davideisinger.com

My personal website
Log | Files | Refs | README

calpaterson-com-gg1ovh.txt (13969B)


      1 Cal Paterson | [1]Home [2]Services [3]About
      2 
      3 Building LLMs is probably not going be a brilliant business
      4 
      5 November 2024
      6 
      7 The Netscapes of AI
      8 
      9 image of early 20th century train advert for Watford Railways improved the
     10 lives of millions - but investors were rewarded with a [4]dramatic bust
     11 
     12 Large language models (LLMs) like Chat-GPT and Claude.ai are whizzy and cool. A
     13 lot of people think that they are going to be The Future. Maybe they are — but
     14 that doesn't mean that building them is going to be a profitable business.
     15 
     16 In the 1960s, airlines were The Future. That is why old films have so many
     17 swish shots of airports in them. Airlines though, turned out to be an
     18 unavoidably rubbish business. I've flown on loads of airlines that have gone
     19 bust: Monarch, WOW Air, Thomas Cook, Flybmi, Zoom. And those are all busts from
     20 before coronavirus - times change but being an airline is always a bad idea.
     21 
     22 That's odd, because other businesses, even ones which seem really stupid, are
     23 much more profitable. Selling fizzy drinks is, surprisingly, an amazing
     24 business. Perhaps the best. Coca-Cola's return on equity has rarely fallen
     25 below 30% in any given year. That seems very unfair because being an airline is
     26 hard work but making coke is pretty easy. It's even more galling because
     27 Coca-Cola don't actually make the coke themselves - that is outsourced to
     28 "bottling companies". They literally just sell it.
     29 
     30 Industry structure - what makes a business good
     31 
     32 If you were to believe LinkedIn you would think a great business is made with
     33 efficiency, hard work, innovation or some other intrinsic reason to do with how
     34 hardworking, or clever, the people in the business are. That simply is not the
     35 case.
     36 
     37 What makes a good business is industry structure.
     38 
     39 Airlines - unfavourable industry structure
     40 
     41 To be an airline is to be in an almost uniquely terrible market position. For
     42 starters, there are only two makers of aeroplanes (Airbus and Boeing). For
     43 reasons of training and staff efficiency, you have to commit to one or the
     44 other, which gives the aeroplane makers very strong pricing power.
     45 
     46 And buyers of airline tickets are incredibly fickle and have no loyalty. They
     47 will switch from one "carrier" to another over even small differences in price.
     48 Annoyingly, there are loads of other airlines and they're all running the same
     49 routes as you!
     50 
     51 Worse yet, starting a new airline is surprisingly easy. Aircraft hold their
     52 value so banks will happily lend against them. There are loads of staff
     53 available that new entrants can hire. So randos will continually enter your
     54 market, often selling tickets below cost for quite a while before they go bust.
     55 And to top it off, there are plenty of substitutes for air travel - from
     56 government-subsidised high speed trains to Zoom calls.
     57 
     58 Airlines that get more efficient, work harder or come up with innovations
     59 aren't going to be able to "capture" the value of what they've done. If you
     60 make more than the bare minimum to survive Airbus will notice that you're being
     61 undercharged and you'll find that the next renewal on your service contract
     62 eats up the difference.
     63 
     64 Fizzy-drinks - very favourable industry structure
     65 
     66 Being the Coca-Cola company is pretty great though.
     67 
     68 Coke is just water, colourant, flavouring, caffeine and sweetener. Those are
     69 all widely available and really cheap. And as I said, you don't even have to
     70 combine them yourselves - bottling companies will do that for you for almost
     71 nothing.
     72 
     73 Handily, consumers are really picky about what goes in their mouth. The
     74 unofficial motto of your main competitor is "Is Pepsi ok?". This is despite the
     75 fact that they are identical in both taste and colour. And a significant
     76 minority of people actually say no!
     77 
     78 And it isn't easy for new competitors to enter the market. They can't call
     79 their new drink "coke" due to trademarks. They have to call it something else.
     80 And consumers will generally refuse it because drinking an alternative is
     81 considered some kind of weird statement.
     82 
     83 What is industry structure?
     84 
     85 Classically, there are five basic parts ("forces") to a company's position:
     86 
     87  1. The power of their suppliers to increase their prices
     88  2. The power of their buyers to reduce your prices
     89  3. The strength of direct competitors
     90  4. The threat of any new entrants
     91  5. The threat of substitutes
     92 
     93 It's industry structure that makes a business profitable or not. Not
     94 efficiency, not hard work and not innovation.
     95 
     96 If none of the forces are very much against you, your business will do ok. If
     97 they are all against you, you'll be in the position of the airlines. And if
     98 they're all in your favour: brill, you're Coca-Cola.
     99 
    100 The industry structure of LLM makers: OpenAI/Anthropic/Gemini/etc
    101 
    102 So is the position of LLM makers any good? I'm afraid it's not good news.
    103 
    104 LLM makers sometimes imply that their suppliers are cloud companies like Amazon
    105 Web Services, Google Cloud, etc. That wouldn't be so bad because you could shop
    106 around and make them compete to cut the huge cost of model training.
    107 
    108 Really though, LLM makers have only one true supplier: NVIDIA. NVIDIA make the 
    109 [5]chips that all models are trained on - regardless of cloud vendor. And that
    110 gives NVIDIA colossal, near total pricing power. NVIDIA are more powerful
    111 relative to Anthropic or OpenAI than Airbus or Boeing could ever dream of
    112 being.
    113 
    114 How much power do buyers have over LLM token prices? So far, it seems fairly
    115 high. Most LLM users seem willing to change from Chat-GPT to Claude, for
    116 example. It doesn't seem like brand loyalty is being built up. And companies
    117 that build AI into their businesses are starting to do so via abstraction
    118 layers that allow them to switch model easily. That makes LLMs interchangeable
    119 - which is bad for those who sell them.
    120 
    121 What's the strength of direct competitors? Again, it is considerable. There are
    122 loads of LLM vendors and pricing [6]appears competitive. Worst of all, Facebook
    123 basically dump their model on the market for no cost. It's [7]reminiscent of
    124 Internet Explorer - not exactly a great portent.
    125 
    126 And it seems fairly easy for new entrants to build brand new models. That is
    127 why there are so many LLM makers. Most of the techniques for making LLMs are
    128 openly published in papers. Even bad models can gain customers if they are
    129 cheap, which allows new entrants to gain a foothold.
    130 
    131 The situation on substitutes is mixed. Instead of having Chat-GPT write some
    132 text you could pay a person to do it instead. That is likely to be much more
    133 expensive but also less likely to hallucinate, which might be important for
    134 some use-cases (law is the field least likely to use LLMs). And then there is
    135 the trend that [8]metadata tends to displace artificial intelligence once
    136 particular application has been proved out - so as soon as you find a solid
    137 use-case you stand to be replaced.
    138 
    139 A single mildly positive point does not make a profitable business. LLM makers
    140 look a lot more like Netscape - who invented graphical web browsers, then went
    141 bust - than Google, who made something good that ran on top of the web
    142 browsers.
    143 
    144 How are they raising so much money?
    145 
    146 If LLM makers seem cursed to an airline-style business destiny, how come they
    147 are able to raise so much money? OpenAI [9]raised $6.6 billion at a valuation
    148 of $157 billion less than two months ago. That might be the biggest VC round of
    149 all time.
    150 
    151 What do they know that I don't? It is a mystery - but let's consider the
    152 options.
    153 
    154 Perhaps they are hoping to develop their own chips to reduce their dependence
    155 on NVIDIA. $6.6 billion is not enough to build a new fab but it might be enough
    156 to get a new chip designed which allows them to migrate off NVIDIA. That would
    157 save them paying so much money for GPU time. But, NVIDIA are actually one of
    158 the investors in the round (although only a fairly small amount) - so it's
    159 unlikely "develop an NVIDIA competitor" was on any of the pitch deck slides.
    160 
    161 Perhaps OpenAI are hoping to build a strong brand so that customers won't
    162 switch to competitors so easily. It's not impossible, there is proof the [10]
    163 branding and lock-in can work in technology - but it seems difficult to manage
    164 given that LLMs themselves generically have a textual interface - meaning that
    165 there is no real API as such - you just send text, and it sends text back.
    166 
    167 Can they do anything about new entrants? Possibly. If investing $6.6bn allows
    168 them to develop a major improvement in their model then that would raise
    169 everyone else's costs considerably and probably force some of their smaller
    170 competitors out of the market. The trouble is that money is the most fungible
    171 of all goods (that is the point, after all) and that $6.6bn is not all that
    172 much of it. So this round wouldn't, by itself, be enough to dissuade others. I
    173 used to work at a bank and I can tell you that individual bond raises can be a
    174 lot more than $6.6bn.
    175 
    176 It's worth saying that even companies that raise huge sums of money sometimes
    177 turn out to have no viable business. WeWork ultimately raised over $10bn at a
    178 valuation of $47bn before it was realised that their business simply did not
    179 make sense. WeWork were valued at just $0.56bn in their most recent financial
    180 restructuring - having lost well over 95% of what was invested.
    181 
    182 Not all AI companies are doomed
    183 
    184 If LLM makers aren't going to be good businesses, does that bode ill for The
    185 Future?
    186 
    187 Firstly, it does not mean the technology will be bad. Whether the technology
    188 ends up being good or not is mostly unrelated to whether Open AI/Anthropic/
    189 Mistral/whoever makes any money off it. Container virtualisation technology is
    190 pretty well developed even though Docker made almost nothing on it. Web
    191 browsers are extremely advanced pieces of software even though making a browser
    192 is such a bad business that most don't usually count it as a business at all.
    193 And CRMs are terrible despite the fact that Salesforce is tremendously
    194 successful. Technology success and business success are mostly unrelated.
    195 
    196 And then: not all AI businesses are building models. Ideally, if I were running
    197 an AI business I would avoid building a model at all costs. Building your own
    198 models looks like an undifferentiated schlep. Using a tiny bit of some
    199 expensively trained model that Anthropic has produced could be very cost
    200 effective and might make some business idea work that wouldn't have 5 years
    201 ago.
    202 
    203 Beware software companies that aren't software companies
    204 
    205 Software companies are really good businesses. You have no real suppliers, your
    206 software is often unique (so no competitors) and the substitute is just users
    207 doing the job themselves. For this reason, software companies tend have really
    208 great margins.
    209 
    210 The problem is that not all technology companies are software companies. If you
    211 have a hugely powerful single supplier like NVIDIA then the economics of your
    212 company are going to look less like Microsoft Office and more like [11]Pan-Am.
    213 
    214 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
    215 
    216 Contact/etc
    217 
    218 Write to me at [12][email protected] about this article, especially if you
    219 disagreed with it.
    220 
    221 See [13]other things I've written or learn more about me on [14]my about page.
    222 
    223 Get an alert when I write something new, by [15]email or [16]RSS[17] rss-logo.
    224 
    225 I am on:
    226 
    227   • [18]Bluesky
    228   • [19]Mastodon
    229   • [20]Twitter
    230   • [21]Github
    231   • and [22]Linkedin.
    232 
    233 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
    234 
    235 Other notes
    236 
    237 The AI safety movement is a fantastic hypeman for LLMs as a technology.
    238 Implying (pretty dubiously) that we are [23]10 minutes from midnight in some
    239 kind of Ghost-in-The-Shell style AI crisis is in fact an extremely effective
    240 form of product marketing. Perhaps that is why OpenAI and others employ so many
    241 AI safety specialists.
    242 
    243 The Coca-Cola company mainly sit back and rake in the megabucks - but they do
    244 spend a little bit of their earnings on research. And a little bit of a lot is
    245 still significant. It's interesting that coke's market research has discovered
    246 that coke works better as a gender segregated product: Coke Zero is Diet Coke,
    247 but for men.
    248 
    249 If you want to read more about industry structure and market strategy, the
    250 place to start is with Michael Porter. He reworked his famous essay [24]The
    251 Five Forces that Shape Corporate Strategy in 2008. It's not the last word, but
    252 it probably should be the first word you read if you want to learn more. And if
    253 you like it, he has a lot more.
    254 
    255 
    256 References:
    257 
    258 [1] https://calpaterson.com/
    259 [2] https://calpaterson.com/services.html
    260 [3] https://calpaterson.com/about.html
    261 [4] https://en.wikipedia.org/wiki/Railway_Mania
    262 [5] https://en.wikipedia.org/wiki/Hopper_(microarchitecture)
    263 [6] https://a16z.com/llmflation-llm-inference-cost/
    264 [7] https://en.wikipedia.org/wiki/Browser_wars#First_browser_war_(1995%E2%80%932001)
    265 [8] https://calpaterson.com/metadata.html
    266 [9] https://www.reuters.com/technology/artificial-intelligence/openai-closes-66-billion-funding-haul-valuation-157-billion-with-investment-2024-10-02/
    267 [10] https://calpaterson.com/amazon-premium.html
    268 [11] https://en.wikipedia.org/wiki/Pan_Am
    269 [12] mailto:[email protected]
    270 [13] https://calpaterson.com/
    271 [14] https://calpaterson.com/about.html
    272 [15] https://calpatersonltd.eo.page/calpaterson
    273 [16] https://calpaterson.com/calpaterson.rss
    274 [17] https://calpaterson.com/calpaterson.rss
    275 [18] https://bsky.app/profile/calpaterson.bsky.social
    276 [19] https://fosstodon.org/@calpaterson
    277 [20] https://twitter.com/cal_paterson
    278 [21] https://github.com/calpaterson/
    279 [22] https://www.linkedin.com/in/calpaterson
    280 [23] https://en.wikipedia.org/wiki/Doomsday_Clock
    281 [24] https://hbr.org/2008/01/the-five-competitive-forces-that-shape-strategy