Building LLMs is probably not going be a brilliant business
建築 LLMs 很可能不會是一個出色的生意

November 2024 2024 年 11 月

The Netscapes of AI
人工智慧的網景

Large language models (LLMs) like Chat-GPT and Claude.ai are whizzy and cool. A lot of people think that they are going to be The Future. Maybe they are — but that doesn't mean that building them is going to be a profitable business.
大型語言模型（LLMs）如 Chat-GPT 和 Claude.ai 非常厲害和酷。許多人認為它們將成為未來。也許它們會 — 但這並不意味著建立它們將是一個有利可圖的業務。

In the 1960s, airlines were The Future. That is why old films have so many swish shots of airports in them. Airlines though, turned out to be an unavoidably rubbish business. I've flown on loads of airlines that have gone bust: Monarch, WOW Air, Thomas Cook, Flybmi, Zoom. And those are all busts from before coronavirus - times change but being an airline is always a bad idea.
在 1960 年代，航空公司是未來。這就是為什麼舊電影中有這麼多關於機場的華麗鏡頭。然而，航空公司最終被證明是一個不可避免的爛生意。我搭過許多破產的航空公司：Monarch、WOW Air、Thomas Cook、Flybmi、Zoom。而這些都是在冠狀病毒之前破產的案例 - 時代在變，但經營航空公司始終是一個壞主意。

That's odd, because other businesses, even ones which seem really stupid, are much more profitable. Selling fizzy drinks is, surprisingly, an amazing business. Perhaps the best. Coca-cola's return on equity has rarely fallen below 30% in any given year. That seems very unfair because being an airline is hard work but making coke is pretty easy. It's even more galling because Coca-cola don't actually make the coke themselves - that is outsourced to "bottling companies". They literally just sell it.
這很奇怪，因為其他企業，甚至那些看起來非常愚蠢的企業，更有利可圖。出售汽水是一個令人驚訝的生意。也許是最好的。可口可樂的股本回報率在任何一年很少低於 30％。這似乎很不公平，因為經營航空公司是一項辛苦的工作，但製作可樂卻相對容易。更令人氣憤的是，可口可樂實際上並不自己製造可樂 - 這是外包給 “裝瓶公司” 的。他們只是賣它。

Industry structure - what makes a business good
產業結構 - 什麼使一家企業優秀

If you were to believe LinkedIn you would think a great business is made with efficiency, hard work, innovation or some other intrinsic reason to do with how hardworking, or clever, the people in the business are. That simply is not the case.
如果您相信 LinkedIn，您會認為一家偉大的企業是通過效率、努力工作、創新或其他內在原因來實現的，這些原因與企業中的人們有多努力或聰明有關。這根本不是事實。

What makes a good business is industry structure.
什麼構成了一個好的企業是行業結構。

Airlines - unfavourable industry structure
航空公司 - 不利的行業結構

To be an airline is to be in an almost uniquely terrible market position. For starters, there are only two makers of aeroplanes (Airbus and Boeing). For reasons of training and staff efficiency, you have to commit to one or the other, which gives the aeroplane makers very strong pricing power.
成為一家航空公司幾乎是處於獨特糟糕的市場位置。首先，只有兩家飛機製造商（空中巴士和波音）。出於培訓和員工效率的原因，您必須選擇其中一家，這使得飛機製造商擁有非常強大的定價權。

And buyers of airline tickets are incredibly fickle and have no loyalty. They will switch from one "carrier" to another over even small differences in price. Annoyingly, there are loads of other airlines and they're all running the same routes as you!
購買航空公司機票的人非常善變，沒有忠誠度。他們會因價格上的微小差異而從一家「航空公司」轉向另一家。令人惱人的是，還有許多其他航空公司，它們都在跑與你相同的航線！

Worse yet, starting a new airline is surprisingly easy. Aircraft hold their value so banks will happily lend against them. There are loads of staff available that new entrants can hire. So randos will continually enter your market, often selling tickets below cost for quite a while before they go bust. And to top it off, there are plenty of substitutes for air travel - from government-subsidised high speed trains to Zoom calls.
更糟的是，開設一家新航空公司竟然很容易。飛機保持其價值，因此銀行樂意提供貸款。有大量的員工可供新進入者僱用。因此，隨機的人將不斷進入您的市場，通常在破產之前相當長一段時間以低於成本的價格出售機票。最糟糕的是，有許多替代品可供選擇，從政府補貼的高速列車到 Zoom 視訊通話。

Airlines that get more efficient, work harder or come up with innovations aren't going to be able to "capture" the value of what they've done. If you make more than the bare minimum to survive Airbus will notice that you're being undercharged and you'll find that the next renewal on your service contract eats up the difference.
航空公司變得更有效率、更努力工作或提出創新的話，將無法「捕捉」他們所做的價值。如果你賺得比最低生存成本更多，空中巴士會注意到你被低估，你會發現下一次服務合同續約將吞噬這個差額。

Fizzy-drinks - very favourable industry structure
碳酸飲料 - 非常有利的行業結構

Being the Coca-cola company is pretty great though.
作為可口可樂公司其實相當不錯。

Coke is just water, colourant, flavouring, caffeine and sweetner. Those are all widely available and really cheap. And as I said, you don't even have to combine them yourselves - bottling companies will do that for you for almost nothing.
可樂只是水、色素、調味劑、咖啡因和甜味劑。這些都是廣泛可得且非常便宜的。而且正如我所說的，你甚至不需要自己組合它們 - 裝瓶公司將為您幾乎免費完成這些工作。

Handily, consumers are really picky about what goes in their mouth. The unofficial motto of your main competitor is "Is Pepsi ok?". This is despite the fact that they are identical in both taste and colour. And a significant minority of people actually say no!
消費者對進入口中的東西非常挑剔。您的主要競爭對手的非官方座右銘是 “百事可樂可以嗎？” 儘管它們在口味和顏色上完全相同。而且有相當大一部分人實際上說不！

And it isn't easy for new competitors to enter the market. They can't call their new drink "coke" due to trademarks. They have to call it something else. And consumers will generally refuse it because drinking an alternative is considered some kind of weird statement.
並且對於新競爭對手進入市場並不容易。他們無法將他們的新飲料稱為 "可樂"，因為商標問題。他們必須給它取個其他名字。消費者通常會拒絕它，因為喝另一種飲料被視為某種奇怪的表態。

What is industry structure?
產業結構是什麼？

Classically, there are five basic parts ("forces") to a company's position:
傳統上，公司位置有五個基本部分（"力量"）

1. The power of their suppliers to increase their prices
   供應商提高價格的權力
2. The power of their buyers to reduce your prices
   買家降低價格的能力
3. The strength of direct competitors
   直接競爭對手的實力
4. The threat of any new entrants
   新進入者的威脅
5. The threat of substitutes
   替代品的威脅

It's industry structure that makes a business profitable or not. Not efficiency, not hard work and not innovation.
行業結構決定一家企業是否有利可圖，而非效率、勤奮或創新。

If none of the forces are very much against you, your business will do ok. If they are all against you, you'll be in the position of the airlines. And if they're all in your favour: brill, you're Coca-cola.
如果沒有任何力量非常反對你，你的生意會做得不錯。如果它們都反對你，你將處於航空公司的位置。如果它們都支持你：太棒了，你就是可口可樂。

The industry structure of LLM makers: OpenAI/Anthropic/Gemini/etc
LLM 製造商的行業結構：OpenAI/Anthropic/Gemini 等

So is the position of LLM makers any good? I'm afraid it's not good news.
這樣，LLM 製造商的位置如何？恐怕這不是個好消息。

LLM makers sometimes imply that their suppliers are cloud companies like Amazon Web Services, Google Cloud, etc. That wouldn't be so bad because you could shop around and make them compete to cut the huge cost of model training.
LLM 製造商有時會暗示他們的供應商是像亞馬遜網絡服務、Google Cloud 等雲公司。這並不壞，因為你可以四處比價，讓他們競爭，降低模型訓練的巨額成本。

Really though, LLM makers have only one true supplier: NVIDIA. NVIDIA make the chips that all models are trained on - regardless of cloud vendor. And that gives NVIDIA colossal, near total pricing power. NVIDIA are more powerful relative to Anthropic or OpenAI than Airbus or Boeing could ever dream of being.
真的，LLM 製造商只有一個真正的供應商：NVIDIA。 NVIDIA 製造所有型號都是基於其訓練的晶片 - 無論雲供應商如何。這使 NVIDIA 擁有巨大的、幾乎完全的定價權力。相對於 Anthropic 或 OpenAI，NVIDIA 比 Airbus 或波音更強大，後者無法想像。

How much power do buyers have over LLM token prices? So far, it seems fairly high. Most LLM users seem willing to change from Chat-GPT to Claude, for example. It doesn't seem like brand loyalty is being built up. And companies that build AI into their businesses are starting to do so via abstraction layers that allow them to switch model easily. That makes LLMs interchangeable - which is bad for those who sell them.
買家對 LLM 代幣價格有多大的影響力？到目前為止，似乎相當高。大多數 LLM 用戶似乎願意從 Chat-GPT 轉換到 Claude，例如。看起來並沒有建立品牌忠誠度。將 AI 整合到他們的業務中的公司開始通過允許他們輕鬆切換模型的抽象層來這樣做。這使 LLMs 可以互換 - 對於那些出售它們的人來說是不好的。

What's the strength of direct competitors? Again, it is considerable. There are loads of LLM vendors and pricing appears competitive. Worst of all, Facebook basically dump their model on the market for no cost. It's reminiscent of Internet Explorer - not exactly a great portent.
直接競爭對手的實力如何？再次，這是相當可觀的。有大量的 LLM 供應商，價格看起來很有競爭力。最糟糕的是，Facebook 基本上免費將他們的模型推向市場。這讓人想起了 Internet Explorer - 不確切是一個偉大的預兆。

And it seems fairly easy for new entrants to build brand new models. That is why there are so many LLM makers. Most of the techniques for making LLMs are openly published in papers. Even bad models can gain customers if they are cheap, which allows new entrants to gain a foothold.
而且對於新進入者來說，建立全新模型似乎相當容易。這就是為什麼有這麼多製造商。製作 LLMs 的大部分技術都是公開發表在論文中的。即使是不好的模型，如果價格便宜，也能吸引顧客，這讓新進入者能夠立足。

The situation on substitutes is mixed. Instead of having Chat-GPT write some text you could pay a person to do it instead. That is likely to be much more expensive but also less likely to hallucinate, which might be important for some usecases (law is the field least likely to use LLMs). And then there is the trend that metadata tends to displace artificial intelligence once particular application has been proved out - so as soon as you find a solid usecase you stand to be replaced.
替代品的情況複雜。與其讓 Chat-GPT 寫一些文本，不如支付一個人來代替。這可能會更昂貴，但也不太可能出現幻覺，這對某些用例可能很重要（法律是最不可能使用 LLMs 的領域）。然後有一個趨勢，即元數據往往會取代一旦特定應用程序已被證明的人工智能 - 所以一旦找到一個堅實的用例，你就有可能被取代。

A single mildly positive point does not make a profitable business. LLM makers look a lot more like Netscape - who invented graphical web browsers, then went bust - than Google, who made something good that ran on top of the the web browsers.
單一輕微正面的點並不能使一家公司盈利。LLM 製造商看起來更像是 Netscape - 發明了圖形網頁瀏覽器，然後破產了 - 而不是 Google，他們創造了一些運行在網頁瀏覽器之上的好東西。

How are they raising so much money?
他們是如何籌集這麼多錢的？

If LLM makers seem cursed to an airline-style business destiny, how come they are able to raise so much money? OpenAI raised $6.6 billion at a valuation of $157 billion less than two months ago. That might be the biggest VC round of all time.
如果 1001 家製造商似乎注定要走向航空公司風格的商業命運，那麼他們是如何能夠籌集如此多的資金呢？OpenAI 在不到兩個月前以 1570 億美元的估值籌集了 66 億美元。這可能是有史以來最大的風險投資輪。

What do they know that I don't? It is a mystery - but let's consider the options.
他們知道什麼我不知道的？這是一個謎 - 但讓我們考慮一下選擇。

Perhaps they are hoping to develop their own chips to reduce their dependance on NVIDIA. $6.6 billion is not enough to build a new fab but it might be enough to get a new chip designed which allows them to migrate off NVIDIA. That would save them paying so much money for GPU time. But, NVIDIA are actually one of the investors in the round (although only a fairly small amount) - so it's unlikely "develop an NVIDIA competitor" was on any of the pitch deck slides.
也許他們希望開發自己的晶片，以減少對 NVIDIA 的依賴。66 億美元不足以建造新的晶圓廠，但可能足夠設計一款新的晶片，讓他們能夠遷移離開 NVIDIA。這將幫助他們節省為 GPU 時間支付如此多錢。不過，NVIDIA 實際上是這輪投資中的一家投資者（雖然只是一筆相當少的金額），因此很可能在任何路演簡報中都不會出現「開發 NVIDIA 競爭對手」的內容。

Perhaps OpenAI are hoping to build a strong brand so that customers won't switch to competitors so easily. It's not impossible, there is proof the branding and lock-in can work in technology - but it seems difficult to manage given that LLMs themselves generically have a textual interface - meaning that there is no real API as such - you just send text, and it sends text back.
也許 OpenAI 希望建立強大品牌，讓客戶不會那麼容易轉向競爭對手。這並非不可能，有證據顯示品牌和鎖定在技術上是有效的 - 但考慮到 LLMs 本身通常具有文本界面，這似乎很難管理 - 這意味著沒有真正的 API - 你只需發送文本，它就會回傳文本。

Can they do anything about new entrants? Possibly. If investing $6.6bn allows them to develop a major improvement in their model then that would raise everyone else's costs considerably and probably force some of their smaller competitors out of the market. The trouble is the money is the most fungible of all goods (that is the point, after all) and that $6.6bn is not all that much of it. So this round wouldn't, by itself, be enough to dissuate others. I used to work at a bank and I can tell you that individual bond raises can be a lot more than $6.6bn.
他們能對新進入者採取任何行動嗎？可能。如果投資 66 億美元使他們能夠在模型中進行重大改進，那將大幅提高其他人的成本，可能迫使一些較小的競爭對手退出市場。問題在於金錢是所有商品中最具替代性的（畢竟這就是重點），而這 66 億美元並不算太多。因此，僅憑這一輪，並不足以阻止其他人。我曾在一家銀行工作，我可以告訴你，單個債券募集的金額可能遠超過 66 億美元。

It's worth saying that even companies that raise huge sums of money sometimes turn out to have no viable business. WeWork ultimately raised over $10bn at a valuation of $47bn before it was realised that their business simply did not make sense. WeWork were valued at just $0.56bn in their most recent financial restructuring - having lost well over 95% of what was invested.
值得一提的是，即使是籌集了大筆資金的公司，有時也可能沒有可行的業務。WeWork 最終在估值為 470 億美元之前籌集了超過 100 億美元，後來才意識到他們的業務根本沒有意義。在最近的財務重組中，WeWork 的估值僅為 0.56 億美元，已經損失了超過 95% 的投資。

Not all AI companies are doomed
並非所有人工智慧公司都注定失敗

If LLM makers aren't going to be good businesses, does that bode ill for The Future?
如果 LLM 製造商不會成為好企業，這對未來不是一個好兆頭嗎？

Firstly, it does not mean the technology will be bad. Whether the technology ends up being good or not is mostly unrelated to whether Open AI/Anthropic/Mistral/whoever makes any money off it. Container virtualisation technology is pretty well developed even though Docker made almost nothing on it. Web browsers are extremely advanced pieces of software even though making a browser is such a bad business that most don't usually count it as a business at all. And CRMs are terrible despite the fact that Salesforce is tremendously successful. Technology success and business success are mostly unrelated.
首先，這並不意味著技術會變差。技術最終是否變好，大多與 Open AI/Anthropic/Mistral/ 任何人是否從中賺錢無關。容器虛擬化技術已經相當成熟，即使 Docker 幾乎沒有從中獲利。網頁瀏覽器是非常先進的軟體，即使製作瀏覽器是一個極為糟糕的生意，大多數人通常不將其視為生意。儘管 Salesforce 非常成功，但 CRM 仍然很糟糕。技術成功和商業成功大多無關。

And then: not all AI businesses are building models. Ideally, if I were running an AI business I would avoid building a model at all costs. Building your own models looks like a undifferentiated schlep. Using a tiny bit of some expensively trained model that Anthropic has produced could be very cost effective and might make some business idea work that wouldn't have 5 years ago.
然後：並非所有人工智慧業務都在建立模型。理想情況下，如果我在經營一家人工智慧業務，我會盡一切努力避免建立模型。建立自己的模型看起來像是一個毫無差異的辛苦工作。使用一點點由 Anthropic 生產的昂貴訓練模型可能非常具成本效益，並且可能使一些在 5 年前無法實現的業務想法得以實現。

Beware software companies that aren't software companies
請小心那些不是軟體公司的軟體公司

Software companies are really good businesses. You have no real suppliers, your software is often unique (so no competitors) and the substitute is just users doing the job themselves. For this reason, software companies tend have really great margins.
軟體公司是非常好的企業。您沒有真正的供應商，您的軟體通常是獨一無二的（因此沒有競爭對手），替代品只是用戶自己完成工作。因此，軟體公司往往有非常好的利潤率。

The problem is that not all technology companies are software companies. If you have a hugely powerful single supplier like NVIDIA then the economics of your company are going to less like Microsoft Office and more like Pan-Am.
問題在於並非所有科技公司都是軟體公司。如果您有像 NVIDIA 這樣的強大單一供應商，那麼您公司的經濟將更像是泛美航空，而不是像 Microsoft Office。

Contact/etc 聯繫 / 等

Other notes 其他筆記

The AI safety movement is a fantastic hypeman for LLMs as a technology. Implying (pretty dubiously) that we are 10 minutes from midnight in some kind of Ghost-in-The-Shell style AI crisis is in fact an extremely effective form of product marketing. Perhaps that is why OpenAI and others employ so many AI safety specialists.
AI 安全運動是 LLMs 作為一種技術的出色炒作者。 暗示（相當可疑地）我們距離某種類型的《攻殼機動隊》風格的 AI 危機只有 10 分鐘的時間，實際上是一種非常有效的產品營銷形式。 也許這就是為什麼 OpenAI 和其他公司雇用了這麼多 AI 安全專家。

The Coca-cola company mainly sit back and rake in the megabucks - but they do spend a little bit of their earnings on research. And a little bit of a lot is still significant. It's interesting that that coke's market research has discovered that coke works better as a gender segregated product: Coke Zero is Diet Coke, but for men.
可口可樂公司主要是坐收巨款，但他們確實會花一點點收入在研究上。而一點點也是有意義的。有趣的是，可口可樂的市場研究發現，可口可樂作為一個性別分離的產品效果更好：Coke Zero 是健怡可樂，但是針對男性。

If you want to read more about industry structure and market strategy, the place to start is with Michael Porter. He reworked his famous essay The Five Forces that Shape Corporate Strategy in 2008. It's not the last word, but it probably should be the first word you read if you want to learn more. And if you like it, he has a lot more.
如果您想要了解更多有關產業結構和市場策略，最好的起點是米高波特（Michael Porter）。他於 2008 年重新修訂了他著名的文章《塑造企業策略的五大力量》。這並非最終答案，但如果您想要深入了解，這可能應該是您閱讀的第一篇文章。如果您喜歡這篇文章，他還有更多相關內容。