider this one of the best businesses among the Magnificent Seven?
It could be a good one. I’m not sure, because they used to call it social media, and now it seems to me that they’re becoming less social every year. Now it’s just media. And if it’s just media, then they may not have much of a competitive moat. But the media business has, of course, shown incredible results in the last few years. Maybe in another five years, though, their platform will simply become a commodity. I don’t know. I’m not saying I’m sure; I’m just presenting a pessimistic scenario.
What about the optimistic scenario?
There is one. The optimistic scenario is that people continue to use them. They become more and more fascinating, more and more interesting, because artificial intelligence makes the content truly better. There’s less garbage there. Maybe that’s even a good thing.
Yes, but the network effect definitely exists, right? If you create content and want to monetize it, you want to connect to a system where there are people. You were talking about data, and Bloomberg’s IB chat, in comparing your Bloomberg terminal to this product and explaining why it can be displaced.
At the institutional level, there’s IB chat: the ability to write to the broker, conclude a deal in the chat, and have everything stored with clear, understandable records. That has some value. But for the average person who doesn’t need that, I understand why data from any source works the same. At the social media level, there’s a network effect. You’re connecting many people to many people, so it’s usually a business with a fairly broad competitive moat.
But intuitively, it doesn’t seem as strong as it was in Facebook’s early days. That’s a pretty easy comparison, right? Back then it was, “If you want to see all your friends and family, you have to be on Facebook.” They wouldn’t be anywhere else. But let’s say some new startup brings in a bunch of artists to start showing their work there, or focuses on some other niche topic. Maybe people will start switching to that new site because it serves that niche.
It doesn’t have to be a huge platform anymore. I don’t know. I feel that, especially with these very short videos, endless scrolling gives them an advantage: they already have so many videos to watch, and the algorithm keeps people hooked. So this can be a pretty serious barrier to entry.
I don’t know exactly what caused it, but I recently left YouTube. I used to listen to music there all the time, but now I use Spotify all the time. Somehow YouTube lost me as a music listener.
Maybe someone could create a better service for short videos. I think TikTok has already done that in a way. But I think my attitude toward music—and maybe that will change in the world you’re describing—is that music is probably the most common art form among people, if you know what I mean. It seems our ability to create music is less rare than the ability to create visual art that interests people.
Music on Spotify is closer to your Bloomberg terminal, right? There are several sources of data that matter. Yes, there are millions of artists, but there are small groups of people who create the main demand for music, like popular artists. So, to me, it sounds a lot more like your argument about the Bloomberg terminal: you simply chose the service that serves you best.
There’s no unknown up-and-coming artist who says, “Oh, I just need to get on Spotify, and then everyone will find my music and my career will start.” You’re more likely to be discovered on Instagram, TikTok, YouTube, or one of those platforms that let you promote yourself. And I don’t think that’s what you’re looking for in your listening experience. Perhaps at this stage of life you’ve already formed your tastes. Maybe you admire new artists, but I think that’s why Spotify is different.
Yes. I think Spotify’s algorithm is simply better. Maybe, because we communicate with artificial intelligence and it’s now learning our preferences, it’ll start recommending things if you dig deeper into the search. You could ask, “Can you create a music playlist for me?” and just swipe left or right—like or dislike. Who knows? Maybe in five years, AI will be great at picking out what I really like, unlike the algorithms on TikTok or Instagram, where you don’t have a choice. They just push you the most addictive content. Even if you don’t necessarily like it, you’ll still watch it.
That’s the thing. They know you’ll like it. I think they know it’s highly addictive. I’m not sure they like it. It’s like with drug addicts.
Yes. This may not be the right way to look at things.
14. Market Exposure Wrap Up
I want to come back to the present, to the markets, to conclude our conversation. Are you thinking about pure exposure now? Are you in a net short position as a fund? What would have to happen for you to change this significantly?
For some reason, my fund-level beta has been negative for the past 30 days. It’s not very negative, but I don’t know if I should trust it because you can never say for sure. I just aim to be close to zero. I try to be in a position where I don’t worry about whether the market is going up or down.
We did quite well thanks to Mike and Rodrigo. I’ve been pretty much in the red this crazy week. But I’m not trying to bet on the direction of the market right now.
I don’t think so. Who knows? No one knows where it will be in a year.
I remember 2022. I think you went into a pretty big short position then, right?
That was a lot, but I think my beta was still close to zero. I was shorting these really bad, just terrible companies, because 2021 created a very simple situation. But trading in artificial intelligence is, of course, completely different. This is a fundamentally real thing. There is a real bullish scenario.
Despite the fact that you believe the market is oversaturated and there has been excessive development, you don’t see this as 2021. So, to be clear, there is a bullish scenario.
I’m going to be wrong at least 40% of the time or so. I don’t… It’s not… That’s why I don’t bet on it. I’m just going to stay out of this. Don’t worry about it.
Well, David, thank you very much for joining us. Where can people follow you?
I’m on X as @orrdavid. We launched a new Substack for Militia. I might start posting there occasionally, and Mike and Rodrigo will start writing there too. The first thing we published was Rodrigo’s third-quarter letter. The URL is content.militia.com. We launched seven days ago.
You have a new follower in me, David. Thank you.
Thank you for subscribing. I hope you enjoyed it.